supercode_interchange/session/

mod.rs

//! Natively load — and continue — real Claude Code and Codex sessions.
//!
//! Both tools persist their conversations as JSONL on disk:
//!
//! - **Claude Code**: `~/.claude/projects/<encoded-cwd>/<session-id>.jsonl`,
//!   one line per event in the Anthropic message format, linked by
//!   `uuid`/`parentUuid`.
//! - **Codex**: `~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl`, where each line
//!   is a `{timestamp, type, payload}` envelope and the `response_item` lines
//!   form the canonical conversation.
//!
//! [`Session::load`] auto-detects the format and normalizes either one into a
//! provider-neutral [`Vec<ChatMessage>`] that can be handed straight back to a
//! model (via OpenRouter or any OpenAI-compatible endpoint) to continue.
//!
//! Provider-internal artifacts that don't replay across vendors — Anthropic
//! `thinking` blocks, Codex `reasoning` items — are dropped during
//! normalization.
//!
//! # Where this sits in supercode's priorities
//!
//! This module is the home of **feature 1 (translate between session formats)**
//! and half of **feature 2 (emulate-to-continue)** — the load/emit surface for
//! each harness ([`SessionFormat`], `from_*_str` loaders, `to_*_jsonl`
//! emitters). See [`AGENTS.md`](../../../AGENTS.md) for the three ranked
//! feature-priorities and the glue-tool positioning; the top priority is
//! **feature 3 (continue losslessly *with massive token reduction*)**, which
//! this fidelity work exists to make trustworthy. `opencode` + `pi` loaders
//! are built against the frozen `docs/interop/opencode-pi-spec.md` contract
//! — OpenCode additionally reads its native SQLite store (`opencode*.db`,
//! PARITY-3/PARITY-16) via `rusqlite`, reconstructing the same envelope form
//! [`Session::from_opencode_str`] already parses for the JSON-tree surfaces.

use serde::{Deserialize, Serialize};
use std::collections::{BTreeMap, HashMap, HashSet};
use std::io::{BufRead, BufReader, Read, Seek, SeekFrom};
use std::path::{Path, PathBuf};

use rusqlite::Connection;
use serde_json::Value;

use crate::{
    ChatMessage, Fidelity, FunctionCall, InterchangeError as Error, Result, Role, ToolCall,
};

mod claude_code;
pub(crate) use claude_code::ClaudeAppendState;
#[doc(hidden)]
pub use claude_code::ClaudeReadIndex;
mod codex;
mod detect;
mod gemini;
mod goose;
mod grok;
mod helpers;
mod hermes;
mod native;
mod openclaw;
mod opencode;
mod pi;
mod residue;

// The per-harness files below are an internal file layout only: every item
// keeps its original `crate::session::…` path through these re-exports, whose
// visibility matches the most-visible item each module holds.
pub(crate) use claude_code::*;
use codex::*;
pub use detect::*;
use gemini::*;
use grok::*;
pub use helpers::*;
pub use hermes::*;
use native::*;
pub use openclaw::*;
pub use opencode::*;
pub use pi::*;
pub use residue::*;

/// Which tool produced a session log.
///
/// This is **read-provenance**: a fact recovered when a log is loaded (stored
/// in [`SessionMeta::source`], filled in by auto-detection in
/// `detect_source`), describing which tool originally wrote the file on
/// disk. It answers "where did this session come from?" — e.g. for
/// `inspect`/`convert` display in the CLI.
///
/// It is deliberately distinct from [`SessionFormat`], even though the two
/// enums' variant lists currently coincide: [`SessionFormat`] selects a
/// serialization codec (what to parse/export *as*), while `SessionSource`
/// records history (what wrote the file). The pair is intentionally kept
/// separate rather than merged — a session loaded from one tool's log can
/// still be exported in the other tool's format, and the two concepts could
/// diverge further (e.g. a format that is readable but not attributable, or
/// multiple versioned formats sharing one source).
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum SessionSource {
    /// A `~/.claude/projects/.../<id>.jsonl` transcript.
    ClaudeCode,
    /// A `~/.codex/sessions/.../rollout-*.jsonl` file.
    Codex,
    /// An OpenCode session — multi-file JSON tree(s) or SQLite `opencode*.db`
    /// (`docs/interop/opencode-pi-spec.md` §1.2). Detection and loading are
    /// wave B; this variant exists now so `SessionSource`/`SessionFormat` stay
    /// 1:1 per the frozen interop spec (§0).
    OpenCode,
    /// A `~/.pi/agent/sessions/--<enc-cwd>--/<iso>_<sessionId>.jsonl`
    /// transcript (`docs/interop/opencode-pi-spec.md` §1.1) — line-oriented
    /// JSONL like Claude Code/Codex, so it shares their byte-lossless native
    /// round-trip property.
    Pi,
    /// A Grok session transcript stored as
    /// `~/.grok/sessions/<percent-encoded-cwd>/<session-id>/chat_history.jsonl`.
    Grok,
    /// A Gemini CLI transcript stored under
    /// `~/.gemini/tmp/<project>/chats/session-*.jsonl`.
    Gemini,
    /// A Goose session exported through `_goose/unstable/session/export`, or
    /// reconstructed from Goose's `sessions/sessions.db` native store.
    Goose,
    /// An OpenClaw agent session (`~/.openclaw/agents/<agentId>/sessions/
    /// <uuid>.jsonl`, openclaw >= 2026.7): pi session-format v3 with
    /// openclaw dialect divergences — `type:"leaf"` navigation-control
    /// entries that REDIRECT the active leaf (pi's last-entry anchor rule is
    /// wrong for them), `appendMode:"side"` entries that never anchor, and
    /// vendor-namespaced `__openclaw` message metadata. READ-ONLY provenance
    /// (UNI-16): there is deliberately no `SessionFormat::OpenClaw` — the
    /// write tier is a permanently skipped direct-DB/store path; loaded
    /// sessions translate OUT through the other formats.
    OpenClaw,
    /// A Hermes Agent session read from its single SQLite store
    /// (`~/.hermes/state.db`, `SCHEMA_VERSION = 22` at the 0.19.0 pin).
    /// READ-ONLY provenance (UNI-15): no `SessionFormat::Hermes` exists —
    /// writing into a live, shared, WAL, single-writer store stays gated by
    /// UNI-22 (not fired; the schema churned 19->22 in one release) — loaded
    /// sessions translate OUT through the other formats.
    Hermes,
    /// P5-3 safety-hardening fix (Fable-5 review, LOW "translation-fidelity
    /// cosmetic"): a session that was never imported from ANY foreign
    /// tool's log at all — authored directly by supercode's own agent loop,
    /// with no foreign-tool prefix (`Session.raw` starts empty). Currently
    /// only `crate::agent::Agent`'s `persist_subagent_transcript` (P5-3,
    /// natively-spawned `spawn_subagent` children) uses this — before this
    /// variant existed, that call site built its blank `Session` via
    /// `Session::from_claude_code_str("")` purely as an "empty parser to
    /// get a blank skeleton" trick, which left `meta.source ==
    /// SessionSource::ClaudeCode` even though nothing Claude-Code-shaped
    /// was ever involved, mislabeling a native supercode spawn as an
    /// imported CC session on disk (and in any `inspect`/`convert` reading
    /// it back). Never produced by auto-detection (`detect_source`) or any
    /// `from_<tool>_str` loader — only by code that explicitly constructs
    /// a `SessionMeta` with this source, so no existing imported-session
    /// path can ever observe this variant appearing where it didn't before.
    Native,
}

