@sema-agent/core 5.20.0 → 5.21.0

package/dist/agents/repair-loop.d.ts

@@ -4,49 +4,275 @@ import type { ExecutionEnv } from "../internal/harness-types.js";
 import type { VerificationResult } from "./verify.js";
 import type { Checkpoint } from "../core/checkpoint-store.js";
 import { type OracleIsolationVerdict } from "../core/oracle-isolation.js";
+/**
+ * SAFE-tier oracle self-repair loop (design/76 D1 命门, design/78 Slice-1). Given an implementation spec
+ * and an **injected, trusted oracle**, this runs the worker, grades it with the oracle, and — on failure —
+ * loops a fix turn (and a bounded clean-restart) until the oracle passes or the loop gives up. The terminal
+ * is then projected through the {@link terminalForTier} table, which **caps every PASS at `candidate_only`**:
+ * this layer NEVER auto-accepts. Auto-accept (`fixed`) requires the OFF Gate-1 out-of-process
+ * `oracleIsolation` boundary that this slice deliberately does not build — SAFE-tier escalates to a human
+ * (`needs_human_oracle`) or surfaces a candidate, but never clears its own work.
+ *
+ * Like {@link verifyCompleted}, this is a **thin composition** over `runner.runTask` (verify.ts:18-21 posture)
+ * — it adds no Runner-core surface, touches no vendored code, and is called by a leader/profile (the sibling
+ * of {@link runWithVerification}). The oracle internals (which gate/judge/property-harness to compose) live
+ * in the PROFILE-injected {@link RepairOracle} closure; core only fixes the {@link OracleResult} shape and the
+ * read-only / identity contract (§1 裁决①). N-parallel candidate fan-out + a robustness-first selection ranker
+ * are PROFILE concerns (§3 裁决②) — core stays single-candidate.
+ */
+/**
+ * The provenance/strength tier of the oracle that produced a verdict. The control plane assigns this — it is
+ * NEVER derived from agent-visible state (not a {@link TaskSpec} field), mirroring design/44 §7 Q4. Drives the
+ * {@link terminalForTier} projection.
+ * - `trusted_hidden` — a hidden, spec-derived held-out oracle (strongest; still candidate_only here).
+ * - `trusted_visible` — a visible trusted oracle (must be paired with an anchor + no property-harness regression + L3).
+ * - `property_harness_weak` — derived property invariants only (a weak signal).
+ * - `l3_judge_advisory` — an L3 read-only judge that advises but does not clear.
+ * - `none` — no oracle available → only a human can adjudicate.
+ */
 export type OracleTier = "trusted_hidden" | "trusted_visible" | "property_harness_weak" | "l3_judge_advisory" | "none";
+/**
+ * The terminal state the repair loop resolves to. SAFE-tier (this slice) can reach every value EXCEPT `fixed`:
+ * `fixed` is reserved for auto-accept, which needs the OFF Gate-1 out-of-process oracleIsolation boundary and
+ * is therefore unreachable here (terminal-by-tier §2 caps PASS at `candidate_only`).
+ * - `fixed` — auto-accepted (NOT reachable SAFE-tier; reserved for a future Gate-1 slice).
+ * - `candidate_only` — the oracle passed; surface the candidate for human acceptance, do not auto-accept.
+ * - `needs_human_oracle` — no usable oracle (`tier: "none"`) → only a human can adjudicate.
+ * - `gave_up` — attempts exhausted without a passing oracle (a first-class abstain, never "satisfy the test").
+ * - `conflict` — the spec and the oracle disagree (escalate; never rewrite the work to satisfy the test).
+ *   Slice-1 core does NOT synthesize this (the test-tampering diff-monitor is §6-deferred to the profile —
+ *   core honestly `gave_up`s rather than under-detect a conflict); it is in the vocabulary for the profile.
+ * - `oracle.unprotected` — fail-closed: the grader env was not isolated from the worker env (identity check).
+ *
+ * `needs_human_oracle` is disjoint from `needs_review` (a future dry-run gate) and `irreversible_ask` (shipped).
+ */
 export type RepairTerminal = "fixed" | "candidate_only" | "needs_human_oracle" | "gave_up" | "conflict" | "oracle.unprotected";
+/**
+ * One oracle verdict. Produced by the injected {@link RepairOracle} closure. `flaky` records that the oracle
+ * was non-deterministic across `retries` (a flaky verdict is NEVER projected to `fixed` — §5 / §6 flakyK).
+ * `trace` is the raw failure output fed back into the next fix turn; it is treated as untrusted worker-adjacent
+ * data (delimited into context, sanitized into the objective) but its INNER text is NEVER rewritten (§5.5 #8a)
+ * — rewriting it would destroy the real-trace feedback lever.
