capsule_runtime/

trace.rs

use std::{
    path::Path,
    time::{Instant, SystemTime, UNIX_EPOCH},
};

use murmur_artifact::{ContainmentClass, LockOrigin, TraceCapture};
use serde::Serialize;
use serde_json::Value;
use tokio::{
    fs::{File, OpenOptions},
    io::{AsyncWriteExt, BufWriter},
};

use crate::{
    agent::DriverUsage,
    containment::ScopeReport,
    gateway_credential::CredentialChange,
    lanes::TaskLane,
    origin::{TaskProvenance, TrustClass},
    trace_blobs::BlobStore,
    types::InstalledArtifactSummary,
};

pub(crate) struct TraceWriter {
    writer: BufWriter<File>,
    session_id: String,
    capsule_name: String,
    capsule_version: String,
    model: String,
    capabilities: Vec<String>,
    /// The session's complete effective grant set, computed once at stage time. Written whole to
    /// `session_start` as `effective_grants`, and also the source of that event's
    /// `containment_declared`/`containment_achieved`/`workdir_exec` summary fields — they are read
    /// off this report rather than passed in beside it, so a trace cannot claim one containment
    /// class at the top level and another inside the report.
    effective_grants: ScopeReport,
    /// How much of each turn's driver request this trace keeps — see [`TraceCapture`]. Decides
    /// whether an `inference` event carries content hashes at all, whether the bodies behind
    /// them reach [`Self::blobs`], and whether `tool_call.output` is written.
    capture: TraceCapture,
    /// The content-addressed bodies this session stored, at `<workdir>/blobs/<sha256>`. Only
    /// written to under [`TraceCapture::Content`]; the directory is created on the first blob.
    blobs: BlobStore,
    /// Where the effective system prompt came from: `"manifest"`, `"cli"` or `"none"`. Derived
    /// once in [`TraceWriter::open`] from the resolved prompt and the override flag, then repeated
    /// on every `session_start` the same way `model` is — the prompt cannot change between the
    /// tasks of one session.
    system_prompt_source: &'static str,
    /// SHA-256 (lowercase hex) of the resolved system prompt, or `None` when no prompt is in
    /// effect. Always written to `session_start` (as `null` when absent) so a trace records *that*
    /// a prompt was in effect and which one, even when the text itself is withheld.
    system_prompt_sha256: Option<String>,
    /// Where the inference credential came from: `"config"`, `"environment"`, `"manifest"` or
    /// `"none"`. Set by [`Self::set_credential_source`] before `session_start` is written.
    credential_source: &'static str,
    /// Every credential gateway the session holds. Set by [`Self::set_gateways`] before
    /// `session_start` is written; empty until then.
    gateways: Vec<SessionGateway>,
    /// What a controller may change on this session. Set by [`Self::set_control`] before
    /// `session_start` is written; `None` for a capsule with no `control:` block.
    control: Option<SessionControl>,
    /// The driver choices. Set by [`Self::set_inference_choices`] before `session_start` is
    /// written; empty for a capsule with no alternates.
    inference_choices: Vec<SessionInferenceChoice>,
    /// `inference.tool_refresh` as `ToolRefresh::wire_name` spells it, or `None` under
    /// `transport: process`. Set by [`Self::set_tool_refresh`] before `session_start` is written.
    tool_refresh: Option<&'static str>,
    /// The staged artifacts whose `murmur.lock` pin a running capsule fetched. Set by
    /// [`Self::set_runtime_artifacts`] before `session_start` is written; empty until then.
    runtime_artifacts: Vec<RuntimeArtifact>,
    /// The session `mur run --resume` continued, verbatim as the operator's address resolved it.
    /// `None` on every ordinary launch. Written to `session_start` on both, so its absence
    /// identifies a trace from a runtime that predates the key.
    resumed_from: Option<String>,
    /// The launch-scoped context id: the `--context` value, or the id `--resume` resolved to.
    /// `None` when each task of this launch mints its own.
    context_id: Option<String>,
    /// The session that spawned this one, and the delegation that created it, read out of the
    /// handle the spawning capsule's launcher injected. Both `None` for a capsule nobody
    /// delegated, and both omitted from `session_start` when absent rather than written as null.
    spawned_by: Option<String>,
    spawn_delegation_id: Option<String>,
    session_start_time: Instant,
    /// The session node of the event tree: the `event_id` `session_start` carries, and the
    /// `parent_id` every launch-scoped event names. Minted in [`TraceWriter::open`] rather than
    /// at write time so [`ResourceTraceAppender`], which is opened before the frame is written,
    /// can be handed the same value.
    session_event_id: String,
    session_started: bool,
    session_ended: bool,
    /// The task node: the `event_id` of the active task's `task_start`. Set by
    /// [`Self::write_task_start`], cleared by [`Self::write_task_end`], `None` between tasks.
    task_event_id: Option<String>,
    /// The turn node: the `event_id` of the active turn's own agent-loop `inference`. Every
    /// event a turn produces hangs off it. Cleared at both task boundaries so a new task never
    /// parents its first events to the previous task's last turn.
    turn_event_id: Option<String>,
    // Running totals — updated as events are written, used for fallback session_end on error exit
    total_turns: u32,
    total_input_tokens: u64,
    total_output_tokens: u64,
    total_tool_calls: u32,
    total_shell_calls: u32,
    // Per-task counters — reset on each write_task_start; consumed by write_task_end
    task_turns: u32,
    task_input_tokens: u64,
    task_output_tokens: u64,
    task_tool_calls: u32,
    task_shell_calls: u32,
    task_start_instant: Option<Instant>,
    pub(crate) active_task_id: Option<String>,
    /// How many `task_failed` lines this writer has written, so a caller can tell whether one run
    /// wrote its own without comparing reasons.
    task_failures_written: u64,
    /// The `reason` of the last `task_failed` line, as written (already capped).
    last_task_failure: Option<String>,
    /// The spend ceilings `session_start` records. Set by [`Self::set_spend_ceilings`].
    max_session_tokens: Option<u64>,
    machine_tokens_per_day: Option<u64>,
}

/// Provenance of an inference record that did **not** come from the agent
/// loop's own turn — today, only a hook's `run-inference` call. `None` at a
/// `write_inference`/`emit_inference` call site means "ordinary agent-loop
/// turn", which serializes exactly as it did before this existed.
#[derive(Debug, Clone)]
pub(crate) struct InferenceOrigin {
    /// `hook:<manifest name>` of the hook that made the call.
    pub(crate) source: String,
    /// Model string actually sent (the attempted model, on failure).
    pub(crate) model: String,
}

/// `compaction_declined.reason` when the threshold was crossed but no bound hook returned
/// `replace-context`.
pub(crate) const COMPACTION_DECLINED_NO_HOOK_REPLACEMENT: &str = "no_hook_replacement";

/// `compaction_declined.reason` when a hook's replacement was rejected because its tool calls
/// and tool results did not pair up.
pub(crate) const COMPACTION_DECLINED_UNRESOLVED_TOOL_CALL: &str = "unresolved_tool_call";

/// `context_seed.outcome` when the whole proposal fit the budget and was committed as-is.
pub(crate) const SEED_OUTCOME_SEEDED: &str = "seeded";

/// `context_seed.outcome` when the proposal was over budget and the oldest messages were
/// dropped from its front until the rest fit.
pub(crate) const SEED_OUTCOME_TRIMMED: &str = "trimmed";

/// `context_seed.outcome` when the overflowing front was handed to the compaction hook and
/// its summary became the seed's first message.
pub(crate) const SEED_OUTCOME_COMPACTED: &str = "compacted";

/// `context_seed.outcome` when nothing was committed. Always paired with a `reason`.
pub(crate) const SEED_OUTCOME_REJECTED: &str = "rejected";

/// `context_seed.reason` when one message alone is wider than the whole budget. Trimming
/// cannot help — dropping everything else still leaves it over — so nothing is committed.
pub(crate) const SEED_REJECTED_MESSAGE_OVER_BUDGET: &str = "message_over_budget";

/// `context_seed.reason` when the proposal overflows the budget by more multiples of it than
/// the runtime will summarize. A seed that far over is a broken hook, not a full memory.
pub(crate) const SEED_REJECTED_OVERFLOW_OVER_LIMIT: &str = "overflow_over_limit";

/// `context_seed.reason` when the capsule declares no `context.max_tokens`, so
/// `context.seed_budget` is a fraction of nothing and there is no ceiling to enforce.
pub(crate) const SEED_REJECTED_NO_BUDGET: &str = "no_budget";

/// `context_seed.reason` when the session runs `inference.transport: process`, which owns its
/// own context and has no host-side message list to seed.
pub(crate) const SEED_REJECTED_UNSUPPORTED_TRANSPORT: &str = "unsupported_transport";

/// `retention.store` for the session directories under a workdir.
pub(crate) const RETENTION_STORE_SESSIONS: &str = "sessions";

/// `retention.store` for the conversation records under `~/.murmur/conversations/`.
pub(crate) const RETENTION_STORE_RECORDS: &str = "records";

/// `retention.reason` when `trace.retain.max_sessions` put a session outside the newest N.
pub(crate) const RETENTION_REASON_MAX_SESSIONS: &str = "max_sessions";

/// `retention.reason` when `trace.retain.max_age` or `context.retain.max_age` put a session or a
/// record outside the window.
pub(crate) const RETENTION_REASON_MAX_AGE: &str = "max_age";

/// `retention.reason` when `context.retain.max_messages` truncated the front of a record.
pub(crate) const RETENTION_REASON_MAX_MESSAGES: &str = "max_messages";

/// Mint one event identity: `evt_` + a UUID v7 in simple form, the same scheme `ses_`, `tsk_`,
/// `ctx_`, `dep_` and `req_` use. Ids therefore sort by mint time and carry their own
/// millisecond timestamp, so a trace can be ordered and time-bounded without reading any
/// payload field.
///
/// Called once per line at the moment of the write. An id is never reused, never derived from
/// the event's content, and never reconstructed from anything else in the file.
pub(crate) fn new_event_id() -> String {
    format!("evt_{}", uuid::Uuid::now_v7().simple())
}

// ── Event structs (Serialize → JSONL lines) ──────────────────────────────────
//
// Every struct opens with `event_type`, `event_id`, `parent_id`, `session_id` and `timestamp`,
// in that order, so a human scanning the file reads a line's identity before its payload.
// `parent_id` is the `event_id` of the event this one hangs off — see [`TraceWriter`]'s
// `session_parent`/`task_parent`/`turn_parent` for which node each event names, and
// [`new_event_id`] for the id format.
//
// The turn-level events (`inference`, `tool_call`, `skill_call`, `shell`, `compaction`,
// `compaction_declined`) additionally carry `task_id`, taken from the writer's active task and
// always written — `null` when no task is in scope. `task_start`/`task_end` carry their own and
// are not duplicated.

#[derive(Serialize)]
struct SessionStartEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    capsule_name: String,
    capsule_version: String,
    model: String,
    max_turns: u32,
    /// `inference.max_session_tokens` as this session enforces it. Always written, `null` when
    /// no session ceiling applies.
    max_session_tokens: Option<u64>,
    /// `spend.machine_tokens_per_day` when this session keeps the machine ledger. Always written,
    /// `null` when the machine ceiling does not cover this session.
    machine_tokens_per_day: Option<u64>,
    capabilities: Vec<String>,
    tools_declared: Vec<String>,
    /// Strongest containment class any source asked for. A *requirement*, not an observation.
    containment_declared: ContainmentClass,
    /// What this session actually ran under: the probed enforcement tier's ceiling, capped by
    /// `workdir_exec` below. Reading these two together is how a session's trace shows whether the
    /// capsule ran under the containment its operator asked for.
    containment_achieved: ContainmentClass,
    /// `capabilities.filesystem.workdir_exec`. Always written, including `false`, so its absence
    /// identifies a trace from a runtime that predates the key rather than a capsule that declined
    /// it.
    ///
    /// `true` is the reason a trace can show `containment_achieved: advisory` on a host whose
    /// `session_start` would otherwise say `scoped` — and it is the record that this session's
    /// `capabilities.shell.allow` was advisory too, since anything in the workdir could run.
    workdir_exec: bool,
    /// Where this host's permission to create an unprivileged user namespace came from, as
    /// `UsernsGrant::wire_name` — `null` only off Linux, where AppArmor does not exist.
    ///
    /// Mirrored to the top level from `effective_grants` for the same reason `workdir_exec` is:
    /// an auditor should not have to descend into a nested object to answer it. Recorded because
    /// `containment_achieved: sealed` reached through the shipped AppArmor profile and the same
    /// class reached on a host whose unprivileged-userns hardening is switched off for every
    /// binary are two very different records, and this key is the only thing separating them.
    ///
    /// Always written, on the same terms as `workdir_exec`: its absence identifies a trace from a
    /// runtime that predates the key.
    userns_grant: Option<crate::sealed::UsernsGrant>,
    /// The complete grant set this session ran under — every destination, binary, path and
    /// environment variable the policy actually opened, plus the probed enforcement tier — in the
    /// exact shape `mur run --explain-scope --json` prints for the same policy on the same host.
    ///
    /// `capabilities` above stays what it always was: a list of category *names*. It answers "did
    /// this session have network access?" but not "to where?", which is the question an auditor
    /// reading a finished trace actually has. This field answers it without re-parsing the
    /// manifest, which by then may have changed or moved.
    effective_grants: ScopeReport,
    /// Where the system prompt this session ran with came from: `"manifest"` when the manifest's
    /// own `inference.system_prompt`/`system_prompt_file`/`system_prompt_artifact` supplied it,
    /// `"cli"` when `mur run --system-prompt` overrode it (including when the override was empty
    /// and therefore cleared it), `"none"` when no prompt was in effect at all.
    ///
    /// Always written, on the same terms as `workdir_exec`: its absence identifies a trace from a
    /// runtime that predates the key, not a session that had no prompt.
    system_prompt_source: &'static str,
    /// SHA-256 (lowercase hex) of the resolved prompt — the text as `resolve_system_prompt`
    /// returned it, before the `[Capsule]` identity block is prepended — or `null` when no prompt
    /// was in effect. Always written, so two sessions can be compared for prompt equality without
    /// either trace having to carry the prompt itself.
    ///
    /// Deliberately not the same value as `inference.system_sha`, which covers the augmented
    /// prompt that actually went on the wire: this one names the prompt the operator wrote.
    /// Under `trace.capture: content` those resolved bytes are also the blob this hash names.
    system_prompt_sha256: Option<String>,
    /// Where the inference credential this session attaches came from: `"config"`,
    /// `"environment"`, `"manifest"` or `"none"`. Never the value.
    credential_source: &'static str,
    /// Every credential gateway the session holds, the inference gateway first. Always written,
    /// and empty for a session with none, so an auditor can tell an unmetered third-party
    /// upstream from spend the ceilings cover.
    gateways: Vec<SessionGateway>,
    /// What a controller holding this session's control token may change: the setting names and
    /// secret names `control:` declares. Omitted entirely for a capsule with no `control:` block.
    #[serde(skip_serializing_if = "Option::is_none")]
    control: Option<SessionControl>,
    /// Every driver choice the agent loop may be switched to, the primary first, each with its
    /// credential's source and whether it resolved at launch. Omitted entirely for a capsule that
    /// declares no `inference.alternates`.
    #[serde(skip_serializing_if = "Vec::is_empty")]
    inference_choices: Vec<SessionInferenceChoice>,
    /// `inference.tool_refresh`: `"compaction"` or `"immediate"` under `transport: http`, `null`
    /// under `transport: process`, where the harness owns its tool list. Always written.
    tool_refresh: Option<&'static str>,
    /// Every staged artifact whose `murmur.lock` origin is `runtime` — pinned by a running
    /// capsule's `manage.pull()`, never vetted by the operator — in staging order. Each reaches
    /// the model marked untrusted. Always written, `[]` when there are none.
    runtime_artifacts: Vec<RuntimeArtifact>,
    /// The session `mur run --resume` continued, or `null` on an ordinary launch. Together with
    /// `context_id` below it is what makes a resumed conversation followable back through the
    /// sessions that built it.
    ///
    /// Always written, on the same terms as `workdir_exec`: its absence identifies a trace from a
    /// runtime that predates the key.
    resumed_from: Option<String>,
    /// The context id every task of this launch runs under — the `mur run --context` value, or
    /// the id `--resume` resolved to — and `null` when each task mints its own. `task_start`
    /// carries the id a task actually ran under either way.
    ///
    /// Always written, on the same terms as `resumed_from`.
    context_id: Option<String>,
    /// The `ses_` id of the session that spawned this one, for a capsule launched by another
    /// capsule's `delegate-task`. Omitted from the record entirely for every other launch: the
    /// field is absent, not null, on the same terms as `task_start.delegation_id`.
    #[serde(skip_serializing_if = "Option::is_none")]
    spawned_by: Option<String>,
    /// The `dlg_` id of the delegation that created this session, character-identical to the id
    /// in the spawning session's own `delegation_start`. Present exactly when `spawned_by` is.
    #[serde(skip_serializing_if = "Option::is_none")]
    delegation_id: Option<String>,
}

#[derive(Serialize)]
struct InferenceEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    turn: u32,
    task_id: Option<String>,
    /// The input this turn cost, and the number the task and session totals accumulate.
    ///
    /// Under `transport: http` it is the runtime's own tiktoken estimate of the request, counted
    /// before the request was sent, and the provider's own count sits beside it in
    /// `input_tokens_actual`. Under `transport: process` the runtime estimates nothing — the
    /// harness holds the conversation — so this *is* the harness's own reported count and
    /// `input_tokens_actual` is absent: there is one measurement on that transport and it is
    /// recorded once.
    ///
    /// Absent means no count was made at all, which only a `process` driver reporting no usage
    /// produces. Never conflate it with `0`, which is a count of none.
    #[serde(skip_serializing_if = "Option::is_none")]
    input_tokens: Option<u64>,
    /// The output this turn cost, on the same terms as [`Self::input_tokens`]: the runtime's own
    /// tiktoken estimate of the raw driver response under `http`, the harness's own reported
    /// count under `process`.
    #[serde(skip_serializing_if = "Option::is_none")]
    output_tokens: Option<u64>,
    decision: String,
    /// The provider's own stop reason for this turn, verbatim as the agent loop dispatched on
    /// it. Present on every agent-loop turn, including one whose driver reported nothing — an
    /// empty string then, because that is what the loop dispatched on. Absent means no driver
    /// response was parsed at all: a hook's `run-inference` and the `process` transport both
    /// name their `decision` locally and never saw a provider stop reason.
    #[serde(skip_serializing_if = "Option::is_none")]
    stop_reason: Option<String>,
    tool_name: Option<String>,
    /// The provider's own counts for this call, each written only when the driver reported
    /// it. Absent means the driver reported nothing, never zero.
    #[serde(skip_serializing_if = "Option::is_none")]
    input_tokens_actual: Option<u64>,
    #[serde(skip_serializing_if = "Option::is_none")]
    output_tokens_actual: Option<u64>,
    #[serde(skip_serializing_if = "Option::is_none")]
    cached_tokens: Option<u64>,
    #[serde(skip_serializing_if = "Option::is_none")]
    cache_write_tokens: Option<u64>,
    /// The part of the output the provider or harness reports as reasoning — a subset of
    /// `output_tokens`, never an addition to it. Absent when nothing reported one.
    #[serde(skip_serializing_if = "Option::is_none")]
    thinking_tokens: Option<u64>,
    /// Where this inference came from. Absent for an ordinary agent-loop turn,
    /// so every pre-existing consumer sees a byte-identical record; `"hook:<name>"`
    /// for a completion that hook ran through `murmur:runtime/inference`'s
    /// `run-inference`.
    #[serde(skip_serializing_if = "Option::is_none")]
    origin: Option<String>,
    /// The driver choice that served an agent-loop turn, by name. Written only for a capsule that
    /// declares `inference.alternates`, beside `model`.
    #[serde(skip_serializing_if = "Option::is_none")]
    driver_choice: Option<String>,
    /// Model string actually sent for this call (the attempted model, on failure). Written
    /// alongside `origin` for a hook's completion, and alongside `driver_choice` for an agent-loop
    /// turn of a capsule that declares alternates. Any other turn's model is the session-start
    /// record's and is not repeated per turn.
    #[serde(skip_serializing_if = "Option::is_none")]
    model: Option<String>,
    /// Identities of the messages this request embedded, in the order they sat in it. Under an
    /// active driver continuation only the tail past the acked length is sent, and this names
    /// exactly that tail. Empty — and therefore absent — for a hook's own `run-inference` and for
    /// the `process` transport, neither of which sends a message list the runtime minted.
    #[serde(skip_serializing_if = "Vec::is_empty")]
    message_ids: Vec<String>,
    /// SHA-256 (lowercase hex) of the UTF-8 bytes of this request's `system` string — the
    /// augmented prompt with the `[Capsule]` block already in it.
    ///
    /// These are the bytes Murmur **sent**, not what the model **saw**: provider-side injection,
    /// tokenizer differences and safety layers all happen past the wire and are invisible to the
    /// runtime. Absent under `trace.capture: none`, and for a record the runtime did not build
    /// the payload for — a hook's `run-inference`, and the `process` transport.
    #[serde(skip_serializing_if = "Option::is_none")]
    system_sha: Option<String>,
    /// SHA-256 (lowercase hex) of this request's serialized `tools` array. The bytes Murmur
    /// sent, on the same terms as [`Self::system_sha`].
    #[serde(skip_serializing_if = "Option::is_none")]
    tools_sha: Option<String>,
    /// SHA-256 (lowercase hex) of the raw driver response body, as the loop read it before
    /// parsing. What the driver returned to Murmur, byte for byte.
    #[serde(skip_serializing_if = "Option::is_none")]
    response_sha: Option<String>,
    /// SHA-256 (lowercase hex) of each message this request embedded, in send order — one entry
    /// per `message_ids` entry, over the same messages after the runtime's identity keys are
    /// stripped. The bytes Murmur sent, on the same terms as [`Self::system_sha`].
    ///
    /// Not redundant with `message_ids`: an id names an entity and is freshly minted every run,
    /// so comparing two runs' id arrays only says every id differs. Comparing these says where
    /// the two prompts stopped agreeing — the divergence index is the first unequal position.
    #[serde(skip_serializing_if = "Vec::is_empty")]
    message_shas: Vec<String>,
}

/// The bodies one driver request put on the wire, split into the pieces an `inference` event
/// hashes.
///
/// Built from the payload [`crate::agent::build_driver_payload`] produced and the raw response
/// string the driver returned, so what the trace names and what the driver received cannot
/// drift: there is no second serialization for them to drift apart in.
///
/// Every piece is the bytes Murmur **sent** (or, for the response, was handed back), never what
/// the model saw.
pub(crate) struct WireCapture {
    /// The payload's `system` string value as UTF-8 text — the augmented prompt, `[Capsule]`
    /// block included. Stored as text so its blob `cat`s as a readable prompt.
    system: Vec<u8>,
    /// The payload's `tools` array, serialized.
    tools: Vec<u8>,
    /// Each element of the payload's `messages` array, serialized, in send order. That array is
    /// the `wire_messages` slice after `strip_message_identity`, so no message identity key can
    /// reach a hash or a blob.
    messages: Vec<Vec<u8>>,
    /// The raw driver response body, exactly as the loop read it.
    response: Vec<u8>,
}

impl WireCapture {
    /// Split a built driver payload and its raw response into the four bodies the trace hashes.
    ///
    /// `payload` must be the value that was serialized and handed to the driver; taking the
    /// pieces out of it, rather than rebuilding them from the loop's own state, is what makes a
    /// blob byte-identical to the substring of the request it came from.
    pub(crate) fn from_driver_payload(payload: &Value, response: &str) -> Self {
        Self {
            system: payload
                .get("system")
                .and_then(Value::as_str)
                .unwrap_or_default()
                .as_bytes()
                .to_vec(),
            tools: payload
                .get("tools")
                .map(|tools| serde_json::to_vec(tools).unwrap_or_default())
                .unwrap_or_default(),
            messages: payload
                .get("messages")
                .and_then(Value::as_array)
                .map(|messages| {
                    messages
                        .iter()
                        .map(|message| serde_json::to_vec(message).unwrap_or_default())
                        .collect()
                })
                .unwrap_or_default(),
            response: response.as_bytes().to_vec(),
        }
    }
}

#[derive(Serialize)]
struct ToolCallEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    turn: u32,
    task_id: Option<String>,
    tool_name: String,
    /// The provider's own id for this call, as it appeared on the driver response block that
    /// asked for it. Recorded verbatim and never parsed — the same contract `resource_id` has.
    /// `null` when the provider named none, which is how every host-synthesized call reads.
    tool_call_id: Option<String>,
    input: Value,
    input_bytes: u64,
    #[serde(skip_serializing_if = "Option::is_none")]
    output: Option<String>,
    output_bytes: u64,
    duration_ms: u64,
    status: String,
    /// The tool's self-declared `state_effect` for this call (`read`/`mutate`),
    /// lifted verbatim from `tool-result.metadata`. Absent when the tool declared
    /// nothing; consumers treat absence conservatively. See `wit/tool.wit`.
    #[serde(skip_serializing_if = "Option::is_none")]
    state_effect: Option<String>,
    /// The resource this call addressed, as declared by the tool and lifted verbatim
    /// from `tool-result.metadata`. An opaque, tool-defined string — never parsed here.
    /// Absent when the tool declared nothing, in which case consumers fall back to
    /// guessing the resource from the call's input. See `wit/tool.wit`.
    #[serde(skip_serializing_if = "Option::is_none")]
    resource_id: Option<String>,
}

#[derive(Serialize)]
struct SkillCallEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    turn: u32,
    task_id: Option<String>,
    skill_name: String,
    output_bytes: u64,
    duration_ms: u64,
    status: String,
    /// The skill's `murmur.lock` origin: `"operator"` or `"runtime"`. Always written.
    origin: &'static str,
    /// `"trusted"` for an operator pin, `"untrusted"` for a runtime one, whose guidance reached
    /// the model fenced. Always written, in the spelling `task_start.trust` uses.
    trust: &'static str,
}