/// An on-disk session format supercode can both read and write.
///
/// Like an image editor that opens and exports several file formats, supercode
/// keeps one canonical in-memory model ([`Session`]) and converts to/from each
/// supported format on the edges.
///
/// This is a **write-target** / codec selector: a caller's request, passed to
/// [`Session::load_str`], [`Session::to_jsonl`], and [`Session::save`],
/// choosing which on-disk dialect to parse or emit. It answers "what format
/// should I read/write?" — as opposed to [`SessionSource`], which records the
/// provenance fact of what actually produced a loaded file. The two enums are
/// intentionally kept separate (provenance fact vs. serialization choice) and
/// should not be unified, even though their variants currently match
/// one-to-one.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum SessionFormat {
    /// Claude Code transcript JSONL.
    ClaudeCode,
    /// Codex rollout JSONL.
    Codex,
    /// OpenCode export-document / envelope JSONL (wave B; see
    /// [`SessionSource::OpenCode`]).
    OpenCode,
    /// Pi session JSONL (see [`SessionSource::Pi`]).
    Pi,
    /// Grok `chat_history.jsonl` transcript.
    Grok,
    /// Gemini CLI session JSONL.
    Gemini,
    /// Goose native session-export JSON.
    Goose,
}

impl SessionFormat {
    /// The [`SessionSource`] a file of this format reports.
    ///
    /// This is the deliberate one-way bridge between the two concepts: a file
    /// saved in this format will, when reloaded, report this provenance (see
    /// `crates/harness/tests/session_saving.rs`), making the relationship
    /// discoverable from the method itself.
    pub fn source(self) -> SessionSource {
        match self {
            SessionFormat::ClaudeCode => SessionSource::ClaudeCode,
            SessionFormat::Codex => SessionSource::Codex,
            SessionFormat::OpenCode => SessionSource::OpenCode,
            SessionFormat::Pi => SessionSource::Pi,
            SessionFormat::Grok => SessionSource::Grok,
            SessionFormat::Gemini => SessionSource::Gemini,
            SessionFormat::Goose => SessionSource::Goose,
        }
    }

    /// The format a session from `source` is written in, when it has one of its own.
    pub fn for_source(source: SessionSource) -> Option<Self> {
        [
            SessionFormat::ClaudeCode,
            SessionFormat::Codex,
            SessionFormat::OpenCode,
            SessionFormat::Pi,
            SessionFormat::Grok,
            SessionFormat::Gemini,
            SessionFormat::Goose,
        ]
        .into_iter()
        .find(|format| format.source() == source)
    }
}

/// Metadata recovered from a session log.
pub use crate::ontology::surface::{
    CrossSurface, Recurrence, SurfaceKey, Trigger, WorkspaceKind, WorkspaceRef,
};

/// How a Codex session started, from its `session_meta`: a subagent's (`source.subagent`) is its parent's;
/// `codex exec` (`codex_exec`) and supercode's runtimes (`supercode`, `supercode_*`) are programs. A TUI
/// start makes no claim (agents run the TUI in panes too), nor do other originators. `source` alone cannot
/// tell programs apart: its `vscode` covers every app-server client, supercode's included.
pub fn codex_start_trigger(payload: &serde_json::Value) -> Option<Trigger> {
    if payload
        .get("source")
        .and_then(|source| source.get("subagent"))
        .is_some()
    {
        return Some(Trigger::Parent);
    }
    match payload
        .get("originator")
        .and_then(serde_json::Value::as_str)?
    {
        "codex_exec" => Some(Trigger::Api),
        other if other == "supercode" || other.starts_with("supercode_") => Some(Trigger::Api),
        _ => None,
    }
}

/// ORCH-6: the ORCH-3 conversation nouns as one additive wire block, carried
/// by `harness.v1.sessions.discover` / `sessions.load` rows and by
/// [`crate::catalog::SessionDescriptor`]. Every field is optional so an older
/// client sees exactly the shape it already knows.
#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize)]
pub struct OrchestrationNouns {
    /// Why the session exists.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub trigger: Option<Trigger>,
    /// Where the conversation is reached.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub surface: Option<SurfaceKey>,
    /// Routed config home (Hermes profile / OpenClaw agent).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub profile: Option<String>,
    /// The job a recurring session belongs to.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub recurrence: Option<Recurrence>,
    /// Moved-to-another-surface state.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub cross_surface: Option<CrossSurface>,
    /// Typed workspace (the D2 precedence result).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub workspace: Option<WorkspaceRef>,
}

impl OrchestrationNouns {
    /// Read the nouns off a loaded session's metadata. `trigger` and
    /// `workspace` always resolve — through [`SessionMeta::trigger_or_default`]
    /// and [`SessionMeta::workspace`], never through a second derivation.
    pub fn from_meta(meta: &SessionMeta) -> Self {
        Self {
            trigger: Some(meta.trigger_or_default()),
            surface: meta.surface.clone(),
            profile: meta.profile.clone(),
            recurrence: meta.recurrence.clone(),
            cross_surface: meta.cross_surface.clone(),
            workspace: Some(meta.workspace_ref()),
        }
    }
}

