@sema-agent/core 7.0.0 → 7.0.2

package/CHANGELOG.md

@@ -1,5 +1,37 @@
 # Changelog
 
+## 7.0.2 — 2026-08-30
+
+### Added
+- **`MemoryEngine.sessionMemoryStatus(sessionId)`** (#511①, design/383 §S-7 — the signed-design read face missed at build): `{captureOptedOut?, committedCount?, foldedCount?, optOutSource?, lastCaptureAt?}`. Every key optional — absence means that fact's source was unreadable, never a coined false/0 stand-in; a store fault answers `optOutSource:"fault"` with `captureOptedOut` ABSENT (the indeterminate axis projected); `foldedCount` counts evidence-only input rows over the full product blocks; the face never throws. New exported type `SessionMemoryStatus` (export surface 1841→1842). Server projects the wire endpoint as a pure derivation ([5786]).
+- **`SessionCaptureRecordStore` dual form** (#511②): the three legs widen to `T | Promise<T>` — a SQL/async backend plugs in with Promise signatures while every existing sync implementation compiles untouched (the additive mechanical proof). The dual form propagates: sync deployments stay sync-transparent end to end; the sync-facing public seats (`MemoryEngineSession.captureOptOut.*`, `ToolExecuteContext.memoryCaptureOptedOut/…Indeterminate`, `parentMemoryCaptureState`) go dual with the await law documented — per-leg failure mapping preserved (mark rejection ⇒ unpersisted, read rejection ⇒ the indeterminate fault axis).
+- **`best` model alias** (#511③, CC 250 $L parity): with `RunnerDeps.tiers` configured, `"best"` resolves to the first DIRECTLY bound tier walking `ultra→max→pro→flash→lite` from the head ("best available", deliberately not a fixed alias-table row). Explicit spellings win over the dynamic walk: a deployment catalog key named `best`, or an explicit `tiers.best` binding. No tiers ⇒ the word stays an unknown ref (refused, never coined).
+
+### Changed — narrowing (named)
+- **Agent tool `model` argument: non-string values on the schema-外 direct-call lane refuse typed** (#511④, test [5781] codex observation): a number/array/boolean/garbage-object now refuses `model.invalid` (closed-set ADDITION beside `model.empty`/`model.unknown` — downstream pinning the details.error set gains a row) instead of crashing or silently no-oping; a legal Model-shape object (string id+api) passes through as the per-call model; strings keep every existing arm byte-identical.
+- **`modelNote` fork-ignored and roster-less arms defuse the model-authored word** (#511⑤, census P2): both arms now run the same `inlineUntrusted` fold (newline fold + fence/envelope-tag neutralization + 80 cap) the family's other three arms already ran — a hostile `model` word can no longer carry raw newlines or fence sentinels into a top-level report line. Clean short words render byte-identical.
+
+## 7.0.1 — 2026-08-30
+
+### Added
+- **`task_progress.model` + `BackgroundChildEvent.model`** (#508①, cli [5725] item-B): every spawn-progress tick and background child event now carries the SERVED model id (PREPARED semantics — the id the child's prepare actually resolved, not the word the caller wrote). The fleet row publishes the same served id, so row/tick/card lanes state one fact; a replayed workflow leg reports the cached model on every face.
+- **`apiFailure` on the completed-agent details card and both notification lanes** (#506 residual): terminal API-failure evidence now reaches every reader of a settled child, not only the host-face `TaskResult`.
+
+### Changed — narrowing (named)
+- **Unknown/empty delegation `model` argument: soft note → typed reject** (#508②, CC $L closed-set anchored): `model.unknown` (echoes the accepted catalog) and `model.empty` replace the old accept-with-note behavior. A caller that previously leaned on the note-and-serve-default path now gets a refusal naming the legal values. Positive control: every cataloged id serves exactly as before. **Two lanes are exempt** (pre-ship rescan, the caller-census duty): the tier-3 revival lane feeds the durable row's MACHINE-minted recorded word back through this gate, so it keeps the soft-degrade (loud `modelNote` + inherited model — a catalog rotation must not permanently strand a named teammate and its queued mail), and its fallback chain skips string candidates the live catalog cannot resolve (a degrade that re-selects the retired word would revive a child dead on arrival); `subagent_type: "fork"` keeps the schema's own "model is ignored" promise for the empty string exactly as for any other word. Membership itself hardened: the catalog read is own-property (`Object.hasOwn`), so `Object.prototype` words (`constructor`, `toString`, …) refuse as `model.unknown` instead of passing the gate and crashing the child.
+- **Per-call model binding is resolved AT the gate** (rescan): a judged per-call word now rides the child spec as the resolved `Model` object, so the judged roster is the authoritative binding — an explicit `opts.models` table naming a key the runner's own catalog lacks now serves on the judged binding instead of passing the gate and dying inside the child's prepare. `BackgroundChildEvent.model` / roster-row docs restated to the served-id-first rule (display label is the unresolvable-ref fallback only); the residual one-generation lag window for definition/tool-level string refs after `swapModels` is documented, not claimed away. The background roster record gains optional `modelFallback` (additive; fork lane already carried it).
+- **File-history mode domain, both doors** (security): the import door refuses a version-row mode that is not an integer in [0, 0o7777] and refuses set-ID modes (setuid/setgid) BY NAME; on replay, `applyRecordedMode` STRIPS the set-ID bits and applies the remainder (`mode & 0o1777`, sticky preserved) under the loud `modeNotPreserved` account — set-ID bits never reach chmod (capture may legitimately record one; replay is where the escalation lives). A file already at exactly the recorded mode answers `unchanged` with no account (the equality fast path precedes the escalation arm), and `previewFileDelta` compares against what replay would actually apply, so a set-ID-recorded drift converges instead of being promised forever.
+
+### Changed — widening (named)
+- **The delegation tool's model roster is read at JUDGEMENT time, not frozen at mount** (#508 residual): after `Runner.swapModels`, a long-lived mount no longer hard-rejects the models the deployment just started serving (the old refusal echoed the RETIRED generation as the accepted values). Explicit `opts.models` keeps snapshot semantics; only the runner-derived roster re-reads live. The schema/announce projection remains a mount-time snapshot — the refusal therefore carries its own accepted-values echo.
+
+### Fixed
+- Flaky-guard ledger: the three new pin files registered as bounded effect-polling (gate was red since the #508 batch landed).
+
+### Docs
+- #509 core half: the subagent transcript LANDING law (three store shapes + the CC 2.1.250 unconditional-landing anchor) spelled out at `SubagentToolOptions.runner`, the background placement mint gate, and `resolveSubagentTranscriptTier`; reproduced-and-located, wiring fix owned by the delegation runner's store choice.
+- 6.0.0 release-criteria acceptance corrections (test [5719]/[5721]): G1.a added-count 65→77, F7 absent-seat claim restated.
+
 ## 7.0.0 — 2026-08-30
 