#[derive(Serialize)]
struct ShellEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    turn: u32,
    task_id: Option<String>,
    /// The program that ran — a canonical absolute path when the host `PATH` resolved
    /// the invoked name, else the bare name. `command` carries the argument list alone,
    /// so this is the only field that says *what* ran.
    binary: String,
    command: String,
    exit_code: i32,
    stdout_bytes: u64,
    stderr_bytes: u64,
    duration_ms: u64,
    /// The `capabilities.resources` field this subprocess was killed for exceeding, when the
    /// kernel's own evidence names exactly one (`SIGXCPU`/`SIGXFSZ`, or a cgroup
    /// `memory.events`/`pids.events` counter that moved). Omitted from the JSONL entirely
    /// otherwise — including for a subprocess that died for a reason no single limit can be
    /// pinned to, which must not read as "no limit was involved".
    #[serde(skip_serializing_if = "Option::is_none")]
    resource_limit: Option<String>,
}

/// A shell command that outran its grace period and moved to the background. Hangs off the turn
/// that started it, so `mur trace steps` shows it where the model saw it.
#[derive(Serialize)]
struct ShellDetachedEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    turn: u32,
    task_id: Option<String>,
    work_id: String,
    binary: String,
    command: String,
    /// The `lifecycle.shell_grace_secs` in force, in milliseconds — what this command outran.
    grace_ms: u64,
}

/// A detached command finishing, written where the task loop drains it. Hangs off the session
/// rather than a turn: by the time it lands, the turn that started the command is over.
#[derive(Serialize)]
struct ShellCompletedEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    work_id: String,
    binary: String,
    command: String,
    exit_code: i32,
    duration_ms: u64,
    /// Workdir-relative path of the file holding the command's full stdout and stderr.
    output_path: String,
    output_bytes: u64,
    /// Attributed the same way [`ShellEvent::resource_limit`] is, and omitted on the same terms.
    #[serde(skip_serializing_if = "Option::is_none")]
    resource_limit: Option<String>,
    /// `"ok"` or `"error"` — the field that tells a failed background command from a successful
    /// one without reading its prose.
    status: String,
    /// The `completion`-origin task this completion was enqueued as, which is what makes "which
    /// command produced this task" answerable from `trace.jsonl` alone.
    completion_task_id: String,
}

/// A detached command the session ended without carrying its result back. No task will ever report
/// it; this record and the operator report beside it are what keep the loss visible.
///
/// The command was either still running at the sweep or finished during teardown, and the second
/// case knows more: [`Self::exit_code`], [`Self::output_path`] and [`Self::output_bytes`] are
/// written together for it and omitted together for the first. A command still running therefore
/// produces a line byte-identical in shape to one written before those fields existed — omitted,
/// never `null`, because `null` would read as a known-absent exit code rather than an unknown one.
#[derive(Serialize)]
struct ShellAbandonedEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    work_id: String,
    binary: String,
    command: String,
    /// How long the command had been running when the session gave up on it. For a command that
    /// finished during teardown, its full duration from spawn to exit.
    running_ms: u64,
    /// `128 + signal` for a signal kill, as on `shell_completed`. Absent for a command still
    /// running.
    #[serde(skip_serializing_if = "Option::is_none")]
    exit_code: Option<i32>,
    /// Workdir-relative path of the file holding the command's full stdout and stderr, always
    /// `logs/<work_id>.log`. Absent for a command still running, for which that file was never
    /// written and never will be.
    #[serde(skip_serializing_if = "Option::is_none")]
    output_path: Option<String>,
    /// Size of that file. Present exactly when [`Self::output_path`] is.
    #[serde(skip_serializing_if = "Option::is_none")]
    output_bytes: Option<u64>,
}

/// A demoted command a later launch found unaccounted for, appended to the `trace.jsonl` of the
/// session that started it.
///
/// Carries no `exit_code`, no `status`, no `duration_ms`, no `output_path` and no `output_bytes`:
/// none of them exists for a command whose runtime died, and a record that named any of them
/// would be readable as a result. The event type is the whole discriminator against
/// [`ShellCompletedEvent`].
///
/// Its presence is also what stops the same work id being reported by a second resume, so this
/// line and the `shell_detached` it answers live in the same file and are pruned together by
/// [`crate::retention::prune_sessions`].
#[derive(Serialize)]
struct ShellLostEvent {
    event_type: &'static str,
    event_id: String,
    /// The prior session's own `session_start` node. Omitted from the line when that record could
    /// not be read back — a line must never name a parent the file does not hold.
    #[serde(skip_serializing_if = "Option::is_none")]
    parent_id: Option<String>,
    /// The prior session's id, not the resuming one's, so the file stays internally consistent.
    session_id: String,
    timestamp: u64,
    work_id: String,
    binary: String,
    command: String,
    /// The `shell_detached` line's own timestamp.
    detached_at_ms: u64,
    /// The session that found the work unaccounted for and reported it.
    reconciled_by_session: String,
    /// The `completion`-origin task that session enqueued to report it, which is what makes "which
    /// task reported this loss" answerable from the two files alone.
    reconciled_task_id: String,
}

/// A demotion whose own `shell_detached` line could not be written.
///
/// Best-effort and usually absent: it goes to the file that just failed. When it does land, it is
/// the record that the work id below has no `shell_detached` line and so will never be reported
/// as lost.
#[derive(Serialize)]
struct ShellDetachUnrecordedEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    turn: u32,
    task_id: Option<String>,
    work_id: String,
    binary: String,
    /// The write error, as the operating system reported it.
    reason: String,
}

#[derive(Serialize)]
struct CompactionEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    turn: u32,
    task_id: Option<String>,
    tokens_before: u64,
    tokens_after: u64,
}

/// The compaction threshold was crossed and the context was left as it was. The session
/// continues over budget, so the record of *why* is the only thing separating this from a
/// session that never needed compacting at all.
#[derive(Serialize)]
struct CompactionDeclinedEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    turn: u32,
    task_id: Option<String>,
    /// Context occupancy at the moment of the decline — the same measurement `compaction`
    /// records as `tokens_before`, and the budget the session went on running over.
    tokens: u64,
    /// `"no_hook_replacement"` when no bound hook returned `replace-context`;
    /// `"unresolved_tool_call"` when the replacement's tool calls and tool results did not pair.
    reason: String,
}

/// An `on-task-start` hook proposed context and the runtime decided what to do with it.
///
/// Written once per task that had a seeding hook return something, whatever the decision —
/// including a rejection, where `tokens` is `0`. A seed that was silently discarded is
/// indistinguishable from a capsule with no memory hook at all, which is the failure this
/// record exists to make impossible.
#[derive(Serialize)]
struct ContextSeedEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: Option<String>,
    /// Manifest name of the hook that returned the `seed-context`.
    hook_name: String,
    /// Tokens actually committed to the head of the context: `0` for a rejection, the
    /// proposal's own count for `seeded`, and the surviving count for `trimmed`/`compacted`.
    tokens: u64,
    /// Tokens the hook proposed, before any trim or summarization. Together with
    /// `budget_tokens` this is what makes a rejection readable without the hook's own log.
    proposed_tokens: u64,
    /// The ceiling in force: `floor(context.max_tokens * context.seed_budget)`, or `0` when
    /// the capsule declares no `context.max_tokens`.
    budget_tokens: u64,
    /// `"seeded"`, `"trimmed"`, `"compacted"` or `"rejected"` — see the `SEED_OUTCOME_*`
    /// constants.
    outcome: String,
    /// Why nothing was committed, on `"rejected"` only — see the `SEED_REJECTED_*`
    /// constants. Absent from the JSONL for every other outcome.
    #[serde(skip_serializing_if = "Option::is_none")]
    reason: Option<String>,
    /// The `msg_`-prefixed id of every message committed, in the order they were placed.
    /// Empty on a rejection. These are the ids the runtime minted; they never reach the
    /// driver payload, so this record is the only place they are visible.
    message_ids: Vec<String>,
}

/// Retention deleted something. One line per (store, reason) pair that removed anything, in the
/// trace of the session that performed the deletion — the only place the pruning of session N's
/// predecessors can go.
///
/// Session-scoped, and written immediately after `session_start`: a session directory that
/// vanishes with no explanation makes "where did my trace go" unanswerable, and the answer has to
/// be in the trace of the run that did it.
#[derive(Serialize)]
struct RetentionEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    /// [`RETENTION_STORE_SESSIONS`] or [`RETENTION_STORE_RECORDS`].
    store: String,
    /// [`RETENTION_REASON_MAX_SESSIONS`], [`RETENTION_REASON_MAX_AGE`] or
    /// [`RETENTION_REASON_MAX_MESSAGES`].
    reason: String,
    /// How many units this reason removed from this store: session directories, context
    /// directories, or — for a truncation — the one record that was rewritten. Never zero;
    /// nothing removed means no event.
    removed: u32,
    /// What went: `ses_` directory names, or context ids.
    targets: Vec<String>,
    /// Messages dropped from the front of the record. Written for
    /// [`RETENTION_REASON_MAX_MESSAGES`] only, and absent from the JSONL otherwise.
    #[serde(skip_serializing_if = "Option::is_none")]
    messages_dropped: Option<u64>,
}

#[derive(Serialize)]
struct A2aTaskReceivedEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: String,
    context_id: String,
    message_id: String,
    traceparent_from_caller: Option<String>,
}

#[derive(Serialize)]
struct A2aSendEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    peer_url: String,
    message_id: String,
    task_id: String,
    context_id: String,
    traceparent: Option<String>,
    /// The class this runtime stamped on the outgoing request, so a chain of capsules reads the
    /// same way from the sending end as the receiving end's `task_start` reads it.
    trust: String,
}

/// A successful `manage.pull()`: the artifact a running capsule fetched and pinned in
/// `murmur.lock`. A refused or failed pull writes nothing.
#[derive(Serialize)]
struct ArtifactPulledEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    name: String,
    version: String,
    /// `"skill"` or `"tool"`, the only roles a pull installs.
    runtime: &'static str,
    /// The origin the lock entry holds after the pull: `"runtime"`, or `"operator"` when the
    /// operator had already pinned the same version.
    origin: &'static str,
    /// The session the lock entry names as the puller — this one, unless an earlier pull of the
    /// same pin recorded another. `null` for an `"operator"` entry.
    session: Option<String>,
    trust: &'static str,
}

#[derive(Serialize)]
struct SessionEndEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    total_turns: u32,
    total_input_tokens: u64,
    total_output_tokens: u64,
    total_tool_calls: u32,
    total_shell_calls: u32,
    duration_ms: u64,
    exit_status: String,
}

#[derive(Serialize)]
struct TaskStartEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: String,
    context_id: String,
    /// Which door the task came through. `source` and `origin` answer different questions —
    /// `"a2a"` covers both a trusted peer and an untyped webhook.
    source: String,
    origin: String,
    trust: String,
    /// The queue lane this task waited in, derived from `origin`. Recorded so the order two
    /// tasks ran in is answerable from the trace alone.
    lane: String,
    /// The delegation whose completion this task is, for a `completion`-origin task that arrived
    /// from a child this capsule launched. Omitted from the record entirely for every other task:
    /// the field is absent, not null.
    #[serde(skip_serializing_if = "Option::is_none")]
    delegation_id: Option<String>,
    message_parts_bytes: u64,
}

#[derive(Serialize)]
struct TaskEndEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: String,
    exit_status: String,
    duration_ms: u64,
    turns: u32,
    input_tokens: u64,
    output_tokens: u64,
    tool_calls: u32,
    shell_calls: u32,
    /// How many times this task's agent loop was reopened by an `on-task-end` hook
    /// before the terminal outcome. `0` for a task that ran once (the common case).
    reopen_count: u32,
}

/// A person stopped this task. Written where the loop actually stopped, so the record names the
/// wait that was interrupted and everything that was still running at that moment.
///
/// Distinct from the `task_end` that follows it: `task_end` says the task is over and with which
/// `exit_status`, this says a person made that happen and what it left behind. A task cancelled
/// while still queued has no `task_start` and gets no `task_end`, so this is its only record.
#[derive(Serialize)]
struct TaskCanceledEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: String,
    /// The turn that was in flight, 0-based. Absent for a task cancelled before it ever ran.
    #[serde(skip_serializing_if = "Option::is_none")]
    turn: Option<u32>,
    /// Where cancellation landed: `queued`, `turn`, `inference`, `input`, `delegation` or
    /// `harness`.
    phase: String,
    /// Demoted shell commands still running when the loop stopped. They keep their own
    /// lifecycle — nothing here was killed.
    detached_work_ids: Vec<String>,
    /// Delegations still in flight when the loop stopped, left running exactly as a delegation
    /// deadline leaves them.
    delegation_ids: Vec<String>,
}

/// `task_rejected.cause` when the session ended on its own — close-out, `after_task: exit`, a
/// failed launch task or the idle timeout — with the task still queued.
pub(crate) const TASK_REJECTED_SESSION_ENDED: &str = "session_ended";

/// `task_rejected.cause` when `mur stop` (`SIGTERM`) ended the session with the task still queued.
pub(crate) const TASK_REJECTED_SESSION_STOPPED: &str = "session_stopped";

/// The session stopped taking work while this task was still queued, so it was refused rather
/// than run. Written once per refused task, after the task loop has ended and before
/// `session_end`.
///
/// A refused task never started, so it has no `task_start`, no `task_end` and never a
/// `task_failed`: this is its only record. It hangs off the session node.
#[derive(Serialize)]
struct TaskRejectedEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: String,
    context_id: String,
    /// The `IncomingTask::source` the task was enqueued with: `a2a`, `detached_shell` or
    /// `detached_lost`.
    source: String,
    /// One of the `TASK_REJECTED_*` constants.
    cause: String,
    /// The `status.message` the task's final `rejected` frame carries.
    reason: String,
}

/// `task_failed.cause` when the driver returned an error status or a response whose
/// `stop_reason` is `"error"`, and no credential rejection is pending.
pub(crate) const TASK_FAILED_DRIVER_ERROR: &str = "driver_error";

/// `task_failed.cause` for a driver error while the gateway holds a credential rejection: the
/// provider keeps refusing the key, and the reason is the `E-RUN-027` message.
pub(crate) const TASK_FAILED_CREDENTIAL_REJECTED: &str = "credential_rejected";

/// `task_failed.cause` for a response the loop cannot act on: `stop_reason: "tool_call"` with no
/// `tool_call` block, or a `stop_reason` that is missing or unsupported.
pub(crate) const TASK_FAILED_MALFORMED_RESPONSE: &str = "malformed_response";

/// `task_failed.cause` when a bound compaction hook returned `Err`.
pub(crate) const TASK_FAILED_COMPACTION_HOOK: &str = "compaction_hook";

/// `task_failed.cause` when a `request-input` wait passed `lifecycle.input_timeout_secs`.
pub(crate) const TASK_FAILED_INPUT_TIMEOUT: &str = "input_timeout";

/// `task_failed.cause` when an `on-task-end` hook still asked to reopen and
/// `lifecycle.max_task_reopens` or `inference.max_turns` refused it.
pub(crate) const TASK_FAILED_REOPEN_BUDGET_EXHAUSTED: &str = "reopen_budget_exhausted";

/// `task_failed.cause` for an attempt that ended in a runtime error no other cause names. The
/// reason is the error's own text.
pub(crate) const TASK_FAILED_RUNTIME_ERROR: &str = "runtime_error";

/// The most `task_failed.reason` carries, in bytes. A driver's error text is provider-supplied
/// and unbounded; the whole of it stays in `out/result.txt`.
pub(crate) const MAX_TASK_FAILURE_REASON_BYTES: usize = 2_000;

/// A task attempt failed, and why. Written once per failing attempt, before its task's
/// `task_end`, by whichever part of the runtime saw the failure.
///
/// `task_end.exit_status` says `failed`; this names the cause from a closed vocabulary (the
/// `TASK_FAILED_*` constants) and carries the reason in words. Never written for a task that
/// ended `ok`, `max_turns_reached`, `spend_ceiling_reached` or `canceled`: those have their own
/// records.
#[derive(Serialize)]
struct TaskFailedEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    /// The task in progress. Absent for the run a launch makes from `input.txt` when no task
    /// arrived, which has no `task_start`.
    #[serde(skip_serializing_if = "Option::is_none")]
    task_id: Option<String>,
    /// The turn that failed, 0-based. Absent when the failure is the task's rather than one
    /// turn's.
    #[serde(skip_serializing_if = "Option::is_none")]
    turn: Option<u32>,
    cause: String,
    /// Capped at [`MAX_TASK_FAILURE_REASON_BYTES`] on a character boundary.
    reason: String,
}

/// `task_reopened.attempt_context` for a next attempt that continues the task's own
/// conversation, with the hook's feedback as the one message it adds.
pub(crate) const REOPEN_CONTEXT_CONTINUED: &str = "continued";

/// `task_reopened.attempt_context` for a next attempt that starts from a fresh context built
/// from the rewritten `task.md`, because the previous attempt left no conversation to continue.
pub(crate) const REOPEN_CONTEXT_RESTARTED: &str = "restarted";

/// One `on-task-end` hook reopened the task: its agent loop is about to run again with
/// the hook's feedback. Written between two agent-loop attempts for the same task, so
/// `mur trace show` can show which hook drove each reopen, why, and what the next attempt
/// starts from.
#[derive(Serialize)]
struct TaskReopenedEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: String,
    /// Manifest name of the `on-task-end` hook that returned `reopen-task`.
    hook_name: String,
    /// Feedback text the hook asked to inject into the reopened task content.
    reason: String,
    /// 1-based ordinal of this reopen within the task (first reopen = 1).
    reopen_number: u32,
    /// [`REOPEN_CONTEXT_CONTINUED`] or [`REOPEN_CONTEXT_RESTARTED`].
    attempt_context: &'static str,
    /// The turns the next attempt is handed: `inference.max_turns` less every turn the task's
    /// attempts have spent so far.
    turns_remaining: u32,
}

/// A spend ceiling refused a driver call before it was sent. No `inference` line accompanies it,
/// because no call was made.
#[derive(Serialize)]
struct SpendCeilingReachedEvent<'a> {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    turn: u32,
    task_id: Option<String>,
    #[serde(flatten)]
    refusal: &'a crate::spend::SpendRefusal,
    /// `hook:<name>` for a hook's `run-inference` call; absent for an agent-loop turn.
    #[serde(skip_serializing_if = "Option::is_none")]
    origin: Option<&'a str>,
}

/// What a `transport: process` session spawned, and what the driver planned for it.
///
/// Separate from [`HarnessStartEvent`] so the runner builds one value rather than passing a dozen
/// positional arguments. Flattened into the event, so the record has no nested object.
#[derive(Serialize)]
pub(crate) struct HarnessStart {
    /// The process driver artifact and its version.
    pub driver: String,
    pub driver_version: String,
    /// What the driver said it drives, and the executable actually spawned.
    pub harness: String,
    /// The absolute path spawned — not the name the manifest or the driver asked for.
    pub binary: String,
    /// `inference.command`, or the process driver's `describe()`.
    pub binary_source: String,
    /// What the harness printed for its version arguments, or `null` when it could not be read.
    pub harness_version: Option<String>,
    /// Whether [`Self::harness_version`] matched one of the driver's `tested-versions`.
    pub version_tested: bool,
    /// The harness session id the runtime minted for this run, and whether it started or resumed.
    pub harness_session_id: String,
    pub session_mode: String,
    /// How many arguments the plan carried. Never the arguments themselves: a driver is free to
    /// put the bridge's bearer token in one.
    pub args_count: usize,
    /// The names of the variables actually delivered, sorted. Never their values.
    pub env_names: Vec<String>,
    /// The names of the files written into the run's private directory. Never their contents.
    pub files: Vec<String>,
    /// The capsule tool names the bridge offered, bare. Empty when no bridge was bound.
    pub bridge_tools: Vec<String>,
    /// How many bytes were written to the harness's stdin, and whether it stayed open after.
    pub stdin_bytes: usize,
    pub keep_stdin_open: bool,
}

#[derive(Serialize)]
struct HarnessStartEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: Option<String>,
    #[serde(flatten)]
    start: HarnessStart,
}

/// A `W-*` warning the run raised. The same text goes to stderr; this is the durable copy, so a
/// run read back later still says what it warned about.
#[derive(Serialize)]
struct HarnessWarningEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: Option<String>,
    code: String,
    message: String,
}

/// The session the harness itself reports, which is not this session: the runtime's `session_id`
/// names the capsule run, `harness_session_id` names the conversation inside the harness.
#[derive(Serialize)]
struct HarnessSessionEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: Option<String>,
    harness_session_id: String,
    /// How the harness says it is billing. Anything but `subscription` is also warned about.
    auth: String,
    /// The model the harness says it is running, when it says.
    model: Option<String>,
}

/// The harness session a person asked this capsule to drop, recorded because forgetting is not
/// the same as never having known: the turn that follows starts a new conversation under a context
/// that had one, and this is what says so.
#[derive(Serialize)]
struct HarnessSessionForgottenEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: Option<String>,
    /// The context whose entry was dropped.
    context_id: String,
    /// The id that was dropped, which the harness is no longer asked to continue.
    harness_session_id: String,
    /// Which surface carried the request: `cli` or `a2a`.
    requested_by: String,
}

#[derive(Serialize)]
struct HarnessRetryEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: Option<String>,
    attempt: u32,
    reason: String,
}

#[derive(Serialize)]
struct HarnessNoteEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: Option<String>,
    text: String,
}

#[derive(Serialize)]
struct HarnessFailedEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: Option<String>,
    /// The driver's `failure-kind`, spelled as the WIT spells it, or `max-turns` from the runtime.
    kind: String,
    message: String,
    /// `harness` when the harness reported it, `runtime` when this runtime ended the turn.
    source: String,
}

/// How the runtime tried to stop a harness a person cancelled the task of.
#[derive(Serialize)]
struct HarnessInterruptEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: Option<String>,
    /// The driver's `interrupt-method`, spelled as the WIT spells it.
    method: String,
    /// Whether a graceful interrupt actually reached the harness.
    delivered: bool,
    /// How long the harness has to end on its own before it is killed. `0` when it is killed at
    /// once, which is every ending where `delivered` is false.
    grace_ms: u64,
}

/// How one spawned harness ended. Written once per spawn, on every path out of the run.
#[derive(Serialize)]
pub(crate) struct HarnessExit {
    /// The process's exit code, or `null` when a signal ended it.
    pub code: Option<i32>,
    /// The signal that ended the process, or `null` when it exited on its own.
    pub signal: Option<i32>,
    /// What ended the run: `terminal`, `eof`, `inactivity`, `max_turns`, `driver_error` or
    /// `canceled`.
    pub cause: String,
    /// Whether the runtime had to kill the process rather than wait for it.
    pub killed: bool,
    /// Wall time from spawn to reaped.
    pub duration_ms: u64,
}

#[derive(Serialize)]
struct HarnessExitEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    task_id: Option<String>,
    #[serde(flatten)]
    exit: HarnessExit,
}

/// A policy hook refused a call before it ran. The one record that says a call the model asked
/// for never happened — there is no `tool_call` or `shell` line for a denied call, because
/// nothing ran.
#[derive(Serialize)]
struct CallDeniedEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    turn: u32,
    /// The gated lifecycle function whose decision point refused: `"on-shell"` or
    /// `"on-tool-call"`.
    event: String,
    /// Manifest name of the policy hook that refused.
    hook_name: String,
    /// What was refused: the resolved executable path for a shell call, the tool name
    /// otherwise.
    target: String,
    /// The hook's own reason, or the runtime's when the hook supplied none it could use.
    reason: String,
}

/// The manifest's own read-only rule refused a call before it ran. Distinct from
/// [`CallDeniedEvent`] because it names a different authority and a different subject: no hook was
/// asked, and what it carries is the path, the declared rule and the evidence that identified the
/// call as a write.
#[derive(Serialize)]
struct ProtectedPathDeniedEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    turn: u32,
    /// Which dispatch path was refused: `"shell"` or `"tool"`.
    call: String,
    /// What was refused: the resolved executable path for a shell call, the tool name otherwise.
    target: String,
    /// The resolved workdir-relative path, never the string the model typed, so two spellings of
    /// one file produce one comparable record.
    path: String,
    /// The `capabilities.filesystem.read_only` entry that covers `path`, as declared.
    rule: String,
    /// What identified the call as a write — a redirection operator, a write-verb argument
    /// position, or the tool-input key pairing.
    signal: String,
    /// The same text the model was given, so the trace and the agent agree on why.
    reason: String,
}

#[derive(Serialize)]
struct HookDispatchErrorEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    hook_name: String,
    event: String,
    arm: String,
}

#[derive(Serialize)]
struct ResourceListEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    /// `exports.files.root` verbatim, or `""` when the capsule declares no export — a refusal on
    /// an undeclared plane is still a complete record of what was asked for.
    root: String,
    entry_count: usize,
    total_bytes: u64,
    /// Which completed turn the listing is as of.
    generation: u64,
    containment_achieved: ContainmentClass,
    /// `"ok"`, or the wire error code the request was refused with.
    outcome: String,
    /// `null` on `ok`, one short sentence otherwise.
    reason: Option<String>,
}

#[derive(Serialize)]
struct ResourceReadEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    /// The requested path *after* percent-decoding, so `%2e%2e%2f` and `../` read as the same
    /// attempt rather than as two unrelated findings.
    path: String,
    outcome: String,
    /// `null` on every non-`ok` outcome.
    bytes: Option<u64>,
    /// `null` on every non-`ok` outcome.
    sha256: Option<String>,
    generation: u64,
    containment_achieved: ContainmentClass,
    reason: Option<String>,
}