#[derive(Debug, Clone)]
#[non_exhaustive]
pub struct SessionMeta {
    /// The tool that wrote the log.
    pub source: SessionSource,
    /// The session/rollout id.
    pub session_id: Option<String>,
    /// Native recorded session end, in RFC 3339 form. Absence is unknown;
    /// it must not be replaced by a transcript quiet-time estimate.
    pub ended_at: Option<String>,
    /// Source-native reason accompanying the recorded session end.
    pub end_reason: Option<String>,
    /// The model the session was running.
    pub model: Option<String>,
    /// The working directory the session ran in.
    pub cwd: Option<PathBuf>,
    /// The system / base-instructions prompt, when the log records it.
    pub system_prompt: Option<String>,
    /// Verbatim source-format header records (the Codex `session_meta` /
    /// `turn_context` lines), preserved so re-export can replay the exact header
    /// the original tool expects rather than guessing its required fields.
    pub codex_headers: Vec<Value>,
    /// Exact source lines for Codex execution/provenance records that affect
    /// continuation semantics but must not be replayed as active events after
    /// a foreign-format hop. Each entry records its original physical-line
    /// index, discriminant, and verbatim JSONL text. Foreign writers carry the
    /// list in a namespaced extension; a later Codex export restores headers
    /// from it while keeping compaction/rollback/review records non-operative,
    /// avoiding a second rollback or compaction of the already-normalized view.
    pub codex_provenance: Vec<Value>,
    /// PARITY-23: generalized source-native residue records for NON-codex
    /// sources — `{record_index, kind, raw}` entries captured at load time
    /// (or restored from a portable v2 envelope) so cross-format hops can
    /// return them exactly. Codex keeps its original dedicated store above.
    pub native_residue: Vec<Value>,
    /// The source format `native_residue` belongs to (e.g. `claude_code`).
    pub native_residue_source: Option<String>,
    /// For each message as loaded, the index in `raw` of the record that created it, when the
    /// loader knows it — the provenance the portable residue cuts segments on
    /// (`docs/plans/portable-residue.md`). Empty when the loader does not record it.
    pub message_records: Vec<Option<usize>>,
    /// The OpenCode analogue of [`Self::codex_headers`]
    /// (`docs/interop/opencode-pi-spec.md` §1.2/§2.1): the verbatim
    /// `SessionInfo` record (always element 0, or `Value::Null` if somehow
    /// absent), plus any captured `session_diff`/`todo` side-records — each
    /// wrapped as `{"key": [...], "value": ...}`, mirroring the envelope
    /// shape `raw` uses, so a consumer can tell which storage key a header
    /// record belongs to. These replay only via the direct-write fallback
    /// (`Session::to_opencode_direct_write`); `opencode import` has no
    /// ingestion path for `session_diff`/`todo` (S5).
    pub opencode_headers: Vec<Value>,
    /// Goose's native session-export object with `conversation` removed.
    /// Goose stores sessions in SQLite but defines this JSON object as its
    /// official import/export boundary. Keeping the shell lets an unchanged
    /// direct round-trip remain byte exact while appended turns are spliced
    /// into a stock-importable artifact without guessing native metadata.
    pub goose_header: Option<Value>,
    /// For a Claude Code subagent session: its `agentId` (the `agent-<id>` file
    /// stem). `None` for top-level sessions.
    pub agent_id: Option<String>,
    /// For a subagent session: the `tool_use_id` of the parent `Task` call that
    /// spawned it, recovered from the parent transcript's tool result. Best
    /// effort — `None` if the link could not be established.
    pub parent_tool_use_id: Option<String>,
    /// Cross-file lineage keys for multi-file/multi-agent sessions (Codex
    /// `parent_thread_id`, `forked_from_id`, `thread_source`, and the
    /// `source.subagent.thread_spawn` fields `agent_role` / `agent_nickname` /
    /// `depth`). Empty for a plain top-level session. Used by
    /// [`Session::reconstruct_tree`] to nest children under their parents.
    pub lineage: std::collections::BTreeMap<String, String>,
    /// ORCH-3: why the session exists, when the source says.
    pub trigger: Option<Trigger>,
    /// ORCH-3: the conversation's surface identity, when it has one.
    pub surface: Option<SurfaceKey>,
    /// ORCH-3: routed config home (Hermes profile / OpenClaw agent / Codex profile).
    pub profile: Option<String>,
    /// ORCH-3: the job a recurring session belongs to.
    pub recurrence: Option<Recurrence>,
    /// ORCH-3: moved-to-another-surface state.
    pub cross_surface: Option<CrossSurface>,
}

impl SessionMeta {
    pub(crate) fn new(source: SessionSource) -> Self {
        SessionMeta {
            source,
            session_id: None,
            ended_at: None,
            end_reason: None,
            model: None,
            cwd: None,
            system_prompt: None,
            codex_headers: Vec::new(),
            codex_provenance: Vec::new(),
            native_residue: Vec::new(),
            native_residue_source: None,
            message_records: Vec::new(),
            opencode_headers: Vec::new(),
            goose_header: None,
            agent_id: None,
            parent_tool_use_id: None,
            lineage: std::collections::BTreeMap::new(),
            trigger: None,
            surface: None,
            profile: None,
            recurrence: None,
            cross_surface: None,
        }
    }

    /// The trigger, defaulting from what the loaders already know: a spawned
    /// child (`agent_id` / a delegate lineage) is `Parent`; otherwise `Human`.
    pub fn trigger_or_default(&self) -> Trigger {
        if let Some(t) = self.trigger {
            return t;
        }
        let delegate = self
            .lineage
            .get("hermes_lineage_kind")
            .map(|k| k == "delegate")
            .unwrap_or(false);
        if self.agent_id.is_some() || self.parent_tool_use_id.is_some() || delegate {
            Trigger::Parent
        } else {
            Trigger::Human
        }
    }

    /// UNI-9 workspace with the D2 precedence: `repo` when a cwd exists, else
    /// `channel` when the surface is a channel, else `none`. Derived, never stored.
    pub fn workspace(&self) -> (WorkspaceKind, Option<String>) {
        if let Some(cwd) = &self.cwd {
            return (
                WorkspaceKind::Repo,
                Some(cwd.to_string_lossy().into_owned()),
            );
        }
        if let Some(surface) = self.surface.as_ref().filter(|s| s.is_channel()) {
            let label = match (&surface.platform, &surface.chat_id) {
                (Some(p), Some(c)) => format!("{p}:{c}"),
                (Some(p), None) => p.clone(),
                _ => String::new(),
            };
            return (WorkspaceKind::Channel, Some(label));
        }
        (WorkspaceKind::None, None)
    }

    /// [`Self::workspace`] as the wire value. Naming only — the precedence
    /// stays in `workspace()`.
    pub fn workspace_ref(&self) -> WorkspaceRef {
        let (kind, value) = self.workspace();
        WorkspaceRef { kind, value }
    }
}

