@sema-agent/core 5.60.0 → 5.61.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +79 -0
- package/dist/agents/subagent.d.ts +4 -2
- package/dist/agents/subagent.js +9 -9
- package/dist/core/checkpoint-store.d.ts +56 -6
- package/dist/core/governance-codes.d.ts +1 -1
- package/dist/core/governance-codes.js +2 -2
- package/dist/core/mcp.d.ts +4 -5
- package/dist/core/memory-engine/consolidation-driver.d.ts +19 -1
- package/dist/core/memory-engine/consolidation-driver.js +75 -3
- package/dist/core/memory-engine/consolidation.d.ts +52 -5
- package/dist/core/memory-engine/consolidation.js +3 -1
- package/dist/core/memory-engine/distiller.d.ts +89 -1
- package/dist/core/memory-engine/distiller.js +94 -5
- package/dist/core/memory-engine/engine.d.ts +13 -6
- package/dist/core/memory-engine/engine.js +51 -8
- package/dist/core/memory-engine/index.d.ts +1 -1
- package/dist/core/memory-engine/index.js +1 -1
- package/dist/core/runner/prepare-task.d.ts +6 -3
- package/dist/core/runner/prepare-task.js +1 -0
- package/dist/core/runner/runtask.js +113 -31
- package/dist/core/task-notification.d.ts +50 -23
- package/dist/core/task-notification.js +20 -4
- package/dist/core/tool-policy.d.ts +8 -2
- package/dist/core/types.d.ts +76 -19
- package/dist/engine/harness/agent-harness.d.ts +58 -2
- package/dist/engine/harness/agent-harness.js +115 -5
- package/dist/engine/loop/agent-loop.js +153 -15
- package/dist/engine/loop/types.d.ts +32 -0
- package/dist/index.d.ts +3 -3
- package/dist/index.js +3 -3
- package/dist/orchestration/run-workflow-tool.d.ts +7 -2
- package/dist/orchestration/run-workflow-tool.js +1 -1
- package/dist/tools/monitor.d.ts +3 -3
- package/dist/tools/monitor.js +1 -1
- package/package.json +1 -1
- package/test/export-surface.snapshot.json +11 -1
package/dist/core/types.d.ts
CHANGED
|
@@ -4177,6 +4177,12 @@ export type TaskEvent = ({
|
|
|
4177
4177
|
*/
|
|
4178
4178
|
type: "task_notification";
|
|
4179
4179
|
notification: TaskNotificationPayload;
|
|
4180
|
+
/** design/373 §3.6 — the frame's injection tier as DELIVERED (additive; consumers tolerate
|
|
4181
|
+
* absence): on the live lane, the tier the routing used (`next`/`now` → steer lane, `later`
|
|
4182
|
+
* → followUp lane); on the turn-open redelivery lane, the tier the park CARRIED. ABSENT is a
|
|
4183
|
+
* fact, not a default: a pre-373 producer event, or a parked record whose tier was not
|
|
4184
|
+
* carried (the terminal sweep's re-pend) — never read absence as `"later"`. */
|
|
4185
|
+
priority?: import("./task-notification.js").SystemInjectionPriority;
|
|
4180
4186
|
} & TaskEventIdentity) | ({
|
|
4181
4187
|
/**
|
|
4182
4188
|
* design/99 MF-10 — a SUBAGENT PROGRESS TICK: the delegated sub-run's usage ACCRUING
|
|
@@ -4380,11 +4386,43 @@ export interface TaskStream extends AsyncIterable<TaskEvent> {
|
|
|
4380
4386
|
* into this run's resume prompt) is NOT in it and WOULD inject again. A deployment that needs
|
|
4381
4387
|
* cross-leg or cross-process idempotency owns that half (its own key ledger), the same division
|
|
4382
4388
|
* of labor `notify`'s park-window dedup states.
|
|
4389
|
+
*
|
|
4390
|
+
* design/373 (additive) — `priority` is the steer face's injection tier, default `"next"`
|
|
4391
|
+
* (= the pre-373 behavior, byte-identical):
|
|
4392
|
+
* - `"next"` — the running turn's next boundary (unchanged);
|
|
4393
|
+
* - `"later"` — never folded into work in progress: delivered at the run's natural
|
|
4394
|
+
* would-otherwise-stop seat (its own closing turn); a run that ends before that seat re-parks
|
|
4395
|
+
* the input via the existing durable/undrained lanes (receipt semantics unchanged — "queued"
|
|
4396
|
+
* still means accepted-at-this-moment, never "consumed"); a `"later"` steer rides the followUp
|
|
4397
|
+
* lane, so its stranded account is `task.user_followup_undrained` (the accounts are LANE-keyed
|
|
4398
|
+
* — same two-key split the settled frame uses);
|
|
4399
|
+
* - `"now"` — `"next"`'s delivery guarantee PLUS the boundary is manufactured early: the frame
|
|
4400
|
+
* takes the queue's class head and, when a turn is in flight, the engine CUTS that turn at a
|
|
4401
|
+
* reconciled boundary (finished tool calls keep their real results; never-started ones settle
|
|
4402
|
+
* as paired interrupted results; the CC-verbatim interruption marker lands) and the run
|
|
4403
|
+
* CONTINUES with this input — never the run-level `interrupt()`. Best-effort accelerator:
|
|
4404
|
+
* with nothing in flight (idle tail, between turns) delivery is exactly `"next"`'s. Every
|
|
4405
|
+
* real cut is announced (`task.turn_interrupted` notice) and counts toward `maxTurns` —
|
|
4406
|
+
* deliberately (the hard cap IS the interrupt storm's bound), which means a `now` issued at
|
|
4407
|
+
* the run's FINAL allowed turn spends the remaining budget on the cut and the run then ends
|
|
4408
|
+
* at the limit before the steer's own turn: the input takes the terminal lanes like any
|
|
4409
|
+
* accepted-but-undelivered steer (durable park when a resource-suspend seat is eligible —
|
|
4410
|
+
* redeemed on resume; otherwise the loud `task.user_steer_undrained` account). Same for the
|
|
4411
|
+
* other terminal gates (budget/walltime) firing at the cut's own `turn_end`. The receipt
|
|
4412
|
+
* never promised consumption; no budget is reserved for the interruptor.