/// One `share-file` call, served or refused, written at the moment of the mint.
#[derive(Serialize)]
struct PeerHandleMintEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    /// `null` on every non-`ok` outcome: a mint that was refused produced no token, so there is
    /// no handle to identify.
    handle_id: Option<String>,
    /// The path relative to `exports.peer_files.root`, canonicalised, or the path the agent asked
    /// for when the mint was refused. Never a host path and never the workdir-relative form.
    path: String,
    audience: String,
    /// `null` on every non-`ok` outcome.
    expires_at_ms: Option<u64>,
    outcome: String,
    reason: Option<String>,
}

/// One redeem against `/resources/peer/<handle>`, written by the listener concurrently with a
/// possibly-running task.
#[derive(Serialize)]
struct PeerHandleRedeemEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    handle_id: String,
    /// `null` until the MAC has verified. A payload that failed the MAC is caller-controlled and
    /// must not be copied into this capsule's own audit record as if it were fact.
    path: Option<String>,
    /// The runtime's own current counter, never a value taken from the token — so it is always
    /// present and always true.
    generation: u64,
    /// The `x-murmur-audience` header exactly as asserted, or `null` when none was sent.
    audience_asserted: Option<String>,
    /// `null` on every non-`ok` outcome.
    bytes: Option<u64>,
    /// `null` on every non-`ok` outcome.
    sha256: Option<String>,
    outcome: String,
    reason: Option<String>,
}

/// One `fetch-peer-file` call on the ingesting side, served or refused.
#[derive(Serialize)]
struct PeerFileFetchEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    peer: String,
    handle_id: String,
    /// Where the bytes landed, relative to the accessible workdir. `null` on every non-`ok`
    /// outcome.
    stored_path: Option<String>,
    bytes: Option<u64>,
    sha256: Option<String>,
    outcome: String,
    reason: Option<String>,
}

/// One launched child, written the moment it is up.
///
/// Two surfaces launch children — the agent's `delegate-task` tool and a plan's `capsule` step —
/// and both write this line, off the same struct and off the same session node. Which one made a
/// given delegation is not recorded, because it is not a fact about the delegation.
///
/// Separate from the `delegation` line below because the two answer different questions and land
/// at different times: this one reaches disk while the delegation is still in flight, so a child
/// that hangs, crashes or is ended at its deadline is attributable from the parent's side whatever
/// happens next. A delegation the daemon refused writes none of these — nothing was launched.
#[derive(Serialize)]
struct DelegationStartEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    /// The `dlg_` id the launcher minted, character-identical to the one on this delegation's
    /// terminal `delegation` line and in the child's own `session_start`.
    delegation_id: String,
    capsule: String,
    version: String,
    /// The session id the child's runtime minted for itself.
    child_session_id: String,
    /// The child's directory, relative to this capsule's accessible workdir — join the two to
    /// reach the child's own `trace.jsonl`.
    child_workdir: String,
}

impl DelegationStartEvent {
    /// The one place this record is built, so the two appenders that write it cannot drift into
    /// two shapes. Which surface launched the child is not an argument, because it is not a field.
    fn new(
        session_id: &str,
        session_event_id: &str,
        delegation_id: &str,
        capsule: &str,
        version: &str,
        child_session_id: &str,
        child_workdir: &str,
    ) -> Self {
        Self {
            event_type: "delegation_start",
            event_id: new_event_id(),
            parent_id: Some(session_event_id.to_string()),
            session_id: session_id.to_string(),
            timestamp: timestamp_ms(),
            delegation_id: delegation_id.to_string(),
            capsule: capsule.to_string(),
            version: version.to_string(),
            child_session_id: child_session_id.to_string(),
            child_workdir: child_workdir.to_string(),
        }
    }
}

/// One delegation, written when it ends.
///
/// One line per call whatever happened, including a call the daemon refused: a delegation that was
/// asked for and not made is as much a fact of the run as one that produced an answer. When it is
/// written depends on how far the delegation got — a `delegate-task` call that started a child
/// writes it as that child's completion arrives, out of the child's own `completion.json`; one that
/// never started a child writes it within the call. A plan `capsule` step waits for its child's
/// answer, so it always writes this line within the step. Carries neither the task text nor the
/// child's answer — both are the agent's own conversation, which the `tool_call` line for the
/// same call already records under the session's `trace.capture` setting.
#[derive(Serialize)]
struct DelegationEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    /// The sub-capsule the agent named, and the exact version it named.
    capsule: String,
    version: String,
    /// The `dlg_` id naming this delegation. `null` on a `refused` outcome, which made none.
    delegation_id: Option<String>,
    /// The child's own session id, so its trace is findable. `null` when no child ran.
    child_session_id: Option<String>,
    /// How long the child ran, for a delegation that started; how long the call took, for one that
    /// never started.
    duration_ms: u64,
    /// How the delegation ended. Which vocabulary it is drawn from depends on whether the
    /// delegating call waited for the answer itself:
    ///
    /// - A delegation whose outcome arrived later as a completion reads `ok`, `error`, `crashed`
    ///   or `terminated` from [`crate::delegation::DelegationStatus`], or `unknown` when the
    ///   completion left no readable file.
    /// - One the caller waited out reads `completed` or `timed_out` from
    ///   [`crate::delegation_plane::DelegationStatus`].
    /// - One that never started reads `failed` or `refused`, from that same enum, either way.
    outcome: String,
    /// `null` where nothing needs saying. Otherwise the completion's `detail`, or the same
    /// sentence the model was given.
    reason: Option<String>,
}

impl DelegationEvent {
    /// The one place this record is built, so the two appenders that write it cannot drift into
    /// two shapes. Which surface launched the child is not an argument, because it is not a field.
    #[allow(clippy::too_many_arguments)]
    fn new(
        session_id: &str,
        session_event_id: &str,
        capsule: &str,
        version: &str,
        delegation_id: Option<String>,
        child_session_id: Option<String>,
        duration_ms: u64,
        outcome: &str,
        reason: Option<String>,
    ) -> Self {
        Self {
            event_type: "delegation",
            event_id: new_event_id(),
            parent_id: Some(session_event_id.to_string()),
            session_id: session_id.to_string(),
            timestamp: timestamp_ms(),
            capsule: capsule.to_string(),
            version: version.to_string(),
            delegation_id,
            child_session_id,
            duration_ms,
            outcome: outcome.to_string(),
            reason,
        }
    }
}

// ── TraceWriter impl ─────────────────────────────────────────────────────────

impl TraceWriter {
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn open(
        workdir: &Path,
        session_id: String,
        capsule_name: String,
        capsule_version: String,
        model: String,
        capabilities: Vec<String>,
        effective_grants: ScopeReport,
        capture: TraceCapture,
        system_prompt: Option<String>,
        system_prompt_overridden: bool,
        resumed_from: Option<String>,
        context_id: Option<String>,
        spawned_by: Option<String>,
        spawn_delegation_id: Option<String>,
    ) -> std::io::Result<Self> {
        let file = OpenOptions::new()
            .create(true)
            .append(true)
            .open(workdir.join("trace.jsonl"))
            .await?;
        // `"cli"` wins over `"none"` when the override cleared the prompt: the operator passed
        // `--system-prompt ""`, which is a decision worth recording, not an absence.
        let system_prompt_source = match (system_prompt_overridden, system_prompt.is_some()) {
            (true, _) => "cli",
            (false, true) => "manifest",
            (false, false) => "none",
        };
        let system_prompt_sha256 = system_prompt
            .as_ref()
            .map(|prompt| murmur_artifact::sha256_hex(prompt.as_bytes()));
        let blobs = BlobStore::new(workdir);
        // The prompt the operator wrote, kept under the hash `session_start` already carried.
        // Written here rather than at `session_start` because a session writes that frame once
        // per task and the blob is the same one every time.
        if let (true, Some(prompt), Some(sha)) = (
            capture.captures_content(),
            system_prompt.as_ref(),
            system_prompt_sha256.as_deref(),
        ) {
            blobs.put(sha, prompt.as_bytes()).await?;
        }
        Ok(Self {
            writer: BufWriter::new(file),
            session_id,
            capsule_name,
            capsule_version,
            model,
            capabilities,
            effective_grants,
            capture,
            blobs,
            system_prompt_source,
            system_prompt_sha256,
            credential_source: "none",
            gateways: Vec::new(),
            control: None,
            inference_choices: Vec::new(),
            tool_refresh: None,
            runtime_artifacts: Vec::new(),
            resumed_from,
            context_id,
            spawned_by,
            spawn_delegation_id,
            session_start_time: Instant::now(),
            session_event_id: new_event_id(),
            session_started: false,
            session_ended: false,
            task_event_id: None,
            turn_event_id: None,
            total_turns: 0,
            total_input_tokens: 0,
            total_output_tokens: 0,
            total_tool_calls: 0,
            total_shell_calls: 0,
            task_turns: 0,
            task_input_tokens: 0,
            task_output_tokens: 0,
            task_tool_calls: 0,
            task_shell_calls: 0,
            task_start_instant: None,
            active_task_id: None,
            task_failures_written: 0,
            last_task_failure: None,
            max_session_tokens: None,
            machine_tokens_per_day: None,
        })
    }

    /// The spend ceilings this session runs under, written into every later `session_start`.
    pub(crate) fn set_spend_ceilings(
        &mut self,
        max_session_tokens: Option<u64>,
        machine_tokens_per_day: Option<u64>,
    ) {
        self.max_session_tokens = max_session_tokens;
        self.machine_tokens_per_day = machine_tokens_per_day;
    }

    /// The session node of this writer's event tree — the `event_id` its `session_start` will
    /// carry, available before that line is written so [`ResourceTraceAppender`] can be opened
    /// against the same node.
    pub(crate) fn session_event_id(&self) -> &str {
        &self.session_event_id
    }

    /// Whether this session's `inference` records carry content hashes — everything but
    /// `trace.capture: none`. The agent loop asks before splitting a payload into a
    /// [`WireCapture`], so `none` costs no serialization at all.
    pub(crate) fn captures_wire(&self) -> bool {
        self.capture.captures_hashes()
    }

    /// Parent for a launch-scoped event: the session node, or `None` on a writer that has not
    /// written `session_start`. A trace line must never name a parent that has no line behind
    /// it, and the script-capsule `a2a_send` drain opens a writer over a file with no session
    /// frame at all.
    fn session_parent(&self) -> Option<String> {
        self.session_started.then(|| self.session_event_id.clone())
    }

    /// Parent for a task-scoped event: the task node, falling back to the session node when no
    /// task is active.
    fn task_parent(&self) -> Option<String> {
        self.task_event_id.clone().or_else(|| self.session_parent())
    }

    /// Parent for a turn-scoped event: the turn node, falling back to the task node and then to
    /// the session node. The fallbacks are what keeps a tool call written outside any turn — a
    /// hook's, or the process transport's, before its first `inference` — attached to the tree.
    fn turn_parent(&self) -> Option<String> {
        self.turn_event_id.clone().or_else(|| self.task_parent())
    }

    /// Records where the inference credential came from, as
    /// [`crate::gateway_credential::CredentialSource::trace_name`] names it. `"none"` until set.
    pub(crate) fn set_credential_source(&mut self, source: &'static str) {
        self.credential_source = source;
    }

    /// Records the session's credential gateways for `session_start.gateways`.
    pub(crate) fn set_gateways(&mut self, gateways: Vec<SessionGateway>) {
        self.gateways = gateways;
    }

    /// Records what `control:` declares for `session_start.control`.
    pub(crate) fn set_control(&mut self, control: Option<SessionControl>) {
        self.control = control;
    }

    /// Records the driver choices for `session_start.inference_choices`. Left empty, the key is
    /// not written.
    pub(crate) fn set_inference_choices(&mut self, choices: Vec<SessionInferenceChoice>) {
        self.inference_choices = choices;
    }

    /// Records the runtime-origin entries of `installed` for `session_start.runtime_artifacts`,
    /// in `installed`'s order. Operator-declared entries are left out.
    pub(crate) fn set_runtime_artifacts(&mut self, installed: &[InstalledArtifactSummary]) {
        self.runtime_artifacts = installed
            .iter()
            .filter_map(|artifact| match &artifact.origin {
                LockOrigin::Runtime { session } => Some(RuntimeArtifact {
                    name: artifact.name.clone(),
                    version: artifact.version.clone(),
                    origin: artifact.origin.as_str(),
                    session: session.clone(),
                    trust: crate::origin::artifact_trust(&artifact.origin).as_str(),
                }),
                LockOrigin::Operator => None,
            })
            .collect();
    }

    /// Records that this inference call is the first to use a setting value a controller set:
    /// `turn` is the call's zero-based turn, the same number its `inference` line carries, and
    /// `change_id` the `event_id` of the `control_change` that set `value`.
    ///
    /// Written before the call is sent, so it precedes the turn's `inference` line and hangs off
    /// the task.
    pub(crate) async fn write_control_applied(
        &mut self,
        turn: u32,
        name: &str,
        value: Value,
        change_id: &str,
    ) -> std::io::Result<()> {
        let event = ControlAppliedEvent {
            event_type: "control_applied",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            turn,
            task_id: self.active_task_id.clone(),
            name,
            value,
            change_id,
        };
        self.write_event(&event).await
    }

    /// Records `inference.tool_refresh` for `session_start.tool_refresh`: the wire name under
    /// `transport: http`, `None` under `transport: process`.
    pub(crate) fn set_tool_refresh(&mut self, tool_refresh: Option<&'static str>) {
        self.tool_refresh = tool_refresh;
    }

    /// Records that the tool array the http agent loop sends changed before this call:
    /// `trigger` is why (`"immediate"` or `"compaction"`), `added` and `removed` the sorted
    /// names that entered and left it, and `tools` every name it now offers, in array order.
    ///
    /// Written before the call is sent, so it precedes the turn's `inference` line and hangs off
    /// the task, as `control_applied` does.
    pub(crate) async fn write_tools_refreshed(
        &mut self,
        turn: u32,
        trigger: &str,
        added: &[String],
        removed: &[String],
        tools: &[String],
    ) -> std::io::Result<()> {
        let event = ToolsRefreshedEvent {
            event_type: "tools_refreshed",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            turn,
            task_id: self.active_task_id.clone(),
            trigger,
            added,
            removed,
            tools,
        };
        self.write_event(&event).await
    }

    pub(crate) async fn write_session_start(
        &mut self,
        max_turns: u32,
        tools_declared: Vec<String>,
    ) -> std::io::Result<()> {
        let event = SessionStartEvent {
            event_type: "session_start",
            event_id: self.session_event_id.clone(),
            parent_id: None,
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            capsule_name: self.capsule_name.clone(),
            capsule_version: self.capsule_version.clone(),
            model: self.model.clone(),
            max_turns,
            max_session_tokens: self.max_session_tokens,
            machine_tokens_per_day: self.machine_tokens_per_day,
            capabilities: self.capabilities.clone(),
            tools_declared,
            containment_declared: self.effective_grants.declared_containment,
            containment_achieved: self.effective_grants.achieved_containment,
            workdir_exec: self.effective_grants.workdir_exec,
            userns_grant: self.effective_grants.userns_grant,
            effective_grants: self.effective_grants.clone(),
            system_prompt_source: self.system_prompt_source,
            system_prompt_sha256: self.system_prompt_sha256.clone(),
            credential_source: self.credential_source,
            gateways: self.gateways.clone(),
            control: self.control.clone(),
            inference_choices: self.inference_choices.clone(),
            tool_refresh: self.tool_refresh,
            runtime_artifacts: self.runtime_artifacts.clone(),
            resumed_from: self.resumed_from.clone(),
            context_id: self.context_id.clone(),
            spawned_by: self.spawned_by.clone(),
            delegation_id: self.spawn_delegation_id.clone(),
        };
        self.write_event(&event).await?;
        self.session_started = true;
        Ok(())
    }

    /// `input_tokens`/`output_tokens` are what the session and task totals accumulate: the
    /// runtime's own tiktoken estimates under `transport: http`, and the counts the harness's
    /// driver reported under `transport: process`, where the runtime measures nothing itself.
    /// `None` is a turn nothing counted, which only a process driver reporting no usage
    /// produces, and it adds nothing to the totals.
    ///
    /// `usage` is written verbatim beside them and accumulated into nothing. Under `http` it is
    /// the provider's own report for the same call; under `process` it carries only the cache and
    /// thinking counts, because the harness's input and output are already the two above.
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn write_inference(
        &mut self,
        turn: u32,
        input_tokens: Option<u64>,
        output_tokens: Option<u64>,
        decision: String,
        stop_reason: Option<&str>,
        tool_name: Option<String>,
        origin: Option<&InferenceOrigin>,
        usage: Option<&DriverUsage>,
        message_ids: Vec<String>,
        wire: Option<&WireCapture>,
    ) -> std::io::Result<()> {
        self.write_inference_event(
            turn,
            input_tokens,
            output_tokens,
            decision,
            stop_reason,
            tool_name,
            origin,
            usage,
            message_ids,
            wire,
            None,
        )
        .await
    }

    /// [`Self::write_inference`] for an agent-loop turn, naming the driver choice that served it
    /// as `driver_choice` and `model`. `choice` is `None` on a capsule with no alternates, whose
    /// line is exactly what [`Self::write_inference`] writes.
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn write_turn_inference(
        &mut self,
        turn: u32,
        input_tokens: Option<u64>,
        output_tokens: Option<u64>,
        decision: String,
        stop_reason: Option<&str>,
        tool_name: Option<String>,
        usage: Option<&DriverUsage>,
        message_ids: Vec<String>,
        wire: Option<&WireCapture>,
        choice: Option<&crate::driver_choice::DriverChoice>,
    ) -> std::io::Result<()> {
        self.write_inference_event(
            turn,
            input_tokens,
            output_tokens,
            decision,
            stop_reason,
            tool_name,
            None,
            usage,
            message_ids,
            wire,
            choice,
        )
        .await
    }

    #[allow(clippy::too_many_arguments)]
    async fn write_inference_event(
        &mut self,
        turn: u32,
        input_tokens: Option<u64>,
        output_tokens: Option<u64>,
        decision: String,
        stop_reason: Option<&str>,
        tool_name: Option<String>,
        origin: Option<&InferenceOrigin>,
        usage: Option<&DriverUsage>,
        message_ids: Vec<String>,
        wire: Option<&WireCapture>,
        choice: Option<&crate::driver_choice::DriverChoice>,
    ) -> std::io::Result<()> {
        let event_id = new_event_id();
        // The agent loop's own inference *is* the turn node — there is no separate turn line —
        // so it hangs off the task and everything the turn goes on to produce hangs off it. A
        // hook's `run-inference` (`origin: Some(_)`) is one of those products, not a new turn.
        let parent_id = if origin.is_none() {
            let parent = self.task_parent();
            self.turn_event_id = Some(event_id.clone());
            parent
        } else {
            self.turn_parent()
        };
        let (system_sha, tools_sha, response_sha, message_shas) =
            match wire.filter(|_| self.capture.captures_hashes()) {
                Some(wire) => self.record_wire(wire).await?,
                None => (None, None, None, Vec::new()),
            };
        let event = InferenceEvent {
            event_type: "inference",
            event_id,
            parent_id,
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            turn,
            task_id: self.active_task_id.clone(),
            input_tokens,
            output_tokens,
            decision,
            stop_reason: stop_reason.map(str::to_string),
            tool_name,
            input_tokens_actual: usage.and_then(|u| u.input_tokens),
            output_tokens_actual: usage.and_then(|u| u.output_tokens),
            cached_tokens: usage.and_then(|u| u.cached_tokens),
            cache_write_tokens: usage.and_then(|u| u.cache_write_tokens),
            thinking_tokens: usage.and_then(|u| u.thinking_tokens),
            origin: origin.map(|o| o.source.clone()),
            driver_choice: choice.map(|choice| choice.name.clone()),
            model: origin
                .map(|o| o.model.clone())
                .or_else(|| choice.map(|choice| choice.model.clone())),
            message_ids,
            system_sha,
            tools_sha,
            response_sha,
            message_shas,
        };
        self.write_event(&event).await?;
        // A turn nothing counted contributes nothing to the totals. The totals are plain `u64`
        // because a session's spend is a sum, and the sum of no measurements is zero; the
        // absent/zero distinction lives on the record itself.
        let (input_tokens, output_tokens) = (input_tokens.unwrap_or(0), output_tokens.unwrap_or(0));
        self.total_turns = self.total_turns.saturating_add(1);
        self.total_input_tokens = self.total_input_tokens.saturating_add(input_tokens);
        self.total_output_tokens = self.total_output_tokens.saturating_add(output_tokens);
        self.task_turns = self.task_turns.saturating_add(1);
        self.task_input_tokens = self.task_input_tokens.saturating_add(input_tokens);
        self.task_output_tokens = self.task_output_tokens.saturating_add(output_tokens);
        Ok(())
    }

    /// Hash the four bodies of one request and, under [`TraceCapture::Content`], store each
    /// behind its hash.
    ///
    /// The hash is computed in every capture mode that reaches here; only the blob write is
    /// gated, so `meta` and `content` name the same bytes and differ solely in whether those
    /// bytes are also on disk.
    async fn record_wire(
        &mut self,
        wire: &WireCapture,
    ) -> std::io::Result<(Option<String>, Option<String>, Option<String>, Vec<String>)> {
        let system_sha = murmur_artifact::sha256_hex(&wire.system);
        let tools_sha = murmur_artifact::sha256_hex(&wire.tools);
        let response_sha = murmur_artifact::sha256_hex(&wire.response);
        let message_shas: Vec<String> = wire
            .messages
            .iter()
            .map(|message| murmur_artifact::sha256_hex(message))
            .collect();

        if self.capture.captures_content() {
            // Verbatim, unredacted. `write_tool_call` redacts peer handles out of its own summary
            // fields; a blob is not a summary, and redacting one would make its hash name bytes
            // no file holds — destroying the only property this store exists to provide.
            self.blobs.put(&system_sha, &wire.system).await?;
            self.blobs.put(&tools_sha, &wire.tools).await?;
            self.blobs.put(&response_sha, &wire.response).await?;
            for (sha, body) in message_shas.iter().zip(&wire.messages) {
                self.blobs.put(sha, body).await?;
            }
        }

        Ok((
            Some(system_sha),
            Some(tools_sha),
            Some(response_sha),
            message_shas,
        ))
    }

    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn write_tool_call(
        &mut self,
        turn: u32,
        tool_name: String,
        tool_call_id: Option<String>,
        mut input: Value,
        input_bytes: u64,
        output: &str,
        output_bytes: u64,
        duration_ms: u64,
        status: String,
        state_effect: Option<String>,
        resource_id: Option<String>,
    ) -> std::io::Result<()> {
        // A peer handle reaches this event by an ordinary route — it is an argument the model
        // passed to `fetch-peer-file`, and every call's arguments are recorded. It is also a
        // credential, and the trace is durable, so the two must not meet: the `handle_id` goes in
        // instead, which is what correlates the record with the mint and the redeem anyway.
        crate::peer_handoff::redact_handles_in_json(&mut input);
        let event = ToolCallEvent {
            event_type: "tool_call",
            event_id: new_event_id(),
            parent_id: self.turn_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            turn,
            task_id: self.active_task_id.clone(),
            tool_name,
            // An absent or empty provider id records as `null`: `""` would read as an id the
            // provider actually issued.
            tool_call_id: tool_call_id.filter(|id| !id.is_empty()),
            input,
            input_bytes,
            output: self
                .capture
                .captures_content()
                .then(|| crate::peer_handoff::redact_handle_tokens(output).into_owned()),
            output_bytes,
            duration_ms,
            status,
            state_effect,
            resource_id,
        };
        self.write_event(&event).await?;
        self.total_tool_calls = self.total_tool_calls.saturating_add(1);
        self.task_tool_calls = self.task_tool_calls.saturating_add(1);
        Ok(())
    }

    /// Records a skill invocation as a `skill_call` event. Skill calls are NOT counted in
    /// `total_tool_calls` — they appear in a separate `── Skill calls ──` section in `mur trace show`.
    ///
    /// `origin` is the skill's `murmur.lock` origin; the record's `trust` is derived from it
    /// through [`crate::origin::artifact_trust`], the same rule that decides the skill's fence.
    pub(crate) async fn write_skill_call(
        &mut self,
        turn: u32,
        skill_name: String,
        origin: &LockOrigin,
        output_bytes: u64,
        duration_ms: u64,
        status: String,
    ) -> std::io::Result<()> {
        let event = SkillCallEvent {
            event_type: "skill_call",
            event_id: new_event_id(),
            parent_id: self.turn_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            turn,
            task_id: self.active_task_id.clone(),
            skill_name,
            output_bytes,
            duration_ms,
            status,
            origin: origin.as_str(),
            trust: crate::origin::artifact_trust(origin).as_str(),
        };
        self.write_event(&event).await
        // Intentionally does not increment total_tool_calls or task_tool_calls.
    }

    // One positional parameter per JSONL column, as every other `write_*` here does; the
    // sibling `write_tool_call` carries the same allowance.
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn write_shell(
        &mut self,
        turn: u32,
        binary: String,
        command: String,
        exit_code: i32,
        stdout_bytes: u64,
        stderr_bytes: u64,
        duration_ms: u64,
        resource_limit: Option<String>,
    ) -> std::io::Result<()> {
        let event = ShellEvent {
            event_type: "shell",
            event_id: new_event_id(),
            parent_id: self.turn_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            turn,
            task_id: self.active_task_id.clone(),
            binary,
            command,
            exit_code,
            stdout_bytes,
            stderr_bytes,
            duration_ms,
            resource_limit,
        };
        self.write_event(&event).await?;
        self.total_shell_calls = self.total_shell_calls.saturating_add(1);
        self.task_shell_calls = self.task_shell_calls.saturating_add(1);
        Ok(())
    }