/// A normalized, replayable conversation loaded from a tool's session log.
#[derive(Debug, Clone)]
pub struct Session {
    /// Recovered metadata.
    pub meta: SessionMeta,
    /// The conversation, normalized to the OpenAI chat-completions shape.
    pub messages: Vec<ChatMessage>,
    /// Subagent (Task) sub-conversations. Claude Code stores these as separate
    /// `<session>/subagents/agent-*.jsonl` files; loading a session by path now
    /// discovers and attaches them here (each is a full [`Session`] whose
    /// `meta.agent_id` / `meta.parent_tool_use_id` link it back to its spawn).
    pub subagents: Vec<Session>,
    /// Every original JSONL line of the source log, STRICT-VERBATIM (IX-1):
    /// captured via `split_lines_verbatim`, not the blank-skipping/trimming
    /// `non_empty_lines` parse view, so a blank line, a CRLF (`\r\n`)
    /// terminator, or trailing whitespace on a line all survive bit-for-bit
    /// rather than being dropped/normalized away. Normalization into
    /// `messages` is still lossy by design (it targets the OpenAI replay
    /// shape), but these raw lines retain *everything* — including records
    /// with no canonical representation (e.g. Claude `file-history-snapshot`)
    /// — so a round-trip through the supercode-native format
    /// ([`Session::to_native_jsonl`]) is byte-lossless for the line-oriented
    /// formats (Claude Code/Codex/Pi), for ANY input (see
    /// [`Self::raw_trailing_newline`] for the one piece of information a line
    /// list alone can't carry).
    pub raw: Vec<String>,
    /// Whether the source text `raw` was captured from ended with a trailing
    /// `\n`. `raw`'s line list alone can't distinguish a source ending with a
    /// trailing newline from one that doesn't (both split into the same
    /// lines) — this flag carries that fact out-of-band so
    /// [`Self::to_native_jsonl`]/[`Self::from_native_str`] can reproduce the
    /// original source bytes exactly, including the presence/absence of a
    /// final newline. `true` for a `Session` whose `raw` isn't captured
    /// verbatim from real source text (e.g. OpenCode's re-synthesized
    /// export-document `raw`, or a `Session` assembled programmatically) —
    /// matching the historical always-terminated-by-newline behavior for
    /// those cases.
    pub raw_trailing_newline: bool,
    /// How many of `messages` (and, symmetrically, of `raw` — see below) came
    /// from parsing the imported log, as opposed to being appended after
    /// import. Set once, at the end of [`Self::from_claude_code_str`] /
    /// [`Self::from_codex_str`], to `messages.len()` at that moment — i.e.
    /// before [`Self::from_native_str`]'s subsequent loop reattaches any
    /// appended [`crate::sidecar::NativeTurn`] records onto `messages`/`raw`.
    /// That loop pushes exactly one `raw` line and one message per appended
    /// turn, so the two lists grow in lockstep from here on: the raw-prefix
    /// boundary A12's [`Self::to_jsonl_spliced`] needs is always recoverable
    /// as `raw.len() - (messages.len() - imported_message_count)`, without a
    /// second counter. `None` only when a `Session` is constructed some other
    /// way than through those two loaders — splicing then has no boundary to
    /// honor and treats every message as imported (equivalent to
    /// `Some(messages.len())`).
    /// A bounded display-history projection uses this field for the total
    /// number of normalized messages observed before its in-memory window was
    /// applied. Such a semantic view is never a continuation source, and all
    /// splice callers clamp the value to `messages.len()`.
    pub imported_message_count: Option<usize>,
    /// Whether `raw` was captured strict-verbatim from real source text
    /// (`true`) or re-synthesized by this crate (`false`) — the fact
    /// [`Self::raw_verbatim`]'s callers need to know before claiming a
    /// same-format `convert` is byte-identical (PARITY-AUDIT.md P006/P007).
    /// `true` for every line-oriented loader (`from_claude_code_str`,
    /// `from_codex_str`, `from_pi_str`) and OpenCode's own ENVELOPE read
    /// surface (`from_opencode_str`'s per-line loop) — each of those splits
    /// `raw` directly out of the source text via `split_lines_verbatim`, so
    /// replaying it reproduces the original bytes exactly. `false` for
    /// OpenCode's EXPORT-DOCUMENT read surface
    /// (`Session::from_opencode_export_doc`): a pretty-printed
    /// `{info, messages:[...]}` document has no per-line envelope structure
    /// of its own, so `raw` there is one envelope line RE-SYNTHESIZED per
    /// record — faithful in value, but not the original document's bytes.
    /// A `Session` assembled programmatically (not through a `from_*_str`
    /// loader) also defaults to `false` — no real source text was captured
    /// at all.
    pub raw_is_verbatim: bool,
    /// PARITY-15: how many non-empty lines of the source text FAILED to
    /// deserialize at all (a genuinely malformed/truncated JSON line — not
    /// a well-formed-but-unmodeled record type, which is a normal,
    /// intentional "skip", tracked separately by `crate::audit`). Every
    /// line-oriented loader tolerates a stray corrupt line rather than
    /// hard-failing the whole load (a single bad line must not make an
    /// otherwise-healthy multi-thousand-line session unloadable) — but that
    /// tolerance used to be completely invisible: `Session::load` returned
    /// `Ok` either way, with no signal that anything was skipped. This
    /// count is what lets a caller (the CLI, `inspect`/`convert`) surface
    /// that loss loudly instead of silently. `0` for a cleanly-parsed file,
    /// and for a `Session` assembled programmatically.
    pub parse_error_lines: usize,
    /// Named degradations a [`Fidelity::Semantic`] load accepted instead of
    /// failing — the same "say exactly what was given up" residue list
    /// `harness.v1.sessions.export` already reports for artifacts.
    ///
    /// ALWAYS empty for a lossless load: every stricter fidelity refuses a
    /// transcript it cannot reconstruct exactly, which is what keeps
    /// continuation/transfer/export guarantees intact. A non-empty list means
    /// this session is a read-only VIEW ([`Session::load_with_fidelity`] with
    /// [`Fidelity::Semantic`]) and must not be used as a continuation source.
    pub load_residue: Vec<String>,
}

impl Session {
    /// The fidelity this reconstruction actually achieved.
    ///
    /// Same rule the export path applies to an artifact: named residue means
    /// [`Fidelity::Semantic`]; otherwise a verbatim source capture is
    /// [`Fidelity::ByteLossless`] and a re-synthesized one is
    /// [`Fidelity::ValueLossless`]. A subagent's residue counts as this
    /// session's: the whole reconstruction is only as faithful as its least
    /// faithful part, and each child still reports its own residue where it
    /// was measured.
    pub fn load_fidelity(&self) -> Fidelity {
        let own = if !self.load_residue.is_empty() {
            Fidelity::Semantic
        } else if self.raw_is_verbatim {
            Fidelity::ByteLossless
        } else {
            Fidelity::ValueLossless
        };
        if own != Fidelity::Semantic
            && self
                .subagents
                .iter()
                .any(|subagent| subagent.load_fidelity() == Fidelity::Semantic)
        {
            return Fidelity::Semantic;
        }
        own
    }

    /// Assemble a session from supercode's own flat store transcript (one
    /// [`ChatMessage`] per JSONL line). These files are the native working
    /// format written by Supercode's native session store, not a foreign
    /// harness log, so routing them through format auto-detection would
    /// misclassify them as an empty Claude Code session.
    pub fn from_native_messages(messages: Vec<ChatMessage>) -> Session {
        Session {
            meta: SessionMeta::new(SessionSource::Native),
            messages,
            subagents: Vec::new(),
            raw: Vec::new(),
            raw_trailing_newline: true,
            imported_message_count: None,
            raw_is_verbatim: false,
            parse_error_lines: 0,
            load_residue: Vec::new(),
        }
    }

