Struct Agent 

pub struct Agent { /* private fields */ }

Implementations

impl Agent

pub fn new(config: Config) -> Result<Agent, Error>

Build an agent backed by an OpenAI-compatible endpoint (OpenRouter by default). The API key is taken from Config::api_key, then Config::api_key_cmd (P4: a credential-helper command, run via the shell — see run_api_key_cmd), then the configured environment variable (Config::api_key_env).

pub fn with_provider(config: Config, provider: Box<dyn Provider>) -> Agent

Build an agent with an explicit provider and the built-in tools. Handy for tests (inject a mock provider) or custom transports.

pub fn with_parts( config: Config, provider: Box<dyn Provider>, registry: ToolRegistry, ) -> Agent

Build an agent from all three parts.

pub fn provider_arc(&self) -> Arc<dyn Provider>

A handle to this agent’s model transport, for sharing with subagents.

pub fn config(&self) -> &Config

Read-only access to this agent’s resolved Config — e.g. so a caller (crates/cli’s attach_mcp) can consult Config::module_registry/Config::module_activation AFTER construction without having to separately thread the config through every call site that builds an Agent and later needs it again. Same trust boundary as every other already-public Agent accessor (history, provider_arc) — the caller is the same process that built this Config in the first place, not a new exposure surface.

pub fn checkpoint_observer(&self) -> Option<&CheckpointObserver>

P5-9 (§2 module 20 checkpoint): this agent’s checkpoint engine, if Config::checkpoint_enabled is true and the shadow store opened successfully — None otherwise (the default-off case, or a graceful-degrade after an I/O failure). A caller (CLI/TUI/embedder) uses this to list/turn_diff/restore WITHOUT re-deriving the shadow-store root itself. Deliberately named checkpoint_observer, not checkpoint — Self::checkpoint already names the unrelated in-memory conversation-position marker (see that method’s doc comment).

pub fn lsp_manager(&self) -> Option<&LspManager>

P5-11 (§2 module 28 lsp): this agent’s LSP server registry, if Config::lsp_enabled is true — None otherwise (the default-off case). impl Drop for Agent already covers production teardown via crate::lsp::LspManager::kill_all_sync (a real, group-killing OS process kill — see crate::lsp’s module doc). This accessor exists for an OPTIONAL caller (CLI/TUI/embedder) that manages its own Agent lifecycle and additionally wants to reach crate::lsp::LspManager::shutdown_all for a graceful LSP shutdown/exit handshake BEFORE dropping the agent — nothing calls shutdown_all automatically today.

pub async fn run_subagent( &self, system: impl Into<String>, task: impl Into<String>, ) -> Result<String, Error>

Spawn a subagent that shares this agent’s model transport, runs task to completion with its own fresh conversation (seeded with system), and returns its final answer. The analog of Agent / spawn_agent.

pub fn with_provider_arc(config: Config, provider: Arc<dyn Provider>) -> Agent

Like Self::with_provider but sharing an existing transport handle.

pub fn run_in_background( self, prompt: impl Into<String>, ) -> JoinHandle<(Agent, Result<String, Error>)>

where Agent: Send + 'static,

Run a prompt on a background task, returning a handle that resolves to the final answer (and the agent, so the caller can continue it). The analog of background/async agent runs.

pub fn resume(config: Config, session: Session) -> Result<Agent, Error>

Build an agent and seed it with a previously-recorded Session so it can continue where Claude Code or Codex left off.

pub fn resume_recorded( config: Config, session: Session, sidecar_path: &Path, ) -> Result<Agent, Error>

Like Self::resume, but also begins recording (A2/A3): a fresh native-v2 sidecar is created at sidecar_path from session (header + session.raw verbatim — the imported prefix’s own fidelity), and every subsequent turn this agent produces is appended to it at full fidelity, independent of whatever cap_tool_output/maybe_compact (D6) do to history.

Invariant this establishes ONLY once Self::set_reduction_policy is also called (the D6/A7 supersession gate, Self::run_loop): at any instant, Session::from_native_str(sidecar).messages equals session.messages (the imported prefix) followed by every message appended since — i.e. self.history()[1..] (history[0] is this agent’s own system prompt, per Self::load_session; it is never part of session and is never written to the sidecar). Recording alone (no policy) leaves the gate off: cap_tool_output still runs on oversized tool results, and history can diverge from the sidecar for them — honestly, via the notice’s “full output in session sidecar” label, never silently.

pub fn set_recorder(&mut self, w: SidecarWriter)

Install (or replace) this agent’s sidecar recorder (A3).

pub fn set_journal(&mut self, journal: SessionJournal)

Install (or replace) this agent’s append-only session journal — the durable, flush-per-record log of every message it produces plus every queue/rewind/plan operation performed on it. Installing one alone changes nothing about the conversation; it only makes the session survive a crash mid-turn.

pub fn has_journal(&self) -> bool

Whether an append-only journal is installed.

pub fn journal_checkpoint(&self, messages: usize)

Declare the durable view caught up: <name>.jsonl now holds messages messages and every journal record before this point is already in it. Everything journaled AFTER the last such record is exactly what a crash would have lost — see crate::session_journal::JournalState::unpersisted.

pub fn journal_usage(&self, record: &UsageRecord)

BP-13: record one per-turn usage entry in the append-only journal.

pub fn journal_model_change(&self, record: &ModelChangeRecord)

BP-13: record one mid-session model change in the append-only journal — the ONE persisted home for a routing record (BP-8’s journal), never a second file.

pub fn session_tree(&self) -> Option<&SessionTree>

BP-8 (catalog:151): this session’s conversation tree, when the module is on.

pub fn set_session_tree(&mut self, tree: SessionTree)

Install a tree loaded from the store (a resume), replacing whatever this agent built. A no-op when the module is off — a session whose preset does not enable session_tree must not acquire one through the back door of an old sidecar.