    /// A command demoted to the background, written from the agent loop at the moment of
    /// demotion.
    ///
    /// Counts as this session's shell call, exactly as a foreground one does — the pair with
    /// [`Self::write_shell_completed`], which deliberately does not count, so `mur trace show`
    /// reports each shell command once whichever way it ran.
    pub(crate) async fn write_shell_detached(
        &mut self,
        turn: u32,
        work_id: &str,
        binary: &str,
        command: &str,
        grace_ms: u64,
    ) -> std::io::Result<()> {
        let event = ShellDetachedEvent {
            event_type: "shell_detached",
            event_id: new_event_id(),
            parent_id: self.turn_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            turn,
            task_id: self.active_task_id.clone(),
            work_id: work_id.to_string(),
            binary: binary.to_string(),
            command: command.to_string(),
            grace_ms,
        };
        self.write_event(&event).await?;
        self.total_shell_calls = self.total_shell_calls.saturating_add(1);
        self.task_shell_calls = self.task_shell_calls.saturating_add(1);
        Ok(())
    }

    /// A detached command finishing, written by the task loop as it turns the completion into a
    /// task. `completion_task_id` is that task's id, so the two records join.
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn write_shell_completed(
        &mut self,
        work_id: &str,
        binary: &str,
        command: &str,
        exit_code: i32,
        duration_ms: u64,
        output_path: &str,
        output_bytes: u64,
        resource_limit: Option<String>,
        status: &str,
        completion_task_id: &str,
    ) -> std::io::Result<()> {
        let event = ShellCompletedEvent {
            event_type: "shell_completed",
            event_id: new_event_id(),
            parent_id: self.session_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            work_id: work_id.to_string(),
            binary: binary.to_string(),
            command: command.to_string(),
            exit_code,
            duration_ms,
            output_path: output_path.to_string(),
            output_bytes,
            resource_limit,
            status: status.to_string(),
            completion_task_id: completion_task_id.to_string(),
        };
        self.write_event(&event).await
    }

    /// A detached command the session ended without waiting for.
    ///
    /// `recovered` carries the exit code, the output path and that file's byte count for a command
    /// that finished during teardown, and is `None` for one still running. The three travel as one
    /// argument because they exist together: a command with an exit code has a written log, and
    /// one without has neither.
    pub(crate) async fn write_shell_abandoned(
        &mut self,
        work_id: &str,
        binary: &str,
        command: &str,
        running_ms: u64,
        recovered: Option<(i32, &str, u64)>,
    ) -> std::io::Result<()> {
        let event = ShellAbandonedEvent {
            event_type: "shell_abandoned",
            event_id: new_event_id(),
            parent_id: self.session_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            work_id: work_id.to_string(),
            binary: binary.to_string(),
            command: command.to_string(),
            running_ms,
            exit_code: recovered.map(|(exit_code, _, _)| exit_code),
            output_path: recovered.map(|(_, path, _)| path.to_string()),
            output_bytes: recovered.map(|(_, _, bytes)| bytes),
        };
        self.write_event(&event).await
    }

    /// A demotion the trace could not record, written where the failure was seen.
    pub(crate) async fn write_shell_detach_unrecorded(
        &mut self,
        turn: u32,
        work_id: &str,
        binary: &str,
        reason: &str,
    ) -> std::io::Result<()> {
        let event = ShellDetachUnrecordedEvent {
            event_type: "shell_detach_unrecorded",
            event_id: new_event_id(),
            parent_id: self.turn_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            turn,
            task_id: self.active_task_id.clone(),
            work_id: work_id.to_string(),
            binary: binary.to_string(),
            reason: reason.to_string(),
        };
        self.write_event(&event).await
    }

    pub(crate) async fn write_a2a_task_received(
        &mut self,
        task_id: &str,
        context_id: &str,
        message_id: &str,
        traceparent_from_caller: Option<&str>,
    ) -> std::io::Result<()> {
        let event = A2aTaskReceivedEvent {
            event_type: "a2a_task_received",
            event_id: new_event_id(),
            parent_id: self.session_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: task_id.to_string(),
            context_id: context_id.to_string(),
            message_id: message_id.to_string(),
            traceparent_from_caller: traceparent_from_caller.map(str::to_string),
        };
        self.write_event(&event).await
    }

    pub(crate) async fn write_a2a_send(
        &mut self,
        peer_url: &str,
        message_id: &str,
        task_id: &str,
        context_id: &str,
        traceparent: Option<&str>,
        trust: TrustClass,
    ) -> std::io::Result<()> {
        let event = A2aSendEvent {
            event_type: "a2a_send",
            event_id: new_event_id(),
            parent_id: self.session_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            peer_url: peer_url.to_string(),
            message_id: message_id.to_string(),
            task_id: task_id.to_string(),
            context_id: context_id.to_string(),
            traceparent: traceparent.map(str::to_string),
            trust: trust.as_str().to_string(),
        };
        self.write_event(&event).await
    }

    /// Records a successful `manage.pull()` of `artifact`, as the store summarised it after the
    /// lock was written. Parented to the session node when `session_start` was written through
    /// this writer, and to nothing otherwise — the script path drains these through a writer it
    /// opens after the guest returns, as it does its buffered `a2a_send` lines.
    pub(crate) async fn write_artifact_pulled(
        &mut self,
        artifact: &InstalledArtifactSummary,
    ) -> std::io::Result<()> {
        let session = match &artifact.origin {
            LockOrigin::Runtime { session } => Some(session.clone()),
            LockOrigin::Operator => None,
        };
        let event = ArtifactPulledEvent {
            event_type: "artifact_pulled",
            event_id: new_event_id(),
            parent_id: self.session_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            name: artifact.name.clone(),
            version: artifact.version.clone(),
            runtime: artifact.runtime.as_str(),
            origin: artifact.origin.as_str(),
            session,
            trust: crate::origin::artifact_trust(&artifact.origin).as_str(),
        };
        self.write_event(&event).await
    }

    pub(crate) async fn write_task_start(
        &mut self,
        task_id: &str,
        context_id: &str,
        source: &str,
        provenance: TaskProvenance,
        delegation_id: Option<&str>,
        message_parts_bytes: u64,
    ) -> std::io::Result<()> {
        // Reset per-task counters for this task
        self.task_turns = 0;
        self.task_input_tokens = 0;
        self.task_output_tokens = 0;
        self.task_tool_calls = 0;
        self.task_shell_calls = 0;
        self.task_start_instant = Some(Instant::now());
        self.active_task_id = Some(task_id.to_string());

        let event_id = new_event_id();
        self.task_event_id = Some(event_id.clone());
        self.turn_event_id = None;

        let event = TaskStartEvent {
            event_type: "task_start",
            event_id: event_id.clone(),
            parent_id: self.session_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: task_id.to_string(),
            context_id: context_id.to_string(),
            source: source.to_string(),
            origin: provenance.origin().as_str().to_string(),
            trust: provenance.trust().as_str().to_string(),
            // One rule for every call site, so the recorded lane cannot disagree with the lane
            // `LaneQueue` filed the task under — both read the same provenance.
            lane: TaskLane::for_origin(provenance.origin())
                .as_str()
                .to_string(),
            delegation_id: delegation_id.map(str::to_string),
            message_parts_bytes,
        };
        self.write_event(&event).await
    }

    pub(crate) async fn write_task_end(
        &mut self,
        task_id: &str,
        exit_status: &str,
        reopen_count: u32,
    ) -> std::io::Result<()> {
        let duration_ms = self
            .task_start_instant
            .take()
            .map(|t| t.elapsed().as_millis().try_into().unwrap_or(u64::MAX))
            .unwrap_or(0);

        let event = TaskEndEvent {
            event_type: "task_end",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: task_id.to_string(),
            exit_status: exit_status.to_string(),
            duration_ms,
            turns: self.task_turns,
            input_tokens: self.task_input_tokens,
            output_tokens: self.task_output_tokens,
            tool_calls: self.task_tool_calls,
            shell_calls: self.task_shell_calls,
            reopen_count,
        };
        self.active_task_id = None;
        self.task_event_id = None;
        self.turn_event_id = None;
        self.write_event(&event).await
    }

    /// Record that a person stopped this task, at the point the loop stopped.
    ///
    /// Leaves `active_task_id` and the task's parent id alone: for a running task the terminal
    /// `task_end` still follows and closes the frame. A task cancelled while queued was never
    /// started, so this record hangs off the session.
    pub(crate) async fn write_task_canceled(
        &mut self,
        task_id: &str,
        turn: Option<u32>,
        phase: &str,
        detached_work_ids: Vec<String>,
        delegation_ids: Vec<String>,
    ) -> std::io::Result<()> {
        let event = TaskCanceledEvent {
            event_type: "task_canceled",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: task_id.to_string(),
            turn,
            phase: phase.to_string(),
            detached_work_ids,
            delegation_ids,
        };
        self.write_event(&event).await
    }

    /// Record that a queued task was refused when the session stopped taking work, with `cause`
    /// from the `TASK_REJECTED_*` vocabulary. Written whatever `trace.capture` is.
    ///
    /// Leaves `active_task_id` alone: the refused task never started, so there is no task frame
    /// to open or close, and with no task active the record hangs off the session.
    pub(crate) async fn write_task_rejected(
        &mut self,
        task_id: &str,
        context_id: &str,
        source: &str,
        cause: &str,
        reason: &str,
    ) -> std::io::Result<()> {
        let event = TaskRejectedEvent {
            event_type: "task_rejected",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: task_id.to_string(),
            context_id: context_id.to_string(),
            source: source.to_string(),
            cause: cause.to_string(),
            reason: reason.to_string(),
        };
        self.write_event(&event).await
    }

    /// Record that the task attempt in progress failed, with `cause` from the `TASK_FAILED_*`
    /// vocabulary. Written whatever `trace.capture` is.
    ///
    /// Leaves the task frame open, as [`Self::write_task_canceled`] does: the `task_end` that
    /// follows closes it. The reason, capped, is kept for [`Self::last_task_failure`].
    pub(crate) async fn write_task_failed(
        &mut self,
        turn: Option<u32>,
        cause: &str,
        reason: &str,
    ) -> std::io::Result<()> {
        let reason = crate::agent::truncate_on_char_boundary(reason, MAX_TASK_FAILURE_REASON_BYTES)
            .to_string();
        let event = TaskFailedEvent {
            event_type: "task_failed",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: self.active_task_id.clone(),
            turn,
            cause: cause.to_string(),
            reason: reason.clone(),
        };
        self.task_failures_written += 1;
        self.last_task_failure = Some(reason);
        self.write_event(&event).await
    }

    /// The `reason` of the last `task_failed` line this writer wrote, or `None` before the first.
    pub(crate) fn last_task_failure(&self) -> Option<&str> {
        self.last_task_failure.as_deref()
    }

    /// How many `task_failed` lines this writer has written. Compared across one run to tell
    /// whether that run wrote its own.
    pub(crate) fn task_failures_written(&self) -> u64 {
        self.task_failures_written
    }

    /// Record that an `on-task-end` hook reopened the task. Written by the runtime's
    /// per-task reopen loop between two agent-loop attempts, once per reopen, before
    /// the terminal `task_end` record. `reopen_number` is 1-based.
    pub(crate) async fn write_task_reopened(
        &mut self,
        task_id: &str,
        hook_name: &str,
        reason: &str,
        reopen_number: u32,
        attempt_context: &'static str,
        turns_remaining: u32,
    ) -> std::io::Result<()> {
        let event = TaskReopenedEvent {
            event_type: "task_reopened",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: task_id.to_string(),
            hook_name: hook_name.to_string(),
            reason: reason.to_string(),
            reopen_number,
            attempt_context,
            turns_remaining,
        };
        self.write_event(&event).await
    }

    /// Cumulative turns consumed by the active task so far, across however many
    /// agent-loop attempts have run since the last [`Self::write_task_start`]. Read by
    /// the reopen loop to compute the remaining `max_turns` budget for the next attempt
    /// so reopening never grants turns past the capsule's ceiling.
    pub(crate) fn task_turns(&self) -> u32 {
        self.task_turns
    }

    pub(crate) async fn write_compaction(
        &mut self,
        turn: u32,
        tokens_before: u64,
        tokens_after: u64,
    ) -> std::io::Result<()> {
        let event = CompactionEvent {
            event_type: "compaction",
            event_id: new_event_id(),
            parent_id: self.turn_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            turn,
            task_id: self.active_task_id.clone(),
            tokens_before,
            tokens_after,
        };
        self.write_event(&event).await
    }

    /// Record that compaction was attempted and declined, leaving the context over budget.
    /// `tokens` is the occupancy at the moment of the decline; `reason` is
    /// [`COMPACTION_DECLINED_NO_HOOK_REPLACEMENT`] or
    /// [`COMPACTION_DECLINED_UNRESOLVED_TOOL_CALL`].
    pub(crate) async fn write_compaction_declined(
        &mut self,
        turn: u32,
        tokens: u64,
        reason: &str,
    ) -> std::io::Result<()> {
        let event = CompactionDeclinedEvent {
            event_type: "compaction_declined",
            event_id: new_event_id(),
            parent_id: self.turn_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            turn,
            task_id: self.active_task_id.clone(),
            tokens,
            reason: reason.to_string(),
        };
        self.write_event(&event).await
    }

    /// Record what the runtime did with an `on-task-start` hook's `seed-context`.
    ///
    /// Written for every outcome, committed or not: `tokens` is what reached the context
    /// (`0` on a rejection), `proposed_tokens` is what the hook returned, and
    /// `budget_tokens` is the ceiling it was measured against. `reason` is `Some` only for
    /// [`SEED_OUTCOME_REJECTED`]. Task-scoped, like `task_start`/`task_end`.
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn write_context_seed(
        &mut self,
        hook_name: &str,
        tokens: u64,
        proposed_tokens: u64,
        budget_tokens: u64,
        outcome: &str,
        reason: Option<&str>,
        message_ids: Vec<String>,
    ) -> std::io::Result<()> {
        let event = ContextSeedEvent {
            event_type: "context_seed",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: self.active_task_id.clone(),
            hook_name: hook_name.to_string(),
            tokens,
            proposed_tokens,
            budget_tokens,
            outcome: outcome.to_string(),
            reason: reason.map(str::to_string),
            message_ids,
        };
        self.write_event(&event).await
    }

    /// Record one store's deletion, naming the reason and everything that went.
    ///
    /// Session-scoped: `parent_id` is the session node, so the deletion hangs off the launch that
    /// performed it rather than off whatever task happened to be open. `messages_dropped` is
    /// `Some` only for [`RETENTION_REASON_MAX_MESSAGES`].
    pub(crate) async fn write_retention(
        &mut self,
        store: &str,
        reason: &str,
        targets: Vec<String>,
        messages_dropped: Option<u64>,
    ) -> std::io::Result<()> {
        let event = RetentionEvent {
            event_type: "retention",
            event_id: new_event_id(),
            parent_id: self.session_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            store: store.to_string(),
            reason: reason.to_string(),
            removed: u32::try_from(targets.len()).unwrap_or(u32::MAX),
            targets,
            messages_dropped,
        };
        self.write_event(&event).await
    }

    /// Record that a policy hook refused a call before it ran. The only account of a call the
    /// model asked for and never got: nothing ran, so no `tool_call` or `shell` record
    /// accompanies it.
    pub(crate) async fn write_call_denied(
        &mut self,
        turn: u32,
        event: &str,
        hook_name: &str,
        target: &str,
        reason: &str,
    ) -> std::io::Result<()> {
        let event = CallDeniedEvent {
            event_type: "call_denied",
            event_id: new_event_id(),
            parent_id: self.turn_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            turn,
            event: event.to_string(),
            hook_name: hook_name.to_string(),
            target: target.to_string(),
            reason: reason.to_string(),
        };
        self.write_event(&event).await
    }

    /// Record that a spend ceiling refused a driver call before it was sent. `origin` is the
    /// `hook:<name>` of a hook's `run-inference` call, `None` for an agent-loop turn.
    pub(crate) async fn write_spend_ceiling_reached(
        &mut self,
        turn: u32,
        refusal: &crate::spend::SpendRefusal,
        origin: Option<&str>,
    ) -> std::io::Result<()> {
        let event = SpendCeilingReachedEvent {
            event_type: "spend_ceiling_reached",
            event_id: new_event_id(),
            parent_id: self.turn_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            turn,
            task_id: self.active_task_id.clone(),
            refusal,
            origin,
        };
        self.write_event(&event).await
    }

    /// Record that the manifest's `capabilities.filesystem.read_only` refused a call before it
    /// ran. Like [`Self::write_call_denied`] it is the only account of a call the model asked for
    /// and never got — nothing ran, so no `tool_call` or `shell` record accompanies it.
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn write_protected_path_denied(
        &mut self,
        turn: u32,
        call: &str,
        target: &str,
        path: &str,
        rule: &str,
        signal: &str,
        reason: &str,
    ) -> std::io::Result<()> {
        let event = ProtectedPathDeniedEvent {
            event_type: "protected_path_denied",
            event_id: new_event_id(),
            parent_id: self.turn_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            turn,
            call: call.to_string(),
            target: target.to_string(),
            path: path.to_string(),
            rule: rule.to_string(),
            signal: signal.to_string(),
            reason: reason.to_string(),
        };
        self.write_event(&event).await
    }

    /// Record that a hook returned a `hook-output` arm the lifecycle event does not
    /// honor. Written by the agent loop from a buffered [`crate::hooks::DispatchFault`]
    /// drained just before `session_end`. Non-fatal: the session already continued as
    /// if the hook had returned `none`; this only makes the discard visible to
    /// `mur trace show` and anything reading the session trace.
    pub(crate) async fn write_hook_dispatch_error(
        &mut self,
        hook_name: &str,
        event: &str,
        arm: &str,
    ) -> std::io::Result<()> {
        let event = HookDispatchErrorEvent {
            event_type: "hook_dispatch_error",
            event_id: new_event_id(),
            parent_id: self.session_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            hook_name: hook_name.to_string(),
            event: event.to_string(),
            arm: arm.to_string(),
        };
        self.write_event(&event).await
    }

    // ── The harness a `transport: process` session drives ────────────────────────────────
    //
    // Eight events account for one harness run: what was spawned, what it said about itself,
    // what it warned or noted, how it failed, how it was stopped, and how it ended. None of them
    // ever carries an argument value, an environment value or a file's contents — the bridge's
    // bearer token reaches the harness through exactly those three, and `trace.jsonl` is durable.

    /// Record the harness about to be spawned: written after the driver planned the run and
    /// before the process exists, so a spawn that goes wrong still has its plan on record.
    pub(crate) async fn write_harness_start(&mut self, start: HarnessStart) -> std::io::Result<()> {
        let event = HarnessStartEvent {
            event_type: "harness_start",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: self.active_task_id.clone(),
            start,
        };
        self.write_event(&event).await
    }

    /// Record a warning the run raised, beside the line it printed to stderr.
    pub(crate) async fn write_harness_warning(
        &mut self,
        code: &str,
        message: &str,
    ) -> std::io::Result<()> {
        let event = HarnessWarningEvent {
            event_type: "harness_warning",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: self.active_task_id.clone(),
            code: code.to_string(),
            message: message.to_string(),
        };
        self.write_event(&event).await
    }

    /// Record the session the harness reports it opened, and how it is billing for it.
    pub(crate) async fn write_harness_session(
        &mut self,
        harness_session_id: &str,
        auth: &str,
        model: Option<&str>,
    ) -> std::io::Result<()> {
        let event = HarnessSessionEvent {
            event_type: "harness_session",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: self.active_task_id.clone(),
            harness_session_id: harness_session_id.to_string(),
            auth: auth.to_string(),
            model: model.map(str::to_string),
        };
        self.write_event(&event).await
    }

    /// Record the harness session a person asked this capsule to forget, and who asked.
    pub(crate) async fn write_harness_session_forgotten(
        &mut self,
        context_id: &str,
        harness_session_id: &str,
        requested_by: &str,
    ) -> std::io::Result<()> {
        let event = HarnessSessionForgottenEvent {
            event_type: "harness_session_forgotten",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: self.active_task_id.clone(),
            context_id: context_id.to_string(),
            harness_session_id: harness_session_id.to_string(),
            requested_by: requested_by.to_string(),
        };
        self.write_event(&event).await
    }

    /// Record that the harness is retrying a request to its own provider.
    pub(crate) async fn write_harness_retry(
        &mut self,
        attempt: u32,
        reason: &str,
    ) -> std::io::Result<()> {
        let event = HarnessRetryEvent {
            event_type: "harness_retry",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: self.active_task_id.clone(),
            attempt,
            reason: reason.to_string(),
        };
        self.write_event(&event).await
    }

    /// Record a diagnostic the driver produced for the trace alone — a line it could not read,
    /// or a result the runtime could not pair with a call.
    pub(crate) async fn write_harness_note(&mut self, text: &str) -> std::io::Result<()> {
        let event = HarnessNoteEvent {
            event_type: "harness_note",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: self.active_task_id.clone(),
            text: text.to_string(),
        };
        self.write_event(&event).await
    }

    /// Record a failed turn. `source` is `"harness"` when the harness reported the failure and
    /// `"runtime"` when this runtime ended the turn itself.
    pub(crate) async fn write_harness_failed(
        &mut self,
        kind: &str,
        message: &str,
        source: &str,
    ) -> std::io::Result<()> {
        let event = HarnessFailedEvent {
            event_type: "harness_failed",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: self.active_task_id.clone(),
            kind: kind.to_string(),
            message: message.to_string(),
            source: source.to_string(),
        };
        self.write_event(&event).await
    }

    /// Record the interrupt a cancelled task sent its harness: the method the driver declared,
    /// whether it went out, and how long the harness has before it is killed.
    pub(crate) async fn write_harness_interrupt(
        &mut self,
        method: &str,
        delivered: bool,
        grace_ms: u64,
    ) -> std::io::Result<()> {
        let event = HarnessInterruptEvent {
            event_type: "harness_interrupt",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: self.active_task_id.clone(),
            method: method.to_string(),
            delivered,
            grace_ms,
        };
        self.write_event(&event).await
    }

    /// Record how the harness ended: once per spawned process, whatever ended it.
    pub(crate) async fn write_harness_exit(&mut self, exit: HarnessExit) -> std::io::Result<()> {
        let event = HarnessExitEvent {
            event_type: "harness_exit",
            event_id: new_event_id(),
            parent_id: self.task_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            task_id: self.active_task_id.clone(),
            exit,
        };
        self.write_event(&event).await
    }

    pub(crate) async fn write_session_end(&mut self, exit_status: &str) -> std::io::Result<()> {
        let duration_ms = self
            .session_start_time
            .elapsed()
            .as_millis()
            .try_into()
            .unwrap_or(u64::MAX);
        let event = SessionEndEvent {
            event_type: "session_end",
            event_id: new_event_id(),
            parent_id: self.session_parent(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            total_turns: self.total_turns,
            total_input_tokens: self.total_input_tokens,
            total_output_tokens: self.total_output_tokens,
            total_tool_calls: self.total_tool_calls,
            total_shell_calls: self.total_shell_calls,
            duration_ms,
            exit_status: exit_status.to_string(),
        };
        self.write_event(&event).await?;
        self.session_ended = true;
        Ok(())
    }

    /// Writes session_end with the accumulated totals if session_start was written but
    /// session_end was not (error exit paths that bypass the normal Ok() return sites).
    pub(crate) async fn write_session_end_if_not_ended(
        &mut self,
        exit_status: &str,
    ) -> std::io::Result<()> {
        if self.session_started && !self.session_ended {
            self.write_session_end(exit_status).await?;
        }
        Ok(())
    }

    pub(crate) async fn flush(&mut self) -> std::io::Result<()> {
        self.writer.flush().await
    }

    /// Writes one event as one line, and flushes.
    ///
    /// The flush is not an optimisation choice: [`ResourceTraceAppender`] appends to this same
    /// `trace.jsonl` while a task runs, so a line left half-buffered here would be split around
    /// whatever the appender wrote in between and the file would stop parsing as JSONL. The
    /// two writers agree on one rule — a complete line per write, then flush — and that rule is
    /// what makes concurrent reads during a running task safe to record.
    async fn write_event(&mut self, event: &impl Serialize) -> std::io::Result<()> {
        let mut line = serde_json::to_string(event).map_err(std::io::Error::other)?;
        line.push('\n');
        self.writer.write_all(line.as_bytes()).await?;
        self.writer.flush().await
    }
}

/// One element of `session_start.gateways`: an artifact's credential gateway. Never the key.
#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
pub(crate) struct SessionGateway {
    /// The artifact whose entry declares the gateway.
    pub(crate) artifact: String,
    /// The upstream's host, with its port when `gateway.endpoint` wrote one.
    pub(crate) host: String,
    /// `"config"`, `"environment"`, `"manifest"`, or `"none"` when the entry names no key.
    pub(crate) credential_source: &'static str,
    /// Whether the gateway is admitted against the spend meter: true only for the configured
    /// `transport: http` driver's.
    pub(crate) metered: bool,
}

/// One element of `session_start.runtime_artifacts`: a staged artifact whose `murmur.lock` pin a
/// running capsule fetched.
#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
pub(crate) struct RuntimeArtifact {
    name: String,
    version: String,
    /// Always `"runtime"`: only runtime-origin entries are listed.
    origin: &'static str,
    /// The session the lock entry names as the puller.
    session: String,
    /// Always `"untrusted"`, derived from `origin`.
    trust: &'static str,
}

/// `session_start.control`: the names a capsule's `control:` block declares, never a value.
#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
pub(crate) struct SessionControl {
    /// Controllable setting names, as `control.settings` lists them.
    pub(crate) settings: Vec<&'static str>,
    /// Controller-supplied secret names, as `control.secrets` lists them.
    pub(crate) secrets: Vec<String>,
    /// Setting names the agent may change, as `control.agent_settings` lists them. Omitted when
    /// the block grants the agent nothing.
    #[serde(skip_serializing_if = "Vec::is_empty")]
    pub(crate) agent_settings: Vec<&'static str>,
}

/// One element of `session_start.inference_choices`: a driver choice and whether its credential
/// resolved at launch. Never a key.
#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
pub(crate) struct SessionInferenceChoice {
    pub(crate) name: String,
    pub(crate) driver: String,
    pub(crate) model: String,
    /// Where the driver's gateway credential came from, as `session_start.gateways` names it, or
    /// `"none"` for one found nowhere.
    pub(crate) credential_source: &'static str,
    /// Whether the credential resolved at launch. An injected credential is `true` here and is
    /// checked again when a switch asks for it.
    pub(crate) available: bool,
}