|
|
4413
|
+
* ⚠️ POWER FACE (§5-4b): passing `priority` through to a third-party caller GRANTS it the power
|
|
4414
|
+
* to cut this run's in-flight turns (cancelling that turn's in-flight tool work). A deployment
|
|
4415
|
+
* relaying steer must gate this option at its own face — relay IS authorization.
|
|
4416
|
+
* An unknown value throws typed `steering.invalid_content` (bad-value loudness). The replay
|
|
4417
|
+
* identity (`inputId` idempotency) includes the NORMALIZED tier: the same text replayed under
|
|
4418
|
+
* the same id at a different tier refuses `steering.duplicate_input_id` — an "idempotent
|
|
4419
|
+
* success" that silently skipped the interrupt would be a disposition lie.
|
|
4383
4420
|
*/
|
|
4384
4421
|
steer(text: string, options?: {
|
|
4385
4422
|
trusted?: boolean;
|
|
4386
4423
|
actor?: ActorAssertion;
|
|
4387
4424
|
inputId?: string;
|
|
4425
|
+
priority?: import("./task-notification.js").SystemInjectionPriority;
|
|
4388
4426
|
}): Promise<void>;
|
|
4389
4427
|
/**
|
|
4390
4428
|
* design/144 §2 — inject an EXTERNAL structured event into this run's task-notification lane, as a
|
|
@@ -4392,19 +4430,37 @@ export interface TaskStream extends AsyncIterable<TaskEvent> {
|
|
|
4392
4430
|
* the payload (`task_type` is cast to `"external"` — a caller can never impersonate an internal lane)
|
|
4393
4431
|
* and every field is sanitized as UNTRUSTED data on the model face. Delivery rides the EXISTING
|
|
4394
4432
|
* notification machinery, so the semantics are identical to internal producers:
|
|
4395
|
-
* -
|
|
4396
|
-
*
|
|
4397
|
-
*
|
|
4398
|
-
*
|
|
4433
|
+
* - design/373 — the injection LADDER is live on this lane (BREAKING vs the flat era, named):
|
|
4434
|
+
* `"next"` = the running turn's next boundary (mid-work included — the previous behavior of
|
|
4435
|
+
* every frame, now this tier's); `"later"` = never folded into work in progress — delivered at
|
|
4436
|
+
* the run's natural would-otherwise-stop seat (its own closing turn), and a run that ends
|
|
4437
|
+
* before that seat parks the frame per session for the NEXT run's turn-open (delivery then
|
|
4438
|
+
* depends on a next run happening — a recorded structural honesty note, this engine owns no
|
|
4439
|
+
* idle process); `"now"` is REFUSED typed (`notify.invalid_priority`) — turn-interrupt
|
|
4440
|
+
* authority belongs to the caller-provenance steer face, never to the notification lane
|
|
4441
|
+
* (callers wanting earliest-boundary delivery say `"next"`). ⚠️ The parameterless DEFAULT is
|
|
4442
|
+
* `"later"` (the CC pending-notification default): an omitted priority moved from
|
|
4443
|
+
* next-boundary to the closing seat with this design — callers that need boundary delivery
|
|
4444
|
+
* must say `"next"` explicitly. NOTE for the flat era's callers: `"later"` was silently inert
|
|
4445
|
+
* then (delivered at the boundary with no runtime signal saying otherwise) — the version is
|
|
4446
|
+
* the ONLY discriminator of which behavior a build has. Delivery order within a tier is
|
|
4447
|
+
* ARRIVAL order; consecutive buffered notifications drain as one boundary batch. `priority`
|
|
4448
|
+
* rides the park path (pend records carry it; the turn-open batch is priority-major) and the
|
|
4449
|
+
* uplink path;
|
|
4399
4450
|
* - duplicate events dedup on `task_id:status[:seq]` — pass a fresh `seq` per repeat event. The
|
|
4400
4451
|
* external lane keys in its OWN dedup domain (design/144 X1): an external event can never collide
|
|
4401
|
-
* with (or pre-occupy) an internal lane's key, and vice versa
|
|
4452
|
+
* with (or pre-occupy) an internal lane's key, and vice versa. The key deliberately excludes
|
|
4453
|
+
* `priority` (tier = delivery metadata, not event identity): a repeat under an in-flight key
|
|
4454
|
+
* folds WHATEVER tier it names — the standing entry keeps the FIRST accept's tier — so
|
|
4455
|
+
* re-tiering requires a fresh `seq` (a new occurrence), never a same-key resend;
|
|
4402
4456
|
* - a notify landing AFTER the run ended PARKS per session and the session's NEXT run delivers it
|
|
4403
4457
|
* (unlike {@link steer}, a finished task does not reject — the park lane is the contract). The park
|
|
4404
4458
|
* leg dedups on the same key (design/144 X4) with an HONEST WINDOW LIMIT: the dedup set spans ONE
|
|
4405
4459
|
* parked batch — it does not reach back across the live-queue/park boundary, and it resets when the
|
|
4406
4460
|
* next run drains the batch (a repeat sent after that drain is deliverable again);
|
|
4407
|
-
* Throws typed `notify.invalid_payload` on a malformed input
|
|
4461
|
+
* Throws typed `notify.invalid_payload` on a malformed input, `notify.invalid_priority` on
|
|
4462
|
+
* `priority: "now"` (see above — a KNOWN value this lane refuses, distinct from an unknown one),
|
|
4463
|
+
* and `notify.not_running` when the run
|
|
4408
4464
|
* never built its notification lane (prepare failed). Everything upstream of the event reaching this
|
|
4409
4465
|
* process — transport, webhook/daemon wiring, inbound authn, durable parking for idle sessions — is
|
|
4410
4466
|
* the DEPLOYMENT half (design/43: zero daemon semantics in core).
|
|
@@ -4992,19 +5048,20 @@ export interface EngineNotice {
|
|
|
4992
5048
|
* — what is announced is the remainder that could NOT be carried, so a fully-migrated park says
|
|
4993
5049
|
* nothing at all and a queue-full / no-longer-pending row still says exactly what was lost.
|
|
4994
5050
|
*
|
|
4995
|
-
* - `"task.