    /// Load a session, auto-detecting whether it's a Claude Code or Codex log
    /// — or, when `path` looks like a SQLite database, a real OpenCode
    /// `opencode*.db` store (PARITY-3/PARITY-16): that check runs BEFORE any
    /// UTF-8 text read, so a binary `.db` file is routed to
    /// [`Self::from_opencode_sqlite`] instead of failing on a raw "stream
    /// did not contain valid UTF-8" error (the confirmed footgun these items
    /// close — see [`looks_like_sqlite`] and the UTF-8 diagnostic reader).
    ///
    /// A DIRECTORY is also accepted directly: `path` is probed with
    /// [`detect_opencode_storage_surface`] BEFORE the SQLite/UTF-8 file
    /// checks below (both of which assume a file and would otherwise surface
    /// a cryptic "Is a directory" `io::Error` — the confirmed footgun this
    /// closes). This lets `inspect`/`convert`/`resume` accept an OpenCode
    /// DATA-ROOT directly (e.g. `~/.local/share/opencode`), matching what
    /// `audit --format opencode` already does. A resolved `Sqlite` surface
    /// loads exactly like pointing `load` at that `opencode*.db` file
    /// directly (most-recently-updated top-level session). The legacy
    /// `JsonTreeA`/`JsonTreeB` surfaces are classifier-only (see
    /// [`OpenCodeStorageSurface`]) — there's no direct-JSON-tree loader, so
    /// that case returns a clear error naming the `.db` file / `audit` as the
    /// way in, rather than silently doing nothing or crashing.
    pub fn load(path: impl AsRef<Path>) -> Result<Session> {
        Self::load_with_fidelity(path, Fidelity::ByteLossless)
    }

    /// Load a session at a declared [`Fidelity`].
    ///
    /// [`Fidelity::Semantic`] is the READ-ONLY VIEW mode: a transcript whose
    /// record graph cannot be reconstructed exactly (the everyday case for a
    /// Claude Code session that has been compacted or resumed across files,
    /// where a live record's `parentUuid` names a record that was pruned)
    /// still loads, stitched best-effort in transcript order, and names what
    /// it gave up in [`Session::load_residue`]. Every stricter level keeps
    /// the historical behavior — refuse loudly — because a continuation,
    /// transfer or export built on a guessed graph is exactly the loss
    /// supercode exists to prevent. Callers that go on to RESUME a session
    /// must therefore use [`Session::load`].
    pub fn load_with_fidelity(path: impl AsRef<Path>, fidelity: Fidelity) -> Result<Session> {
        Self::load_with_fidelity_and_subagents(path, fidelity, true)
    }

    /// Load only the selected session's own transcript at a declared fidelity.
    ///
    /// This is the read-only frontend path: Claude Code can place hundreds of
    /// child transcripts beside a parent, but a chat viewport displaying the
    /// parent must not eagerly parse and transport that entire child tree.
    /// Translation, continuation, export, and the ordinary [`Self::load`]
    /// path keep attaching every subagent unchanged.
    #[doc(hidden)]
    pub fn load_parent_with_fidelity(
        path: impl AsRef<Path>,
        fidelity: Fidelity,
    ) -> Result<Session> {
        Self::load_with_fidelity_and_subagents(path, fidelity, false)
    }

    /// Load a bounded, parent-only transcript for human display.
    ///
    /// Unlike the continuation loader, Codex compaction records do not erase
    /// earlier visible assistant turns here: the native rollout still holds
    /// those records, and a scrollback view should show what the human saw,
    /// not only the compacted context the next model call will receive.
    #[doc(hidden)]
    pub fn load_display_view(
        path: impl AsRef<Path>,
        fidelity: Fidelity,
        message_limit: usize,
    ) -> Result<Session> {
        let path = path.as_ref();
        if path.is_dir() || looks_like_sqlite(path) {
            let mut session = Self::load_parent_with_fidelity(path, fidelity)?;
            truncate_session_messages(&mut session, message_limit);
            return Ok(session);
        }
        let mut read_limit = message_limit.max(1);
        let mut previous_window_len = 0usize;
        let (mut session, omitted_prefix) = loop {
            let (source, text, omitted_prefix) = read_display_jsonl(path, read_limit)?;
            let mut candidate = match source {
                Some(SessionSource::Codex) => Self::from_codex_display_str(&text, message_limit)?,
                Some(SessionSource::Gemini) => {
                    let mut session = Self::from_gemini_str(&text)?;
                    session.raw_is_verbatim = false;
                    session.load_residue.push(
                        "display history is a bounded native-record projection, not a complete Gemini artifact"
                            .to_string(),
                    );
                    session
                }
                Some(SessionSource::Pi) => Self::from_pi_str(&text)?,
                Some(SessionSource::Grok) => {
                    let mut session = Self::from_grok_str(&text)?;
                    session.capture_grok_path_metadata(path);
                    session
                }
                Some(SessionSource::OpenCode) => Self::from_opencode_str(&text)?,
                _ => Self::from_claude_code_str_with_fidelity(&text, fidelity)?,
            };
            let observed_messages = candidate
                .imported_message_count
                .unwrap_or(candidate.messages.len())
                .max(candidate.messages.len());
            let human_turns = candidate
                .messages
                .iter()
                .filter(|message| message.role == Role::User)
                .count();
            let window_len = text.len();
            let sufficient =
                !omitted_prefix || (observed_messages > message_limit.max(1) && human_turns >= 2);
            // 16 KiB/message with a 64 MiB ceiling means 4096 is the first
            // read limit that cannot grow the native byte window further.
            // Smaller repeated lengths can be the intentional 4 MiB floor;
            // keep doubling through that plateau instead of declaring a
            // false pagination end.
            let byte_window_exhausted = window_len <= previous_window_len && read_limit >= 4096;
            if sufficient || byte_window_exhausted {
                if omitted_prefix {
                    // The prefix is known to contain more native history even
                    // when this bounded window cannot cheaply normalize its
                    // exact size. Never turn that into a false end-of-history.
                    candidate.imported_message_count =
                        Some(observed_messages.max(message_limit.max(1).saturating_add(1)));
                }
                break (candidate, omitted_prefix);
            }
            previous_window_len = window_len;
            read_limit = read_limit.saturating_mul(2);
        };
        if omitted_prefix {
            session.load_residue.push(
                "older native records remain outside this bounded display window".to_string(),
            );
        }
        truncate_session_messages(&mut session, message_limit);
        Ok(session)
    }