/// One accepted setting or secret change, as [`ResourceTraceAppender::write_control_change`]
/// records it. `event_id` is minted by the caller, because the setting it changed carries it
/// forward to the `control_applied` that names it.
pub(crate) struct ControlChange<'a> {
    pub(crate) event_id: String,
    /// Who made the change.
    pub(crate) principal: crate::control_plane::Principal,
    /// The control token's id for a controller's change; `None` for the agent's, which presents
    /// no token.
    pub(crate) token_id: Option<&'a str>,
    /// `"setting"` or `"secret"`.
    pub(crate) kind: &'static str,
    pub(crate) name: &'a str,
    /// `"set"` or `"forget"`.
    pub(crate) action: &'static str,
    /// A setting's value before and after. Always `None` for a secret.
    pub(crate) previous: Option<Value>,
    pub(crate) value: Option<Value>,
    /// When a setting change takes effect. `None` for a secret.
    pub(crate) applies_from: Option<&'static str>,
    /// For a secret `set`: whether it replaced a held value.
    pub(crate) replaced: Option<bool>,
}

/// `control_change`: a controller changed a setting or a secret. A secret's line names it and
/// carries nothing of its value.
#[derive(Serialize)]
struct ControlChangeEvent<'a> {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    principal: &'static str,
    #[serde(skip_serializing_if = "Option::is_none")]
    token_id: Option<&'a str>,
    kind: &'static str,
    name: &'a str,
    action: &'static str,
    #[serde(skip_serializing_if = "Option::is_none")]
    previous: Option<Value>,
    #[serde(skip_serializing_if = "Option::is_none")]
    value: Option<Value>,
    #[serde(skip_serializing_if = "Option::is_none")]
    applies_from: Option<&'static str>,
    #[serde(skip_serializing_if = "Option::is_none")]
    replaced: Option<bool>,
}

/// `control_refused`: the control plane refused a request, or the runtime refused the agent's
/// change. `kind`, `name` and `token_id` are written only once a controller's request
/// authenticated, so a `401` names nothing; the agent's refusal names its setting and no token.
#[derive(Serialize)]
struct ControlRefusedEvent<'a> {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    principal: &'static str,
    status: u16,
    reason: &'static str,
    #[serde(skip_serializing_if = "Option::is_none")]
    kind: Option<&'static str>,
    #[serde(skip_serializing_if = "Option::is_none")]
    name: Option<&'a str>,
    #[serde(skip_serializing_if = "Option::is_none")]
    token_id: Option<&'a str>,
}

/// `control_applied`: the first inference call to use a setting value a controller set.
#[derive(Serialize)]
struct ControlAppliedEvent<'a> {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    turn: u32,
    task_id: Option<String>,
    name: &'a str,
    value: Value,
    change_id: &'a str,
}

/// `tools_refreshed`: the first inference call to send a tool array rebuilt after an install.
#[derive(Serialize)]
struct ToolsRefreshedEvent<'a> {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    turn: u32,
    task_id: Option<String>,
    trigger: &'a str,
    added: &'a [String],
    removed: &'a [String],
    tools: &'a [String],
}

/// `gateway_credential`: the `inference_credential` event for the credential of any artifact other
/// than the configured inference driver, with the artifact's name.
#[derive(Serialize)]
struct GatewayCredentialEvent<'a> {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    artifact: &'a str,
    source: &'static str,
    #[serde(skip_serializing_if = "Option::is_none")]
    credential: Option<&'a str>,
    change: &'static str,
    #[serde(skip_serializing_if = "Option::is_none")]
    trigger: Option<&'static str>,
    #[serde(skip_serializing_if = "Option::is_none")]
    status: Option<u16>,
    #[serde(skip_serializing_if = "Option::is_none")]
    retried: Option<bool>,
    #[serde(skip_serializing_if = "Option::is_none")]
    reason: Option<&'static str>,
}

/// `inference_credential`: a rotation, rejection or unreadable source for the session's inference
/// credential. `trigger` is written only with `change: "rotated"`, `status` and `retried` only with
/// `"rejected"`, `reason` only with `"unreadable"`, and `credential` whenever the manifest named one.
#[derive(Serialize)]
struct InferenceCredentialEvent<'a> {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    source: &'static str,
    #[serde(skip_serializing_if = "Option::is_none")]
    credential: Option<&'a str>,
    change: &'static str,
    #[serde(skip_serializing_if = "Option::is_none")]
    trigger: Option<&'static str>,
    #[serde(skip_serializing_if = "Option::is_none")]
    status: Option<u16>,
    #[serde(skip_serializing_if = "Option::is_none")]
    retried: Option<bool>,
    #[serde(skip_serializing_if = "Option::is_none")]
    reason: Option<&'static str>,
}

/// A credential change as the `change`, `trigger`, `status`, `retried` and `reason` fields of an
/// `inference_credential` or `gateway_credential` event.
#[allow(clippy::type_complexity)]
fn credential_change_fields(
    change: CredentialChange,
) -> (
    &'static str,
    Option<&'static str>,
    Option<u16>,
    Option<bool>,
    Option<&'static str>,
) {
    match change {
        CredentialChange::Rotated { trigger } => ("rotated", Some(trigger), None, None, None),
        CredentialChange::Rejected { status, retried } => {
            ("rejected", None, Some(status), Some(retried), None)
        }
        CredentialChange::Unreadable { reason } => ("unreadable", None, None, None, Some(reason)),
    }
}

// ── Resource-plane appender ───────────────────────────────────────────────────

/// A second, independent `O_APPEND` handle to the session's `trace.jsonl`, used by the resource
/// plane, by the peer plane's listener, and by the peer-handoff tools in the agent loop.
///
/// [`TraceWriter`] is `&mut`-owned by the agent loop and cannot be shared, and the motivating
/// case for a resource-plane event — a gateway reading a finished-but-alive capsule — happens
/// after `session_end` has already been written. So the plane gets its own handle rather than a
/// borrow of the loop's, and every event is written at the moment of the request rather than
/// deferred to a task boundary: a denied read that only surfaced at the next task end would be a
/// record of the wrong thing at the wrong time.
///
/// Interleaving is safe because both writers emit exactly one complete line per `write_all` and
/// flush it, and `O_APPEND` makes each such write atomic against the other's.
pub struct ResourceTraceAppender {
    /// Async rather than a `std::sync::Mutex`: this is held across an `await` on the write.
    file: tokio::sync::Mutex<File>,
    session_id: String,
    /// The session node every line this appender writes hangs off — [`TraceWriter`]'s own
    /// [`TraceWriter::session_event_id`], handed over at open time. The plane has no turn or
    /// task of its own: a read served after `session_end` belongs to the launch, not to
    /// whatever task happened to run last.
    session_event_id: String,
}

impl ResourceTraceAppender {
    pub async fn open(
        workdir: &Path,
        session_id: String,
        session_event_id: String,
    ) -> std::io::Result<Self> {
        let file = OpenOptions::new()
            .create(true)
            .append(true)
            .open(workdir.join("trace.jsonl"))
            .await?;
        Ok(Self {
            file: tokio::sync::Mutex::new(file),
            session_id,
            session_event_id,
        })
    }

    /// Records one `list`. Failures are swallowed: a trace that cannot be written must not turn
    /// a served read into an error, and the alternative — refusing the read — would let anyone
    /// who can fill the disk take the plane down.
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn write_resource_list(
        &self,
        root: &str,
        entry_count: usize,
        total_bytes: u64,
        generation: u64,
        containment_achieved: ContainmentClass,
        outcome: &str,
        reason: Option<String>,
    ) {
        let event = ResourceListEvent {
            event_type: "resource_list",
            event_id: new_event_id(),
            parent_id: Some(self.session_event_id.clone()),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            root: root.to_string(),
            entry_count,
            total_bytes,
            generation,
            containment_achieved,
            outcome: outcome.to_string(),
            reason,
        };
        self.append(&event).await;
    }

    /// Records one `read`, served or refused. `bytes` and `sha256` are `None` on every
    /// non-`ok` outcome — a refusal must not carry a hash of something it did not serve.
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn write_resource_read(
        &self,
        path: &str,
        outcome: &str,
        bytes: Option<u64>,
        sha256: Option<String>,
        generation: u64,
        containment_achieved: ContainmentClass,
        reason: Option<String>,
    ) {
        let event = ResourceReadEvent {
            event_type: "resource_read",
            event_id: new_event_id(),
            parent_id: Some(self.session_event_id.clone()),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            path: path.to_string(),
            outcome: outcome.to_string(),
            bytes,
            sha256,
            generation,
            containment_achieved,
            reason,
        };
        self.append(&event).await;
    }

    /// Records one mint. `handle_id` and `expires_at_ms` are `None` on every non-`ok` outcome —
    /// a refused mint produced no token, and an audit record must not imply one exists.
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn write_peer_handle_mint(
        &self,
        handle_id: Option<String>,
        path: &str,
        audience: &str,
        expires_at_ms: Option<u64>,
        outcome: &str,
        reason: Option<String>,
    ) {
        let event = PeerHandleMintEvent {
            event_type: "peer_handle_mint",
            event_id: new_event_id(),
            parent_id: Some(self.session_event_id.clone()),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            handle_id,
            path: path.to_string(),
            audience: audience.to_string(),
            expires_at_ms,
            outcome: outcome.to_string(),
            reason,
        };
        self.append(&event).await;
    }

    /// Records one redeem, served or refused. The token itself is never written — only its
    /// `handle_id`.
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn write_peer_handle_redeem(
        &self,
        handle_id: &str,
        path: Option<String>,
        generation: u64,
        audience_asserted: Option<String>,
        bytes: Option<u64>,
        sha256: Option<String>,
        outcome: &str,
        reason: Option<String>,
    ) {
        let event = PeerHandleRedeemEvent {
            event_type: "peer_handle_redeem",
            event_id: new_event_id(),
            parent_id: Some(self.session_event_id.clone()),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            handle_id: handle_id.to_string(),
            path,
            generation,
            audience_asserted,
            bytes,
            sha256,
            outcome: outcome.to_string(),
            reason,
        };
        self.append(&event).await;
    }

    /// Records one fetch on the ingesting side, served or refused.
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn write_peer_file_fetch(
        &self,
        peer: &str,
        handle_id: &str,
        stored_path: Option<String>,
        bytes: Option<u64>,
        sha256: Option<String>,
        outcome: &str,
        reason: Option<String>,
    ) {
        let event = PeerFileFetchEvent {
            event_type: "peer_file_fetch",
            event_id: new_event_id(),
            parent_id: Some(self.session_event_id.clone()),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            peer: peer.to_string(),
            handle_id: handle_id.to_string(),
            stored_path,
            bytes,
            sha256,
            outcome: outcome.to_string(),
            reason,
        };
        self.append(&event).await;
    }

    /// Records one launched child, while its delegation is still running.
    ///
    /// Every field is known by the time the child has reported its session id, so none is
    /// optional: a delegation with nothing to say here was never launched and writes no line.
    pub(crate) async fn write_delegation_start(
        &self,
        delegation_id: &str,
        capsule: &str,
        version: &str,
        child_session_id: &str,
        child_workdir: &str,
    ) {
        let event = DelegationStartEvent::new(
            &self.session_id,
            &self.session_event_id,
            delegation_id,
            capsule,
            version,
            child_session_id,
            child_workdir,
        );
        self.append(&event).await;
    }

    /// Records one delegation ending. `delegation_id` and `child_session_id` are `None` when no
    /// child was launched — a delegation that never started names no id, and an audit record must
    /// not imply one exists.
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn write_delegation(
        &self,
        capsule: &str,
        version: &str,
        delegation_id: Option<String>,
        child_session_id: Option<String>,
        duration_ms: u64,
        outcome: &str,
        reason: Option<String>,
    ) {
        let event = DelegationEvent::new(
            &self.session_id,
            &self.session_event_id,
            capsule,
            version,
            delegation_id,
            child_session_id,
            duration_ms,
            outcome,
            reason,
        );
        self.append(&event).await;
    }

    /// Records one change to the session's inference credential. Carries where the credential
    /// comes from and what happened to it, never the value.
    pub(crate) async fn write_inference_credential(
        &self,
        source: &'static str,
        credential: Option<&str>,
        change: CredentialChange,
    ) {
        let (change_name, trigger, status, retried, reason) = credential_change_fields(change);
        let event = InferenceCredentialEvent {
            event_type: "inference_credential",
            event_id: new_event_id(),
            parent_id: Some(self.session_event_id.clone()),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            source,
            credential,
            change: change_name,
            trigger,
            status,
            retried,
            reason,
        };
        self.append(&event).await;
    }

    /// Writes a `gateway_credential` event for `artifact`'s credential: the fields
    /// [`Self::write_inference_credential`] writes, plus `artifact`.
    pub(crate) async fn write_gateway_credential(
        &self,
        artifact: &str,
        source: &'static str,
        credential: Option<&str>,
        change: CredentialChange,
    ) {
        let (change_name, trigger, status, retried, reason) = credential_change_fields(change);
        let event = GatewayCredentialEvent {
            event_type: "gateway_credential",
            event_id: new_event_id(),
            parent_id: Some(self.session_event_id.clone()),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            artifact,
            source,
            credential,
            change: change_name,
            trigger,
            status,
            retried,
            reason,
        };
        self.append(&event).await;
    }

    /// Records one accepted control-plane change, under the `event_id` the caller minted.
    pub(crate) async fn write_control_change(&self, change: ControlChange<'_>) {
        let event = ControlChangeEvent {
            event_type: "control_change",
            event_id: change.event_id,
            parent_id: Some(self.session_event_id.clone()),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            principal: change.principal.as_str(),
            token_id: change.token_id,
            kind: change.kind,
            name: change.name,
            action: change.action,
            previous: change.previous,
            value: change.value,
            applies_from: change.applies_from,
            replaced: change.replaced,
        };
        self.append(&event).await;
    }

    /// Records one refused change. `kind`, `name` and `token_id` are `None` for a controller's
    /// request that did not authenticate.
    pub(crate) async fn write_control_refused(
        &self,
        principal: crate::control_plane::Principal,
        status: u16,
        reason: &'static str,
        kind: Option<&'static str>,
        name: Option<&str>,
        token_id: Option<&str>,
    ) {
        let event = ControlRefusedEvent {
            event_type: "control_refused",
            event_id: new_event_id(),
            parent_id: Some(self.session_event_id.clone()),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            principal: principal.as_str(),
            status,
            reason,
            kind,
            name,
            token_id,
        };
        self.append(&event).await;
    }

    async fn append(&self, event: &impl Serialize) {
        let Ok(mut line) = serde_json::to_string(event) else {
            return;
        };
        line.push('\n');
        let mut file = self.file.lock().await;
        if file.write_all(line.as_bytes()).await.is_ok() {
            let _ = file.flush().await;
        }
    }
}

// -- Prior-session appender ---------------------------------------------------

/// An `O_APPEND` handle to a *previous* session's `trace.jsonl`, opened by a resuming launch to
/// account for demoted commands that session left unreported.
///
/// The marker goes in the file that holds the unmatched `shell_detached` line rather than in the
/// resuming session's own trace, which keeps clearing O(1) instead of a scan of every sibling
/// session, and means [`crate::retention::prune_sessions`] removes a marker and the work it
/// accounts for together.
///
/// Writing to another session's trace is safe because that session's writer is dead — that is the
/// precondition for there being anything unaccounted — and because this obeys the same rule
/// [`ResourceTraceAppender`] does: one complete line per `write_all`, then flush, which `O_APPEND`
/// makes atomic against any other writer.
pub(crate) struct PriorSessionTraceAppender {
    file: File,
    /// The prior session's id, which is what its own lines carry.
    session_id: String,
    /// That session's `session_start` node, read back from the same file.
    session_event_id: Option<String>,
    /// Whether the file's last record is unterminated, which a writer killed mid-`write_all`
    /// leaves behind. The first appended line opens with a newline when it is, so the marker is
    /// a line of its own rather than spliced onto the torn one.
    terminate_torn_tail: bool,
}

impl PriorSessionTraceAppender {
    /// `trace_path` is the prior session's `trace.jsonl`. Errors when it cannot be opened for
    /// appending, which the caller reports rather than reporting work it could not mark.
    pub(crate) async fn open(
        trace_path: &Path,
        session_id: String,
        session_event_id: Option<String>,
    ) -> std::io::Result<Self> {
        let terminate_torn_tail = !ends_with_newline(trace_path)?;
        let file = OpenOptions::new().append(true).open(trace_path).await?;
        Ok(Self {
            file,
            session_id,
            session_event_id,
            terminate_torn_tail,
        })
    }

    /// Records one demoted command as lost, and reports whether the line landed.
    ///
    /// The result is not swallowed: a work id whose marker did not land must be dropped from the
    /// report rather than reported unmarked, because an unmarked work id is reported again by
    /// every later resume of the same session.
    pub(crate) async fn write_shell_lost(
        &mut self,
        work: &crate::detached::LostWork,
        reconciled_by_session: &str,
        reconciled_task_id: &str,
    ) -> std::io::Result<()> {
        let event = ShellLostEvent {
            event_type: "shell_lost",
            event_id: new_event_id(),
            parent_id: self.session_event_id.clone(),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            work_id: work.work_id.clone(),
            binary: work.binary.clone(),
            command: work.command.clone(),
            detached_at_ms: work.detached_at_ms,
            reconciled_by_session: reconciled_by_session.to_string(),
            reconciled_task_id: reconciled_task_id.to_string(),
        };
        let mut line = if self.terminate_torn_tail {
            "\n".to_string()
        } else {
            String::new()
        };
        line.push_str(&serde_json::to_string(&event).map_err(std::io::Error::other)?);
        line.push('\n');
        self.file.write_all(line.as_bytes()).await?;
        self.file.flush().await?;
        self.terminate_torn_tail = false;
        Ok(())
    }
}

/// Whether the file's last byte is a newline. An empty file counts as terminated: there is no
/// record for an appended line to run into.
fn ends_with_newline(path: &Path) -> std::io::Result<bool> {
    use std::io::{Read, Seek, SeekFrom};

    let mut file = std::fs::File::open(path)?;
    let length = file.metadata()?.len();
    if length == 0 {
        return Ok(true);
    }
    file.seek(SeekFrom::End(-1))?;
    let mut last = [0u8; 1];
    file.read_exact(&mut last)?;
    Ok(last[0] == b'\n')
}

pub(crate) fn timestamp_ms() -> u64 {
    SystemTime::now()
        .duration_since(UNIX_EPOCH)
        .unwrap_or_default()
        .as_millis() as u64
}

// ── Plan-scheduler events ────────────────────────────────────────────────────

/// One step of a plan's DAG as authored, written into `plan_start`.
///
/// The shape is recorded once, up front, so the run stays legible for a step that never ran:
/// the trace names every step the plan declared whether or not the scheduler reached it.
#[derive(Serialize, Clone)]
pub(crate) struct PlanStepShape {
    pub(crate) step_id: String,
    /// `"tool"`, `"shell"` or `"capsule"`; `"unknown"` for a step whose kind cannot be inferred,
    /// which is a step the plan validator refuses.
    pub(crate) kind: &'static str,
    pub(crate) depends_on: Vec<String>,
    /// Whether the step carries an `if`, and so may settle without ever being dispatched.
    pub(crate) has_condition: bool,
}

/// The plan's structure, written once before any step runs. Its `event_id` is the plan node
/// every other line of the run hangs off.
#[derive(Serialize)]
struct PlanStartEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    plan_id: String,
    step_count: usize,
    steps: Vec<PlanStepShape>,
}

/// One dispatched step, written as the scheduler hands it to a worker.
///
/// A step that settled without ever being dispatched — an `if` that evaluated false, a
/// dependency that never resolved — writes no line here, only a terminal `plan_step`. Joins to
/// that line on `(plan_id, step_id)`.
#[derive(Serialize)]
struct PlanStepStartEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    plan_id: String,
    step_id: String,
    kind: &'static str,
    depends_on: Vec<String>,
}

/// A settled step, written by the scheduler loop after the step's `on_error` policy has been
/// applied — which is why the worker thread does not write it: a step that failed under
/// `on_error: skip` settles as skipped in the returned report and must read as skipped here.
#[derive(Serialize)]
struct PlanStepEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    plan_id: String,
    step_id: String,
    kind: &'static str,
    status: &'static str,
    attempts: u32,
    duration_ms: u64,
    #[serde(skip_serializing_if = "Option::is_none")]
    error: Option<String>,
    /// The interpolated input the step dispatched, with peer handle tokens redacted. Tool steps
    /// only — a shell step's command and a capsule step's task text are not carried here.
    #[serde(skip_serializing_if = "Option::is_none")]
    input: Option<Value>,
    /// The tool's self-declared `state_effect` for this step, lifted verbatim from
    /// `tool-result.metadata`. Absent when the tool declared nothing. Feeds the same
    /// redundant-call analysis `tool_call.state_effect` does.
    #[serde(skip_serializing_if = "Option::is_none")]
    state_effect: Option<String>,
    /// The resource this step addressed, as the tool declared it. Opaque and never parsed here.
    #[serde(skip_serializing_if = "Option::is_none")]
    resource_id: Option<String>,
}

/// The run's summary, written once as `plan::execute` returns, whatever ended it.
#[derive(Serialize)]
struct PlanEndEvent {
    event_type: &'static str,
    event_id: String,
    parent_id: Option<String>,
    session_id: String,
    timestamp: u64,
    plan_id: String,
    outcome: &'static str,
    #[serde(skip_serializing_if = "Option::is_none")]
    failed_step: Option<String>,
    steps_total: usize,
    steps_succeeded: usize,
    steps_failed: usize,
    steps_skipped: usize,
    duration_ms: u64,
    /// Why the run ended when the reason was not a step's own failure: a plan that would not
    /// parse, a cgroup scope the host refused, a DAG with nothing left to run.
    #[serde(skip_serializing_if = "Option::is_none")]
    reason: Option<String>,
}

/// Everything the scheduler loop knows about a settled step, as one argument.
pub(crate) struct PlanStepRecord {
    pub(crate) plan_id: String,
    pub(crate) step_id: String,
    pub(crate) kind: &'static str,
    /// `"success"`, `"failed"` or `"skipped"` — the status the returned `ExecutionReport` carries
    /// for this step, so the two can never disagree.
    pub(crate) status: &'static str,
    /// How many times the step was dispatched. `0` for a step that settled without dispatch.
    pub(crate) attempts: u32,
    pub(crate) duration_ms: u64,
    pub(crate) error: Option<String>,
    pub(crate) input: Option<Value>,
    pub(crate) state_effect: Option<String>,
    pub(crate) resource_id: Option<String>,
}

/// Everything the scheduler knows about a finished run, as one argument.
pub(crate) struct PlanEndRecord {
    pub(crate) plan_id: String,
    pub(crate) outcome: &'static str,
    pub(crate) failed_step: Option<String>,
    pub(crate) steps_total: usize,
    pub(crate) steps_succeeded: usize,
    pub(crate) steps_failed: usize,
    pub(crate) steps_skipped: usize,
    pub(crate) duration_ms: u64,
    pub(crate) reason: Option<String>,
}

/// A third `O_APPEND` handle to the session's `trace.jsonl`, held by a plan run.
///
/// Synchronous where [`ResourceTraceAppender`] is async, because [`crate::plan::execute`] is: it
/// runs its ready set on plain threads inside a `thread::scope`, so there is no runtime to await
/// on and no `await` for a lock to be held across. Interleaving with [`TraceWriter`] and with
/// the resource plane is safe on the same terms they interleave with each other — one complete
/// line per `write_all`, then flush, which `O_APPEND` makes atomic against the other writers.
///
/// Handing one of these to a [`crate::plan::SchedulerContext`] is the whole of the opt-in: a
/// context carrying `None` opens no file and writes no line.
pub struct PlanTraceAppender {
    file: std::sync::Mutex<std::fs::File>,
    session_id: String,
    /// The session node a run's `plan_start` hangs off — [`TraceWriter`]'s own
    /// `session_start`, handed over at open time so an appended line names a real parent.
    session_event_id: String,
}

impl PlanTraceAppender {
    pub fn open(
        workdir: &Path,
        session_id: String,
        session_event_id: String,
    ) -> std::io::Result<Self> {
        let file = std::fs::OpenOptions::new()
            .create(true)
            .append(true)
            .open(workdir.join("trace.jsonl"))?;
        Ok(Self {
            file: std::sync::Mutex::new(file),
            session_id,
            session_event_id,
        })
    }

    /// Records the plan's DAG and returns the plan node every later line of this run hangs off.
    pub(crate) fn write_plan_start(&self, plan_id: &str, steps: Vec<PlanStepShape>) -> String {
        let event_id = new_event_id();
        let event = PlanStartEvent {
            event_type: "plan_start",
            event_id: event_id.clone(),
            parent_id: Some(self.session_event_id.clone()),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            plan_id: plan_id.to_string(),
            step_count: steps.len(),
            steps,
        };
        self.append(&event);
        event_id
    }

    pub(crate) fn write_plan_step_start(
        &self,
        plan_node: &str,
        plan_id: &str,
        step_id: &str,
        kind: &'static str,
        depends_on: Vec<String>,
    ) {
        let event = PlanStepStartEvent {
            event_type: "plan_step_start",
            event_id: new_event_id(),
            parent_id: Some(plan_node.to_string()),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            plan_id: plan_id.to_string(),
            step_id: step_id.to_string(),
            kind,
            depends_on,
        };
        self.append(&event);
    }

    pub(crate) fn write_plan_step(&self, plan_node: &str, record: PlanStepRecord) {
        // On the same terms `TraceWriter::write_tool_call` redacts a tool call's input: a peer
        // handle is a credential and this file is durable.
        let input = record.input.map(|mut value| {
            crate::peer_handoff::redact_handles_in_json(&mut value);
            value
        });
        let event = PlanStepEvent {
            event_type: "plan_step",
            event_id: new_event_id(),
            parent_id: Some(plan_node.to_string()),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            plan_id: record.plan_id,
            step_id: record.step_id,
            kind: record.kind,
            status: record.status,
            attempts: record.attempts,
            duration_ms: record.duration_ms,
            error: record.error,
            input,
            state_effect: record.state_effect,
            resource_id: record.resource_id,
        };
        self.append(&event);
    }