|
|
4996
|
-
*
|
|
4997
|
-
*
|
|
4998
|
-
*
|
|
4999
|
-
*
|
|
5000
|
-
*
|
|
5001
|
-
*
|
|
5002
|
-
*
|
|
5003
|
-
*
|
|
5004
|
-
*
|
|
5005
|
-
*
|
|
5006
|
-
*
|
|
5007
|
-
*
|
|
5051
|
+
* - `"task.turn_interrupted"` (design/373) — a caller-provenance steer with `priority: "now"`
|
|
5052
|
+
* ACTUALLY CUT the running turn: the in-flight provider stream / tool batch was aborted at a
|
|
5053
|
+
* manufactured boundary (finished tool calls keep their real results, never-started ones
|
|
5054
|
+
* settle as paired interrupted results, the interruption marker lands) and the run CONTINUES
|
|
5055
|
+
* with the steer at the queue head. One notice per REAL cut (a `now` that found nothing in
|
|
5056
|
+
* flight, or whose frame already rode the imminent boundary, announces nothing — no false
|
|
5057
|
+
* interrupt claims); every cut turn counts toward `maxTurns`, so an interrupt storm's cost is
|
|
5058
|
+
* bounded and each of its cuts is on the record. `detail: { inputId, sessionId, actorId?,
|
|
5059
|
+
* taskId? }` — `inputId` is the steer's own correlation key (the `human_input` frame's id),
|
|
5060
|
+
* `actorId` the caller's asserted identity when one rode the steer. Audience `"user"`: the
|
|
5061
|
+
* person whose input forced the boundary is the one entitled to see that it landed.
|
|
5062
|
+
* RETIRED here (BREAKING, named): `"task.injection_priority_unimplemented"` — the ladder is
|
|
5063
|
+
* implemented, so the unhonored-knob disclosure it carried has no referent; consumers must
|
|
5064
|
+
* judge ladder support by VERSION, never by that code's absence.
|
|
5008
5065
|
*
|
|
5009
5066
|
* - `"memory.session_polluted"` (design/178 §3, #324a; message mode-aware since design/336) —
|
|
5010
5067
|
* this session's memory crossed into the one-way externally-exposed state (a tool classified
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import type { ActorAssertion, AssistantMessage, ImageContent, Model } from "../llm/index.js";
|
|
1
|
+
import type { ActorAssertion, AssistantMessage, ImageContent, Model, UserMessage } from "../llm/index.js";
|
|
2
2
|
import type { AgentMessage, AgentTool, LoopMalformedToolUseRecovery, LoopThinkingOnlyRecovery, LoopTruncatedOutputRecovery, QueueMode, ThinkingLevel } from "../loop/types.js";
|
|
3
3
|
import { type EngineSegment } from "../../core/untrusted-text.js";
|
|
4
4
|
import type { AbortResult, AgentHarnessEvent, AgentHarnessEventResultMap, AgentHarnessOptions, AgentHarnessOwnEvent, AgentHarnessResources, AgentHarnessStreamOptions, ExecutionEnv, PromptTemplate, Skill } from "./types.js";
|
|
@@ -47,6 +47,22 @@ export interface UserMessageProvenance {
|
|
|
47
47
|
* the parked-steer queue needs (pre-framing text + trust + inputId), which only the minting seam
|
|
48
48
|
* knows — hence opaque. */
|
|
49
49
|
parkRecord?: unknown;
|
|
50
|
+
/** design/373 — an IMMEDIATE-class ("now") frame: the steer-lane enqueue inserts it ahead of the
|
|
51
|
+
* first non-immediate frame (FIFO among immediates — the dequeue-priority form), the loop's
|
|
52
|
+
* pre-request re-check sweeps it into the imminent turn, and {@link AgentHarness.interruptTurn}
|
|
53
|
+
* may force the boundary for it. Class metadata only (a sidecar, judged at the mint) — it never
|
|
54
|
+
* rides the message or the wire, and it grants NOTHING by itself: the interrupt is a separate,
|
|
55
|
+
* caller-provenance verb. */
|
|
56
|
+
immediate?: true;
|
|
57
|
+
/** design/373 (#445 saturation prerequisite) — terminal-preference at the engine-note backlog
|
|
58
|
+
* cap: when the target queue's engine-note count is AT the cap, a frame carrying this mark
|
|
59
|
+
* displaces the OLDEST non-preferred engine-note frame instead of being refused — the displaced
|
|
60
|
+
* payload is handed to the undrained sink (the runner's lossless per-session park lane), so a
|
|
61
|
+
* watcher's event storm can delay its own batches but never crowd a terminal completion out of
|
|
62
|
+
* the run that is waiting on it (the pend store's terminal-preferred eviction, mirrored at the
|
|
63
|
+
* queue mouth). Only meaningful WITH `enginePayload`; a preferred frame among only preferred
|
|
64
|
+
* frames still gets the cap refusal (nothing displaceable). */
|
|
65
|
+
capPreferred?: true;
|
|
50
66
|
}
|
|
51
67
|
/**
|
|
52
68
|
* Harness-level recovery wiring (design/118 ④b/⑤). The loop's prompt-too-long seam wants a
|
|
@@ -190,6 +206,9 @@ export declare class AgentHarness<TSkill extends Skill = Skill, TPromptTemplate
|
|
|
190
206
|
private loopRecovery?;
|
|
191
207
|
/** roadmap #5 (CC Stop hook): the runner-wired stop gate — see {@link setStopGate}. */
|
|
192
208
|
private stopGate?;
|
|
209
|
+
/** design/373 — the loop-published turn-scoped AbortController (live exactly while a main turn's
|
|
210
|
+
* work phase is in flight; `undefined` between turns). Read by {@link interruptTurn} only. */
|
|
211
|
+
private turnInterruptSeat?;
|
|
193
212
|
private handlers;
|
|
194
213
|
constructor(options: AgentHarnessOptions<TSkill, TPromptTemplate, TTool>);
|
|
195
214
|
private getHandlers;
|
|
@@ -264,12 +283,49 @@ export declare class AgentHarness<TSkill extends Skill = Skill, TPromptTemplate
|
|
|
264
283
|
* engine lanes each check the halt fact before injecting, so this arm exists for the lane that
|
|
265
284
|
* does not — including the one not written yet. */
|
|
266
285
|
private refuseHeldEngineInjection;
|
|
286
|
+
/** design/373 — returns the minted queue frame (or `undefined` on the empty-injection no-op), so
|
|
287
|
+
* a caller-provenance immediate steer can hand the SAME frame to {@link interruptTurn} (the
|
|
288
|
+
* consumption guard is identity-keyed). Additive: every existing caller awaited `void`. */
|
|
267
289
|
steer(text: string, options?: {
|
|
268
290
|
images?: ImageContent[];
|
|
269
|
-
} & UserMessageProvenance): Promise<
|
|
291
|
+
} & UserMessageProvenance): Promise<UserMessage | undefined>;
|
|
270
292
|
followUp(text: string, options?: {
|
|
271
293
|
images?: ImageContent[];
|
|
272
294
|
} & UserMessageProvenance): Promise<void>;
|
|
295
|
+
/**
|
|
296
|
+
* design/373 §3.4 — force the running turn's boundary FOR a queued immediate-class frame:
|
|
297
|
+
* abort the CURRENT main turn (provider stream + tool batch — never the run, never boundary
|
|
298
|
+
* housekeeping), whose loop-side settlement reconciles the cut and continues the run with the
|
|
299
|
+
* frame at the queue head. Best-effort accelerator by contract: `false` means no interruption
|
|
300
|
+
* happened AND none was needed — the frame is already consumed or already rides the imminent
|
|
301
|
+
* boundary — so delivery semantics are exactly the non-interrupting ones either way.