    fn load_with_fidelity_and_subagents(
        path: impl AsRef<Path>,
        fidelity: Fidelity,
        include_subagents: bool,
    ) -> Result<Session> {
        let path = path.as_ref();
        if path.is_dir() {
            return match detect_opencode_storage_surface(path) {
                Some((OpenCodeStorageSurface::Sqlite, db_path)) => {
                    Self::from_opencode_sqlite(&db_path, None)
                }
                Some((
                    OpenCodeStorageSurface::JsonTreeA | OpenCodeStorageSurface::JsonTreeB,
                    _,
                )) => Err(crate::Error::Other(format!(
                    "{} is an OpenCode data root using a legacy JSON storage tree, which \
                         supercode does not load directly — point `inspect`/`convert`/`resume` \
                         at the store's `opencode*.db` SQLite file if this install has one, or \
                         use `audit --format opencode {}` instead",
                    path.display(),
                    path.display()
                ))),
                None => Err(crate::Error::Other(format!(
                    "{} is a directory, but no session file or OpenCode store was found in it \
                     (expected an `opencode*.db` SQLite file, or an OpenCode legacy JSON storage \
                     tree)",
                    path.display()
                ))),
            };
        }
        if looks_like_sqlite(path) {
            // Two SQLite-backed stores exist: OpenCode's (schema_meta-free
            // key/value envelope db) and Hermes's `state.db` (UNI-15). The
            // fingerprint check is cheap and read-only.
            if let Ok(conn) = Connection::open_with_flags(
                path,
                rusqlite::OpenFlags::SQLITE_OPEN_READ_ONLY
                    | rusqlite::OpenFlags::SQLITE_OPEN_NO_MUTEX,
            ) {
                if hermes_sqlite_fingerprint(&conn) {
                    drop(conn);
                    return Self::from_hermes_sqlite(path, None);
                }
            }
            return Self::from_opencode_sqlite(path, None);
        }
        let text = read_utf8_or_diagnose(path)?;
        match detect_source(&text) {
            Some(SessionSource::Codex) => Self::from_codex_str(&text),
            Some(SessionSource::Pi) => Self::from_pi_str(&text),
            Some(SessionSource::OpenClaw) => {
                let mut session = Self::from_openclaw_str(&text)?;
                if session.meta.profile.is_none() {
                    session.meta.profile = openclaw_agent_id_from_path(path);
                }
                Ok(session)
            }
            Some(SessionSource::Grok) => {
                let mut session = Self::from_grok_str(&text)?;
                session.capture_grok_path_metadata(path);
                Ok(session)
            }
            Some(SessionSource::Gemini) => Self::from_gemini_str(&text),
            Some(SessionSource::Goose) => Self::from_goose_str(&text),
            // IX-3: a detected OpenCode session must route to its own
            // loader, not the Claude Code fallback below
            // (`docs/interop/build-followups.md`).
            Some(SessionSource::OpenCode) => Self::from_opencode_str(&text),
            _ => {
                let mut session = Self::from_claude_code_str_with_fidelity(&text, fidelity)?;
                if include_subagents {
                    session.attach_claude_subagents(path, &text, fidelity)?;
                }
                Ok(session)
            }
        }
    }

    /// Reconstruct multi-file subagent trees from a flat set of loaded sessions.
    ///
    /// Codex stores subagents as separate rollout files linked to their parent
    /// by `lineage["parent_thread_id"]` (→ the parent's `session_id`). Given a
    /// collection of sessions, this nests each child into its parent's
    /// [`Session::subagents`] and returns only the roots. Children whose parent
    /// isn't in the set are returned as roots themselves (best effort).
    pub fn reconstruct_tree(sessions: Vec<Session>) -> Vec<Session> {
        use std::collections::HashMap;
        // Index each session's position by its session_id.
        let mut idx: HashMap<String, usize> = HashMap::new();
        for (i, s) in sessions.iter().enumerate() {
            if let Some(id) = &s.meta.session_id {
                idx.insert(id.clone(), i);
            }
        }
        // Determine each session's parent (by index), if present in the set.
        let parent_of: Vec<Option<usize>> = sessions
            .iter()
            .map(|s| {
                s.meta
                    .lineage
                    .get("parent_thread_id")
                    .and_then(|p| idx.get(p).copied())
            })
            .collect();

        // Move children into parents, deepest-first so chains nest correctly.
        let mut slots: Vec<Option<Session>> = sessions.into_iter().map(Some).collect();
        let mut order: Vec<usize> = (0..slots.len()).collect();
        order.sort_by_key(|&i| std::cmp::Reverse(depth_of(i, &parent_of)));
        for i in order {
            if let Some(p) = parent_of[i] {
                if p != i {
                    if let Some(child) = slots[i].take() {
                        if let Some(parent) = slots[p].as_mut() {
                            parent.subagents.push(child);
                        } else {
                            slots[i] = Some(child); // parent already moved; keep as root
                        }
                    }
                }
            }
        }
        slots.into_iter().flatten().collect()
    }

    /// Parse a session of a known format from an in-memory JSONL string.
    pub fn load_str(jsonl: &str, format: SessionFormat) -> Result<Session> {
        match format {
            SessionFormat::ClaudeCode => Self::from_claude_code_str(jsonl),
            SessionFormat::Codex => Self::from_codex_str(jsonl),
            SessionFormat::Pi => Self::from_pi_str(jsonl),
            SessionFormat::OpenCode => Self::from_opencode_str(jsonl),
            SessionFormat::Grok => Self::from_grok_str(jsonl),
            SessionFormat::Gemini => Self::from_gemini_str(jsonl),
            SessionFormat::Goose => Self::from_goose_str(jsonl),
        }
    }

    /// Serialize this session to JSONL in the given format.
    ///
    /// The conversation is synthesized from the canonical messages, so this
    /// works for sessions loaded from *either* tool as well as ones supercode
    /// built itself. Converting between formats (e.g. Codex → Claude Code) is an
    /// "export": format-specific framing that has no slot in the target may be
    /// dropped, but the user/assistant/tool conversation is preserved.
    pub fn to_jsonl(&self, format: SessionFormat) -> Result<String> {
        match format {
            SessionFormat::ClaudeCode => Ok(self.to_claude_code_jsonl()),
            SessionFormat::Codex => Ok(self.to_codex_jsonl()),
            SessionFormat::Pi => Ok(self.to_pi_jsonl()),
            SessionFormat::OpenCode => self.to_opencode_jsonl(),
            SessionFormat::Grok => Ok(self.to_grok_jsonl()),
            SessionFormat::Gemini => Ok(self.to_gemini_jsonl()),
            SessionFormat::Goose => Ok(self.to_goose_json()),
        }
    }