    /// Records the run's summary. `plan_node` is `None` for a run that never had one — a plan
    /// file that would not parse describes no DAG — and the line hangs off the session instead.
    pub(crate) fn write_plan_end(&self, plan_node: Option<&str>, record: PlanEndRecord) {
        let event = PlanEndEvent {
            event_type: "plan_end",
            event_id: new_event_id(),
            parent_id: Some(
                plan_node
                    .map(str::to_string)
                    .unwrap_or_else(|| self.session_event_id.clone()),
            ),
            session_id: self.session_id.clone(),
            timestamp: timestamp_ms(),
            plan_id: record.plan_id,
            outcome: record.outcome,
            failed_step: record.failed_step,
            steps_total: record.steps_total,
            steps_succeeded: record.steps_succeeded,
            steps_failed: record.steps_failed,
            steps_skipped: record.steps_skipped,
            duration_ms: record.duration_ms,
            reason: record.reason,
        };
        self.append(&event);
    }

    /// Records one child a `capsule` step launched, while its delegation is still running.
    ///
    /// The synchronous twin of [`ResourceTraceAppender::write_delegation_start`]. Both go through
    /// [`DelegationStartEvent::new`], so the two lines are identical but for their ids and
    /// clocks — held by there being one constructor rather than by a test.
    pub(crate) fn write_delegation_start(
        &self,
        delegation_id: &str,
        capsule: &str,
        version: &str,
        child_session_id: &str,
        child_workdir: &str,
    ) {
        let event = DelegationStartEvent::new(
            &self.session_id,
            &self.session_event_id,
            delegation_id,
            capsule,
            version,
            child_session_id,
            child_workdir,
        );
        self.append(&event);
    }

    /// Records one `capsule` step's delegation ending. `delegation_id` and `child_session_id` are
    /// `None` when no child was launched, on the same terms
    /// [`ResourceTraceAppender::write_delegation`] passes them.
    ///
    /// One pair per *attempt*: a step with retries delegates once per attempt, so it writes one
    /// of these per attempt. Two `failed` lines and a `completed` one is what happened, not a
    /// duplicate of one thing.
    #[allow(clippy::too_many_arguments)]
    pub(crate) fn write_delegation(
        &self,
        capsule: &str,
        version: &str,
        delegation_id: Option<String>,
        child_session_id: Option<String>,
        duration_ms: u64,
        outcome: &str,
        reason: Option<String>,
    ) {
        let event = DelegationEvent::new(
            &self.session_id,
            &self.session_event_id,
            capsule,
            version,
            delegation_id,
            child_session_id,
            duration_ms,
            outcome,
            reason,
        );
        self.append(&event);
    }

    /// Writes one complete line and flushes it. Every failure is swallowed, exactly as
    /// [`ResourceTraceAppender::append`] swallows one: a trace that cannot be written must never
    /// turn a step that ran into a step that failed.
    fn append(&self, event: &impl Serialize) {
        use std::io::Write as _;

        let Ok(mut line) = serde_json::to_string(event) else {
            return;
        };
        line.push('\n');
        let Ok(mut file) = self.file.lock() else {
            return;
        };
        if file.write_all(line.as_bytes()).is_ok() {
            let _ = file.flush();
        }
    }
}

// ── Unit tests ────────────────────────────────────────────────────────────────

#[cfg(test)]
mod tests {
    use super::*;
    use crate::{
        containment::scope_report_for_tier, sandbox::EnforcementTier, sealed::UsernsGrant,
        trace_blobs::BLOB_DIR_NAME, types::CapabilityPolicy,
    };
    use murmur_artifact::{ArtifactRuntime, InterpreterRuntimeDir, InterpreterRuntimeGrant};
    use serde_json::Value;

    use crate::origin::TaskOrigin;

    /// The class an A2A task carrying no origin claim resolves to.
    fn event_provenance() -> TaskProvenance {
        TaskProvenance::derive(TaskOrigin::Event, None)
    }

    /// A [`ScopeReport`] built the way production builds one — through
    /// `containment::scope_report_for_tier` with an *explicit* tier, never a live probe, so these
    /// tests assert the same values on a Landlock-capable host and a laptop that has none.
    fn report_for(
        declared: ContainmentClass,
        tier: EnforcementTier,
        workdir_exec: bool,
    ) -> ScopeReport {
        scope_report_for_tier(
            &CapabilityPolicy {
                workdir_exec_allowed: workdir_exec,
                ..CapabilityPolicy::default()
            },
            declared,
            tier,
            None,
            None,
            None,
            Vec::new(),
            Vec::new(),
            Vec::new(),
            crate::cgroup::IoMaxReport::default(),
        )
    }

    fn demotion_info() -> crate::detached::DetachedDispatchInfo {
        crate::detached::DetachedDispatchInfo {
            work_id: "wrk_0a1b2c3d4e5f6a7b".to_string(),
            binary: "/usr/bin/bash".to_string(),
            command: "make -j8".to_string(),
            grace_ms: 1_000,
        }
    }

    /// A demotion whose marker cannot be written stands anyway: the write failure is reported and
    /// swallowed, and the session carries on with the command running in the background.
    #[tokio::test]
    async fn a_demotion_marker_that_cannot_be_written_does_not_fail_the_demotion() {
        let dir = tempfile::tempdir().unwrap();
        // `/dev/full` opens and fails every write with `ENOSPC`, which is the one way to get a
        // trace handle that is valid and unwritable without root.
        std::os::unix::fs::symlink("/dev/full", dir.path().join("trace.jsonl")).unwrap();
        let mut writer = make_writer(dir.path()).await;
        let info = demotion_info();

        assert!(
            writer
                .write_shell_detached(1, &info.work_id, &info.binary, &info.command, info.grace_ms)
                .await
                .is_err(),
            "the sink under this writer must be unwritable for the claim below to mean anything"
        );
        crate::agent::record_demotion(&mut writer, 1, &info).await;
    }

    /// The ordinary path writes one `shell_detached` and nothing else.
    #[tokio::test]
    async fn a_recorded_demotion_writes_only_its_own_marker() {
        let dir = tempfile::tempdir().unwrap();
        let mut writer = make_writer(dir.path()).await;
        crate::agent::record_demotion(&mut writer, 3, &demotion_info()).await;

        let events = read_events(dir.path());
        assert_eq!(events.len(), 1);
        assert_eq!(events[0]["event_type"], "shell_detached");
        assert_eq!(events[0]["work_id"], "wrk_0a1b2c3d4e5f6a7b");
        assert_eq!(events[0]["turn"], 3);
    }

    #[tokio::test]
    async fn an_unrecorded_demotion_names_the_work_id_and_the_write_error() {
        let dir = tempfile::tempdir().unwrap();
        let mut writer = make_writer(dir.path()).await;
        writer
            .write_shell_detach_unrecorded(2, "wrk_0a1b", "/usr/bin/bash", "No space left")
            .await
            .unwrap();

        let events = read_events(dir.path());
        assert_eq!(events.len(), 1);
        assert_eq!(events[0]["event_type"], "shell_detach_unrecorded");
        assert_eq!(events[0]["turn"], 2);
        assert_eq!(events[0]["work_id"], "wrk_0a1b");
        assert_eq!(events[0]["binary"], "/usr/bin/bash");
        assert_eq!(events[0]["reason"], "No space left");
    }

    /// A command still running when the session ended has no exit code, no output file and no
    /// byte count, and its line carries none of the three — omitted rather than `null`, so the
    /// shape is identical to a line written before those fields existed.
    #[tokio::test]
    async fn an_abandoned_command_still_running_omits_the_fields_it_cannot_know() {
        let dir = tempfile::tempdir().unwrap();
        let mut writer = make_writer(dir.path()).await;
        writer
            .write_shell_abandoned(
                "wrk_0a1b2c3d4e5f6a7b",
                "/usr/bin/bash",
                "sleep 45; echo done",
                45_012,
                None,
            )
            .await
            .unwrap();

        let events = read_events(dir.path());
        assert_eq!(events.len(), 1);
        let event = &events[0];
        assert_eq!(event["event_type"], "shell_abandoned");
        assert_eq!(event["work_id"], "wrk_0a1b2c3d4e5f6a7b");
        assert_eq!(event["binary"], "/usr/bin/bash");
        assert_eq!(event["command"], "sleep 45; echo done");
        assert_eq!(event["running_ms"], 45_012);
        for absent in ["exit_code", "output_path", "output_bytes"] {
            assert!(
                event.get(absent).is_none(),
                "{absent} must be omitted, not written as null: {event}"
            );
        }
    }

    /// A command that finished during teardown carries the exit code, the output path and the
    /// byte count that do exist for it, so a trace reader can tell it from one still running
    /// without reading anything else.
    #[tokio::test]
    async fn an_abandoned_command_that_finished_carries_its_result() {
        let dir = tempfile::tempdir().unwrap();
        let mut writer = make_writer(dir.path()).await;
        writer
            .write_shell_abandoned(
                "wrk_0a1b2c3d4e5f6a7b",
                "/usr/bin/bash",
                "sleep 2; echo done",
                2_014,
                Some((0, "logs/wrk_0a1b2c3d4e5f6a7b.log", 91)),
            )
            .await
            .unwrap();

        let event = &read_events(dir.path())[0];
        assert_eq!(event["event_type"], "shell_abandoned");
        assert_eq!(event["running_ms"], 2_014);
        assert_eq!(event["exit_code"], 0);
        assert_eq!(event["output_path"], "logs/wrk_0a1b2c3d4e5f6a7b.log");
        assert_eq!(event["output_bytes"], 91);
    }

    async fn make_writer(dir: &std::path::Path) -> TraceWriter {
        make_writer_with_opts(dir, TraceCapture::Meta).await
    }

    async fn make_writer_with_opts(dir: &std::path::Path, capture: TraceCapture) -> TraceWriter {
        make_writer_with_prompt(dir, capture, None, false).await
    }

    async fn make_writer_with_prompt(
        dir: &std::path::Path,
        capture: TraceCapture,
        system_prompt: Option<&str>,
        system_prompt_overridden: bool,
    ) -> TraceWriter {
        TraceWriter::open(
            dir,
            "test-session-id".to_string(),
            "test-capsule".to_string(),
            "1.0.0".to_string(),
            "claude-test".to_string(),
            vec!["shell".to_string()],
            report_for(
                ContainmentClass::Advisory,
                EnforcementTier::EnvironmentOnly,
                false,
            ),
            capture,
            system_prompt.map(str::to_string),
            system_prompt_overridden,
            None,
            None,
            None,
            None,
        )
        .await
        .unwrap()
    }

    fn read_events(dir: &std::path::Path) -> Vec<Value> {
        let content = std::fs::read_to_string(dir.join("trace.jsonl")).unwrap();
        content
            .lines()
            .filter(|l| !l.is_empty())
            .map(|l| serde_json::from_str(l).unwrap())
            .collect()
    }

    #[tokio::test]
    async fn session_start_fields() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_session_start(10, vec!["bash".to_string()])
            .await
            .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        assert_eq!(events.len(), 1);
        let e = &events[0];
        assert_eq!(e["event_type"], "session_start");
        assert_eq!(e["session_id"], "test-session-id");
        assert_eq!(e["capsule_name"], "test-capsule");
        assert_eq!(e["capsule_version"], "1.0.0");
        assert_eq!(e["model"], "claude-test");
        assert_eq!(e["max_turns"], 10);
        assert_eq!(e["capabilities"], serde_json::json!(["shell"]));
        assert_eq!(e["tools_declared"], serde_json::json!(["bash"]));
        assert_eq!(e["containment_declared"], "advisory");
        assert_eq!(e["containment_achieved"], "advisory");
        assert_eq!(e["workdir_exec"], false);
        assert!(e["timestamp"].as_u64().unwrap() > 0);
    }

    /// The three `system_prompt_source` values, each paired with the hash the same call derives.
    /// A manifest prompt and a CLI override carrying the *same* text differ only in the source
    /// field — which is the whole reason the override flag is threaded through separately rather
    /// than inferred from the resolved prompt.
    #[tokio::test]
    async fn session_start_records_system_prompt_source_and_hash() {
        let expected_sha = murmur_artifact::sha256_hex(b"Be terse.");

        for (prompt, overridden, source, sha) in [
            (Some("Be terse."), false, "manifest", Some(&expected_sha)),
            (Some("Be terse."), true, "cli", Some(&expected_sha)),
            (None, false, "none", None),
            // `--system-prompt ""` cleared the manifest's prompt: no prompt is in effect, but the
            // operator made that call, so the source is the CLI and not `"none"`.
            (None, true, "cli", None),
        ] {
            let dir = tempfile::tempdir().unwrap();
            let mut w =
                make_writer_with_prompt(dir.path(), TraceCapture::Meta, prompt, overridden).await;
            w.write_session_start(1, Vec::new()).await.unwrap();
            w.flush().await.unwrap();

            let e = &read_events(dir.path())[0];
            assert_eq!(e["system_prompt_source"], source, "prompt={prompt:?}");
            match sha {
                Some(sha) => assert_eq!(e["system_prompt_sha256"], sha.as_str()),
                None => assert!(
                    e["system_prompt_sha256"].is_null(),
                    "no prompt in effect must hash to null, got {}",
                    e["system_prompt_sha256"]
                ),
            }
        }
    }

    /// `session_start` never carries the prompt text. The hash is written in every capture
    /// mode; under `content` the bytes behind it are a blob named by that same hash, which is
    /// how a reader gets from the record to the prompt.
    #[tokio::test]
    async fn session_start_records_the_prompt_by_hash_and_stores_it_only_under_content() {
        let sha = murmur_artifact::sha256_hex(b"Be terse.");

        let dir = tempfile::tempdir().unwrap();
        let mut w =
            make_writer_with_prompt(dir.path(), TraceCapture::Meta, Some("Be terse."), false).await;
        w.write_session_start(1, Vec::new()).await.unwrap();
        w.flush().await.unwrap();

        let e = &read_events(dir.path())[0];
        assert!(
            e.get("system_prompt").is_none(),
            "the verbatim prompt is not a session_start field, got {e}"
        );
        assert_eq!(e["system_prompt_sha256"], sha.as_str());
        assert!(
            !dir.path().join(BLOB_DIR_NAME).exists(),
            "meta stores no bodies"
        );

        let opted_in = tempfile::tempdir().unwrap();
        let mut w = make_writer_with_prompt(
            opted_in.path(),
            TraceCapture::Content,
            Some("Be terse."),
            false,
        )
        .await;
        w.write_session_start(1, Vec::new()).await.unwrap();
        w.flush().await.unwrap();

        let event = &read_events(opted_in.path())[0];
        assert!(event.get("system_prompt").is_none());
        assert_eq!(event["system_prompt_sha256"], sha.as_str());
        let blob = opted_in.path().join(BLOB_DIR_NAME).join(&sha);
        assert_eq!(std::fs::read(&blob).unwrap(), b"Be terse.");
    }

    /// Opting in cannot conjure a prompt that was never in effect — and a session with no
    /// prompt and nothing else to store leaves no blob directory at all.
    #[tokio::test]
    async fn content_capture_stores_no_prompt_blob_when_there_is_no_prompt() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer_with_prompt(dir.path(), TraceCapture::Content, None, false).await;
        w.write_session_start(1, Vec::new()).await.unwrap();
        w.flush().await.unwrap();

        let e = &read_events(dir.path())[0];
        assert!(e.get("system_prompt").is_none());
        assert!(e["system_prompt_sha256"].is_null());
        assert!(!dir.path().join(BLOB_DIR_NAME).exists());
    }

    /// The prompt is a session constant, like `model` — every task's `session_start` repeats it.
    #[tokio::test]
    async fn every_task_session_start_repeats_the_same_system_prompt_record() {
        let dir = tempfile::tempdir().unwrap();
        let mut w =
            make_writer_with_prompt(dir.path(), TraceCapture::Meta, Some("Be terse."), true).await;
        w.write_session_start(1, Vec::new()).await.unwrap();
        w.write_session_start(1, Vec::new()).await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        assert_eq!(events.len(), 2);
        assert_eq!(
            events[0]["system_prompt_source"],
            events[1]["system_prompt_source"]
        );
        assert_eq!(
            events[0]["system_prompt_sha256"],
            events[1]["system_prompt_sha256"]
        );
        assert_eq!(events[1]["system_prompt_source"], "cli");
    }

    /// The pairing an operator reads a trace for: `workdir_exec: true` next to the `advisory` it
    /// forces, on a session whose *declared* floor was higher. Without the flag on the record, a
    /// trace showing `achieved: advisory` is indistinguishable from one written on a host with no
    /// Landlock at all.
    #[tokio::test]
    async fn session_start_records_workdir_exec_next_to_the_class_it_forced() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = TraceWriter::open(
            dir.path(),
            "test-session-id".to_string(),
            "test-capsule".to_string(),
            "1.0.0".to_string(),
            "claude-test".to_string(),
            Vec::new(),
            // A host that could back `scoped`, capped to `advisory` by the declaration alone.
            report_for(
                ContainmentClass::Advisory,
                EnforcementTier::KernelFull,
                true,
            ),
            TraceCapture::Meta,
            None,
            false,
            None,
            None,
            None,
            None,
        )
        .await
        .unwrap();
        w.write_session_start(1, Vec::new()).await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        assert_eq!(events[0]["workdir_exec"], true);
        assert_eq!(events[0]["containment_achieved"], "advisory");
    }

    /// A declared floor and an achieved class are recorded independently — the trace shows
    /// what was asked for next to what the host actually gave, not one derived from the other.
    #[tokio::test]
    async fn session_start_records_declared_and_achieved_containment_separately() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = TraceWriter::open(
            dir.path(),
            "test-session-id".to_string(),
            "test-capsule".to_string(),
            "1.0.0".to_string(),
            "claude-test".to_string(),
            Vec::new(),
            report_for(ContainmentClass::Sealed, EnforcementTier::KernelFull, false),
            TraceCapture::Meta,
            None,
            false,
            None,
            None,
            None,
            None,
        )
        .await
        .unwrap();
        w.write_session_start(1, Vec::new()).await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        assert_eq!(events[0]["containment_declared"], "sealed");
        assert_eq!(events[0]["containment_achieved"], "scoped");
    }

    /// The audit property the containment feature exists to provide: two sessions that reached
    /// the *same* achieved class through *different* host permissions must not produce the same
    /// record. A `sealed` result obtained through the shipped AppArmor profile and one obtained on
    /// a host whose unprivileged-userns hardening is switched off for every binary differ in
    /// nothing else the event carries, so `userns_grant` is what keeps the two apart.
    ///
    /// The tier is a literal on both sides, so the only thing that differs is the grant.
    #[tokio::test]
    async fn session_start_distinguishes_two_hosts_with_the_same_achieved_class() {
        async fn event_for(grant: UsernsGrant) -> Value {
            let report = scope_report_for_tier(
                &CapabilityPolicy::default(),
                ContainmentClass::Sealed,
                EnforcementTier::KernelSealed,
                None,
                Some(grant),
                None,
                Vec::new(),
                Vec::new(),
                Vec::new(),
                crate::cgroup::IoMaxReport::default(),
            );
            let dir = tempfile::tempdir().unwrap();
            let mut w = TraceWriter::open(
                dir.path(),
                "test-session-id".to_string(),
                "test-capsule".to_string(),
                "1.0.0".to_string(),
                "claude-test".to_string(),
                Vec::new(),
                report,
                TraceCapture::Meta,
                None,
                false,
                None,
                None,
                None,
                None,
            )
            .await
            .unwrap();
            w.write_session_start(1, Vec::new()).await.unwrap();
            w.flush().await.unwrap();
            let mut event = read_events(dir.path()).remove(0);
            // The one field that legitimately differs between two runs, removed so the assertion
            // below is about the grant and nothing else.
            event.as_object_mut().unwrap().remove("timestamp");
            event
        }

        let through_profile = event_for(UsernsGrant::ProfileConfining).await;
        let host_wide = event_for(UsernsGrant::RestrictionDisabledHostWide).await;

        assert_eq!(through_profile["containment_achieved"], "sealed");
        assert_eq!(host_wide["containment_achieved"], "sealed");
        assert_eq!(through_profile["userns_grant"], "profile_confining");
        assert_eq!(host_wide["userns_grant"], "restriction_disabled_host_wide");
        assert_ne!(
            through_profile, host_wide,
            "two hosts granting the user namespace by different mechanisms must not write the \
             same session_start record"
        );

        // Written for every Linux host, never skipped — its absence identifies a runtime that
        // predates the key rather than a host that was not asked.
        let unprobed = scope_report_for_tier(
            &CapabilityPolicy::default(),
            ContainmentClass::Advisory,
            EnforcementTier::EnvironmentOnly,
            None,
            None,
            None,
            Vec::new(),
            Vec::new(),
            Vec::new(),
            crate::cgroup::IoMaxReport::default(),
        );
        assert!(
            serde_json::to_value(&unprobed).unwrap()["userns_grant"].is_null(),
            "off Linux the key is present and null, not absent"
        );
    }

    /// Central property: `session_start.effective_grants` is the *whole* `ScopeReport`,
    /// serialized byte-for-byte as `mur run --explain-scope --json` prints it — not a re-derived
    /// summary of it. Asserted by comparing against `serde_json::to_value` of the very report
    /// handed to the writer, so any field added to `ScopeReport` later is covered without this
    /// test being touched.
    ///
    /// The tier is a literal, never a probe: this asserts the recording, not the host.
    #[tokio::test]
    async fn session_start_records_the_whole_scope_report_as_effective_grants() {
        let policy = CapabilityPolicy {
            network_allow: vec!["https://api.example.com".to_string()],
            unix_sockets_allowed: false,
            filesystem_scope: Some("workdir".to_string()),
            shell_allow: vec!["python3".to_string()],
            spawn_allow: vec!["helper".to_string()],
            env_allow: vec!["TZ".to_string()],
            shell_interpreter_runtime: vec![InterpreterRuntimeGrant {
                binary: "python3".to_string(),
                dirs: vec![InterpreterRuntimeDir {
                    path: "/usr/lib/python3.11".to_string(),
                    list_dir: true,
                }],
            }],
            ..CapabilityPolicy::default()
        };
        let expected = scope_report_for_tier(
            &policy,
            ContainmentClass::Scoped,
            EnforcementTier::KernelFull,
            None,
            None,
            None,
            // Populated rather than empty so the whole-report assertion below really covers this
            // field: an empty vec would pass whether or not it was written at all.
            vec![crate::containment::StateStoreReport {
                artifact: "notes-tool".to_string(),
                store: "shey".to_string(),
                host_path: "/home/dev/.murmur/state/shey".to_string(),
            }],
            // Populated for the same reason as `state_stores` above.
            vec!["config-echo".to_string()],
            vec![crate::containment::PreopenReport {
                artifact: "notes-tool".to_string(),
                role: "tool".to_string(),
                scope: None,
                surface: crate::containment::PreopenSurface::WholeWorkdir,
            }],
            crate::cgroup::IoMaxReport::default(),
        );

        let dir = tempfile::tempdir().unwrap();
        let mut w = TraceWriter::open(
            dir.path(),
            "test-session-id".to_string(),
            "test-capsule".to_string(),
            "1.0.0".to_string(),
            "claude-test".to_string(),
            vec!["network".to_string(), "shell".to_string()],
            expected.clone(),
            TraceCapture::Meta,
            None,
            false,
            None,
            None,
            None,
            None,
        )
        .await
        .unwrap();
        w.write_session_start(1, Vec::new()).await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        assert_eq!(
            events[0]["effective_grants"],
            serde_json::to_value(&expected).unwrap()
        );

        // The category-name summary that predates this field is unchanged, and the three
        // top-level containment fields agree with the report they are now read from.
        assert_eq!(
            events[0]["capabilities"],
            serde_json::json!(["network", "shell"])
        );
        assert_eq!(
            events[0]["effective_grants"]["state_stores"],
            serde_json::json!([{
                "artifact": "notes-tool",
                "store": "shey",
                "host_path": "/home/dev/.murmur/state/shey",
            }])
        );
        // Names only: what an artifact is configured *with* is operator plaintext and reaches
        // no trace record.
        assert_eq!(
            events[0]["effective_grants"]["configured_artifacts"],
            serde_json::json!(["config-echo"])
        );
        // The per-artifact preopen set reaches the trace with no trace-side plumbing: the report
        // is cloned whole, so `filesystem_scope` above (which narrows no guest) and the surface a
        // guest is actually opened into are both on the record.
        assert_eq!(
            events[0]["effective_grants"]["preopens"],
            serde_json::json!([{
                "artifact": "notes-tool",
                "role": "tool",
                "scope": null,
                "surface": "whole-workdir",
            }])
        );
        assert_eq!(events[0]["containment_declared"], "scoped");
        assert_eq!(events[0]["containment_achieved"], "scoped");
        assert_eq!(events[0]["workdir_exec"], false);
        assert_eq!(
            events[0]["effective_grants"]["network_allow"],
            serde_json::json!(["https://api.example.com"])
        );
        assert_eq!(
            events[0]["effective_grants"]["interpreter_runtime_grants"],
            serde_json::json!(["python3: /usr/lib/python3.11 (list_dir)"])
        );
    }