|
|
302
|
+
*
|
|
303
|
+
* Guards, in order:
|
|
304
|
+
* - an unwinding run (`abort()` latched) owns its ending — never contest it;
|
|
305
|
+
* - CONSUMPTION guard (r4/R3-1): the frame must still sit in the steer queue. Drained/consumed ⇒
|
|
306
|
+
* it is already part of a turn's input — aborting THAT turn would cut the very delivery this
|
|
307
|
+
* verb exists to speed up. Same-tick check-then-abort (single-threaded: no window);
|
|
308
|
+
* - seat liveness: no published turn (idle tail / between turns / boundary housekeeping) ⇒ no-op,
|
|
309
|
+
* the frame rides the next boundary; an already-fired seat ⇒ idempotent no-op (one boundary per
|
|
310
|
+
* turn — later immediates join the same boundary batch).
|
|
311
|
+
*/
|
|
312
|
+
interruptTurn(frame: AgentMessage): boolean;
|
|
313
|
+
/**
|
|
314
|
+
* design/373 §3.3 (the final-commit-point double check) — how many queued injection frames a
|
|
315
|
+
* boundary drain could deliver RIGHT NOW: both lanes, minus frames the current state refuses to
|
|
316
|
+
* drain (an unwinding run drains nothing — those frames belong to the terminal account; a live
|
|
317
|
+
* human-halt hold skips engine-authored frames). Synchronous and side-effect-free by contract:
|
|
318
|
+
* the loop reads it between its last drain and the terminal commit, with zero awaits.
|
|
319
|
+
*/
|
|
320
|
+
pendingInjectionCount(): number;
|
|
321
|
+
/** design/373 §3.4-3 — the pre-request re-check's drain half (see the loop-config seam): sweep
|
|
322
|
+
* queued immediate-class frames into the imminent turn's batch, merged at the class position
|
|
323
|
+
* (after immediates already pending, ahead of everything else — FIFO within the class holds
|
|
324
|
+
* across the drain/re-check split). Same booking discipline as {@link drainQueuedMessages}
|
|
325
|
+
* (in-flight account, queue_update broadcast, consumption sink, throw-rollback); the hold skips
|
|
326
|
+
* engine-authored immediates exactly as the boundary drain would (a theoretical lane today —
|
|
327
|
+
* no engine mint issues immediates — held by construction rather than by remembering). */
|
|
328
|
+
private absorbImmediateSteering;
|
|
273
329
|
nextTurn(text: string, options?: {
|
|
274
330
|
images?: ImageContent[];
|
|
275
331
|
} & UserMessageProvenance): Promise<void>;
|
|
@@ -38,6 +38,8 @@ const engineNotePayloads = new WeakMap();
|
|
|
38
38
|
const engineAuthoredInjections = new WeakSet();
|
|
39
39
|
const userInputParkRecords = new WeakMap();
|
|
40
40
|
const drainedFromLane = new WeakMap();
|
|
41
|
+
const immediateClassInjections = new WeakSet();
|
|
42
|
+
const capPreferredInjections = new WeakSet();
|
|
41
43
|
function isEngineAuthoredInjection(options) {
|
|
42
44
|
if (options === undefined)
|
|
43
45
|
return false;
|
|
@@ -257,6 +259,7 @@ export class AgentHarness {
|
|
|
257
259
|
nextTurnQueue = [];
|
|
258
260
|
loopRecovery;
|
|
259
261
|
stopGate;
|
|
262
|
+
turnInterruptSeat;
|
|
260
263
|
handlers = new Map();
|
|
261
264
|
constructor(options) {
|
|
262
265
|
this.env = options.env;
|
|
@@ -598,6 +601,11 @@ export class AgentHarness {
|
|
|
598
601
|
};
|
|
599
602
|
},
|
|
600
603
|
getSteeringMessages: async () => this.drainQueuedMessages(this.steerQueue, this.steeringQueueMode),
|
|
604
|
+
publishTurnInterruptSeat: (seat) => {
|
|
605
|
+
this.turnInterruptSeat = seat;
|
|
606
|
+
},
|
|
607
|
+
recheckImmediateInjections: async (pending) => this.absorbImmediateSteering(pending),
|
|
608
|
+
pendingInjectionCount: () => this.pendingInjectionCount(),
|
|
601
609
|
getFollowUpMessages: async () => {
|
|
602
610
|
const queued = await this.drainQueuedMessages(this.followUpQueue, this.followUpQueueMode);
|
|
603
611
|
if (queued.length > 0)
|
|
@@ -841,9 +849,16 @@ export class AgentHarness {
|
|
|
841
849
|
}
|
|
842
850
|
async enqueueInjection(queue, text, options) {
|
|
843
851
|
if (AgentHarness.emptyInjection(text, options))
|
|
844
|
-
return;
|
|
852
|
+
return undefined;
|
|
853
|
+
let displaced;
|
|
845
854
|
if (options?.enginePayload !== undefined && queue.filter((q) => engineNotePayloads.has(q)).length >= ENGINE_NOTE_STEER_BACKLOG_CAP) {
|
|
846
|
-
|
|
855
|
+
const victimIdx = options.capPreferred === true
|
|
856
|
+
? queue.findIndex((q) => engineNotePayloads.has(q) && !capPreferredInjections.has(q))
|
|
857
|
+
: -1;
|
|
858
|
+
if (victimIdx === -1) {
|
|
859
|
+
throw new AgentHarnessError("invalid_state", `engine-note backlog at cap (${ENGINE_NOTE_STEER_BACKLOG_CAP}) — park the payload for the session's next run`);
|
|
860
|
+
}
|
|
861
|
+
displaced = { frame: queue.splice(victimIdx, 1)[0], at: victimIdx };
|
|
847
862
|
}
|
|
848
863
|