pub fn rebuild_session_tree_from_history(&mut self)

BP-8: rebuild the tree from the current linear history — used after a resume that loaded a transcript but had no .tree.json to restore (every session recorded before the module was on).

pub fn rewind_conversation(&mut self, keep: usize) -> RewindOutcome

BP-8 (catalog:152 “Rewind/rollback conversation”): move THIS conversation back to an earlier point — the whole row, not the last-exchange special case Self::rewind_to serves and not sessions fork --at, which makes a different session.

keep is a message count (index into history), so keep = 1 leaves only the system message. Three things happen, in this order:

1. the removed tail is pushed onto an undo stack, so Self::undo_rewind can put it back;
2. a crate::session_journal::JournalOp::Rewind record is APPENDED — nothing is deleted from disk, so the rewound-away messages remain recoverable from the log;
3. when the tree module is on, the active branch’s leaf moves to the node at keep, and the old leaf is preserved under a fresh sibling branch — the next message appended forks there rather than overwriting.

Returns what it did. Rewinding to a point at or past the end is a no-op with removed = 0, never an error.

pub fn undo_rewind(&mut self) -> bool

BP-8: invert the most recent Self::rewind_conversation — the messages come back, and the inversion is itself an appended journal record. false when there is nothing to undo.

pub fn append_recovered_messages(&mut self, messages: &[ChatMessage])

BP-8 (catalog:150): append messages recovered from the journal after a crash — they were already recorded, so this deliberately does NOT re-journal them; it puts the live conversation back where the interrupted process left it.

pub fn undoable_rewinds(&self) -> usize

BP-8: how many rewinds are currently undoable.

pub fn restore_rewind_undo(&mut self, stack: Vec<Vec<ChatMessage>>)

BP-8: restore the undo stack a previous process left in the journal, so /rewind undo works across a restart.

pub fn restore_queues(&mut self, steer: &[String], follow_up: &[String])

BP-8 (catalog:154 “Queued-prompt persistence”): re-queue pending inputs recovered from the journal WITHOUT re-recording them — they are already in the log, and journaling them again would double them on the next restart.

pub fn plan(&self) -> Vec<PlanEntry>

BP-8 (catalog:156 “Todos/plan persisted per session”): the session’s current update_plan checklist.

pub fn set_plan(&mut self, steps: Vec<PlanEntry>)

BP-8: restore a plan read back from the store on resume. Marked as already-journaled, so a resume that changes nothing writes nothing.

pub fn set_reduction_policy(&mut self, policy: ReductionPolicy)

Install (or replace) this agent’s reduction policy (A5/A7/A10). Once set, every provider request is built from a projected view of history[1..] (reduce::project_messages) rather than history verbatim — history itself is never shrunk or mutated by this; only the request view does.

pub fn reduction_policy(&self) -> Option<&ReductionPolicy>

This agent’s reduction policy, if one is installed.

pub fn set_schema_tier(&mut self, tier: SchemaTier)

Change the global tool-schema tier (TR-8/T5) mid-session. Takes effect starting with the NEXT request this agent builds. Under CachePlan::ImportedPrefix, the first request built after a change is flagged as a cache-bust event and its cache-control annotation is skipped for that one request (see [provider::tier_change_is_cache_bust], consulted in Self::build_request_messages) — normal annotation resumes on the next request if the tier doesn’t change again.

pub fn set_tool_schema_tier( &mut self, name: impl Into<String>, tier: SchemaTier, )

Override the schema tier for a single tool (TR-8/T5) mid-session, same cache-bust interaction as Self::set_schema_tier.

pub fn set_span_summarizer( &mut self, summarizer: impl SpanSummarizer + Send + Sync + 'static, )

Install (or replace) this agent’s TR-7 span summarizer — the injectable side-call Self::build_request_messages uses to turn an A10 TurnsCleared span into an LLM-written summary paragraph when policy.summarize_cleared_turns is on. Installing one alone changes nothing: ReductionPolicy::summarize_cleared_turns (off by default) is the actual gate, so tests/callers that want the deterministic stub can simply never call this.

pub fn set_span_summarizer_arc( &mut self, summarizer: Arc<dyn SpanSummarizer + Send + Sync>, )

Install an already-shared summarizer — same seam as Self::set_span_summarizer, for callers (and tests) that need to keep their own handle on it.

pub fn prepare_cleared_turns_summary( &self, msgs: &[ChatMessage], policy: &ReductionPolicy, prior: &ReductionLog, ) -> Option<PreparedClearSummary>

Prepare TR-7 metadata with this agent’s installed summarizer for a projection performed by an outer driver before session history/log are loaded (the CLI foreign-resume preflight). None preserves the deterministic fallback when the gate is off, no summarizer exists, the span is below the cost floor, or the side-call fails.

pub fn set_event_sink(&mut self, sink: Box<dyn Fn(AgentEvent) + Send + Sync>)

P5-4: install (or replace) this agent’s crate::EventSink AFTER construction — Config::event_sink is otherwise only set at Config-build time (before Agent::new), which is too early for a tui embedder that only knows it’s activating (and needs to replace whatever print-mode/REPL sink was already installed with one that feeds its own render loop instead of writing straight to stdout) once it already holds a live Agent. Mirrors [Self:: set_permissions_approval_handler]’s “installing one alone changes nothing beyond what already consults Config::event_sink” pattern — this is a plain replacement, not a new activation gate.

pub fn permissions_approval_cache(&self) -> &ApprovalCache

P5-1: install (or replace) this agent’s permissions-engine approval handler — see crate::permissions::PermissionsApprovalHandler. This is the non-interactive decision seam a CLI/TUI/SDK embedder implements for the Ask-tier prompt; the TUI’s actual interactive UI is a separate module (P5 row 4), not built here. Installing one alone changes nothing: Config::permissions_enabled (off by default) is the actual gate — with no handler installed, every Ask-tier decision denies (fail-closed, see that trait’s doc comment).