 > Ships as its own major immediately after 6.0.0 (ruled: the three-dimensional rule scope travels alone so downstream adapts to it in isolation). The 6.0.0 artifact is this same tree with the design/382 P1 surface excised whole (type-driven surgery in an isolated worktree, tag v6.0.0); `npm run handoff:diff v6.0.0 v7.0.0` reproduces exactly this surface.

package/dist/agents/launch-receipt-contract.d.ts

@@ -0,0 +1,34 @@
+/**
+ * The HANDLING contract shared by every tool that launches background work and hands the model a
+ * receipt: the Agent tool's async-launch card and `run_workflow`'s backgrounded launch receipt.
+ *
+ * Why a module rather than two copies. Both receipts state the same two things — "these bytes are
+ * plumbing, do not paste them at the user" and "say what you launched in your own words" — and they
+ * had drifted into two wordings of it (`; ` vs ` — `, "this result" vs "this tool result"). Neither
+ * difference carried meaning; each was just the copy that happened to be typed second. A model reading
+ * one receipt and a model reading the other should be reading the same instruction, and a later edit
+ * to one of them should not be able to leave the other behind.
+ *
+ * What deliberately stays PER-LANE (the earlier ruling, narrowed rather than reversed): each receipt
+ * hands back different artifacts, so each names its own. The Agent card names `task_id`; the workflow
+ * receipt names the ids plus the host filesystem `scriptPath` it may have written. That enumeration is
+ * a fact about the lane and cannot be centralized without one of them lying. So the SCOPE clause is the
+ * caller's, and the sentence around it is this module's. Likewise the envelopes stay as they were —
+ * the Agent lane renders prose, the workflow lane a JSON field — which is why these are string parts
+ * rather than a rendered receipt.
+ */
+/**
+ * The no-quote clause for one lane. `artifactScope` is the caller's own enumeration of what "any part
+ * of it" covers, spliced in verbatim INCLUDING its leading separator — a comma clause
+ * (`", including the task_id below,"`) and a parenthetical (`" (the ids above, and scriptPath when
+ * present)"`) punctuate differently and the caller owns which it uses. Pass `""` for a lane with
+ * nothing to enumerate; the sentence still reads.
+ */
+export declare function launchReceiptNoQuoteClause(artifactScope: string): string;
+/**
+ * The closing instruction, whole and identical on every launch lane: report the launch in the model's
+ * own words instead of echoing the receipt. Callers append their own lane-specific continuation (how
+ * the result will or will not arrive) after a single space; this clause is a complete sentence on its
+ * own so a lane with nothing to add can use it unchanged.
+ */
+export declare const LAUNCH_RECEIPT_OWN_WORDS_CLAUSE = "In your own words, briefly tell the user what you launched \u2014 do not echo this tool result.";