const m = createUserMessage(text, options?.images, options);
|
|
849
864
|
if (options?.enginePayload !== undefined)
|
|
@@ -852,8 +867,43 @@ export class AgentHarness {
|
|
|
852
867
|
engineAuthoredInjections.add(m);
|
|
853
868
|
if (options?.parkRecord !== undefined)
|
|
854
869
|
userInputParkRecords.set(m, options.parkRecord);
|
|
855
|
-
|
|
856
|
-
|
|
870
|
+
if (options?.immediate === true)
|
|
871
|
+
immediateClassInjections.add(m);
|
|
872
|
+
if (options?.capPreferred === true && options?.enginePayload !== undefined)
|
|
873
|
+
capPreferredInjections.add(m);
|
|
874
|
+
if (options?.immediate === true && queue === this.steerQueue) {
|
|
875
|
+
let at = 0;
|
|
876
|
+
while (at < queue.length && immediateClassInjections.has(queue[at]))
|
|
877
|
+
at++;
|
|
878
|
+
queue.splice(at, 0, m);
|
|
879
|
+
}
|
|
880
|
+
else {
|
|
881
|
+
queue.push(m);
|
|
882
|
+
}
|
|
883
|
+
try {
|
|
884
|
+
await this.emitQueueUpdate();
|
|
885
|
+
}
|
|
886
|
+
catch (error) {
|
|
887
|
+
const at = queue.indexOf(m);
|
|
888
|
+
if (at !== -1) {
|
|
889
|
+
queue.splice(at, 1);
|
|
890
|
+
if (displaced !== undefined)
|
|
891
|
+
queue.splice(Math.min(displaced.at, queue.length), 0, displaced.frame);
|
|
892
|
+
throw error;
|
|
893
|
+
}
|
|
894
|
+
}
|
|
895
|
+
if (displaced !== undefined) {
|
|
896
|
+
const victimPayload = engineNotePayloads.get(displaced.frame);
|
|
897
|
+
engineNotePayloads.delete(displaced.frame);
|
|
898
|
+
if (victimPayload !== undefined && this.onUndrainedEngineNotes) {
|
|
899
|
+
try {
|
|
900
|
+
this.onUndrainedEngineNotes([victimPayload]);
|
|
901
|
+
}
|
|
902
|
+
catch {
|
|
903
|
+
}
|
|
904
|
+
}
|
|
905
|
+
}
|
|
906
|
+
return m;
|
|
857
907
|
}
|
|
858
908
|
refuseHeldEngineInjection(options) {
|
|
859
909
|
if (isEngineAuthoredInjection(options) && this.engineInjectionsHeld?.() === true) {
|
|
@@ -865,7 +915,7 @@ export class AgentHarness {
|
|
|
865
915
|
throw new AgentHarnessError("invalid_state", "Cannot steer while idle");
|
|
866
916
|
}
|
|
867
917
|
this.refuseHeldEngineInjection(options);
|
|
868
|
-
await this.enqueueInjection(this.steerQueue, text, options);
|
|
918
|
+
return await this.enqueueInjection(this.steerQueue, text, options);
|
|
869
919
|
}
|
|
870
920
|
async followUp(text, options) {
|
|
871
921
|
if (this.phase === "idle") {
|
|
@@ -874,6 +924,66 @@ export class AgentHarness {
|
|
|
874
924
|
this.refuseHeldEngineInjection(options);
|
|
875
925
|
await this.enqueueInjection(this.followUpQueue, text, options);
|
|
876
926
|
}
|
|
927
|
+
interruptTurn(frame) {
|
|
928
|
+
if (this.aborting)
|
|
929
|
+
return false;
|
|
930
|
+
if (!this.steerQueue.includes(frame))
|
|
931
|
+
return false;
|
|
932
|
+
const seat = this.turnInterruptSeat;
|
|
933
|
+
if (seat === undefined || seat.signal.aborted)
|
|
934
|
+
return false;
|
|
935
|
+
seat.abort();
|
|
936
|
+
return true;
|
|
937
|
+
}
|
|
938
|
+
pendingInjectionCount() {
|
|
939
|
+
if (this.aborting)
|
|
940
|
+
return 0;
|
|
941
|
+
const held = this.engineInjectionsHeld?.() === true;
|
|
942
|
+
const drainable = (q) => held ? q.filter((m) => !engineAuthoredInjections.has(m)).length : q.length;
|
|
943
|
+
return drainable(this.steerQueue) + drainable(this.followUpQueue);
|
|
944
|
+
}
|
|
945
|
+
async absorbImmediateSteering(pending) {
|
|
946
|
+
if (this.aborting)
|
|
947
|
+
return pending;
|
|
948
|
+
const held = this.engineInjectionsHeld?.() === true;
|
|
949
|
+
const taken = [];
|
|
950
|
+
for (let i = 0; i < this.steerQueue.length; i++) {
|
|
951
|
+
const m = this.steerQueue[i];
|
|
952
|
+
if (!immediateClassInjections.has(m))
|
|
953
|
+
break;
|
|
954
|
+
if (held && engineAuthoredInjections.has(m))
|
|
955
|
+
continue;
|
|
956
|
+
taken.push(...this.steerQueue.splice(i, 1));
|
|
957
|
+
i--;
|
|
958
|
+
}
|
|
959
|
+
if (taken.length === 0)
|
|
960
|
+
return pending;
|
|
961
|
+
for (const m of taken)
|
|
962
|
+
drainedFromLane.set(m, "steer");
|
|
963
|
+
this.drainedPendingInjection.push(...taken);
|
|
964
|
+
try {
|
|
965
|
+
await this.emitQueueUpdate();
|
|
966
|
+
for (const m of taken) {
|
|
967
|
+
const payload = engineNotePayloads.get(m);
|
|
968
|
+
if (payload !== undefined) {
|
|
969
|
+
try {
|
|
970
|
+
this.onEngineNoteConsumed?.(payload);
|
|
971
|
+
}
|
|
972
|
+
catch {
|
|
973
|
+
}
|
|
974
|
+
}
|
|
975
|
+
}
|
|
976
|
+
}
|
|
977
|
+
catch (error) {
|
|
978
|
+
this.steerQueue.unshift(...taken);
|
|
979
|
+
this.unbookInFlight(taken);
|
|
980
|
+
throw normalizeHookError(error);
|
|
981
|
+
}
|
|
982
|
+
let at = 0;
|
|
983
|
+
while (at < pending.length && immediateClassInjections.has(pending[at]))
|
|
984
|
+
at++;
|
|
985
|
+