P5-10 (§2 module 12, escalation = "ask"): the SAME handler also backs a sandbox-unenforceable ask decision (crate::sandbox::decide_fs‘s approval parameter) — one installed seam serves both permissions.rules’ Ask tier and permissions.sandbox’s escalation = "ask", rather than requiring an embedder to install two near-identical handlers. Kept in sync on self.ctx (not just self.permissions_approval_handler) because BashTool::execute/PersistentShellTool::execute only ever see &ToolContext, never &Agent — see ToolContext:: sandbox_approval_handler’s doc comment. BP-10 (catalog row “Session approval caching”): this agent’s approval cache — the door an embedder/TUI uses to inspect or REVOKE remembered grants (ApprovalCache::clear forgets every one, in memory and on disk, and the next matching call asks again). Also how a test proves a grant really did survive the process: store_path() names the file a second agent reads back.

pub fn set_permissions_approval_handler( &mut self, handler: impl PermissionsApprovalHandler + 'static, )

pub fn set_user_question_handler( &mut self, handler: Arc<dyn McpElicitationHandler>, )

BP-3 (§2 module 6 tools.question): install the door ask_user asks the human through — the elicitation/create handler the design names as the module’s protocol side. SDK-owned frontends pass the broker-backed handler (crate::server::FrontendRequestBridge:: elicitation_handler), which is what makes the question a real frontend request the turn waits on. None (the default, nothing installed) leaves the tool deny-default: it reports that nobody can be asked instead of blocking.

Installing one alone changes nothing about whether the tool EXISTS — [capabilities.tools_question] is that gate, applied by crate::tools::ToolRegistry::from_config.

pub fn plan_mode(&self) -> &Arc<PlanModeState>

BP-3 (§2 module 8): the shared plan-mode state, so a frontend (the REPL’s /plan, a TUI toggle) can enter or leave the read-only research phase the same tools and permission gate see.

pub fn set_legacy_approval_handler( &mut self, handler: Box<dyn Fn(&ToolCall) -> bool + Send + Sync>, )

Install the compatibility approval seam used when the composable permissions engine is disabled. SDK-owned interactive frontends call this alongside Self::set_permissions_approval_handler so the same authenticated request channel works under either policy engine; the selected engine remains entirely a configuration decision.

pub fn set_subagent_store( &mut self, store: Arc<SessionStore>, session_name: impl Into<String>, )

P5-3 (§2 module 9 D5 “subagent transcripts… persisted + linked”): install a crate::store::SessionStore (+ this agent’s own session name in it) so spawn_subagent persists each child’s transcript (via crate::store::SessionStore::save_subagent_transcript) and lineage record (via crate::store::SessionStore::save_subagent_lineage) once the child finishes. Installing one alone changes nothing about whether spawning WORKS — Config::subagents_enabled is the actual gate; this only controls whether a completed spawn’s transcript additionally lands on disk.

pub fn set_claude_runtime_manifest(&mut self, manifest: ClaudeRuntimeManifest)

Seed the Claude runtime manifest reconstructed during resume.

Installing state enables the matching Claude runtime tool schemas so a disk-reloaded continuation does not lose that vocabulary, but never starts a timer by itself. The supplied execution posture is preserved: an embedding scheduler may deliberately activate before installing it.

pub fn restore_claude_project_agents(&mut self) -> Result<usize, Error>

Reinstall project-scoped Claude named-agent definitions when a Supercode continuation carrying a Claude runtime manifest is reopened from disk. The manifest is the durable capability marker; definitions themselves remain authoritative in <cwd>/.claude/agents/*.md.

pub fn claude_runtime_manifest(&self) -> Option<&ClaudeRuntimeManifest>

Current imported Claude runtime state, including paused mutations made by Cron*/ScheduleWakeup, for persistence by the embedding loop.

pub fn claude_runtime_manifest_mut( &mut self, ) -> Option<&mut ClaudeRuntimeManifest>

Mutable access for an embedding scheduler driver to atomically claim due events and persist the resulting manifest. Merely borrowing this state does not start a timer; execution remains the driver’s explicit responsibility.

pub fn set_child_approval_handler_factory( &mut self, factory: impl Fn(String, Arc<Mutex<Vec<QueuedApproval>>>) -> Arc<dyn PermissionsApprovalHandler> + Send + Sync + 'static, )

P5-4 (tui, closes the P5-3 §2.2 C6 deferred chain): install a factory this agent’s Self::run_spawn_subagent calls (with the fresh child’s own id and this agent’s shared Self::pending_child_approvals queue) to build the PermissionsApprovalHandler a background_prompts = "parent" child gets, INSTEAD of the default crate::subagents::ParentQueueApprovalHandler. Installing one alone changes nothing about whether background spawning works — Config::subagents_background_prompts being crate::subagents::BackgroundPromptsPolicy::Parent is the actual gate that reaches this factory at all; a Parent-policy child spawned before this is installed (or on an agent that never installs it) still gets the immediate-deny default, unchanged.