+ */
 export interface OracleResult {
     tier: OracleTier;
     passed: boolean;
+    /** Raw failure trace (on a failed verdict). Untrusted; fed back verbatim into the fix turn (never inner-rewritten). */
     trace?: string;
+    /** The oracle's verdict was non-deterministic across `retries` re-isolations. A flaky verdict never → `fixed`. */
     flaky: boolean;
+    /** How many re-isolation retries the oracle ran to settle the verdict. */
     retries: number;
+    /**
+     * HRD-AGO-6 — what producing this verdict COST (micro-USD, the oracle's own spend: judge/gate model calls,
+     * re-isolation retries, …). Optional: a mechanical oracle (an exec gate, a property harness) spends nothing
+     * and omits it. When supplied it counts against {@link RepairLoopConfig.costCeilingMicroUsd} exactly like the
+     * worker's spend — an LLM-judge oracle re-run per attempt is real money the ceiling used to be blind to —
+     * and is surfaced cumulatively as {@link RepairResult.oracleCostMicroUsd}. It is NOT folded into
+     * `TaskResult.stats` (those are the WORKER's, and the oracle runs outside `runner.runTask`).
+     */
     costMicroUsd?: number;
 }
+/**
+ * The PROFILE-injected oracle (§1 裁决①). The profile composes the actual grading inside this closure —
+ * `runExecGate(graderEnv, steps)` (exec-gate.ts:185, env is the FIRST param = L2 provenance bound to the
+ * grader) → `verifyCompleted` (verify.ts:212, the read-only L3 judge) → `checkInvariants`
+ * (property-harness.ts:217, explore-only) — and maps the composite to an {@link OracleResult}. Core never
+ * composes these internals (so the oracle type can't grow a long discriminated union); it only fixes the
+ * `OracleResult` shape and the read-only/identity contract. `graderEnv` is the isolated grader env; `evidence`
+ * is the diff/results to judge (recompute it from the post-resume working tree on a resume — verify.ts:328 BUG5).
+ */
 export type RepairOracle = (graderEnv: ExecutionEnv, evidence: string | undefined) => Promise<OracleResult>;
+/**
+ * The durable, JSON-safe repair state carried across an orthogonal durable suspend (§4). Serialized ONLY when
+ * an unrelated resource/HITL suspend interleaves the loop (happy-path is in-memory) and re-seeded on resume so
+ * `attemptCount` advances MONOTONICALLY (never reset). Contains NO functions/Dates (epoch ms if a timestamp is
+ * ever needed) so `structuredClone`/`JSON.stringify` round-trip it on the checkpoint. `baselinePassTests` is
+ * DELIBERATELY ABSENT (§4 MAJOR-C): the anchor is grader-computed out-of-process so a worker can't shrink it.
+ */
 export interface RepairBundle {
+    /** The latest failure trace (untrusted; fed back into the fix turn). */
     failureTrace: string;
+    /** Reflexion-style diagnoses accumulated across attempts (≤3, Reflexion cap — §1). */
     diagnostics: string[];
+    /** Hypotheses tried and rejected (so a re-seed doesn't re-explore them). */
     rejectedHypotheses: string[];
+    /** The in-loop attempt counter — re-seeded MONOTONICALLY on resume, never reset/max/downscaled (§4 / §5). */
     attemptCount: number;
+    /** The tier of the last oracle verdict (for the resumed loop's projection). */
     oracleTier: OracleTier;
+    /** F1 (ruling 2026-08-02) — cumulative loop spend (micro-USD: worker + nested + oracle) at the last
+     *  attempt boundary. Re-seeded on resume so a resumed loop CONTINUES the money account instead of
+     *  restarting it at 0 (a resume must not wash `costCeilingMicroUsd`). Optional: bundles serialized
+     *  before this field existed read as 0 — the pre-F1 behavior, monotonic direction only. Updated at
+     *  attempt boundaries HERE; a mid-attempt park is topped up by the RUNNER (RB-481): the checkpoint mint
+     *  folds the in-flight leg's spend into the serialized copy on a HUMAN-DECISION park (whose resume reports
+     *  its own leg only), and deliberately does not on a resource-slice park (whose resume returns that money
+     *  through its cumulative stats). Either way the parked account is complete, never washed. */
     spentMicroUsd?: number;
+    /** F1 — cumulative ACTIVE loop wall-time (ms) at the last attempt boundary. ACTIVE means time the loop
+     *  was actually running: the suspended interval (a human deciding a HITL gate for days) is excluded by
+     *  construction, so `totalTimeoutMs` bounds loop work, not human latency. Same optionality/monotonic
+     *  contract as {@link RepairBundle.spentMicroUsd}. */
     activeElapsedMs?: number;
 }
 export interface RepairLoopConfig {
+    /** The PROFILE-injected, trusted oracle (§1 裁决①). Core calls it; it never composes the oracle internals. */
     oracle: RepairOracle;
+    /**
+     * The ISOLATED grader env the oracle grades in. MUST be a distinct object from {@link workerEnv}: if they
+     * are the SAME reference the loop fails closed to `oracle.unprotected` BEFORE running anything (§5.1 —
+     * necessary-not-sufficient; the real out-of-process boundary is the OFF Gate-1).