return [...pending.slice(0, at), ...taken, ...pending.slice(at)];
|
|
986
|
+
}
|
|
877
987
|
async nextTurn(text, options) {
|
|
878
988
|
await this.enqueueInjection(this.nextTurnQueue, text, options);
|
|
879
989
|
}
|
|
@@ -1,5 +1,6 @@
|
|
|
1
1
|
import { findToolByName, validateToolArguments } from "../llm/index.js";
|
|
2
2
|
import { truncateError } from "../../core/tool-errors.js";
|
|
3
|
+
import { INTERRUPTED_BY_USER_FOR_TOOL_USE_MARKER, INTERRUPTED_BY_USER_MARKER, } from "../../core/session-reconcile.js";
|
|
3
4
|
import { resolveAgentCoreStreamFn } from "./runtime-deps.js";
|
|
4
5
|
function appendTextDeltaToAssistantMessage(message, contentIndex, delta) {
|
|
5
6
|
const content = [...message.content];
|
|
@@ -158,6 +159,13 @@ async function runLoop(initialContext, newMessages, initialConfig, signal, emit,
|
|
|
158
159
|
if (!state.hasMoreToolCalls && state.pendingMessages.length === 0) {
|
|
159
160
|
const followUpMessages = (await state.config.getFollowUpMessages?.()) || [];
|
|
160
161
|
if (followUpMessages.length === 0) {
|
|
162
|
+
if ((state.config.pendingInjectionCount?.() ?? 0) > 0) {
|
|
163
|
+
state.pendingMessages = (await state.config.getSteeringMessages?.()) || [];
|
|
164
|
+
if (state.pendingMessages.length > 0) {
|
|
165
|
+
trace?.({ kind: "continue", reason: "steer_injected" });
|
|
166
|
+
}
|
|
167
|
+
continue;
|
|
168
|
+
}
|
|
161
169
|
trace?.({ kind: "terminal", reason: "completed" });
|
|
162
170
|
break;
|
|
163
171
|
}
|
|
@@ -193,6 +201,38 @@ async function runSingleTurn(state, signal, emit, streamFn, runtime, trace) {
|
|
|
193
201
|
else {
|
|
194
202
|
state.firstTurn = false;
|
|
195
203
|
}
|
|
204
|
+
const turnController = new AbortController();
|
|
205
|
+
const onRunAbort = () => turnController.abort();
|
|
206
|
+
if (signal !== undefined) {
|
|
207
|
+
if (signal.aborted)
|
|
208
|
+
turnController.abort();
|
|
209
|
+
else
|
|
210
|
+
signal.addEventListener("abort", onRunAbort, { once: true });
|
|
211
|
+
}
|
|
212
|
+
let seatLive = false;
|
|
213
|
+
const publishSeat = () => {
|
|
214
|
+
seatLive = true;
|
|
215
|
+
state.config.publishTurnInterruptSeat?.(turnController);
|
|
216
|
+
};
|
|
217
|
+
const retractSeat = () => {
|
|
218
|
+
if (!seatLive)
|
|
219
|
+
return;
|
|
220
|
+
seatLive = false;
|
|
221
|
+
state.config.publishTurnInterruptSeat?.(undefined);
|
|
222
|
+
};
|
|
223
|
+
try {
|
|
224
|
+
return await runTurnPhases(state, signal, turnController, publishSeat, retractSeat, emit, streamFn, runtime, trace);
|
|
225
|
+
}
|
|
226
|
+
finally {
|
|
227
|
+
retractSeat();
|
|
228
|
+
signal?.removeEventListener("abort", onRunAbort);
|
|
229
|
+
}
|
|
230
|
+
}
|
|
231
|
+
async function runTurnPhases(state, signal, turnController, publishSeat, retractSeat, emit, streamFn, runtime, trace) {
|
|
232
|
+
publishSeat();
|
|
233
|
+
if (state.config.recheckImmediateInjections) {
|
|
234
|
+
state.pendingMessages = await state.config.recheckImmediateInjections(state.pendingMessages);
|
|
235
|
+
}
|
|
196
236
|
if (state.pendingMessages.length > 0) {
|
|
197
237
|
for (const message of state.pendingMessages) {
|
|
198
238
|
await emit({ type: "message_start", message });
|
|
@@ -201,7 +241,7 @@ async function runSingleTurn(state, signal, emit, streamFn, runtime, trace) {
|
|
|
201
241
|
state.newMessages.push(message);
|
|
202
242
|
}
|
|
203
243
|
}
|
|
204
|
-
let executor = new StreamToolExecutor(state.context, state.config, signal, emit);
|
|
244
|
+
let executor = new StreamToolExecutor(state.context, state.config, turnController.signal, emit);
|
|
205
245
|
const staticReasoningCutDowngrade = state.config.recovery?.truncatedOutput !== undefined &&
|
|
206
246
|
state.staticReasoningCutRecoveries < MAX_STATIC_REASONING_CUT_RECOVERIES;
|
|
207
247
|
const ptl = state.config.recovery?.promptTooLong;
|
|
@@ -210,7 +250,7 @@ async function runSingleTurn(state, signal, emit, streamFn, runtime, trace) {
|
|
|
210
250
|
let withhold = withholdEnabled ? createPtlWithholdBuffer(emit, ptlDetect) : undefined;
|
|
211
251
|
let message;
|
|
212
252
|
try {
|
|
213
|
-
message = await streamAssistantResponse(state.context, state.config, signal, withhold?.sink ?? emit, streamFn, runtime, executor, staticReasoningCutDowngrade);
|
|
253
|
+
message = await streamAssistantResponse(state.context, state.config, turnController.signal, withhold?.sink ?? emit, streamFn, runtime, executor, staticReasoningCutDowngrade);
|
|
214
254
|
}
|
|
215
255
|
finally {
|
|
216
256
|
await executor.settle();
|
|
@@ -246,9 +286,9 @@ async function runSingleTurn(state, signal, emit, streamFn, runtime, trace) {
|
|
|
246
286
|
withhold = withholdEnabled ? createPtlWithholdBuffer(emit, ptlDetect) : undefined;