    /// Export back to `format`, replaying the imported `raw` prefix
    /// **verbatim** — original uuids/ids, real timestamps, and
    /// loader-skipped records (e.g. Claude Code `file-history-snapshot`) that
    /// [`Self::to_jsonl`]'s full synthesis discards or fakes — when `format`
    /// is the session's own origin (`format.source() == self.meta.source`,
    /// see [`SessionFormat::source`]) and there is a `raw` prefix to replay.
    /// Only messages appended *after* import (tracked by
    /// [`Self::imported_message_count`]) are synthesized, chained onto the
    /// last original record found in the raw prefix.
    ///
    /// `session_id` of `Some(new)` rewrites the session id on every emitted
    /// line, raw and synthesized alike (`sessionId` for Claude Code,
    /// `session_meta.payload.id` for Codex); `None` leaves ids as recorded.
    ///
    /// Cross-format export (no verbatim prefix exists in the target dialect,
    /// by definition) and a session with no `raw` lines both fall back
    /// unchanged to [`Self::to_jsonl`] — full synthesis, same output as
    /// today. A12 (SPEC.md §6): this turns "export back to origin" from
    /// *semantic* to *near-byte* fidelity for the dominant hop-back case;
    /// cross-format stays at the documented semantic tier.
    pub fn to_jsonl_spliced(
        &self,
        format: SessionFormat,
        session_id: Option<&str>,
    ) -> Result<String> {
        if self.parse_error_lines > 0
            || self
                .subagents
                .iter()
                .any(|subagent| subagent.parse_error_lines > 0)
        {
            return Err(Error::InvalidSession(
                "refusing spliced export because the loaded session contains parse loss"
                    .to_string(),
            ));
        }
        if self.raw.is_empty() || format.source() != self.meta.source {
            if let Some(session_id) = session_id {
                let mut rewritten = self.clone();
                rewritten.meta.session_id = Some(session_id.to_string());
                return rewritten.to_jsonl(format);
            }
            return self.to_jsonl(format);
        }
        match format {
            SessionFormat::ClaudeCode => Ok(self.to_claude_code_jsonl_spliced(session_id)),
            SessionFormat::Codex => Ok(self.to_codex_jsonl_spliced(session_id)),
            SessionFormat::Pi => self.to_pi_jsonl_spliced(session_id),
            SessionFormat::OpenCode => self.to_opencode_jsonl_spliced(session_id),
            SessionFormat::Grok => Ok(self.to_grok_jsonl_spliced()),
            SessionFormat::Gemini => Ok(self.to_gemini_jsonl_spliced(session_id)),
            SessionFormat::Goose => Ok(self.to_goose_json_spliced(session_id)),
        }
    }

    /// Write this session to `path` in the given format.
    pub fn save(&self, path: impl AsRef<Path>, format: SessionFormat) -> Result<()> {
        std::fs::write(path.as_ref(), self.to_jsonl(format)?)?;
        Ok(())
    }

    /// Reconstruct the exact source bytes this `Session` was loaded from,
    /// out of [`Self::raw`] + [`Self::raw_trailing_newline`] (the exact
    /// inverse of the strict-verbatim capture those two fields record — see
    /// `join_lines_verbatim`).
    ///
    /// For a genuinely line-oriented source (Claude Code, Codex, Pi, and an
    /// OpenCode *envelope*-form JSONL), `raw` is captured verbatim from the
    /// original text, so this reproduces the original file byte-for-byte —
    /// the P008/P009 diagonal-convert fix (`convert <file> --to
    /// <same-format>` is byte-identical to `<file>`) is built on exactly
    /// this. The one documented exception is an OpenCode **export-document**
    /// source (a single pretty-printed JSON value, not JSONL): `raw` there
    /// is RE-SYNTHESIZED as one envelope line per record (see
    /// `from_opencode_export_doc`'s contract), so this returns a
    /// verbatim reproduction of THAT captured representation rather than the
    /// original pretty-printed document — a known, narrow residue, not a
    /// silent loss (the same records are all still present).
    pub fn raw_verbatim(&self) -> String {
        join_lines_verbatim(&self.raw, self.raw_trailing_newline)
    }

    /// The session as it stood before its message `keep` (0-based): its first `keep` normalized
    /// messages, for a fork at a turn (edit a sent message, retry a reply). The verbatim source lines
    /// and the side records captured from them (Claude Code's file snapshots and last prompt, Codex's
    /// provenance) describe the whole session and are keyed by source record, not by message, so they
    /// are dropped: every export of the cut is regenerated from the kept messages. A subagent is kept
    /// only when it is linked to a spawning tool call that is kept; an unlinked one cannot be placed
    /// before or after the cut and is dropped. The source session is unchanged. `keep` past the end
    /// keeps every message.
    pub fn cut_before(&self, keep: usize) -> Session {
        let messages: Vec<ChatMessage> = self.messages.iter().take(keep).cloned().collect();
        let spawned: std::collections::HashSet<&str> = messages
            .iter()
            .flat_map(|message| message.tool_calls.iter().flatten())
            .map(|call| call.id.as_str())
            .collect();
        let subagents = self
            .subagents
            .iter()
            .filter(|subagent| {
                subagent
                    .meta
                    .parent_tool_use_id
                    .as_deref()
                    .is_some_and(|id| spawned.contains(id))
            })
            .cloned()
            .collect();
        let mut meta = self.meta.clone();
        meta.native_residue = Vec::new();
        meta.native_residue_source = None;
        meta.codex_provenance = Vec::new();
        Session {
            meta,
            messages,
            subagents,
            raw: Vec::new(),
            raw_trailing_newline: true,
            imported_message_count: None,
            raw_is_verbatim: false,
            parse_error_lines: self.parse_error_lines,
            load_residue: self.load_residue.clone(),
        }
    }

    /// P5-5 (design §2 module 21 `session.tree`, §2.1 D-6 "session.tree →
    /// core.session(tree-addressable transcript)"): materialize this
    /// session's linear [`Self::messages`] into a native in-place
    /// [`crate::session_tree::SessionTree`] — the bridge a caller uses the
    /// FIRST time it wants to run a tree operation (rewind/branch/label)
    /// against an otherwise-linear [`Session`]. `created_at_ms` stamps every
    /// synthesized node (see
    /// [`crate::session_tree::SessionTree::from_linear`]'s doc comment for
    /// why a single timestamp is used: the source linear messages carry no
    /// per-turn timestamp of their own here).
    ///
    /// This does not mutate `self` or persist anything — see
    /// the composition layer's session-store tree writer for persistence, and
    /// [`Self::apply_session_tree`] for the inverse bridge.
    pub fn to_session_tree(&self, created_at_ms: i64) -> crate::session_tree::SessionTree {
        crate::session_tree::SessionTree::from_linear(&self.messages, created_at_ms)
    }

    /// P5-5: the inverse of [`Self::to_session_tree`] — overwrite
    /// [`Self::messages`] with `tree`'s ACTIVE branch's linear projection
    /// (C7's "tree-with-linear-projection": this is exactly what keeps every
    /// existing linear consumer — the agent loop, exporters — working
    /// unchanged after a tree operation runs). Nothing else on `self`
    /// (`meta`, `raw`, ...) is touched.
    ///
    /// **Fail-closed.** Propagates [`crate::session_tree::SessionTree::linear_projection`]'s
    /// `Err` rather than applying anything — a structurally-corrupt tree
    /// (a cycle, a dangling leaf, an active branch pointing at nothing) must
    /// error, not silently overwrite [`Self::messages`] with an empty `Vec`.
    /// `self` is left untouched on `Err` (the assignment only happens after
    /// the projection has already succeeded).
    pub fn apply_session_tree(&mut self, tree: &crate::session_tree::SessionTree) -> Result<()> {
        self.messages = tree.linear_projection()?;
        Ok(())
    }
}