+     */
     graderEnv: ExecutionEnv;
+    /**
+     * The worker's execution env (the env `runner.runTask` runs the impl in), passed by the trusted caller so
+     * the §5.1 identity check can run: `graderEnv === workerEnv` → fail-closed `oracle.unprotected`. `TaskSpec`
+     * deliberately has NO `executionEnv` (the worker env is wired on the Runner deps / `executionEnvFactory`, not
+     * reachable from `implSpec`), so the caller must supply the reference here for the check to be meaningful.
+     * Omit it only when the worker env genuinely can't collide with the grader (e.g. distinct factories) — the
+     * check is then skipped (no reference to compare) and isolation is the caller's deployment contract.
+     */
     workerEnv?: ExecutionEnv;
+    /**
+     * design/77 §1 Gate-1 (oracleIsolation): the paths the oracle/spec lives behind that the worker must NOT be
+     * able to corrupt (e.g. a hidden held-out test dir). When supplied **together with** {@link workerEnv}, the
+     * loop runs the FULL structural {@link assertOracleIsolation} (identity + structural-class + bash write-probe)
+     * instead of the bare reference-identity check, and surfaces {@link RepairResult.isolationClass}. Omit it (the
+     * default) to keep the existing necessary-not-sufficient reference-identity check (the merged Slice-1 posture).
+     *
+     * The real isolated grader env is SERVICE-provided ({@link import("../core/types.js").RunnerDeps.graderEnvFactory});
+     * `graderEnv` here is its `.env`, and the control-plane `provenance` brand is reattached internally for the
+     * assertion. The assertion only matters for the (mandate-OFF) auto-accept path: `isolationClass` caps a
+     * non-`out_of_process` grader at `candidate_only`. SAFE-tier never auto-accepts regardless.
+     */
     immutableOraclePaths?: string[];
+    /**
+     * In-loop attempt ceiling (the loop runs at most this many GENERATE/oracle attempts, PLUS the documented
+     * one-shot CLEAN_RESTART the state machine grants once at exhaustion). Profile validates this to 2-3; core
+     * only enforces it as a ceiling, it does not fence the value — except that a NON-FINITE one falls back
+     * to 1 and is reported on {@link RepairLoopConfig.onWarn} (CLS-A-3/A-7 class: a NaN ceiling makes every
+     * `>=` gate false, i.e. no ceiling at all). This counter is DISTINCT from
+     * `suspendCount`(maxSuspends) and `sliceCount`(maxSlices) — it never borrows those budgets (§5 / §7).
+     *
+     * HRD-AGO-4: enforced BEFORE each dispatch as well as after, so a loop entered with {@link resumeBundle}
+     * whose `attemptCount` already reached the ceiling dispatches NOTHING and returns `gave_up`
+     * (`errorCode: "repair.attempts_exhausted"`) instead of spending one more worker run (plus a restart).
+     */
     maxAttempts: number;
+    /**
+     * On a resume after an orthogonal durable suspend, the restored {@link RepairBundle} (from
+     * `PrepareResume.seed.repairBundle`). `attemptCount` is re-seeded from it MONOTONICALLY (§4) — never reset.
+     */
     resumeBundle?: RepairBundle;
+    /**
+     * Stop the loop once cumulative cost reaches this (verify.ts:278 backstop). HRD-AGO-6 — "cumulative" is
+     * every leg the loop pays for: each attempt's own + nested `runTask` cost PLUS each verdict's
+     * {@link OracleResult.costMicroUsd} (an LLM-judge oracle is charged per attempt too).
+     */
     costCeilingMicroUsd?: number;
+    /** Overall ACTIVE wall-clock ceiling for the whole loop (verify.ts:277 backstop). F1: both ceilings
+     *  are CROSS-RESUME — the carried {@link RepairBundle.spentMicroUsd}/{@link RepairBundle.activeElapsedMs}
+     *  re-seed the accounts, so a durable resume continues the budget rather than restarting it. */
     totalTimeoutMs?: number;
+    /** Per-attempt callback (observability). */
     onAttempt?: (info: {
         attempt: number;
         terminal?: RepairTerminal;
         oracle: OracleResult;
     }) => void;
+    /**
+     * Operational-warning sink (observability). MINOR-2: the §5.1 isolation identity check is SKIPPED when
+     * `workerEnv` is undefined (no reference to compare). An ACCIDENTAL omission — `graderEnv` supplied but
+     * `workerEnv` forgotten — would then disable the check silently. When wired, that case is surfaced here
+     * once at loop start so the gap is observable; the posture stays necessary-not-sufficient (no hard-fail).