Security note. The factory only controls WHICH handler answers an Ask-tier request — it can never widen what gets asked in the first place: crate::permissions::approval::resolve_ask only calls a handler’s ask when the rule engine has already resolved the call to Ask (Deny short-circuits before any handler is consulted; Allow never needs one), so a parent’s “allow” answer here can only grant what the policy already routed to a prompt — never override a Deny the engine already decided.

pub fn pending_child_approvals(&self) -> Vec<QueuedApproval>

P5-3 (§2.2 C6 “parent-surfaced queue”): every approval request a background_prompts = "parent" child has raised so far, oldest first — a read-only audit view, not a mutable queue the caller answers. Under P5-3’s own default handler (no P5-4 TUI factory installed) every entry here WAS already resolved Deny (a background call can’t wait for an answer with no handler installed) — but once a crate::tui::TuiChildApprovalHandler factory is installed (P5-4, Self::set_child_approval_handler_factory), the underlying call genuinely blocks and may resolve Allow/AllowForSession; this method still records the SAME entry for the audit trail either way, so “queued here” no longer implies “was denied” in general — see crate::subagents::QueuedApproval’s doc comment.

pub fn set_session_titler( &mut self, titler: impl SessionTitler + Send + Sync + 'static, )

P4b: install (or replace) this agent’s auto-title side-call — see crate::session_title::SessionTitler. Installing one alone changes nothing: Config::auto_title (off by default) is the actual gate a caller should consult before calling Self::auto_title.

pub fn auto_title(&self) -> Option<String>

P4b: produce a title for this agent’s current conversation via the installed Self::set_session_titler side-call. Returns None (never panics, never blocks longer than the titler itself does) if no titler is installed, or the side-call itself declined (see crate::session_title::auto_title). Does NOT consult Config::auto_title itself — that gate is the caller’s responsibility, matching Self::span_summarizer’s precedent of keeping the mechanism and the policy gate separate.

pub fn usage_records(&self) -> &[UsageRecord]

P4b (§1.6, catalog §4a “persisted per-turn usage records”): every crate::usage_log::UsageRecord this agent has accumulated so far.

pub fn save_usage_log( &self, store: &SessionStore, name: &str, ) -> Result<(), Error>

P4b: persist this agent’s accumulated usage log to store under name — a thin wrapper over crate::store::SessionStore::save_usage_log so callers don’t need to import both types.

pub fn turn_records(&self) -> &[TurnRecord]

BP-7 (catalog §4a “Turn/step bracketing records”): every crate::turn_record::TurnRecord this agent has accumulated — the context/usage/finish brackets of each model round-trip plus the retry, abort, effort and goal markers between them.

pub fn save_turn_records( &self, store: &SessionStore, name: &str, ) -> Result<(), Error>

BP-7: persist the marker log to store under name (<name>.events.jsonl), the same thin-wrapper shape Self::save_usage_log has.

pub fn total_cost_usd(&self) -> f64

BP-7 (catalog §4a “Per-turn cost/usage accounting”): dollars this agent has spent so far. 0.0 when the model is unpriceable — read Self::model_priced to tell “free” from “unknown”.

pub fn model_priced(&self) -> bool

BP-7: whether this build can price this agent’s model, i.e. whether Self::total_cost_usd is a real figure rather than a floor.

pub fn total_steps(&self) -> usize

BP-7 (catalog §4a “Turn/budget caps”): tool calls this agent has executed so far — the counter Config::max_steps bounds.

pub fn note_abort(&mut self, source: &str)

BP-7 (catalog §4a “Interrupt/abort with state preserved”): record that the in-flight turn was interrupted.

Called by whoever owns the cancellation (the CLI’s Ctrl-C race), NOT by the loop itself: a cancelled send future is dropped mid-await, so the loop never runs another line. The partial work already appended to the transcript stands; this marker is what makes the interruption a persisted FACT — the residue the ledger row named — rather than something a reader has to infer from a dangling tool call on reload. Emits AgentEvent::TurnAborted as the live counterpart.

pub fn set_goal(&mut self, objective: impl Into<String>) -> bool

Set (or revise) this session’s standing objective.

Returns false, changing nothing, when capabilities.todos.goals is off — the module gate, not a silent success. A goal restates itself at the tail of every request until Self::clear_goal, and each change appends a goal marker to the turn-record log.

pub fn goal(&self) -> Option<&GoalRecord>

This session’s standing objective, if one is set.

pub fn clear_goal(&mut self) -> bool

Drop the standing objective. true when there was one to drop.

pub fn save_goal(&self, store: &SessionStore, name: &str) -> Result<(), Error>

BP-7: persist (or, when cleared, remove) the standing objective beside the session — same thin-wrapper shape as Self::save_usage_log.

pub fn restore_goal(&mut self, goal: Option<GoalRecord>)

BP-7: adopt a goal loaded from the store (a resumed session picks up exactly where it left off). Bypasses the module gate on purpose: a goal already persisted is data to restore, not a new capability being turned on, and dropping it silently would lose session state.

pub fn effort(&self) -> Option<&str>

The reasoning-effort level in force for the NEXT request, or None when extended thinking is off.

pub fn set_effort(&mut self, effort: Option<String>) -> Option<String>

Change the reasoning-effort level mid-session.

Some(level) sets the level; None turns extended thinking OFF — the on/off toggle the ledger row named as distinct from the level. run_loop reads self.config.effort fresh when it builds each ChatRequest, so this takes effect on the very next request with no other copy to update (the same contract Self::set_model has). The change is appended to the turn-record log as an effort marker, the extended-thinking analog of the model_change log.

Returns the PREVIOUS setting.

pub fn review_prompt(&self, args: &str) -> Option<String>

The purpose-built review turn’s prompt: the code-review template from Config::prompts with {args} replaced by args.

None when the resolved config carries no code-review template — the preset decides whether this harness has a review mode, and the template IS the report format (both parity presets pin one).

pub async fn review(&mut self, args: &str) -> Result<String, Error>

Run the review turn: an ordinary Self::send of Self::review_prompt, so the review’s request, tools, transcript and records are the session’s own — a purpose-built TURN, not a second agent.

pub async fn side_question(&self, question: &str) -> Result<String, Error>

Answer question over this session’s FULL current context without recording anything.