|
|
247
287
|
state.context.messages = replaced;
|
|
248
288
|
trace?.({ kind: "continue", reason: "reactive_compact_retry" });
|
|
249
|
-
executor = new StreamToolExecutor(state.context, state.config, signal, emit);
|
|
289
|
+
executor = new StreamToolExecutor(state.context, state.config, turnController.signal, emit);
|
|
250
290
|
try {
|
|
251
|
-
message = await streamAssistantResponse(state.context, state.config, signal, withhold?.sink ?? emit, streamFn, runtime, executor, staticReasoningCutDowngrade);
|
|
291
|
+
message = await streamAssistantResponse(state.context, state.config, turnController.signal, withhold?.sink ?? emit, streamFn, runtime, executor, staticReasoningCutDowngrade);
|
|
252
292
|
}
|
|
253
293
|
finally {
|
|
254
294
|
await executor.settle();
|
|
@@ -261,6 +301,10 @@ async function runSingleTurn(state, signal, emit, streamFn, runtime, trace) {
|
|
|
261
301
|
state.newMessages.push(message);
|
|
262
302
|
if (message.staticReasoningCut === true)
|
|
263
303
|
state.staticReasoningCutRecoveries++;
|
|
304
|
+
if (message.stopReason === "aborted" && turnController.signal.aborted && !signal?.aborted) {
|
|
305
|
+
retractSeat();
|
|
306
|
+
return await settleInterruptedTurn(state, message, executor, [], signal, emit);
|
|
307
|
+
}
|
|
264
308
|
{
|
|
265
309
|
const degen = state.config.recovery?.degenerateOutput;
|
|
266
310
|
if (degen &&
|
|
@@ -270,6 +314,7 @@ async function runSingleTurn(state, signal, emit, streamFn, runtime, trace) {
|
|
|
270
314
|
executor.admittedCount === 0 &&
|
|
271
315
|
message.content.every((b) => b.type !== "toolCall") &&
|
|
272
316
|
state.degenerateContinues < (degen.maxContinues ?? 2)) {
|
|
317
|
+
retractSeat();
|
|
273
318
|
state.degenerateContinues++;
|
|
274
319
|
await emit({ type: "turn_end", message, toolResults: [] });
|
|
275
320
|
const drained = (await state.config.getSteeringMessages?.()) || [];
|
|
@@ -280,7 +325,8 @@ async function runSingleTurn(state, signal, emit, streamFn, runtime, trace) {
|
|
|
280
325
|
state.degenerateContinues = 0;
|
|
281
326
|
}
|
|
282
327
|
if (message.stopReason === "error" || message.stopReason === "aborted") {
|
|
283
|
-
|
|
328
|
+
retractSeat();
|
|
329
|
+
const errorFinalResults = await harvestExecutorOnErrorFinal(executor, state, message, turnController.signal, emit);
|
|
284
330
|
await emit({ type: "turn_end", message, toolResults: errorFinalResults });
|
|
285
331
|
await emit({ type: "agent_end", messages: state.newMessages });
|
|
286
332
|
return { kind: "terminal", reason: "assistant_error" };
|
|
@@ -289,7 +335,7 @@ async function runSingleTurn(state, signal, emit, streamFn, runtime, trace) {
|
|
|
289
335
|
const toolResults = [];
|
|
290
336
|
state.hasMoreToolCalls = false;
|
|
291
337
|
if (toolCalls.length > 0) {
|
|
292
|
-
const executedToolBatch = await executeToolCalls(state.context, message, state.config, signal, emit, executor);
|
|
338
|
+
const executedToolBatch = await executeToolCalls(state.context, message, state.config, turnController.signal, emit, executor);
|
|
293
339
|
toolResults.push(...executedToolBatch.messages);
|
|
294
340
|
state.hasMoreToolCalls = !executedToolBatch.terminate;
|
|
295
341
|
for (const result of toolResults) {
|
|
@@ -303,6 +349,10 @@ async function runSingleTurn(state, signal, emit, streamFn, runtime, trace) {
|
|
|
303
349
|
await closeOrphanedStreamEntry(orphan, emit);
|
|
304
350
|
}
|
|
305
351
|
}
|
|
352
|
+
retractSeat();
|
|
353
|
+
if (toolCalls.length > 0 && turnController.signal.aborted && !signal?.aborted) {
|
|
354
|
+
return await settleInterruptedTurn(state, message, executor, toolResults, signal, emit);
|
|
355
|
+
}
|
|
306
356
|
await emit({ type: "turn_end", message, toolResults });
|
|
307
357
|
const nextTurnContext = {
|
|
308
358
|
message,
|
|
@@ -311,15 +361,7 @@ async function runSingleTurn(state, signal, emit, streamFn, runtime, trace) {
|
|
|
311
361
|
newMessages: state.newMessages,
|
|
312
362
|
};
|
|
313
363
|
const nextTurnSnapshot = await state.config.prepareNextTurn?.(nextTurnContext);
|
|
314
|
-
|
|
315
|
-
state.context = nextTurnSnapshot.context ?? state.context;
|
|
316
|
-
state.config = Object.assign({}, state.config, {
|
|
317
|
-
model: nextTurnSnapshot.model ?? state.config.model,
|
|
318
|
-
reasoning: nextTurnSnapshot.thinkingLevel === undefined
|
|
319
|
-
? state.config.reasoning
|
|
320
|
-
: nextTurnSnapshot.thinkingLevel,
|
|
321
|
-
});
|
|
322
|
-
}
|
|
364
|
+
adoptNextTurnSnapshot(state, nextTurnSnapshot);
|
|
323
365
|
if (await state.config.shouldStopAfterTurn?.({
|
|
324
366
|
message,
|
|
325
367
|
toolResults,
|
|
@@ -383,6 +425,102 @@ async function runSingleTurn(state, signal, emit, streamFn, runtime, trace) {