+     */
     onWarn?: (warning: Error) => void;
 }
 export interface RepairResult extends VerificationResult {
+    /** The projected terminal (§2). SAFE-tier never returns `fixed`. */
     terminal: RepairTerminal;
+    /** The final repair state (in-memory unless an orthogonal suspend serialized it — §4). */
     bundle: RepairBundle;
+    /**
+     * design/77 §1 Gate-1: the structural isolation class of the grader env, when the full
+     * {@link assertOracleIsolation} ran (i.e. {@link RepairLoopConfig.immutableOraclePaths} + `workerEnv` were
+     * supplied). `"out_of_process"` is the ONLY class the (mandate-OFF) auto-accept path may consider;
+     * `"in_process_probe_only"` CAPS the run at `candidate_only`. Undefined when only the bare reference-identity
+     * check ran. SAFE-tier never auto-accepts regardless of this value.
+     */
     isolationClass?: OracleIsolationVerdict["isolationClass"];
+    /**
+     * HRD-AGO-6 — cumulative ORACLE spend (micro-USD) across every verdict this loop took, summed from
+     * {@link OracleResult.costMicroUsd}. Reported SEPARATELY from `stats` (the worker legs) because the oracle
+     * runs outside `runner.runTask`: `stats.costMicroUsd + stats.nested.costMicroUsd + oracleCostMicroUsd` is the
+     * loop's true total. `0` when no verdict reported a cost (a mechanical oracle).
+     */
     oracleCostMicroUsd: number;
 }
+/**
+ * design/77 §1 Gate-1 invariant — auto-accept (`fixed`) is UNREACHABLE here. The terminal-by-tier projection
+ * (§2) already caps every PASS at `candidate_only`, so `isolationClass` only ever matters for the (OFF)
+ * auto-accept path. This guard makes the invariant load-bearing rather than caller discipline: a non-isolated
+ * (`in_process_probe_only`) or unprotected verdict can NEVER carry a `fixed` terminal — at most
+ * `candidate_only` / `oracle.unprotected`. It is a pure assertion over the projected terminal; it never
+ * upgrades anything (auto-accept stays OFF — only a future Gate-1 slice running on an `out_of_process` grader
+ * may even consider it).
+ */
 export declare function isolationPermitsAutoAccept(verdict: OracleIsolationVerdict): boolean;
+/**
+ * CONSUMER seam (design/78 Slice-1, MAJOR-3 wiring — the re-entry side of the round-trip). Read the durable
+ * {@link RepairBundle} a resumed checkpoint carries, so a leader/profile resuming a repair-interleaved
+ * suspend can re-seed a FRESH {@link runRepairLoop} call with it.
+ *
+ * ## The contract (core vs profile)
+ * - **CORE owns the persistence + restore.** When an orthogonal durable suspend (resource/HITL) interleaves a
+ *   `runRepairLoop` attempt, the Runner serializes the live bundle onto the minted checkpoint
+ *   (`prepareTask.serializeCheckpointState` sources `internals.repairBundle`), and on resume re-seeds it back
+ *   into `prepareTask` from `cp.state.repairBundle`. So `cp.state.repairBundle` is round-trip-true after a
+ *   `runner.resume(token, …)`: it survives the suspend with `attemptCount` intact (never reset — §4 / MAJOR-A).
+ * - **PROFILE owns the re-entry loop.** The leader that resumes the worker (`runner.resume`) is a PROFILE
+ *   concern — core does NOT build the "resume → grade → re-seed `runRepairLoop`" loop. The profile reads the
+ *   restored bundle (via this accessor) off the SAME checkpoint it resumed, and on its NEXT repair attempt
+ *   passes it as {@link RepairLoopConfig.resumeBundle}. `runRepairLoop` then re-seeds `attemptCount`
+ *   MONOTONICALLY from it (`config.resumeBundle` → the loop's `bundle`), so the repair ratchet is preserved
+ *   across the durable suspend rather than starting the attempt budget over.
+ *
+ * Returns `undefined` when the checkpoint carries no repair state (a non-repair task, or a repair attempt that
+ * suspended on its very first GENERATE before any oracle verdict — the re-seed treats `undefined` as a clean
+ * start). The returned object is the checkpoint's own (already an independent `structuredClone` from the
+ * store); pass it straight into `resumeBundle` — `runRepairLoop` copies its arrays defensively on re-seed.