    #[tokio::test]
    async fn inference_fields_and_snake_case() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_inference(
            0,
            Some(100),
            Some(50),
            "tool_call".to_string(),
            Some("tool_call"),
            Some("bash".to_string()),
            None,
            None,
            vec!["msg_0000000000000000000000000000000a".to_string()],
            None,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let e = &events[0];
        assert_eq!(e["event_type"], "inference");
        assert_eq!(
            e["message_ids"],
            serde_json::json!(["msg_0000000000000000000000000000000a"]),
            "an agent-loop turn names the messages its request embedded"
        );
        assert_eq!(e["turn"], 0);
        assert_eq!(e["input_tokens"], 100);
        assert_eq!(e["output_tokens"], 50);
        assert_eq!(e["decision"], "tool_call");
        assert_eq!(e["tool_name"], "bash");
        // Confirm snake_case (not camelCase)
        assert!(e.get("inputTokens").is_none(), "must use snake_case");
        assert!(e.get("toolName").is_none(), "must use snake_case");
        // No driver reported usage, so no key claims one was reported.
        for key in [
            "input_tokens_actual",
            "output_tokens_actual",
            "cached_tokens",
            "cache_write_tokens",
        ] {
            assert!(e.get(key).is_none(), "{key} must be omitted, not null or 0");
        }
    }

    /// A partially-populated `usage` writes only the members the driver reported: a member it
    /// left out stays off the line entirely rather than being recorded as a zero.
    #[tokio::test]
    async fn inference_writes_only_the_reported_usage_members() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        let usage = DriverUsage {
            input_tokens: Some(12043),
            output_tokens: None,
            cached_tokens: Some(0),
            cache_write_tokens: None,
            thinking_tokens: None,
        };
        w.write_inference(
            0,
            Some(100),
            Some(50),
            "end_turn".to_string(),
            Some("end_turn"),
            None,
            None,
            Some(&usage),
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let e = &events[0];
        assert_eq!(e["input_tokens"], 100, "the estimate is untouched");
        assert_eq!(e["input_tokens_actual"], 12043);
        assert_eq!(e["cached_tokens"], 0, "a reported zero is written");
        assert!(e.get("output_tokens_actual").is_none());
        assert!(e.get("cache_write_tokens").is_none());
    }

    #[tokio::test]
    async fn tool_call_fields() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_tool_call(
            1,
            "bash".to_string(),
            None,
            serde_json::json!({"command": "echo hi"}),
            42,
            "hi\n",
            100,
            55,
            "ok".to_string(),
            None,
            None,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let e = &events[0];
        assert_eq!(e["event_type"], "tool_call");
        assert_eq!(e["turn"], 1);
        assert_eq!(e["tool_name"], "bash");
        assert_eq!(e["input_bytes"], 42);
        assert_eq!(e["output_bytes"], 100);
        assert_eq!(e["duration_ms"], 55);
        assert_eq!(e["status"], "ok");
        assert!(
            e.get("state_effect").is_none(),
            "state_effect must be omitted when the tool declared nothing"
        );
    }

    #[tokio::test]
    async fn tool_call_records_declared_state_effect() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_tool_call(
            0,
            "murmur-tool-editor".to_string(),
            None,
            serde_json::json!({"operation": "read_file", "path": "a.txt"}),
            10,
            "hi\n",
            3,
            5,
            "ok".to_string(),
            Some("read".to_string()),
            None,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        assert_eq!(
            events[0]["state_effect"], "read",
            "declared state_effect must be recorded verbatim on the tool_call event"
        );
    }

    #[tokio::test]
    async fn tool_call_records_declared_resource_id() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_tool_call(
            0,
            "murmur-tool-code-graph".to_string(),
            None,
            serde_json::json!({"symbol": "Foo::bar"}),
            10,
            "hi\n",
            3,
            5,
            "ok".to_string(),
            Some("read".to_string()),
            Some("sym:Foo".to_string()),
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        assert_eq!(
            events[0]["resource_id"], "sym:Foo",
            "declared resource_id must reach trace.jsonl verbatim, unparsed"
        );
    }

    #[tokio::test]
    async fn tool_call_omits_undeclared_resource_id() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_tool_call(
            0,
            "bash".to_string(),
            None,
            serde_json::json!({"command": "echo hi"}),
            10,
            "hi\n",
            3,
            5,
            "ok".to_string(),
            None,
            None,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        assert!(
            events[0].get("resource_id").is_none(),
            "resource_id must be omitted entirely (not null) when the tool declared nothing"
        );
    }

    #[tokio::test]
    async fn shell_fields() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_shell(
            0,
            "/usr/bin/echo".to_string(),
            "echo hi".to_string(),
            0,
            7,
            0,
            10,
            None,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let e = &events[0];
        assert_eq!(e["event_type"], "shell");
        // `binary` names what ran; `command` still carries only the argument list, so
        // neither can be derived from the other.
        assert_eq!(e["binary"], "/usr/bin/echo");
        assert_eq!(e["command"], "echo hi");
        assert_eq!(e["exit_code"], 0);
        assert_eq!(e["stdout_bytes"], 7);
        assert_eq!(e["stderr_bytes"], 0);
        assert_eq!(e["duration_ms"], 10);
    }

    #[tokio::test]
    async fn compaction_fields() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_compaction(3, 80000, 20000).await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let e = &events[0];
        assert_eq!(e["event_type"], "compaction");
        assert_eq!(e["turn"], 3);
        assert_eq!(e["tokens_before"], 80000);
        assert_eq!(e["tokens_after"], 20000);
    }

    #[tokio::test]
    async fn hook_dispatch_error_fields() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_hook_dispatch_error("my-hook", "on-tool-call", "write-manifests")
            .await
            .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let e = &events[0];
        assert_eq!(e["event_type"], "hook_dispatch_error");
        assert_eq!(e["session_id"], "test-session-id");
        assert_eq!(e["hook_name"], "my-hook");
        assert_eq!(e["event"], "on-tool-call");
        assert_eq!(e["arm"], "write-manifests");
        assert!(e["timestamp"].as_u64().unwrap() > 0);
        // snake_case checks
        assert!(e.get("hookName").is_none(), "must use snake_case");
    }

    #[tokio::test]
    async fn session_end_fields_and_snake_case() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_inference(
            0,
            Some(100),
            Some(50),
            "tool_call".to_string(),
            Some("tool_call"),
            None,
            None,
            None,
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        w.write_tool_call(
            0,
            "bash".to_string(),
            None,
            serde_json::json!({"command": "ls"}),
            10,
            "file.txt\n",
            20,
            5,
            "ok".to_string(),
            None,
            None,
        )
        .await
        .unwrap();
        w.write_shell(0, "/bin/ls".to_string(), "ls".to_string(), 0, 3, 0, 2, None)
            .await
            .unwrap();
        w.write_session_end("ok").await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let se = events.last().unwrap();
        assert_eq!(se["event_type"], "session_end");
        assert_eq!(se["total_turns"], 1);
        assert_eq!(se["total_input_tokens"], 100);
        assert_eq!(se["total_output_tokens"], 50);
        assert_eq!(se["total_tool_calls"], 1);
        assert_eq!(se["total_shell_calls"], 1);
        assert_eq!(se["exit_status"], "ok");
        assert!(se["duration_ms"].as_u64().is_some());
        // snake_case checks
        assert!(se.get("exitStatus").is_none(), "must use snake_case");
        assert!(se.get("totalTurns").is_none(), "must use snake_case");
    }

    #[tokio::test]
    async fn write_session_end_if_not_ended_is_noop_when_already_ended() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_session_start(10, vec![]).await.unwrap();
        w.write_session_end("ok").await.unwrap();
        w.write_session_end_if_not_ended("failed").await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        assert_eq!(
            events
                .iter()
                .filter(|e| e["event_type"] == "session_end")
                .count(),
            1,
            "session_end should appear exactly once"
        );
    }

    #[tokio::test]
    async fn write_session_end_if_not_ended_writes_when_started_not_ended() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_session_start(10, vec![]).await.unwrap();
        // Simulate error path — session_end_if_not_ended in launch_session
        w.write_session_end_if_not_ended("failed").await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let se = events
            .iter()
            .find(|e| e["event_type"] == "session_end")
            .unwrap();
        assert_eq!(se["exit_status"], "failed");
    }

    #[tokio::test]
    async fn no_session_end_written_when_session_never_started() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        // session_start never written — fallback should be no-op
        w.write_session_end_if_not_ended("failed").await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        assert!(
            events.iter().all(|e| e["event_type"] != "session_end"),
            "session_end must not appear when session_start was never written"
        );
    }

    #[tokio::test]
    async fn session_id_on_every_event() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_session_start(10, vec!["bash".to_string()])
            .await
            .unwrap();
        w.write_inference(
            0,
            Some(10),
            Some(5),
            "end_turn".to_string(),
            Some("end_turn"),
            None,
            None,
            None,
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        w.write_session_end("ok").await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        for e in &events {
            assert_eq!(e["session_id"], "test-session-id");
        }
    }

    #[tokio::test]
    async fn counts_match_per_event_sum() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_session_start(10, vec![]).await.unwrap();
        w.write_inference(
            0,
            Some(50),
            Some(25),
            "tool_call".to_string(),
            Some("tool_call"),
            Some("bash".to_string()),
            None,
            None,
            vec!["msg_0000000000000000000000000000000a".to_string()],
            None,
        )
        .await
        .unwrap();
        w.write_tool_call(
            0,
            "bash".to_string(),
            None,
            serde_json::json!({"command": "echo"}),
            5,
            "ok",
            10,
            3,
            "ok".to_string(),
            None,
            None,
        )
        .await
        .unwrap();
        w.write_shell(
            0,
            "/bin/echo".to_string(),
            "echo".to_string(),
            0,
            4,
            0,
            1,
            None,
        )
        .await
        .unwrap();
        w.write_inference(
            1,
            Some(60),
            Some(30),
            "end_turn".to_string(),
            Some("end_turn"),
            None,
            None,
            None,
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        w.write_session_end("ok").await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let inference_count = events
            .iter()
            .filter(|e| e["event_type"] == "inference")
            .count();
        let tool_count = events
            .iter()
            .filter(|e| e["event_type"] == "tool_call")
            .count();
        let shell_count = events.iter().filter(|e| e["event_type"] == "shell").count();

        let se = events
            .iter()
            .find(|e| e["event_type"] == "session_end")
            .unwrap();
        assert_eq!(
            se["total_turns"].as_u64().unwrap() as usize,
            inference_count
        );
        assert_eq!(
            se["total_tool_calls"].as_u64().unwrap() as usize,
            tool_count
        );
        assert_eq!(
            se["total_shell_calls"].as_u64().unwrap() as usize,
            shell_count
        );
    }

    #[tokio::test]
    async fn tool_call_input_always_present() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await; // TraceCapture::Meta
        w.write_tool_call(
            0,
            "bash".to_string(),
            None,
            serde_json::json!({"command": "ls -la"}),
            20,
            "total 8\n",
            80,
            12,
            "ok".to_string(),
            None,
            None,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let e = &events[0];
        assert_eq!(e["event_type"], "tool_call");
        assert_eq!(e["input"]["command"], "ls -la");
        assert!(e.get("input").is_some(), "input must always be present");
    }

    #[tokio::test]
    async fn tool_call_output_absent_by_default() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await; // TraceCapture::Meta
        w.write_tool_call(
            0,
            "bash".to_string(),
            None,
            serde_json::json!({"command": "ls"}),
            10,
            "file.txt\n",
            90,
            8,
            "ok".to_string(),
            None,
            None,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let e = &events[0];
        assert!(
            e.get("output").is_none(),
            "output must be absent under a capture mode that stores no bodies"
        );
        assert_eq!(e["output_bytes"], 90, "output_bytes must still be recorded");
    }

    #[tokio::test]
    async fn skill_call_writes_skill_call_event() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_skill_call(
            2,
            "my-skill".to_string(),
            &LockOrigin::Operator,
            1024,
            8,
            "ok".to_string(),
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let e = &events[0];
        assert_eq!(e["event_type"], "skill_call");
        assert_eq!(e["turn"], 2);
        assert_eq!(e["skill_name"], "my-skill");
        assert_eq!(e["output_bytes"], 1024);
        assert_eq!(e["duration_ms"], 8);
        assert_eq!(e["status"], "ok");
    }

    #[tokio::test]
    async fn skill_call_does_not_increment_total_tool_calls() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_session_start(10, vec![]).await.unwrap();
        w.write_tool_call(
            0,
            "real-tool".to_string(),
            None,
            serde_json::json!({}),
            2,
            "data",
            4,
            5,
            "ok".to_string(),
            None,
            None,
        )
        .await
        .unwrap();
        w.write_skill_call(
            1,
            "my-skill".to_string(),
            &LockOrigin::Operator,
            512,
            3,
            "ok".to_string(),
        )
        .await
        .unwrap();
        w.write_session_end("ok").await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let se = events
            .iter()
            .find(|e| e["event_type"] == "session_end")
            .unwrap();
        assert_eq!(
            se["total_tool_calls"], 1,
            "skill_call must not inflate total_tool_calls"
        );
    }

    #[tokio::test]
    async fn tool_call_output_present_when_opted_in() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer_with_opts(dir.path(), TraceCapture::Content).await;
        w.write_tool_call(
            0,
            "bash".to_string(),
            None,
            serde_json::json!({"command": "echo hello"}),
            22,
            "hello\n",
            60,
            5,
            "ok".to_string(),
            None,
            None,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let e = &events[0];
        assert_eq!(
            e["output"].as_str().unwrap(),
            "hello\n",
            "output must be present under capture: content"
        );
        assert_eq!(e["output_bytes"], 60);
    }

    #[tokio::test]
    async fn task_end_carries_reopen_count() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_task_start("tsk_1", "ctx_1", "a2a", event_provenance(), None, 3)
            .await
            .unwrap();
        w.write_task_end("tsk_1", "ok", 2).await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let end = events
            .iter()
            .find(|e| e["event_type"] == "task_end")
            .unwrap();
        assert_eq!(end["task_id"], "tsk_1");
        assert_eq!(end["exit_status"], "ok");
        assert_eq!(end["reopen_count"], 2);
    }

    #[tokio::test]
    async fn task_reopened_names_hook_reason_and_ordinal() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_task_start("tsk_1", "ctx_1", "a2a", event_provenance(), None, 3)
            .await
            .unwrap();
        w.write_task_reopened(
            "tsk_1",
            "gatekeeper",
            "tests still fail",
            1,
            REOPEN_CONTEXT_CONTINUED,
            7,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let re = events
            .iter()
            .find(|e| e["event_type"] == "task_reopened")
            .unwrap();
        assert_eq!(re["task_id"], "tsk_1");
        assert_eq!(re["hook_name"], "gatekeeper");
        assert_eq!(re["reason"], "tests still fail");
        assert_eq!(re["reopen_number"], 1);
        assert_eq!(re["attempt_context"], "continued");
        assert_eq!(re["turns_remaining"], 7);
        assert!(re["timestamp"].as_u64().unwrap() > 0);
    }

    /// `task_turns()` reports the cumulative per-task turn count the reopen loop reads to
    /// compute the remaining `max_turns` budget: it advances with each `write_inference`
    /// and resets on `write_task_start`.
    #[tokio::test]
    async fn task_turns_accessor_tracks_and_resets() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_task_start("tsk_1", "ctx_1", "a2a", event_provenance(), None, 3)
            .await
            .unwrap();
        assert_eq!(w.task_turns(), 0);
        w.write_inference(
            0,
            Some(10),
            Some(5),
            "end_turn".to_string(),
            Some("end_turn"),
            None,
            None,
            None,
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        w.write_inference(
            1,
            Some(10),
            Some(5),
            "end_turn".to_string(),
            Some("end_turn"),
            None,
            None,
            None,
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        assert_eq!(w.task_turns(), 2);
        // A new task resets the per-task counter.
        w.write_task_start("tsk_2", "ctx_2", "a2a", event_provenance(), None, 3)
            .await
            .unwrap();
        assert_eq!(w.task_turns(), 0);
    }

    // ── Event identity and parenting ──────────────────────────────────────────

    /// The full tree one task produces, walked the way a reader walks it: every line is
    /// identified, every non-null `parent_id` names a line already written, and the shape is
    /// session → task → turn → {tool_call, shell}.
    #[tokio::test]
    async fn every_event_carries_a_unique_id_and_parents_to_an_earlier_event() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_session_start(10, vec!["bash".to_string()])
            .await
            .unwrap();
        w.write_task_start("tsk_1", "ctx_1", "a2a", event_provenance(), None, 3)
            .await
            .unwrap();
        w.write_inference(
            0,
            Some(10),
            Some(5),
            "tool_call".to_string(),
            Some("tool_call"),
            None,
            None,
            None,
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        w.write_tool_call(
            0,
            "bash".to_string(),
            Some("toolu_1".to_string()),
            serde_json::json!({"command": "echo hi"}),
            10,
            "hi\n",
            3,
            5,
            "ok".to_string(),
            None,
            None,
        )
        .await
        .unwrap();
        w.write_shell(
            0,
            "/bin/bash".to_string(),
            "echo hi".to_string(),
            0,
            3,
            0,
            2,
            None,
        )
        .await
        .unwrap();
        w.write_artifact_pulled(&runtime_summary("pulled-style", ArtifactRuntime::Skill))
            .await
            .unwrap();
        w.write_task_end("tsk_1", "ok", 0).await.unwrap();
        w.write_session_end("ok").await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let mut seen: std::collections::HashSet<String> = std::collections::HashSet::new();
        for (i, event) in events.iter().enumerate() {
            let id = event["event_id"].as_str().unwrap();
            assert!(id.starts_with("evt_") && id.len() == 36, "event {i}: {id}");
            assert!(id[4..].chars().all(|c| c.is_ascii_hexdigit()), "event {i}");
            assert!(seen.insert(id.to_string()), "event {i} reuses {id}");
            assert!(
                event.get("parent_id").is_some(),
                "event {i} has no parent_id"
            );
            if let Some(parent) = event["parent_id"].as_str() {
                assert!(seen.contains(parent), "event {i} names dangling {parent}");
            }
        }

        let by_type = |ty: &str| {
            events
                .iter()
                .find(|e| e["event_type"] == ty)
                .unwrap_or_else(|| panic!("no {ty}"))
                .clone()
        };
        let session = by_type("session_start");
        assert!(session["parent_id"].is_null());
        assert_eq!(session["event_id"], w.session_event_id());

        let task = by_type("task_start");
        assert_eq!(task["parent_id"], session["event_id"]);

        let turn = by_type("inference");
        assert_eq!(turn["parent_id"], task["event_id"]);

        assert_eq!(by_type("tool_call")["parent_id"], turn["event_id"]);
        assert_eq!(by_type("shell")["parent_id"], turn["event_id"]);
        assert_eq!(by_type("task_end")["parent_id"], task["event_id"]);
        assert_eq!(by_type("session_end")["parent_id"], session["event_id"]);
        assert_eq!(by_type("artifact_pulled")["parent_id"], session["event_id"]);
    }

    fn runtime_summary(name: &str, runtime: ArtifactRuntime) -> InstalledArtifactSummary {
        InstalledArtifactSummary {
            name: name.to_string(),
            version: "1.0.0".to_string(),
            runtime,
            implementation: None,
            origin: LockOrigin::Runtime {
                session: "ses_puller".to_string(),
            },
        }
    }

    fn operator_summary(name: &str, runtime: ArtifactRuntime) -> InstalledArtifactSummary {
        InstalledArtifactSummary {
            origin: LockOrigin::Operator,
            ..runtime_summary(name, runtime)
        }
    }

    /// `runtime_artifacts` is always written: `[]` for a session whose every artifact the operator
    /// pinned, including one that never called the setter.
    #[tokio::test]
    async fn runtime_origin_session_start_lists_nothing_for_an_operator_session() {
        for set in [false, true] {
            let dir = tempfile::tempdir().unwrap();
            let mut w = make_writer(dir.path()).await;
            if set {
                w.set_runtime_artifacts(&[
                    operator_summary("house-style", ArtifactRuntime::Skill),
                    operator_summary("fetcher", ArtifactRuntime::Tool),
                ]);
            }
            w.write_session_start(10, Vec::new()).await.unwrap();
            w.flush().await.unwrap();
            let events = read_events(dir.path());
            assert_eq!(events[0]["runtime_artifacts"], serde_json::json!([]));
        }
    }

    /// Only runtime-origin entries are listed, in the order staging recorded them, each naming
    /// its puller.
    #[tokio::test]
    async fn runtime_origin_session_start_lists_runtime_artifacts_in_order() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.set_runtime_artifacts(&[
            runtime_summary("zeta-tool", ArtifactRuntime::Tool),
            operator_summary("house-style", ArtifactRuntime::Skill),
            InstalledArtifactSummary {
                origin: LockOrigin::Runtime {
                    session: "ses_other".to_string(),
                },
                ..runtime_summary("alpha-style", ArtifactRuntime::Skill)
            },
        ]);
        w.write_session_start(10, Vec::new()).await.unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        assert_eq!(
            events[0]["runtime_artifacts"],
            serde_json::json!([
                {"name": "zeta-tool", "version": "1.0.0", "origin": "runtime",
                 "session": "ses_puller", "trust": "untrusted"},
                {"name": "alpha-style", "version": "1.0.0", "origin": "runtime",
                 "session": "ses_other", "trust": "untrusted"},
            ])
        );
    }

    #[tokio::test]
    async fn runtime_origin_skill_call_carries_origin_and_trust() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        for origin in [
            LockOrigin::Operator,
            LockOrigin::Runtime {
                session: "ses_puller".to_string(),
            },
        ] {
            w.write_skill_call(0, "a-skill".to_string(), &origin, 1, 1, "ok".to_string())
                .await
                .unwrap();
        }
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        assert_eq!(events[0]["origin"], "operator");
        assert_eq!(events[0]["trust"], "trusted");
        assert_eq!(events[1]["origin"], "runtime");
        assert_eq!(events[1]["trust"], "untrusted");
    }

    /// Every key is written, `session` included: `null` when the lock entry is the operator's.
    #[tokio::test]
    async fn runtime_origin_artifact_pulled_writes_every_key() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_artifact_pulled(&runtime_summary("pulled-style", ArtifactRuntime::Skill))
            .await
            .unwrap();
        w.write_artifact_pulled(&operator_summary("fetcher", ArtifactRuntime::Tool))
            .await
            .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let runtime = &events[0];
        assert_eq!(runtime["event_type"], "artifact_pulled");
        assert_eq!(runtime["name"], "pulled-style");
        assert_eq!(runtime["version"], "1.0.0");
        assert_eq!(runtime["runtime"], "skill");
        assert_eq!(runtime["origin"], "runtime");
        assert_eq!(runtime["session"], "ses_puller");
        assert_eq!(runtime["trust"], "untrusted");
        assert!(
            runtime["parent_id"].is_null(),
            "no session_start was written"
        );

        let operator = &events[1];
        assert_eq!(operator["runtime"], "tool");
        assert_eq!(operator["origin"], "operator");
        assert!(operator.get("session").is_some_and(Value::is_null));
        assert_eq!(operator["trust"], "trusted");
    }

    /// A hook's `run-inference` record is a product of the turn it ran inside, not a turn of
    /// its own: it parents to the agent loop's inference and never replaces it as the node the
    /// turn's other events hang off.
    #[tokio::test]
    async fn hook_origin_inference_hangs_off_the_turn_rather_than_becoming_one() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_session_start(10, Vec::new()).await.unwrap();
        w.write_task_start("tsk_1", "ctx_1", "a2a", event_provenance(), None, 3)
            .await
            .unwrap();
        w.write_inference(
            0,
            Some(10),
            Some(5),
            "tool_call".to_string(),
            Some("tool_call"),
            None,
            None,
            None,
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        let origin = InferenceOrigin {
            source: "hook:gatekeeper".to_string(),
            model: "claude-test".to_string(),
        };
        w.write_inference(
            0,
            Some(1),
            Some(1),
            "end_turn".to_string(),
            // A hook's `run-inference` saw no provider stop reason, so it records none.
            None,
            None,
            Some(&origin),
            None,
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let turn = &events[2];
        let hook = &events[3];
        assert!(turn["origin"].is_null());
        assert_eq!(hook["origin"], "hook:gatekeeper");
        assert_eq!(hook["parent_id"], turn["event_id"]);
        // The hook sent a message list the runtime never minted, so the key is absent rather
        // than empty.
        assert!(hook.get("message_ids").is_none(), "{hook}");
    }

    /// A writer with no session frame behind it names no parent. This is the script-capsule
    /// `a2a_send` drain, which opens a writer purely to flush buffered sends into a file that
    /// has no `session_start` line — naming one would dangle.
    #[tokio::test]
    async fn a2a_send_parents_to_null_without_a_session_frame_and_to_it_with_one() {
        let bare = tempfile::tempdir().unwrap();
        let mut w = make_writer(bare.path()).await;
        w.write_a2a_send(
            "http://peer",
            "msg_1",
            "tsk_1",
            "ctx_1",
            None,
            TrustClass::Untrusted,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();
        assert!(read_events(bare.path())[0]["parent_id"].is_null());

        let framed = tempfile::tempdir().unwrap();
        let mut w = make_writer(framed.path()).await;
        w.write_session_start(1, Vec::new()).await.unwrap();
        w.write_a2a_send(
            "http://peer",
            "msg_1",
            "tsk_1",
            "ctx_1",
            None,
            TrustClass::Untrusted,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();
        let events = read_events(framed.path());
        assert_eq!(events[1]["parent_id"], events[0]["event_id"]);
    }