package/dist/agents/launch-receipt-contract.js

@@ -0,0 +1,5 @@
+const NO_QUOTE_HEAD = "This tool result is internal metadata — never quote or paste any part of it";
+export function launchReceiptNoQuoteClause(artifactScope) {
+    return `${NO_QUOTE_HEAD}${artifactScope} into a user-facing reply.`;
+}
+export const LAUNCH_RECEIPT_OWN_WORDS_CLAUSE = "In your own words, briefly tell the user what you launched — do not echo this tool result.";

package/dist/agents/subagent.d.ts

@@ -4,6 +4,60 @@ import { type ExecutionEnv } from "../internal/harness.js";
 import type { RunInternals } from "../core/runner/prepare-task.js";
 import type { TaskNotificationPayload } from "../core/task-notification.js";
 import { RETAIN_DEFAULT_TTL_MS, RETAIN_DEFAULT_MAX } from "../config/defaults.js";
+/**
+ * #505 residual — the two GRANULARITY facts a delegation's terminal `usage` was missing beside
+ * tokens/turns/cost, closing a real reporting asymmetry rather than adding decoration: the workflow
+ * lane's completion notification has published `tool_uses` (summed `stats.toolCalls`) and `duration_ms`
+ * (wall clock) since it was written, and the Agent tool's own SYNC report publishes both in its
+ * `<usage>` footer — only the delegation TERMINAL faces (the background/fork/revive notification and its
+ * observer twin) reported neither, so the same child answered "how much work, how long" on one lane and
+ * refused on the other.
+ *
+ * PROVENANCE is the workflow lane's — `stats.toolCalls`, the RUN's own count — because that is the
+ * quantity the other side of the comparison publishes, and because it is available on every terminal
+ * lane, including the revive lane, which has no parent-side forwarding observer at all.
+ *
+ * WHAT THIS DOES NOT CLOSE, stated because the honest scope is narrower than "one name, one meaning":
+ * the Agent tool's SYNC `<usage>` footer keeps its own `tool_uses`, counted by the parent-side
+ * `toolStatsCounter`, and the two numbers differ in BOTH directions for the same child — that counter
+ * EXCLUDES the delegation tool's own calls and INCLUDES descendants' (CC's nested fold), while
+ * `stats.toolCalls` counts the direct child's calls and counts its delegations among them. A child that
+ * itself delegates therefore reports two different `tool_uses` depending on whether it was run
+ * synchronously or in the background. That footer is a byte-pinned CC-anchored face; reconciling the
+ * two is a separate ruling, and until it is made this comment is the place the divergence is written
+ * down rather than a claim that it is gone.
+ *
+ * Knownness follows `delegatedCostField`'s rule exactly: `toolCalls` is optional on the stats face
+ * ("undefined means the gateway reported no usage", and a checkpointed resume folds it per-leg), so the
+ * key is ABSENT rather than zeroed — a fabricated 0 reads as "ran no tools", which is a claim, not a
+ * gap. Duration is always knowable (the caller holds the lane's launch instant), so it is unconditional.
+ *
+ * TWO SPELLINGS, one meaning, matching each face's established vocabulary: the observer
+ * ({@link import("../core/types.js").BackgroundChildEvent}) `usage` face is camelCase and already
+ * DECLARES `toolUses`/`durationMs`; the model-facing `TaskNotificationPayload.usage` is rendered into
+ * `<usage>` as JSON where the workflow lane and the Agent sync footer both spell them snake_case. No new
+ * word is minted on either side.
+ *
+ * Exported at MODULE level (not from the package index — the same posture as `completedAgentCard` and
+ * `classifySubagentError`, and for the same reason): the ABSENT arm is the point of the knownness rule
+ * and it is not reachable through a real child. This engine's runner initializes `stats.toolCalls` to a
+ * number on every path, so only a `TaskResult` from elsewhere — a foreign runner, an older checkpoint —
+ * can carry the optional-and-missing shape the stats face declares. A pin that only ever sees a live
+ * child cannot tell "absent" apart from "zero", which is precisely the confusion being prevented.
+ */
+export declare function delegatedGranularityFields(stats: {
+    toolCalls?: number;
+}, elapsedMs: number): {
+    toolUses?: number;
+    durationMs: number;