+ */
 export declare function repairBundleFromCheckpoint(cp: Checkpoint): RepairBundle | undefined;
+/**
+ * The terminal-by-tier projection (§2) — a PURE function, the core invariant. **SAFE-tier caps every PASS at
+ * `candidate_only`**; `fixed` (auto-accept) is unreachable here because it needs the OFF Gate-1 out-of-process
+ * boundary. Enforced HERE (in the projection table), not as caller discipline, so no consumer can promote a
+ * PASS to `fixed`.
+ *
+ * - any tier, `passed: true` → `candidate_only` (trusted_hidden/trusted_visible/property_harness_weak alike).
+ * - `l3_judge_advisory` → advisory only; even a "pass" can't clear → `candidate_only` (it never self-clears).
+ * - `none` → `needs_human_oracle` (no usable oracle → only a human adjudicates).
+ * - not-passed with a usable tier → not a terminal yet (the loop continues / gives up); represented as
+ *   `undefined` so the caller's state machine drives the `gave_up`/`conflict` path explicitly.
+ */
 export declare function terminalForTier(oracle: OracleResult): RepairTerminal | undefined;
+/**
+ * Run the SAFE-tier oracle self-repair loop over `implSpec` (design/78 §1 state machine — a DETERMINISTIC
+ * switch, no LLM in the control flow):
+ *
+ * `ASSERT_ISOLATION` (identity check → fail-closed `oracle.unprotected`) → `GENERATE` (`runner.runTask`) →
+ * `ORACLE` (the injected closure) → { passed → terminal-by-tier projection } | { failed & attempt<max → REPAIR
+ * (a fix turn) → loop } | { failed & attempt==max → CLEAN_RESTART once, else `gave_up` } | { conflict → terminal }.
+ *
+ * `ASSERT_ISOLATION` is the design/77 §1 Gate-1 seam: when {@link RepairLoopConfig.immutableOraclePaths} +
+ * `workerEnv` are supplied it runs the FULL {@link assertOracleIsolation} (identity + structural-class + bash
+ * write-probe) and surfaces {@link RepairResult.isolationClass}; otherwise it runs the bare reference-identity
+ * check (the merged Slice-1 posture). Either way a failed assertion fails closed to `oracle.unprotected` before
+ * any work runs, and the result NEVER reaches `fixed` — `isolationClass` only gates the (mandate-OFF) auto-accept
+ * path that this SAFE-tier loop does not take.
+ *
+ * SAFE-tier: the projection caps every PASS at `candidate_only` — this NEVER auto-accepts. `runRepairLoop` is a
+ * SIBLING of {@link verifyCompleted}: it composes over `runner.runTask` and never touches the Runner core.
+ */
 export declare function runRepairLoop(runner: Runner, implSpec: TaskSpec, config: RepairLoopConfig): Promise<RepairResult>;

package/dist/agents/retain-ledger.d.ts

@@ -1,78 +1,219 @@
 import type { Runner } from "../core/runner/runtask.js";
 import type { TaskSpec } from "../core/types.js";
 import type { RunInternals } from "../core/runner/prepare-task.js";
+/** design/122 D2 — per-child resume cap (default 8): a child can be revived at most this many times. */
 export declare const SUBAGENT_RESUME_CAP = 8;
+/**
+ * design/122 D1 — one RETAINED child ledger row. `specSnapshot` is a FROZEN plain value-copy of the spawn
+ * childSpec (r1-m5: retaining the live `ctx`/spec-builder closures would pin the parent run's whole prepare
+ * graph — harness ref, tool wrappers — in memory for up to ttl×max); `internalsSnapshot` is the spawn-time
+ * trusted internals copy (parentToolCallId / onForwardEvent / onSubagentSpawn — the grandchild chain stays
+ * unbroken on resume, D2). `release` captures the DELEGATION runner (deployment-scoped, allowed) and
+ * performs the unpin + explicit release that restores throwaway semantics.
+ */
 export interface SubagentRetainEntry {
     childSessionId: string;
+    /**
+     * Dual-runner wake fix — the DELEGATION runner that OWNS the child session (its `sessions` store minted
+     * and holds `childSessionId`, and its run face executes resumes). The Agent tool can be CALLER-mounted
+     * with its own runner (e.g. a server's throwaway in-memory subRunner) while SendMessage/AgentTranscript
+     * are core-mounted with the HOST runner — two DIFFERENT session stores. A resume/transcript acquire on
+     * the host runner's (durable) store then fails `not_found` for a child that only ever lived in the spawn
+     * runner's store ("resumable:true 停机帧 → resume.session_not_found"). Resume MUST go through this runner
+     * when present; the mount runner is only the fallback for pre-existing rows. In-process pointer only —
+     * retain is NOT durable, so cross-process semantics are unchanged (the ledger dies with the process →
+     * resume.evicted, the 1.283 ruling). Lifetime: the entry's `release` closure already captures the same
+     * runner, so this adds no new retention class.