/// Read `path` as UTF-8 text, translating a non-UTF-8 failure into a clear,
/// format-aware diagnostic (PARITY-16) instead of the raw "stream did not
/// contain valid UTF-8" `io::Error` — named path, and what supercode DOES
/// accept there. Binary SQLite input never reaches this function: callers
/// check [`looks_like_sqlite`] first and route to
/// [`Session::from_opencode_sqlite`] instead.
pub(super) fn read_utf8_or_diagnose(path: &Path) -> Result<String> {
    let bytes = read_session_bytes(path)?;
    String::from_utf8(bytes).map_err(|_| {
        crate::Error::Other(format!(
            "{} is not valid UTF-8 text, and is not a recognized OpenCode SQLite store \
             (no `SQLite format 3` header) — supercode reads Claude Code / Codex / Pi / \
             OpenCode session logs as UTF-8 JSONL, or an OpenCode `opencode*.db` SQLite file",
            path.display()
        ))
    })
}

/// Whether `path` is a zstd-compressed transcript (`*.jsonl.zst`, Codex's cold-rollout form).
pub fn is_zstd_session_path(path: &Path) -> bool {
    path.extension().and_then(|e| e.to_str()) == Some("zst")
}

/// Open a session file for reading, decompressing a `*.zst` transcript as it is read.
pub fn open_session_reader(path: &Path) -> std::io::Result<Box<dyn Read>> {
    let file = std::fs::File::open(path)?;
    if !is_zstd_session_path(path) {
        return Ok(Box::new(file));
    }
    let decoder = ruzstd::decoding::StreamingDecoder::new(file).map_err(|error| {
        std::io::Error::new(
            std::io::ErrorKind::InvalidData,
            format!("{} is not a readable zstd stream: {error}", path.display()),
        )
    })?;
    Ok(Box::new(decoder))
}

/// Every byte of a session file, decompressed when it is `*.zst`.
pub fn read_session_bytes(path: &Path) -> std::io::Result<Vec<u8>> {
    if !is_zstd_session_path(path) {
        return std::fs::read(path);
    }
    let mut bytes = Vec::new();
    open_session_reader(path)?.read_to_end(&mut bytes)?;
    Ok(bytes)
}

/// Read only the portion of a JSONL transcript a bounded scrollback can use.
///
/// The first record carries durable session metadata (especially for Codex),
/// while the trailing window carries the messages the viewport will render.
/// Full lossless loaders intentionally continue to read every byte.
pub(super) fn read_display_jsonl(
    path: &Path,
    message_limit: usize,
) -> Result<(Option<SessionSource>, String, bool)> {
    const MIN_TAIL_BYTES: u64 = 4 * 1024 * 1024;
    const MAX_TAIL_BYTES: u64 = 64 * 1024 * 1024;
    const BYTES_PER_MESSAGE: u64 = 16 * 1024;

    // A compressed transcript cannot be windowed by seeking; it is read whole.
    if is_zstd_session_path(path) {
        let text = read_utf8_or_diagnose(path)?;
        return Ok((detect_source(&text), text, false));
    }
    let mut first = String::new();
    BufReader::new(std::fs::File::open(path)?).read_line(&mut first)?;
    let source = detect_source(&first);
    if !matches!(
        source,
        Some(SessionSource::ClaudeCode | SessionSource::Codex | SessionSource::Gemini)
    ) {
        let text = read_utf8_or_diagnose(path)?;
        return Ok((detect_source(&text), text, false));
    }

    let mut file = std::fs::File::open(path)?;
    let file_len = file.metadata()?.len();
    let requested = (message_limit.max(1) as u64)
        .saturating_mul(BYTES_PER_MESSAGE)
        .clamp(MIN_TAIL_BYTES, MAX_TAIL_BYTES);
    if file_len <= requested {
        let text = read_utf8_or_diagnose(path)?;
        return Ok((source, text, false));
    }

    let start = file_len - requested;
    file.seek(SeekFrom::Start(start))?;
    let mut bytes = Vec::with_capacity(requested as usize);
    file.read_to_end(&mut bytes)?;
    // The window normally starts in the middle of a JSON record. Discard that
    // partial prefix so every line passed to the existing parsers is valid.
    if let Some(newline) = bytes.iter().position(|byte| *byte == b'\n') {
        bytes.drain(..=newline);
    }
    let mut tail = String::from_utf8(bytes).map_err(|_| {
        crate::Error::Other(format!(
            "{} contains non-UTF-8 data in its display window",
            path.display()
        ))
    })?;
    if start > 0 {
        // Always recover the human boundary immediately before the byte
        // window, even when the window already contains newer prompts. A
        // long run of large tool records can otherwise make the numeric tail
        // begin in one old turn while its only retained users belong to much
        // newer turns. The display projector then (correctly) hides the
        // orphaned activity, making pagination appear inert.
        //
        // Search backward independently of the render window and retain only
        // two complete human JSONL records. The search grows geometrically but
        // never reads more than the same 64 MiB hard ceiling as the display
        // window, and none of the intervening tool bytes are normalized or
        // sent over RPC.
        let max_search_bytes = start.min(MAX_TAIL_BYTES);
        let mut search_bytes = requested.min(max_search_bytes);
        let anchors = loop {
            let search_start = start - search_bytes;
            file.seek(SeekFrom::Start(search_start))?;
            let mut search = Vec::with_capacity(search_bytes as usize);
            (&mut file).take(search_bytes).read_to_end(&mut search)?;
            if search_start > 0 {
                if let Some(newline) = search.iter().position(|byte| *byte == b'\n') {
                    search.drain(..=newline);
                } else {
                    search.clear();
                }
            }
            // `start` normally cuts the record whose remainder the tail
            // reader discarded. Exclude its incomplete prefix here too.
            if let Some(newline) = search.iter().rposition(|byte| *byte == b'\n') {
                search.truncate(newline + 1);
            } else {
                search.clear();
            }
            let anchors = std::str::from_utf8(&search)
                .ok()
                .map(|search| {
                    let mut found = search
                        .lines()
                        .rev()
                        .filter(|line| native_display_human_line(line, source))
                        .take(2)
                        .map(str::to_string)
                        .collect::<Vec<_>>();
                    found.reverse();
                    found
                })
                .unwrap_or_default();
            if anchors.len() >= 2 || search_start == 0 || search_bytes == max_search_bytes {
                break anchors;
            }
            search_bytes = search_bytes.saturating_mul(2).min(max_search_bytes);
        };
        if !anchors.is_empty() {
            tail = format!("{}\n{tail}", anchors.join("\n"));
        }
    }
    let text = if matches!(source, Some(SessionSource::Codex | SessionSource::Gemini)) {
        format!("{first}{tail}")
    } else {
        tail
    };
    Ok((source, text, true))
}