    /// `task_rejected` carries every field and, with no task active, hangs off `session_start`.
    #[tokio::test]
    async fn task_rejected_names_the_task_source_cause_and_reason_under_the_session() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_session_start(10, Vec::new()).await.unwrap();
        w.write_task_start("tsk_ran", "ctx_ran", "a2a", event_provenance(), None, 3)
            .await
            .unwrap();
        w.write_task_end("tsk_ran", "ok", 0).await.unwrap();
        w.write_task_rejected(
            "tsk_1",
            "ctx_1",
            "detached_shell",
            TASK_REJECTED_SESSION_ENDED,
            "task rejected: the session ended before this task started",
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let session = events
            .iter()
            .find(|e| e["event_type"] == "session_start")
            .unwrap();
        let rejected: Vec<_> = events
            .iter()
            .filter(|e| e["event_type"] == "task_rejected")
            .collect();
        assert_eq!(rejected.len(), 1);
        let rejected = rejected[0];
        assert_eq!(rejected["parent_id"], session["event_id"], "{rejected}");
        assert_eq!(rejected["session_id"], session["session_id"]);
        assert_eq!(rejected["task_id"], "tsk_1");
        assert_eq!(rejected["context_id"], "ctx_1");
        assert_eq!(rejected["source"], "detached_shell");
        assert_eq!(rejected["cause"], "session_ended");
        assert_eq!(
            rejected["reason"],
            "task rejected: the session ended before this task started"
        );
        assert!(rejected["event_id"].is_string());
        assert!(rejected["timestamp"].is_u64());
        assert_eq!(rejected.as_object().unwrap().len(), 10, "{rejected}");
    }

    /// `task_failed` hangs off the task it failed, and a reason past the cap is cut on a character
    /// boundary. Both optional fields are absent, not null, when there is nothing to say.
    #[tokio::test]
    async fn task_failed_names_the_task_turn_cause_and_capped_reason() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_session_start(10, Vec::new()).await.unwrap();
        assert_eq!(w.last_task_failure(), None);
        w.write_task_failed(None, TASK_FAILED_RUNTIME_ERROR, "before any task")
            .await
            .unwrap();
        w.write_task_start("tsk_1", "ctx_1", "a2a", event_provenance(), None, 3)
            .await
            .unwrap();
        let long = "é".repeat(MAX_TASK_FAILURE_REASON_BYTES);
        w.write_task_failed(Some(2), TASK_FAILED_DRIVER_ERROR, &long)
            .await
            .unwrap();
        w.write_task_end("tsk_1", "failed", 0).await.unwrap();
        w.flush().await.unwrap();

        assert_eq!(w.task_failures_written(), 2);
        let events = read_events(dir.path());
        let failed: Vec<_> = events
            .iter()
            .filter(|e| e["event_type"] == "task_failed")
            .collect();
        assert_eq!(failed.len(), 2);

        assert!(failed[0].get("task_id").is_none(), "{}", failed[0]);
        assert!(failed[0].get("turn").is_none(), "{}", failed[0]);
        assert_eq!(failed[0]["cause"], "runtime_error");
        assert_eq!(failed[0]["reason"], "before any task");

        let start = events
            .iter()
            .find(|e| e["event_type"] == "task_start")
            .unwrap();
        assert_eq!(failed[1]["task_id"], "tsk_1");
        assert_eq!(failed[1]["turn"], 2);
        assert_eq!(failed[1]["cause"], "driver_error");
        assert_eq!(failed[1]["parent_id"], start["event_id"]);
        assert_eq!(failed[1]["session_id"], start["session_id"]);
        assert!(failed[1]["event_id"].is_string());
        assert!(failed[1]["timestamp"].is_u64());
        let reason = failed[1]["reason"].as_str().unwrap();
        assert!(reason.len() <= MAX_TASK_FAILURE_REASON_BYTES);
        assert!(reason.len() > MAX_TASK_FAILURE_REASON_BYTES - 2);
        assert_eq!(w.last_task_failure(), Some(reason));
    }

    /// Turn-level events carry the enclosing task's id, and `null` once the task has ended —
    /// `task_id` is a record of scope, not a value the writer holds on to.
    #[tokio::test]
    async fn turn_level_events_carry_the_active_task_id() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_session_start(10, Vec::new()).await.unwrap();
        w.write_task_start("tsk_1", "ctx_1", "a2a", event_provenance(), None, 3)
            .await
            .unwrap();
        w.write_inference(
            0,
            Some(10),
            Some(5),
            "tool_call".to_string(),
            Some("tool_call"),
            None,
            None,
            None,
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        w.write_skill_call(
            0,
            "house-style".to_string(),
            &LockOrigin::Operator,
            12,
            1,
            "ok".to_string(),
        )
        .await
        .unwrap();
        w.write_compaction(0, 100, 40).await.unwrap();
        w.write_task_end("tsk_1", "ok", 0).await.unwrap();
        w.write_inference(
            1,
            Some(1),
            Some(1),
            "end_turn".to_string(),
            Some("end_turn"),
            None,
            None,
            None,
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        for event in events
            .iter()
            .filter(|e| {
                matches!(
                    e["event_type"].as_str(),
                    Some("inference" | "skill_call" | "compaction")
                )
            })
            .take(3)
        {
            assert_eq!(event["task_id"], "tsk_1", "{event}");
        }
        let after = events.last().unwrap();
        assert_eq!(after["event_type"], "inference");
        assert!(
            after["task_id"].is_null(),
            "an inference outside any task carries a null task_id, not the last task's"
        );
        // Once the task has ended, its events fall back to the session node.
        assert_eq!(after["parent_id"], events[0]["event_id"]);
    }

    /// The provider's id is recorded verbatim; an absent or empty one records as `null`, never
    /// as an id the provider never issued.
    #[tokio::test]
    async fn tool_call_records_the_provider_id_verbatim_and_empty_as_null() {
        for (given, expected) in [
            (Some("toolu_1".to_string()), Some("toolu_1")),
            (Some(String::new()), None),
            (None, None),
        ] {
            let dir = tempfile::tempdir().unwrap();
            let mut w = make_writer(dir.path()).await;
            w.write_tool_call(
                0,
                "bash".to_string(),
                given.clone(),
                serde_json::json!({}),
                2,
                "",
                0,
                1,
                "ok".to_string(),
                None,
                None,
            )
            .await
            .unwrap();
            w.flush().await.unwrap();
            let e = &read_events(dir.path())[0];
            match expected {
                Some(id) => assert_eq!(e["tool_call_id"], id, "given={given:?}"),
                None => assert!(e["tool_call_id"].is_null(), "given={given:?}"),
            }
        }
    }

    /// `tools_refreshed` carries the turn, the task, the trigger and three sorted name lists,
    /// hangs off the task node rather than the previous turn, and `session_start` records the
    /// declared trigger — `null` under `transport: process`.
    #[tokio::test]
    async fn tools_refreshed_event_records_trigger_names_and_task_parent() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.set_tool_refresh(Some("immediate"));
        w.write_session_start(10, Vec::new()).await.unwrap();
        w.write_task_start("tsk_1", "ctx_1", "a2a", event_provenance(), None, 3)
            .await
            .unwrap();
        w.write_inference(
            0,
            Some(10),
            Some(5),
            "tool_call".to_string(),
            Some("tool_call"),
            None,
            None,
            None,
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        w.write_tools_refreshed(
            1,
            "immediate",
            &["aaa-late-skill".to_string()],
            &[],
            &[
                "aaa-late-skill".to_string(),
                "zzz-existing-tool".to_string(),
            ],
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        assert_eq!(events[0]["tool_refresh"], "immediate");
        let task = events
            .iter()
            .find(|e| e["event_type"] == "task_start")
            .unwrap();
        let refreshed: Vec<&Value> = events
            .iter()
            .filter(|e| e["event_type"] == "tools_refreshed")
            .collect();
        assert_eq!(refreshed.len(), 1);
        let r = refreshed[0];
        assert!(r["event_id"].as_str().unwrap().starts_with("evt_"));
        assert_eq!(r["parent_id"], task["event_id"]);
        assert_eq!(r["session_id"], "test-session-id");
        assert!(r["timestamp"].as_u64().is_some());
        assert_eq!(r["turn"], 1);
        assert_eq!(r["task_id"], "tsk_1");
        assert_eq!(r["trigger"], "immediate");
        assert_eq!(r["added"], serde_json::json!(["aaa-late-skill"]));
        assert_eq!(r["removed"], serde_json::json!([]));
        assert_eq!(
            r["tools"],
            serde_json::json!(["aaa-late-skill", "zzz-existing-tool"])
        );

        let process_dir = tempfile::tempdir().unwrap();
        let mut process = make_writer(process_dir.path()).await;
        process.set_tool_refresh(None);
        process.write_session_start(10, Vec::new()).await.unwrap();
        process.flush().await.unwrap();
        let start = &read_events(process_dir.path())[0];
        assert!(start.as_object().unwrap().contains_key("tool_refresh"));
        assert!(start["tool_refresh"].is_null());
    }

    /// A declined compaction is a full record of the decline, not a bare marker: it names the
    /// turn, the task, the occupancy the session went on running over, and the reason.
    #[tokio::test]
    async fn compaction_declined_records_turn_task_tokens_and_reason() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_session_start(10, Vec::new()).await.unwrap();
        w.write_task_start("tsk_1", "ctx_1", "a2a", event_provenance(), None, 3)
            .await
            .unwrap();
        w.write_inference(
            2,
            Some(10),
            Some(5),
            "tool_call".to_string(),
            Some("tool_call"),
            None,
            None,
            None,
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        w.write_compaction_declined(2, 4321, COMPACTION_DECLINED_UNRESOLVED_TOOL_CALL)
            .await
            .unwrap();
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        let declined: Vec<&Value> = events
            .iter()
            .filter(|e| e["event_type"] == "compaction_declined")
            .collect();
        assert_eq!(declined.len(), 1);
        let d = declined[0];
        assert_eq!(d["reason"], "unresolved_tool_call");
        assert_eq!(d["turn"], 2);
        assert_eq!(d["task_id"], "tsk_1");
        assert_eq!(d["tokens"], 4321);
        assert!(d["event_id"].as_str().unwrap().starts_with("evt_"));
        assert_eq!(
            d["parent_id"],
            events
                .iter()
                .find(|e| e["event_type"] == "inference")
                .unwrap()["event_id"]
        );
    }

    /// `evt_` ids sort by mint time and carry their own millisecond timestamp, the same
    /// property `mur trace list --since` reads off a `ses_` id.
    #[tokio::test]
    async fn event_ids_sort_by_mint_time() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        for turn in 0..5 {
            w.write_inference(
                turn,
                Some(1),
                Some(1),
                "end_turn".to_string(),
                Some("end_turn"),
                None,
                None,
                None,
                Vec::new(),
                None,
            )
            .await
            .unwrap();
        }
        w.flush().await.unwrap();

        let ids: Vec<String> = read_events(dir.path())
            .iter()
            .map(|e| e["event_id"].as_str().unwrap().to_string())
            .collect();
        let mut sorted = ids.clone();
        sorted.sort();
        assert_eq!(ids, sorted, "evt_ ids must sort into mint order");
    }

    // ── Wire capture: hashes, blobs and the capture gate ─────────────────────

    /// A payload in the shape `build_driver_payload` produces, with `n` messages.
    fn wire_payload(system: &str, messages: usize) -> Value {
        let messages: Vec<Value> = (0..messages)
            .map(|i| serde_json::json!({"role": "user", "content": format!("m{i}")}))
            .collect();
        serde_json::json!({
            "model": "claude-test",
            "max_tokens": 1024,
            "messages": messages,
            "tools": [{"name": "bash"}],
            "params": {},
            "system": system,
        })
    }

    async fn write_turn(w: &mut TraceWriter, turn: u32, payload: &Value, response: &str) {
        let wire = WireCapture::from_driver_payload(payload, response);
        w.write_inference(
            turn,
            Some(10),
            Some(5),
            "end_turn".to_string(),
            Some("end_turn"),
            None,
            None,
            None,
            vec!["msg_0000000000000000000000000000000a".to_string()],
            Some(&wire),
        )
        .await
        .unwrap();
    }

    fn blob_path(dir: &std::path::Path, sha: &str) -> std::path::PathBuf {
        dir.join(BLOB_DIR_NAME).join(sha)
    }

    /// Every hash an `inference` event names under `content` is the filename of a real file whose
    /// own digest is that filename — the whole point of the store.
    #[tokio::test]
    async fn content_capture_names_blobs_that_exist_and_rehash_to_their_names() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer_with_opts(dir.path(), TraceCapture::Content).await;
        let payload = wire_payload("You are a capsule.", 3);
        write_turn(&mut w, 0, &payload, r#"{"stop_reason":"end_turn"}"#).await;
        w.flush().await.unwrap();

        let e = &read_events(dir.path())[0];
        let mut named: Vec<String> = ["system_sha", "tools_sha", "response_sha"]
            .iter()
            .map(|key| {
                e[*key]
                    .as_str()
                    .unwrap_or_else(|| panic!("{key} missing from {e}"))
                    .to_string()
            })
            .collect();
        let shas = e["message_shas"].as_array().expect("message_shas");
        assert_eq!(shas.len(), 3);
        named.extend(shas.iter().map(|s| s.as_str().unwrap().to_string()));

        for sha in &named {
            let path = blob_path(dir.path(), sha);
            let bytes =
                std::fs::read(&path).unwrap_or_else(|err| panic!("blob {sha} must exist: {err}"));
            assert_eq!(murmur_artifact::sha256_hex(&bytes), *sha);
            // A bare sha256 names content; `evt_`/`msg_`/`ses_` name entities. Neither a prefix
            // nor an extension may creep onto a blob name.
            assert_eq!(sha.len(), 64, "{sha}");
            assert!(
                sha.chars()
                    .all(|c| c.is_ascii_hexdigit() && !c.is_ascii_uppercase()),
                "{sha}"
            );
        }
    }

    /// The hashed system body is the payload's `system` string, as text — so the blob reads as a
    /// prompt rather than as a JSON-quoted one.
    #[tokio::test]
    async fn the_system_blob_holds_the_prompt_as_readable_text() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer_with_opts(dir.path(), TraceCapture::Content).await;
        let payload = wire_payload("[Capsule] you are cap.\nBe terse.", 1);
        write_turn(&mut w, 0, &payload, "{}").await;
        w.flush().await.unwrap();

        let sha = read_events(dir.path())[0]["system_sha"]
            .as_str()
            .unwrap()
            .to_string();
        assert_eq!(
            std::fs::read_to_string(blob_path(dir.path(), &sha)).unwrap(),
            "[Capsule] you are cap.\nBe terse."
        );
    }

    /// `meta` is the default and hashes exactly what `content` hashes; the difference is only
    /// whether the bodies are on disk.
    #[tokio::test]
    async fn meta_writes_the_same_hashes_and_no_bodies() {
        let payload = wire_payload("You are a capsule.", 2);
        let response = r#"{"stop_reason":"end_turn"}"#;

        let meta = tempfile::tempdir().unwrap();
        let mut w = make_writer_with_opts(meta.path(), TraceCapture::Meta).await;
        write_turn(&mut w, 0, &payload, response).await;
        w.flush().await.unwrap();

        let content = tempfile::tempdir().unwrap();
        let mut w = make_writer_with_opts(content.path(), TraceCapture::Content).await;
        write_turn(&mut w, 0, &payload, response).await;
        w.flush().await.unwrap();

        let meta_event = read_events(meta.path()).remove(0);
        let content_event = read_events(content.path()).remove(0);
        for key in ["system_sha", "tools_sha", "response_sha", "message_shas"] {
            assert_eq!(meta_event[key], content_event[key], "{key}");
        }
        assert!(
            !meta.path().join(BLOB_DIR_NAME).exists(),
            "meta must not create the blob directory"
        );
        assert!(content.path().join(BLOB_DIR_NAME).is_dir());
    }

    /// The default a writer gets when the manifest declares no `trace:` block.
    #[tokio::test]
    async fn the_default_capture_mode_is_meta() {
        assert_eq!(TraceCapture::default(), TraceCapture::Meta);
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer_with_prompt(dir.path(), TraceCapture::default(), None, false).await;
        write_turn(&mut w, 0, &wire_payload("s", 1), "{}").await;
        w.flush().await.unwrap();

        let e = &read_events(dir.path())[0];
        assert!(e.get("system_sha").is_some());
        assert!(!dir.path().join(BLOB_DIR_NAME).exists());
    }

    /// Under `none` not one of the four hash keys is present, and every other field is.
    #[tokio::test]
    async fn none_omits_every_hash_and_leaves_the_rest_of_the_event_intact() {
        let payload = wire_payload("You are a capsule.", 2);

        let off = tempfile::tempdir().unwrap();
        let mut w = make_writer_with_opts(off.path(), TraceCapture::None).await;
        write_turn(&mut w, 3, &payload, "{}").await;
        w.flush().await.unwrap();

        let e = read_events(off.path()).remove(0);
        for key in ["system_sha", "tools_sha", "response_sha", "message_shas"] {
            assert!(
                e.get(key).is_none(),
                "{key} must be absent under none, got {e}"
            );
        }
        assert!(!off.path().join(BLOB_DIR_NAME).exists());
        for key in [
            "event_type",
            "event_id",
            "parent_id",
            "session_id",
            "timestamp",
            "turn",
            "task_id",
            "input_tokens",
            "output_tokens",
            "decision",
            "tool_name",
            "message_ids",
        ] {
            assert!(
                e.get(key).is_some(),
                "{key} must survive under none, got {e}"
            );
        }
        assert_eq!(e["turn"], 3);
        assert_eq!(e["message_ids"].as_array().unwrap().len(), 1);
    }

    /// A prompt that does not change across ten turns is one blob, written once and never
    /// rewritten — the file's bytes are replaced by hand after the first turn and must survive.
    #[tokio::test]
    async fn an_unchanged_system_prompt_is_stored_once() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer_with_opts(dir.path(), TraceCapture::Content).await;
        let payload = wire_payload("You are a capsule.", 1);
        write_turn(&mut w, 0, &payload, "{}").await;

        let sha = {
            w.flush().await.unwrap();
            read_events(dir.path())[0]["system_sha"]
                .as_str()
                .unwrap()
                .to_string()
        };
        let path = blob_path(dir.path(), &sha);
        std::fs::write(&path, b"sentinel").unwrap();

        for turn in 1..10 {
            write_turn(&mut w, turn, &payload, "{}").await;
        }
        w.flush().await.unwrap();

        let events = read_events(dir.path());
        assert_eq!(events.len(), 10);
        for e in &events {
            assert_eq!(e["system_sha"], sha.as_str());
        }
        assert_eq!(
            std::fs::read(&path).unwrap(),
            b"sentinel",
            "an existing blob is never rewritten"
        );
        let matching = std::fs::read_dir(dir.path().join(BLOB_DIR_NAME))
            .unwrap()
            .filter(|entry| entry.as_ref().unwrap().file_name() == sha.as_str())
            .count();
        assert_eq!(matching, 1, "exactly one file bears that name");
    }

    /// A record written for a completion the runtime parsed no driver response for names no stop
    /// reason: the key is absent, never an empty string. Absence means "no driver response was
    /// parsed here", so a reader can tell it apart from a driver that reported nothing — which is
    /// recorded as `""`.
    #[tokio::test]
    async fn stop_reason_is_absent_when_no_driver_response_was_parsed() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer(dir.path()).await;
        w.write_inference(
            0,
            Some(10),
            Some(5),
            "end_turn".to_string(),
            None,
            None,
            None,
            None,
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let event = read_events(dir.path()).remove(0);
        assert!(
            event.get("stop_reason").is_none(),
            "stop_reason must be absent, got {event}"
        );
    }

    /// A hook's `run-inference` sends a payload the runtime never built, so its record names no
    /// hashes — absent from the JSON, not empty strings.
    #[tokio::test]
    async fn a_hook_origin_record_carries_no_hashes() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer_with_opts(dir.path(), TraceCapture::Content).await;
        let origin = InferenceOrigin {
            source: "hook:gatekeeper".to_string(),
            model: "claude-test".to_string(),
        };
        w.write_inference(
            0,
            Some(1),
            Some(1),
            "end_turn".to_string(),
            // A hook's `run-inference` saw no provider stop reason, so it records none.
            None,
            None,
            Some(&origin),
            None,
            Vec::new(),
            None,
        )
        .await
        .unwrap();
        w.flush().await.unwrap();

        let e = read_events(dir.path()).remove(0);
        assert_eq!(e["origin"], "hook:gatekeeper");
        for key in ["system_sha", "tools_sha", "response_sha", "message_shas"] {
            assert!(e.get(key).is_none(), "{key} must be absent, got {e}");
        }
        assert!(!dir.path().join(BLOB_DIR_NAME).exists());
    }

    /// Two runs sharing a prefix agree on that prefix's `message_shas` and disagree from the
    /// changed message onwards — the divergence index is the first unequal position.
    #[tokio::test]
    async fn message_shas_locate_the_index_two_runs_diverge_at() {
        async fn shas_for(payload: &Value) -> Vec<String> {
            let dir = tempfile::tempdir().unwrap();
            let mut w = make_writer_with_opts(dir.path(), TraceCapture::Meta).await;
            write_turn(&mut w, 0, payload, "{}").await;
            w.flush().await.unwrap();
            read_events(dir.path())[0]["message_shas"]
                .as_array()
                .unwrap()
                .iter()
                .map(|s| s.as_str().unwrap().to_string())
                .collect()
        }

        let first = wire_payload("You are a capsule.", 4);
        let mut second = first.clone();
        second["messages"][2]["content"] = serde_json::json!("diverged");

        let a = shas_for(&first).await;
        let b = shas_for(&second).await;
        assert_eq!(a.len(), 4);
        let divergence = a.iter().zip(&b).position(|(x, y)| x != y);
        assert_eq!(divergence, Some(2));
        assert_eq!(a[3], b[3], "only the changed message changes its own sha");
    }

    /// A message repeated inside one request hashes to one name and one file. Ids never repeat;
    /// shas repeat exactly when content does, and that is what makes them comparable.
    #[tokio::test]
    async fn identical_messages_share_one_sha_and_one_blob() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer_with_opts(dir.path(), TraceCapture::Content).await;
        let payload = serde_json::json!({
            "messages": [{"role": "user", "content": "same"}, {"role": "user", "content": "same"}],
            "tools": [],
            "system": "s",
        });
        write_turn(&mut w, 0, &payload, "{}").await;
        w.flush().await.unwrap();

        let shas = read_events(dir.path())[0]["message_shas"]
            .as_array()
            .unwrap()
            .iter()
            .map(|s| s.as_str().unwrap().to_string())
            .collect::<Vec<_>>();
        assert_eq!(shas[0], shas[1]);
        assert!(blob_path(dir.path(), &shas[0]).is_file());
    }

    /// Blobs are the bytes as sent, so a peer handle a tool call would have had redacted out of
    /// its summary stays in the body a `content` capture stores.
    #[tokio::test]
    async fn a_message_blob_is_the_body_verbatim() {
        let dir = tempfile::tempdir().unwrap();
        let mut w = make_writer_with_opts(dir.path(), TraceCapture::Content).await;
        let message = serde_json::json!({"role": "user", "content": "hello"});
        let payload = serde_json::json!({
            "messages": [message.clone()],
            "tools": [],
            "system": "s",
        });
        write_turn(&mut w, 0, &payload, "{}").await;
        w.flush().await.unwrap();

        let sha = read_events(dir.path())[0]["message_shas"][0]
            .as_str()
            .unwrap()
            .to_string();
        assert_eq!(
            std::fs::read(blob_path(dir.path(), &sha)).unwrap(),
            serde_json::to_vec(&message).unwrap()
        );
    }

    /// A delegation record a plan's `capsule` step wrote and one `delegate-task` wrote are the
    /// same record.
    ///
    /// Identical key sets and identical values on everything but the two fields that are
    /// per-line by construction — the event id and the clock. That is what makes a trace
    /// unable to say which surface launched a child, and it holds because both appenders
    /// serialize the same struct rather than because this test watches them.
    #[tokio::test]
    async fn a_plan_written_delegation_record_matches_a_delegate_task_one() {
        let plan_dir = tempfile::tempdir().unwrap();
        let tool_dir = tempfile::tempdir().unwrap();
        let session = "ses_0000000000004000800000000000plan".to_string();
        let node = "evt_0000000000004000800000000000node".to_string();

        let plan = PlanTraceAppender::open(plan_dir.path(), session.clone(), node.clone()).unwrap();
        let tool = ResourceTraceAppender::open(tool_dir.path(), session.clone(), node.clone())
            .await
            .unwrap();

        plan.write_delegation_start("dlg_abc123", "worker", "0.1.0", "ses_child", ".murmur/c");
        tool.write_delegation_start("dlg_abc123", "worker", "0.1.0", "ses_child", ".murmur/c")
            .await;
        plan.write_delegation(
            "worker",
            "0.1.0",
            Some("dlg_abc123".to_string()),
            Some("ses_child".to_string()),
            42,
            "completed",
            None,
        );
        tool.write_delegation(
            "worker",
            "0.1.0",
            Some("dlg_abc123".to_string()),
            Some("ses_child".to_string()),
            42,
            "completed",
            None,
        )
        .await;

        let read = |dir: &Path| -> Vec<serde_json::Map<String, Value>> {
            std::fs::read_to_string(dir.join("trace.jsonl"))
                .unwrap()
                .lines()
                .map(|line| serde_json::from_str::<serde_json::Map<_, _>>(line).unwrap())
                .collect()
        };
        let from_plan = read(plan_dir.path());
        let from_tool = read(tool_dir.path());
        assert_eq!(from_plan.len(), 2);
        assert_eq!(from_tool.len(), 2);

        for (planned, called, event_type) in [
            (&from_plan[0], &from_tool[0], "delegation_start"),
            (&from_plan[1], &from_tool[1], "delegation"),
        ] {
            assert_eq!(planned["event_type"], event_type);
            assert_eq!(called["event_type"], event_type);
            assert_eq!(
                planned.keys().collect::<Vec<_>>(),
                called.keys().collect::<Vec<_>>(),
                "{planned:?} vs {called:?}"
            );
            assert_eq!(planned["parent_id"], node.as_str());
            assert_eq!(called["parent_id"], node.as_str());
            for (key, value) in planned {
                if key == "event_id" || key == "timestamp" {
                    continue;
                }
                assert_eq!(value, &called[key], "{key}: {planned:?} vs {called:?}");
            }
            for key in ["plan_id", "step_id", "surface", "kind"] {
                assert!(!planned.contains_key(key), "{key} is on {planned:?}");
                assert!(!called.contains_key(key), "{key} is on {called:?}");
            }
        }
    }
}