+     */
     runner?: Runner;
     agentName?: string;
+    /** fix-C5: PER-ENTRY retain TTL override (ms). A SESSION-scoped ledger is shared by two purposes with
+     *  different windows (failure-evidence retain = FAILED_SESSION_RETAIN_TTL_MS vs a session-scoped bg
+     *  child's opt-in retain = the run's configured ttl); the ledger-level `ttlMs` is fixed by whichever
+     *  purpose CREATED the ledger first, which silently truncated (or stretched) the other's window.
+     *  Expiry (lazy sweep + active timer) uses this when present, the ledger default otherwise. */
     ttlMs?: number;
     specSnapshot: Readonly<TaskSpec>;
     internalsSnapshot: RunInternals;
+    /** True while the initial run OR a resume run is in flight (resume re-entry rejects `steering.still_running`). */
     running: boolean;
+    /** True once the initial run settled at least once. */
     settled: boolean;
+    /** Last settle timestamp (epoch ms) — the retain-TTL base. */
     settledAt: number;
     resumeCount: number;
+    /** design/144 §1.1 — the child's STOP-CYCLE number (207-equivalent per-stop notification identity).
+     *  Minted 1 at registration (the spawn run's settle notifies cycle 1); every resume that actually
+     *  LAUNCHES bumps it, so the resumed run's settle notifies 2, 3, … The background_agent notification
+     *  producers ride it as `TaskNotificationPayload.seq`, which uniquifies the `task_id:status[:seq]`
+     *  dedup key per stop cycle — without it a stop→wake→stop chain's second `completed` frame collided
+     *  with the first and was silently swallowed by the SystemInjectionQueue/inbox/shell dedup layers. */
     cycleSeq: number;
+    /** design/144 §1.2 (child gate) — the PARKED completion notification of this child's spawn settle:
+     *  present while a `completed` frame sits deferred behind the child's still-running background tasks
+     *  (owner key = childSessionId). `seq` is the parked frame's stop-cycle number (park does not change
+     *  cycle semantics); `cancel` unsubscribes the registry quiescence watcher and voids the frame.
+     *  RULING (recorded in design/144 §1.2): a resume/wake VOIDS the parked frame — the wake opens the
+     *  next stop cycle, whose own settle produces this child's next frame; delivering the stale parked
+     *  "completed (idle)" after the child is running again would break "one cycle, exactly one frame".
+     *  1.284 C2 RULING (宁发勿丢): entry EVICTION (TTL / LRU / abandon / disposeAll / session reap) voids
+     *  the park PROMISE, not the frame — `flush` delivers the park-time snapshot IMMEDIATELY (summary
+     *  undecorated, honest ungated degradation) and unsubscribes the quiescence watcher; the pre-fix
+     *  leak left the watcher alive (a late quiescence poke delivered a ZOMBIE notification for an entry
+     *  long gone, and a never-settling grandchild pinned watcher + captured frame forever). */
     deferredNotify?: {
         seq?: number;
         cancel: () => void;
         flush: () => void;
     };
+    /** The in-flight resume run's abort controller (parent-terminal cleanup aborts it — r1-M2). */
     activeAbort?: AbortController;
+    /** Unpin + explicitly release the retained session (best-effort; the TTL reaper is the backstop). */
     release: () => Promise<void>;
 }
+/**
+ * HRD-AGO-5 — which DETACHED (fire-and-forget) cleanup failed. The ledger frees slots synchronously and
+ * releases the session in the background; every one of those background calls can reject, because
+ * {@link SubagentRetainEntry.release} is caller-supplied and reaches a session store / pin.
+ */
 export type RetainEvictionTrigger = "ttl_sweep" | "ttl_timer" | "lru_overflow" | "session_ledger_lru" | "session_reap";
+/** HRD-AGO-5 — deployment-side sinks for the retain ledger's background failures. */
 export interface RetainLedgerHooks {
+    /**
+     * A detached eviction's `release` (unpin + session release) FAILED. The row is already gone from the ledger
+     * (containment is synchronous), so this is not recoverable state — it is the honest signal that a retained
+     * session may still be pinned in the session store, which was previously either an `unhandledRejection` on
+     * the host process or nothing at all. A throwing sink is swallowed (it cannot strand the reaper).