+};
+/** {@link delegatedGranularityFields}'s snake_case twin for the model-facing notification `usage`. */
+export declare function delegatedGranularityFieldsWire(stats: {
+    toolCalls?: number;
+}, elapsedMs: number): {
+    tool_uses?: number;
+    duration_ms: number;
+};
 export { RETAIN_DEFAULT_TTL_MS, RETAIN_DEFAULT_MAX };
 import { SubagentRetainLedger } from "./retain-ledger.js";
 import { type PeerSelfRef } from "./peer-admission.js";
@@ -268,6 +322,17 @@ export declare function completedAgentCard(child: {
     blockedReason?: string;
     errorMessage?: string;
     errorCode?: string;
+    /** The child run's provider-boundary fault assertion, verbatim from its `TaskResult.apiFailure`.
+     *  PRESENCE is the claim ("this child's terminal failure came from the transport/provider, not from
+     *  this deployment refusing to send, a limit, or unusable model output"); the members are whatever
+     *  the failing attempt stated about itself (`status` when the provider answered with one, `requestId`
+     *  when it named one). An EMPTY object is therefore meaningful and must not be normalized away: it
+     *  says "provider fault, unlabelled", which is a different statement from absence ("not a provider
+     *  fault"). Absent on every non-provider terminal and on every completed child. */
+    apiFailure?: {
+        status?: number;
+        requestId?: string;
+    };
     retryAfterMs?: number;
     degraded?: unknown;
     structuredOutput?: unknown;
@@ -514,9 +579,61 @@ export declare function delegationEntryLedgerFootprint(registry: object): {
  * catalog) after mounting must REBUILD the ToolSpec — call `createSubagentTool` again with the fresh
  * roster and serve the new tool — for the change to take effect; already-running tasks keep the roster
  * they were prepared with.
+ *
+ * The MODEL roster ({@link SubagentToolOptions.models}) deliberately does NOT follow that rule when it
+ * is left to the runner: a hot-swapped catalog is a first-class runner verb (`Runner.swapModels`) with
+ * no rebuild step, so the tool re-reads it at each judgement instead of requiring a remount. Only the
+ * schema/announce projection of it stays a mount-time snapshot (cache-prefix stability). Passing
+ * `models` explicitly opts back into snapshot semantics for that table — see the field.
  */
 export interface SubagentToolOptions {
-    /** Runner used to execute child tasks. */
+    /**
+     * Runner used to execute child tasks.
+     *
+     * **This runner's session store decides where a child TRANSCRIPT lives — and therefore whether one
+     * exists at all after the fact.** A SYNC or BACKGROUND child mints its session through
+     * `runner.sessions`, never through the parent leg's store, so a deployment that mounts the Agent
+     * tool on a DIFFERENT runner than the one serving root tasks has split the transcript lane away
+     * from the session lane. (The FORK lane is the deliberate exception and shows the split is a known
+     * shape: a fork branches the CALLER's session through `ctx.hostSessionFork` — the host Runner's
+     * own store — precisely because `runner` may be a split child-execution runner whose store never
+     * held the source. A fork child therefore follows the HOST store's fate, an ordinary child this
+     * runner's, and the same deployment can land one on disk and lose the other.) Three shapes for the
+     * ordinary lanes, all reachable and all silent:
+     *  - durable store on THIS runner whose `release` PRESERVES history (`evict:"forget"`, the
+     *    `FileStorageBackend` default), no {@link SubagentToolOptions.background}.`agentStore` — the
+     *    child transcript lands UNPLACED and survives the settle leg's eager release (a cache drop),
+     *    so post-hoc forensics work even though nothing can ADDRESS it through an `a*` handle. The
+     *    qualifier is load-bearing, not decorative: durability and release semantics are INDEPENDENT
+     *    declarations, and a store that declares itself durable while deleting on release
+     *    (`new FileStorageBackend({ evict: "delete" })`, or any custom store with the same policy)
+     *    destroys the transcript at settle — the settle leg calls `release`, and there is no placement