Three properties, all load-bearing and all asserted by this build’s tests: the request carries the whole conversation as the next turn would see it; it advertises NO tools, so the model can only answer; and neither history, the sidecar recorder, the usage log nor the turn-record log is touched — &self, not &mut self, is the type system saying so. cc’s /btw and cx’s /side.

pub fn queue_steer(&self, message: impl Into<String>)

P4b (§1.7, pi§3 semantics): queue a mid-turn steering message — delivered “after current tool calls” (pi’s phrasing): at the top of Self::run_loop’s NEXT iteration, before the next model request is built, regardless of whether this turn is still mid-flight with pending tool calls. Drained per Config::steering_mode.

pub fn queue_follow_up(&mut self, message: impl Into<String>)

P4b: queue a follow-up message — delivered “at idle” (pi’s phrasing): only once Self::run_loop would otherwise return a final answer (no more tool calls pending). Drained per Config::follow_up_mode.

pub fn queued_steer_count(&self) -> usize

P4b: how many steering messages are currently queued (mid-turn + follow-up combined) — mostly for tests/diagnostics.

pub fn reduction_log(&self) -> &ReductionLog

The accumulating reduction log (A5) — every reduction applied to any projected request view so far. Combined with a full-fidelity sidecar Session, this is enough to reduce::invert any projected view back to the exact original.

pub fn set_context_limit(&mut self, limit: u64)

PARITY-18 D4 — arm the per-send context guard: Self::run_loop will refuse (via Error::ContextLimitExceeded) to build and issue ANY request — the first or any later turn — whose crate::tokens::context_guard verdict is “does not fit” against limit. Call this once the target model’s context-window size is known (resume --reduced’s preflight already computes it). Leaving this unset (the default) is a no-op: no guard runs, exactly today’s pre-PARITY-18 behavior.

pub fn context_limit(&self) -> Option<u64>

This agent’s armed context limit, if Self::set_context_limit has been called.

pub fn model(&self) -> &str

The model identifier this agent sends on its next request (Config::model, as of construction/resume or the last Self::set_model call).

pub fn set_model(&mut self, model: impl Into<String>)

UX-30 dev/02 — switch the model this agent sends, starting with the NEXT request it builds (and every one after, until changed again). Self::run_loop reads self.config.model fresh on every request (see its ChatRequest construction), so this alone is enough — there is no cached/baked-in copy anywhere else to also update. Takes effect immediately; safe to call only between turns (the REPL’s /model picker runs at the prompt, never mid-turn). Touches nothing else: history, the sidecar, and reduction state are exactly as untouched as Self::set_schema_tier leaves them for a mid-session tier change.

P4c-review note: this is the LOW-LEVEL primitive — it swaps Config::model and nothing else. It does NOT run dep 8’s reasoning-artifact filter (reduce::rehydrate::filter_reasoning_artifacts) and does NOT create a crate::model_change::ModelChangeRecord, so calling it directly for a mid-session handoff between two DIFFERENT models leaves model-A’s reasoning artifacts in history for model-B to inherit. Self::switch_model is the safe superset — gated by Config::model_switch_allow_switch, it filters and records the switch before delegating to this method — and is what callers performing a governed mid-session model switch should use instead.

pub fn switch_model(&mut self, model: impl Into<String>)

P4c (§1.10/§3.1 core.model_switch.allow_switch, D9 row, dep 8, design’s “core NEW-significant” item): the mid-session model switch — a superset of Self::set_model gated by Config::model_switch_allow_switch.

allow_switch = false (the default): EXACTLY Self::set_model — same single field write, nothing else touched, no crate::model_change::ModelChangeRecord created. Byte-identical to calling set_model directly.

allow_switch = true: additionally, before the swap takes effect, runs reduce::rehydrate::filter_reasoning_artifacts over Self::history — model-A’s reasoning/thinking artifacts (any crate::message::ChatMessage::metadata key in reduce::rehydrate::REASONING_METADATA_KEYS, any content_parts block whose "type" is in reduce::rehydrate::REASONING_CONTENT_PART_TYPES) are stripped BEFORE model-B ever builds a request from this history — then appends a typed, translatable crate::model_change::ModelChangeRecord to Self::model_change_records (persist it via Self::save_model_change_log). A switch TO the current model (model == Self::model()) is treated as a no-op — still exactly set_model’s mechanics, no record for a switch that didn’t actually change anything (and nothing to filter FOR, since there was no handoff).

pub fn record_model_change( &mut self, from: &str, to: &str, reason: Option<&str>, )

BP-13 — the ONE place a mid-session model change is performed and recorded, shared by Self::switch_model (a user asked) and the run loop’s fallback pass (a provider failed).

It does four things, in this order, and nothing else: strips model-A reasoning artifacts out of the live history (dep 8 — model B must never inherit them), moves Config::model, appends the typed crate::model_change::ModelChangeRecord, and writes that same record into the append-only session journal (BP-8) — which is where every persisted routing record lives; there is no second file. The change is also EMITTED, so a surface that renders events shows the switch instead of silently answering as a different model.

pub fn set_service_tier(&mut self, tier: Option<String>)

BP-13 (catalog D9 “Fast mode / service tiers”): set or clear the session-level service-tier override. Some(tier) WINS over the [capabilities.model_catalog] service_tier rule for every subsequent request (it is the live toggle the user just pulled); None puts the configured rule back in charge. Takes effect on the next request the loop builds, like Self::set_model.

pub fn model_change_records(&self) -> &[ModelChangeRecord]

P4c: every crate::model_change::ModelChangeRecord this agent has accumulated so far (via Self::switch_model with allow_switch on). Empty when the knob is off or no switch has happened yet.

pub fn save_model_change_log( &self, store: &SessionStore, name: &str, ) -> Result<(), Error>

P4c: persist this agent’s accumulated model-change log to store under name — the crate::model_change::ModelChangeRecord analog of Self::save_usage_log.

pub fn git_metadata(&self) -> Option<&GitMetadataRecord>

P4e (§1.6/§3.1 core.session.git_metadata, catalog:331): this agent’s captured git provenance, if Config::session_git_metadata was on at construction and the best-effort probe found a repo.

pub fn save_git_metadata( &self, store: &SessionStore, name: &str, ) -> Result<(), Error>