+     */
     onEvictionError?: (info: {
         parentToolCallId?: string;
         trigger: RetainEvictionTrigger;
         error: Error;
     }) => void;
 }
+/**
+ * design/122 D1 — the parent-run-scoped RETAIN ledger (`TaskSpec.retainSubagentSessions`). Created by
+ * `prepareTask` when the spec opts in, threaded to delegation tools as the TRUSTED `ctx.subagentRetain`,
+ * and disposed by the Runner in the task's terminal `finally` (D4: abort in-flight resumes + unpin +
+ * release every retained session — retain is NOT durable; resume is reachable only while the parent run
+ * lives, codex-B3). Leaks are double-bounded by `max` + `ttlMs` (codex-B2).
+ */
 export declare class SubagentRetainLedger {
     readonly ttlMs: number;
     readonly max: number;
     private entries;
+    /** Tombstones: ids that WERE (or would have been) retained but got evicted — their resume rejects `resume.evicted`.
+     *  fix-C2: capped FIFO (see `tombstone`) — a long-lived SESSION ledger kept producing unique failed
+     *  toolCallIds, so an uncapped set grew without bound for the session's whole life. */
     private evictedIds;
+    /** fidelity R4-1: per-entry unref'd TTL timers, armed at `markSettled` — ACTIVE expiry enforcement.
+     *  The lazy `sweepExpired` used to be the only reaper, so a ledger nobody touched again (no later
+     *  spawn/resume on it) kept its pinned sessions until process end. Cleared on evict/abandon/dispose;
+     *  the lazy sweep stays as the backstop (and the only reaper for absurd over-clamp TTLs). */
     private ttlTimers;
     private isDisposed;
     private readonly hooks;
     private failedEvictions;
     private lastFailure;
+    /** RB-463: the shared host-callback isolation scope (this ledger's own `onEvictionError` sink lives on it).
+     *  The EVICTION counters above are a different quantity and stay: they count evictions that failed, not
+     *  invocations of the sink that threw. */
     private readonly notifier;
     constructor(config: true | {
         ttlMs?: number;
         max?: number;
     }, hooks?: RetainLedgerHooks);
+    /** HRD-AGO-5 — how many DETACHED evictions failed to release their session on this ledger. Non-zero means
+     *  some retained session may still be pinned in the session store. Observable even with no sink wired. */
     get evictionFailureCount(): number;
+    /** HRD-AGO-5 — the most recent detached-eviction failure (see {@link evictionFailureCount}). */
     get lastEvictionError(): Error | undefined;
+    /** RB-463 diagnostics — how many times the deployment's `onEvictionError` SINK itself threw (distinct from
+     *  {@link evictionFailureCount}, which counts the evictions being reported). */
     get evictionSinkFailureCount(): number;
+    /**
+     * HRD-AGO-5 — record a failure from a DETACHED call this ledger made (or that a module-level helper made on
+     * its behalf, e.g. the session-ledger LRU disposal). Counts it, keeps the last error, and offers it to the
+     * sink; a throwing sink is swallowed so a broken observer cannot strand the reaper.
+     */
     noteDetachedFailure(trigger: RetainEvictionTrigger, error: unknown, parentToolCallId?: string): void;
+    /**
+     * HRD-AGO-5 — the ONLY way this class issues a background eviction. `evict` removes the row + tombstones it
+     * synchronously and only the `release` is async, so the caller never waits; but a bare `void evict(...)` made
+     * a rejecting `release` an `unhandledRejection` on the host process (a library killing its embedder over a
+     * best-effort cleanup) and left no trace that a session went unreleased. Every detached site routes here.
+     */
     private evictDetached;
+    /** HRD-AGO-5 — `disposeAll` for a caller that cannot await it (the module-level session-ledger LRU). Same
+     *  contract as {@link evictDetached}: containment is synchronous, the failure is reported, never thrown. */
     disposeAllDetached(trigger: RetainEvictionTrigger): void;
+    /** fix-C2: hard cap on tombstones. A tombstone is only an anti-resurrection sentinel (an evicted id's
+     *  resume reads the honest `resume.evicted` instead of the generic not-retained text) — dropping the
+     *  OLDEST one merely degrades that id's error text, so FIFO eviction is a safe bound. */
     private static readonly TOMBSTONES_MAX;
+    /** fix-C2: record an eviction sentinel, FIFO-bounded (re-tombstoning an id refreshes its position). */
     private tombstone;
+    /** fix-C5: an entry's own retain window when it carries one, the ledger default otherwise. */
     private effectiveTtlMs;
     get disposed(): boolean;
     get size(): number;
     get(parentToolCallId: string): SubagentRetainEntry | undefined;
     wasEvicted(parentToolCallId: string): boolean;
+    /** codex impl-review MAJOR-1 — LAZY TTL sweep: evict every settled idle entry whose retain TTL
+     *  elapsed (unpin + release + tombstone), so an expired session is freed at the NEXT ledger touch
+     *  (every spawn-registration and resume entry call this) rather than only when its own resume is tried.