+     *    to make the deletion refuse. "Durable" is a statement about restarts, never about release;
+     *  - durable store + `agentStore` — the placed partition (see the mint gate in the background
+     *    lane): addressable across a restart, list-excluded, joint-reaped with its row. This arm is
+     *    also the one immune to the evict policy above (the settle leg keeps a confirmed row's
+     *    transcript instead of releasing it);
+     *  - transient/process-local store on THIS runner (e.g. a throwaway sub-task store, or a routing
+     *    store that sends ordinary child acquires to a private cache) — the child transcript NEVER
+     *    reaches disk whatever the root leg's store is. Nothing in the engine compensates: the
+     *    completion notification's recent steps and the in-memory registry are the only readers, and
+     *    both die with the process.
+     * The asymmetry this produces is worth naming, because it reads as a bug from the outside: the
+     * `run_workflow` tool is mounted by the ENGINE on the run's own runner (prepare-task's
+     * `runnerSelf`), so a workflow node's child session always follows the ROOT store, while this tool
+     * is mounted by the DEPLOYMENT and follows whatever runner it was handed. One data root can
+     * therefore hold every workflow node's transcript and none of the Agent tool's.
+     * There is deliberately NO manifest field that answers "does a child transcript reach disk here":
+     * `WiringManifest.fleet.subagentTranscripts` is the ADDRESSABILITY tier and folds to `"none"` on
+     * the absent-row-store arm before it ever looks at a session store, so it cannot separate the
+     * first shape above from the third (the landing pins carry that as an explicit arm). The facts
+     * that DO answer it are this runner's store's own declarations — `sessions.durability` and its
+     * release/evict policy — read together, which is why they are named here rather than folded into
+     * a word. Upstream anchor (CC 2.1.250): a sub-agent transcript is written unconditionally to
+     * its own file beside the parent session (`projects/<projectKey>/<sessionId>/subagents/agent-<id>
+     * .jsonl` + a `.meta.json` sidecar), so "the child conversation is readable afterwards" is a
+     * property CC never makes conditional on a store tier.
+     */
     runner: Runner;
     /**
      * design/115 P3 — background sub-agents (CC `run_in_background`). When set, the tool exposes the
@@ -599,7 +716,22 @@ export interface SubagentToolOptions {
      * wins — fork-shadowing precedent).
      */
     builtinAgents?: boolean;
-    /** Model catalog, used ONLY to validate an agent's string `model` ref at assembly time (fail-fast). */
+    /**
+     * Model catalog the tool validates model words against — an agent definition's string `model` ref at
+     * assembly time (fail-fast), and the per-call `model` argument at call time.
+     *
+     * OMITTED is the normal deployment shape: the tool then reads the runner's CURRENT catalog
+     * (`Runner.agentCatalog.models`) at each of those moments, so a catalog hot-swapped after mount is
+     * the one a delegation is judged against. Passing this explicitly is a deliberate NARROWING and
+     * pins the tool to exactly this table for the tool's whole life — the runner's generation changes
+     * no longer reach it. And this table is AUTHORITATIVE for the per-call word, not merely advisory:
+     * a word judged against it leaves the gate as this table's resolved `Model` object and the child
+     * runs that binding — including a key the runner's own catalog does not carry, and including a
+     * same-name key the runner binds differently. (The judged-then-re-resolved split this replaced
+     * admitted such words and then killed the child inside its own prepare, or ran a binding the spawn
+     * frame never named.) A definition's string `model` ref is different: it is validated here at
+     * assembly time but still resolves at the child's own prepare through the runner's live table.
+     */
     models?: Record<string, Model>;
     /** Tools available to the child (a deliberately narrowed subset). */
     tools?: ToolSpec[];