P4e: persist this agent’s captured git metadata to store under name — a thin wrapper over crate::store::SessionStore::save_git_metadata, the crate::git_metadata::GitMetadataRecord analog of Self::save_usage_log. A no-op (Ok(()), nothing written) when Self::git_metadata is None.

pub fn session_persist(&self) -> bool

P4e DEFECT-FIX (independent Fable-5 review of P4e: core.session.persist had a Config field and CLI plumbing at ConfigProfile → Config but no consumer at all): whether a CLI caller’s session-store save sites (persist_session, persist_full_view) should actually write to disk. true (the default) is byte-identical to pre-fix behavior — every session persists. false makes a session ephemeral: it runs exactly as before, but no <name>.jsonl/sidecar family is ever written for it. A plain getter, same posture as Self::model — this crate itself never reads or enforces it; the CLI’s save sites do.

pub fn session_name(&self) -> Option<&str>

P4e DEFECT-FIX (independent Fable-5 review of P4e: core.session.name had a Config field and CLI plumbing but no consumer): the caller-configured session name, if [core.session] name was set. None (the default) leaves session naming exactly as before — mint_session_name’s auto-generated <tag>-<adjective>-<noun> shape. A plain getter, same posture as Self::session_persist.

pub fn request_issued(&self) -> bool

PARITY-18 D3 — whether this agent has actually issued at least one live request to its Provider so far (set the instant Self::run_loop reaches its real send site, regardless of whether that call then succeeds or fails). Callers should report “request sent” from THIS, never from having merely passed the context guard or having called Self::send — either of those can happen with zero requests actually issued (a guard refusal, an interactive session quit before any turn completes).

pub fn cache_established(&self) -> bool

P5-2 (§2.2 C2): whether this agent currently considers its CachePlan::ImportedPrefix cache entry warm — mirrors Self::request_issued’s read-only-observability precedent, so a caller (or a test) can confirm Self::register_tool’s C2 invalidation actually took effect without reaching into private state.

pub fn imported_prefix_len(&self) -> Option<usize>

B7: length of the imported-prefix protected by CachePlan::ImportedPrefix (this agent’s own system message plus every message of a previously-imported session), set by Self::load_session. None until a session has been loaded.

pub fn set_reduction_log(&mut self, log: ReductionLog)

Replace this agent’s accumulating reduction log (C4: /expand//reduce mutate the log directly via reduce::invert_one/reduce::project_messages and must feed the result back here so the next request build or persist sees the updated state instead of silently recomputing from an empty log). Also lets a caller (resume_cmd, C1) seed the log with the initial projection it already computed for the entry banner, so reduction_log() reflects reality even before this agent’s first send() (which is otherwise the only place build_request_messages populates it).

pub fn load_session(&mut self, session: Session)

Replace the conversation with a loaded session, keeping this agent’s own system prompt at the front. The session’s own system/developer turns are preserved after it for context.

pub fn save_transcript(&self, path: impl AsRef<Path>) -> Result<(), Error>

Persist the live conversation to path as JSONL (one ChatMessage per line) so the session can be resumed later — supercode’s own sessions become first-class, resumable artifacts.

pub fn load_transcript(&mut self, path: impl AsRef<Path>) -> Result<(), Error>

Restore a conversation previously written with Self::save_transcript, replacing the current history.

pub fn checkpoint(&self) -> usize

Take a checkpoint of the current conversation position. Pass it to Self::rewind_to to discard everything sent since (the rewind/undo analog of fork/checkpoint).

pub fn rewind_to(&mut self, checkpoint: usize)

Rewind the conversation to a Self::checkpoint, discarding later turns.

pub async fn send_with_files( &mut self, text: impl Into<String>, files: &[PathBuf], ) -> Result<String, Error>

Send a message with file inputs attached — the --file / -i analog. Each file’s contents are injected into the prompt: UTF-8 text inline, binary (e.g. images) noted with a size marker. (Native image vision would additionally require multimodal content parts.)

pub async fn send_with_images( &mut self, text: impl Into<String>, image_urls: &[String], ) -> Result<String, Error>

Send a message with image inputs to a vision model — the -i/--image analog. image_urls may be https://… links or data:image/…;base64,… URLs; they’re attached as multimodal image_url content parts.

pub fn expand_prompt(&self, input: &str) -> String

Expand a /<name> <args> slash command against the registered prompt templates ({args} is replaced with the trailing text). Non-matching input is returned unchanged. BP-6 additionally resolves SKILL.md invocations here, after the template table misses: /skill:name args (pi§2 “Skill commands”), /name args when the config follows Claude Code (cc§7: “a SKILL.md in a directory = a /name command”), and $slug mentions (cx§7). $ARGUMENTS in the body is replaced with the trailing text.

pub fn skills(&self) -> &[LoopSkill]

The SKILL.md packages this agent discovered (frontmatter only) — the exact set its prompt index lists and its skill tool can load.

pub async fn expand_prompt_async(&self, input: &str) -> String

P5-2 (§2 module 15 D7 row 4 “prompts-as-commands”): like Self::expand_prompt, but also consults MCP-server-sourced prompts registered via Self::register_mcp_prompt when the local Config::prompts table has no match — a live prompts/get round-trip, which is why this is async and Self::expand_prompt itself stays synchronous (its public sync signature is unchanged, for every existing caller that doesn’t need MCP prompts).