|
|
|
383
425
|
}
|
|
384
426
|
return { kind: "ran" };
|
|
385
427
|
}
|
|
428
|
+
function adoptNextTurnSnapshot(state, nextTurnSnapshot) {
|
|
429
|
+
if (!nextTurnSnapshot)
|
|
430
|
+
return;
|
|
431
|
+
state.context = nextTurnSnapshot.context ?? state.context;
|
|
432
|
+
state.config = Object.assign({}, state.config, {
|
|
433
|
+
model: nextTurnSnapshot.model ?? state.config.model,
|
|
434
|
+
reasoning: nextTurnSnapshot.thinkingLevel === undefined
|
|
435
|
+
? state.config.reasoning
|
|
436
|
+
: nextTurnSnapshot.thinkingLevel,
|
|
437
|
+
});
|
|
438
|
+
}
|
|
439
|
+
const TURN_INTERRUPTED_TOOL_TEXT = "[INTERRUPTED] The user interrupted this turn before this tool call started. It was never executed and " +
|
|
440
|
+
"had no side effects. Address the user's interjection first; re-issue the call afterwards if it is still needed.";
|
|
441
|
+
async function settleInterruptedTurn(state, message, executor, committedResults, signal, emit) {
|
|
442
|
+
const toolResults = [...committedResults];
|
|
443
|
+
toolResults.push(...(await harvestExecutorOnErrorFinal(executor, state, message, signal, emit)));
|
|
444
|
+
const turnCalls = message.content.filter((c) => c.type === "toolCall");
|
|
445
|
+
const answered = new Set(toolResults.map((r) => r.toolCallId));
|
|
446
|
+
for (const call of turnCalls) {
|
|
447
|
+
if (answered.has(call.id))
|
|
448
|
+
continue;
|
|
449
|
+
await emit({
|
|
450
|
+
type: "tool_execution_start",
|
|
451
|
+
toolCallId: call.id,
|
|
452
|
+
toolName: call.name,
|
|
453
|
+
args: call.arguments,
|
|
454
|
+
});
|
|
455
|
+
const finalized = {
|
|
456
|
+
toolCall: call,
|
|
457
|
+
result: {
|
|
458
|
+
content: [{ type: "text", text: TURN_INTERRUPTED_TOOL_TEXT }],
|
|
459
|
+
details: { errorKind: "interrupted_never_started" },
|
|
460
|
+
},
|
|
461
|
+
isError: true,
|
|
462
|
+
};
|
|
463
|
+
await emit({
|
|
464
|
+
type: "tool_execution_end",
|
|
465
|
+
toolCallId: call.id,
|
|
466
|
+
toolName: call.name,
|
|
467
|
+
result: finalized.result,
|
|
468
|
+
isError: true,
|
|
469
|
+
notExecuted: true,
|
|
470
|
+
});
|
|
471
|
+
const toolResultMessage = createToolResultMessage(finalized);
|
|
472
|
+
await emitToolResultMessage(toolResultMessage, emit);
|
|
473
|
+
state.context.messages.push(toolResultMessage);
|
|
474
|
+
state.newMessages.push(toolResultMessage);
|
|
475
|
+
toolResults.push(toolResultMessage);
|
|
476
|
+
}
|
|
477
|
+
if (turnCalls.length === 0 &&
|
|
478
|
+
isBlankFailureContent(message) &&
|
|
479
|
+
state.context.messages[state.context.messages.length - 1] === message) {
|
|
480
|
+
state.context.messages.pop();
|
|
481
|
+
if (state.newMessages[state.newMessages.length - 1] === message)
|
|
482
|
+
state.newMessages.pop();
|
|
483
|
+
}
|
|
484
|
+
const marker = {
|
|
485
|
+
role: "user",
|
|
486
|
+
content: [
|
|
487
|
+
{
|
|
488
|
+
type: "text",
|
|
489
|
+
text: turnCalls.length > 0 ? INTERRUPTED_BY_USER_FOR_TOOL_USE_MARKER : INTERRUPTED_BY_USER_MARKER,
|
|
490
|
+
},
|
|
491
|
+
],
|
|
492
|
+
provenance: "engine-note",
|
|
493
|
+
timestamp: Date.now(),
|
|
494
|
+
};
|
|
495
|
+
await emit({ type: "message_start", message: marker });
|
|
496
|
+
await emit({ type: "message_end", message: marker });
|
|
497
|
+
state.context.messages.push(marker);
|
|
498
|
+
state.newMessages.push(marker);
|
|
499
|
+
state.hasMoreToolCalls = false;
|
|
500
|
+
await emit({ type: "turn_end", message, toolResults });
|
|
501
|
+
if (signal?.aborted) {
|
|
502
|
+
await emit({ type: "agent_end", messages: state.newMessages });
|
|
503
|
+
return { kind: "terminal", reason: "aborted_before_stream" };
|
|
504
|
+
}
|
|
505
|
+
const nextTurnSnapshot = await state.config.prepareNextTurn?.({
|
|
506
|
+
message,
|
|
507
|
+
toolResults,
|
|
508
|
+
context: state.context,
|
|
509
|
+
newMessages: state.newMessages,
|
|
510
|
+
});
|
|
511
|
+
adoptNextTurnSnapshot(state, nextTurnSnapshot);
|
|
512
|
+
if (await state.config.shouldStopAfterTurn?.({
|
|
513
|
+
message,
|
|
514
|
+
toolResults,
|
|
515
|
+
context: state.context,
|
|
516
|
+
newMessages: state.newMessages,
|
|
517
|
+
})) {
|
|
518
|
+
await emit({ type: "agent_end", messages: state.newMessages });
|
|
519
|
+
return { kind: "terminal", reason: "stop_requested" };
|
|
520
|
+
}
|
|
521
|
+
state.pendingMessages = (await state.config.getSteeringMessages?.()) || [];
|
|
522
|
+
return { kind: "ran" };
|
|
523
|
+
}
|
|
386
524
|
async function streamAssistantResponse(context, config, signal, emit, streamFn, runtime, executor, staticReasoningCutDowngrade) {
|
|
387
525
|
let messages = context.messages;
|
|
388
526
|
if (config.transformContext) {
|