+     *  fidelity R4-1: no longer the only reaper — `markSettled` also arms a per-entry ACTIVE timer, so
+     *  expiry fires by time even on a ledger nobody touches again. This sweep stays as the backstop.
+     *  The parent-terminal `disposeAll` stays the final backstop. `evict` removes the row + tombstones
+     *  synchronously; the release itself is fire-and-forget best-effort. */
     sweepExpired(now?: number): void;
+    /** Register a child at SPAWN (running until `markSettled`). Over `max`: the oldest settled idle entry is
+     *  evicted (unpin+release); when every slot is still running, the NEW child is NOT retained (tombstoned →
+     *  its resume rejects `resume.evicted`) and `undefined` is returned — the caller falls back to throwaway. */
     register(parentToolCallId: string, entry: Omit<SubagentRetainEntry, "running" | "settled" | "settledAt" | "resumeCount" | "cycleSeq" | "activeAbort" | "deferredNotify">): SubagentRetainEntry | undefined;
+    /** Flip a child to settled (idle, resumable) — called when the initial run's stream drains.
+     *  fidelity R4-1: also (re-)arms the entry's own TTL timer, so the retain window expires by TIME,
+     *  not only at the next ledger touch. */
     markSettled(parentToolCallId: string): void;
+    /** fidelity R4-1 — arm/re-arm the ACTIVE per-entry TTL reaper (unref'd: never keeps the process
+     *  alive). Idempotent with the lazy sweep: the callback re-checks the same settled/idle/elapsed
+     *  predicate, so a resume in flight (or an entry already evicted/re-settled) is left alone. */
     private armTtlTimer;
     private clearTtlTimer;
+    /** 1.284 C2 — an entry leaving the ledger takes its park promise with it: deliver the parked frame
+     *  NOW (undecorated; 宁发勿丢 ruling — see the deferredNotify field doc) and unsubscribe the registry
+     *  quiescence watcher. Without this, EVERY release path (TTL timer / lazy sweep / LRU register
+     *  overflow / abandon / disposeAll / session reap) leaked the watcher: a late quiescence poke then
+     *  delivered a zombie notification for a long-gone entry, and a never-settling grandchild pinned the
+     *  watcher + captured frame forever. */
     private flushDeferredNotify;
+    /** Drop a row WITHOUT releasing (durable-pause posture: a committed checkpoint owns the session/pin —
+     *  releasing here would orphan it, mirroring the delegation tool's own no-release guard). */
     abandon(parentToolCallId: string): void;
+    /** Evict one row: flush its parked notification (C2), abort its in-flight resume (if any), unpin +
+     *  release the session, tombstone the id. */
     evict(parentToolCallId: string): Promise<void>;
+    /** design/122 D4 — parent-run terminal cleanup: abort EVERY in-flight resume + unpin + release EVERY
+     *  retained session (throwaway semantics restored; no orphan run burns tokens — same posture as
+     *  `abortBackgroundAgentsForOwner`). Idempotent; further register/resume attempts are refused. */
     disposeAll(): Promise<void>;
 }
+/** Look up (never create) the session-scoped retain ledger for `sessionId` — SendMessage's read side.
+ *  fix-C1: a read hit refreshes `touchedAt` (TRUE LRU) — a session kept alive ONLY by SendMessage resumes
+ *  otherwise looked stale to the cap eviction below and got its ledger disposed (aborting in-flight
+ *  resumes) while 63+ other sessions churned. */
 export declare function getSessionRetainLedger(sessionId: string): SubagentRetainLedger | undefined;
 export declare function getOrCreateSessionRetainLedger(sessionId: string, config: {
     ttlMs: number;
     max: number;
 }, hooks?: RetainLedgerHooks): SubagentRetainLedger;
+/** Anchor ② — dispose the session's retain ledger (abort in-flight resumes, unpin + release every retained
+ *  child session) and drop it. Idempotent; wired to `TaskRegistry.onSessionReap` at retain time. */
 export declare function releaseSessionRetainLedger(sessionId: string): Promise<void>;
 export declare function ensureSessionReapHook(registry: import("../core/task-registry.js").TaskRegistry): void;
+/**
+ * design/122 D2 — the SINGLE resume-prompt builder: a constant TRUSTED frame core owns (operator authority +
+ * an unpredictable correlation marker) around the operator content, which stays fenced DATA (same posture as
+ * steer — content is never authority, and `delimitUntrusted` neutralizes fence/tag break-out sentinels).
+ */
 export declare function createResumePrompt(marker: string, content: string, origin?: "operator" | "peer"): string;