Argument mapping (a scope decision, not a protocol requirement — the MCP spec leaves “how does free CLI text become named prompt arguments” to the client): a prompt with zero or one declared arguments gets the whole trailing text (empty string if the prompt takes no arguments and none was given); a prompt with two or more declared arguments expects key=value pairs, whitespace-separated (/mcp__server__prompt lang=rust topic=async) — an unparseable pair (no =) is simply skipped, never a hard error (matches this method’s “non-matching input passes through” fail-open posture for the LOCAL-prompt case above).

pub fn register_mcp_prompt( &mut self, command_name: impl Into<String>, source: impl SdkPromptSource + 'static, )

P5-2 (§2 module 15 D7 row 4): register an MCP server’s prompt as a slash-command source — command_name MUST already be the namespaced mcp__<server>__<prompt> form (crate::mcp::McpServerHandle::prompts produces exactly that shape); this method does not re-namespace or validate it, so a caller that hands it a bare name defeats the collision protection crate::mcp::McpPromptSource’s doc comment describes. Overwrites any prior registration under the same command name (re-attaching the same server replaces its own earlier prompt list; this can never touch a NON-mcp__-prefixed key, i.e. never a local Config::prompts entry).

pub fn append_system_note(&mut self, text: &str)

P5-2 (§2 module 15 D7 row 5 “instructions”): fold an MCP server’s initialize-time instructions (or any other free-text note) into this agent’s system message — the context-assembly site every other core.*/capabilities.* prompt-section append already uses (Self::with_parts), except this one fires AFTER construction (attaching MCP servers happens once the agent already exists — see crates/cli/src/main.rs’s attach_mcp). A no-op if history is somehow empty or its first message isn’t a system message (never true for an Agent built via Self::new/Self::with_parts, but checked rather than assumed).

pub fn refresh_env_context(&mut self) -> bool

BP-4 (catalog:90, cx§2 <environment_context> “re-emitted on change”): re-derive the # Environment block and, if anything in it moved — cwd, the approval/sandbox policy, the git branch or its dirty state, the date — replace the stale copy in the system message with the fresh one. Returns whether the block changed.

A no-op (and free — no git subprocess) when core.env_context is off, which is the default and every non-parity config. Replacing in place rather than appending a second block is deliberate: two # Environment sections disagreeing about cwd is worse context than one stale one, and the system message is re-sent on every request, so the rewrite IS the re-emission the model sees.

pub fn model_input(&mut self) -> ChatRequest

BP-5 (catalog D2 “Prompt-input debugging”: render the exact model-visible input for inspection; cx§2 codex debug prompt-input, which “renders the exact model-visible input list as JSON”): the request this agent would send next.

Built by the SAME two calls the loop makes (Self::build_request_messages, Self::tool_schemas) and assembled by the SAME Self::chat_request — it is the real request, not a reconstruction of one. &mut self because build_request_messages is: rendering the input is exactly as stateful as building it for a send.

pub async fn model_input_for(&mut self, prompt: &str) -> ChatRequest

BP-5: Self::model_input for a turn that has not been sent — prompt is expanded exactly as Self::send would expand it (slash templates, skills, @path mentions, MCP prompts) and appended to the conversation IN MEMORY, then the request is rendered.

Deliberately not recorded: this door inspects an input, it does not take a turn. Nothing is written to the session store, no journal entry is made, and no request is issued.

pub fn render_model_input(req: &ChatRequest) -> Value

BP-5: a ChatRequest as the JSON a human (or jq) inspects — the system prompt, every message in order, and every advertised tool schema, plus the sampling controls that travel with them.

pub fn base_prompt(&self) -> &str

BP-5 (catalog D2 “Per-model-family base-prompt selection”): the base system prompt currently in force for this agent’s model.

pub fn inject_context_block( &mut self, name: impl Into<String>, content: impl Into<String>, ) -> bool

BP-4 (catalog:91 “Synthetic context-injection blocks”): splice one named ambient block into the live context — the seam a hook’s additionalContext, a frontend nudge or an orchestrator’s brief enters through, mid-session, after construction.

Requires core.context_injections (returns false otherwise): the gate governs the whole registry, not just its startup half. The block is appended to the system message and remembered, so it is carried by every later request and re-rendered by crate::context_injection::assemble wherever the prompt is rebuilt.

pub fn spliced_context_blocks(&self) -> &[ContextInjectionBlock]

The blocks spliced in since construction — see Self::inject_context_block.

pub fn maybe_compact(&mut self) -> bool

Compact the conversation if it has grown past the configured threshold.

Re-founded (A10): with a ReductionPolicy installed (Self::set_reduction_policy), this no longer touches self.history at all. It derives policy.clear_turns_older_than from compact_after_messages so the next projected request view (reduce::project_messages, built in Self::run_loop) collapses the old turns into one reversible TurnsCleared stub instead — history() and the sidecar keep every message forever; only the view shrinks. Returns whether the live (unreduced) history currently exceeds the threshold, i.e. whether a clearing will actually be visible in the next projected view.

Legacy path (no policy) — LOSSY, kept only for byte-identical backward compatibility (D6): destructively rewrites self.history, permanently discarding the dropped middle turns (replaced by a single non-reversible summary marker that becomes their SOLE remaining copy — exactly the lossy compaction this reduction layer differentiates against). Once a sidecar/recorder or a ReductionPolicy is in play, prefer installing a policy so this method takes the re-founded path above instead.

pub fn compact_now(&mut self, focus: Option<&str>) -> bool

BP-4 (catalog:98 “Manual compact with focus instructions”, cc§2 / cx§2 /compact [instructions]): compact NOW, regardless of whether either automatic trigger has fired — the mechanism behind the REPL’s /compact [focus].

focus is this invocation’s steering text: it overrides the standing core.compaction.focus_instructions for this compaction only, is carried into the SUMMARIZER’s input (so the model-written summary preserves what the user asked for), and is stated on the marker. An empty/whitespace focus falls back to the configured standing value, which is what a bare /compact means.

Returns whether anything was compacted (false when the history is already at or below the keep-window, or when a ReductionPolicy is installed — under a policy the reversible A10 path owns clearing, and a manual compact would be the lossy one).

pub fn new_context( &mut self, objective: &str, keep_recent: Option<usize>, ) -> usize

BP-4 (catalog:106 “Handoff (fresh objective + curated keep-set)”, cx§1 new_context): reset the live working view to a fresh objective plus a curated keep-set, in-session.

The new view is: the system prompt (plus any imported prefix a CachePlan::ImportedPrefix config protects — same clamp compaction uses), then a handoff marker stating the objective and what was set aside, then the most recent keep_recent messages (None = the token-budget-derived count core.compaction.keep_recent_tokens already governs, so the keep-set is curated by the same budget the rest of the compaction machinery uses, not by a magic number). The keep-set never begins on a tool result, so no tool message is left orphaned from its originating assistant turn.

Returns how many messages were set aside. Like compaction, the marker is PERSISTED through the recorder, so a resumed session can see where the handoff happened; and like compaction, the set-aside messages remain in the transcript sidecar whenever one is attached.

Scope note: this is the in-session new_context mechanism, NOT Config::handoff_enabled’s reversible ReductionLog snapshot (the offline supercode handoff projection) — that one is the reduction module’s, and stays there.

pub fn register_tool(&mut self, tool: impl Tool + 'static)

Register an additional tool (e.g. your own capability).

P5-2 (§2.2 C2 “connect invalidates cache prefix”): registering a tool AFTER this agent has already issued a request (Self::request_issued) changes the tools schema every subsequent request carries — the exact prefix-churn shape C2 describes, MCP-sourced or not. Resets Self::cache_established so the next cache-warmth check (provider::cache_cold_reason) doesn’t wrongly assume the entry is still warm. A no-op call before the first request (the common case: attach_mcp registers tools once at startup, before any turn runs) changes nothing — byte-identical to today.

pub fn history(&self) -> &[ChatMessage]

The current conversation, including the system prompt.

pub async fn send( &mut self, user_input: impl Into<String>, ) -> Result<String, Error>

Send a user message and run the loop until the model produces a final answer (text with no tool calls) or the iteration budget is exhausted.

pub fn context_usage(&self) -> ContextUsage

BP-4 (catalog:109 “Context-usage introspection”, cc§2 /context grid, cx§8 /status + get_context_remaining): the LIVE context-window accounting for this session — the same numbers resume --dry-run’s preflight already computes (tokens::estimate_request_tokens / tokens::context_guard), read out mid-session instead of only before one.

Pure: it projects the request view exactly as Self::guard_candidate_message does (reduction stubs included, cache annotation included) without mutating the reduction log, so asking “how full am I?” can never change what the next request carries.

pub fn tool_schemas(&self) -> Vec<ToolSchema>

The tool-schema array this agent would advertise on its NEXT request, exactly as Self::run_loop computes it. Public (PARITY-18 D1) so a caller can measure the real request-token cost of an agent’s tool surface — including the current crate::config::ToolAdvertising mode’s core/deferred split and the synthetic tool_search/expand_reduction/sidecar_search schemas — BEFORE ever calling Self::send, e.g. for a preflight context-guard check.

pub fn child_config_for_agent_type(&self, agent_type: &str) -> Option<Config>

BP-7 (catalog §4a “Named agent definitions as data”): the child Config a spawn_subagent of agent_type would build — the resolved posture a named definition actually produces, including its crate::subagents::AgentPermissions bundle applied through the tightening-only rules. None when no definition of that name is registered or discovered.

Exposed so a caller (and this build’s tests) can ask what a named agent WOULD run as without spawning it and paying for a turn.

pub fn reaped_subagent_ids(&self) -> Vec<String>

BP-7 (catalog §4a “Background subagents + resume”): the ids of children that have finished and been reaped, and can therefore be continued with [SUBAGENT_RESUME].

pub fn set_lifecycle_hook( &mut self, hook: Box<dyn Fn(&LifecycleEvent) + Send + Sync>, )

Installs the lifecycle observer (compaction and subagent boundaries).

pub fn turn_count(&self) -> usize

Number of non-system messages exchanged so far.

pub fn total_output_tokens(&self) -> u64

Cumulative output (completion) tokens reported by the provider across every send on this agent. Zero if the provider reports no usage.

Trait Implementations

impl Drop for Agent

P5-3 safety hardening (Fable-5 review, MEDIUM-LOW “orphaned billed spend”): a dropped parent must not leave a detached background child running against a REAL provider. Without this, a parent dropped mid-run (the caller’s own process exits the scope, panics, or simply stops polling) leaves every still-running BackgroundSubagent::handle as an orphaned tokio::spawn task: nothing had ever awaited or aborted it, so it runs to its own (max_iterations-bounded) completion regardless — bounded but real provider spend nobody is paying attention to.

.abort() on a tokio::task::JoinHandle is safe to call unconditionally, including on an ALREADY-finished task (a documented no-op there — see tokio’s JoinHandle::abort docs) — so this never needs to distinguish “still running” from “already done”; a background child that already finished and is merely awaiting a subagent_status reap is untouched in practice (aborting a finished task changes nothing observable). For a task still mid-flight, tokio cancels it at its next .await point, which drops that future in place — including the _guard: ConcurrencyGuard moved into it at spawn time (see Self::run_spawn_subagent’s tokio::spawn body) — so the concurrency-gauge slot is released exactly the same way a normal completion releases it (ConcurrencyGuard’s own Drop, in crate::subagents). No separate cleanup call site to forget.

Deliberately does NOT touch [Self::pending_child_approvals] or Self::subagent_store — this is purely “stop burning provider calls on behalf of a caller who’s gone”, not a transcript-persistence path (a child aborted mid-flight has no finished result to persist; see this build’s named residual on abort-time transcript loss).

fn drop(&mut self)

Executes the destructor for this type.

fn pin_drop(self: Pin<&mut Self>)

🔬This is a nightly-only experimental API. (pin_ergonomics)

Execute the destructor for this type, but different to Drop::drop, it requires self to be pinned.

impl From<Agent> for SdkAgent

fn from(agent: Agent) -> SdkAgent

Converts to this type from the input type.