@try-works/dsh-recursive-mode 0.4.4 → 0.4.5
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/lib/errors.d.ts +12 -0
- package/lib/goals-projection.d.ts +34 -4
- package/lib/index.js +471 -41
- package/lib/recursive_ask.tool.d.ts +60 -0
- package/lib/recursive_init.tool.d.ts +15 -0
- package/lib/run-start.d.ts +55 -0
- package/lib/runtime.d.ts +137 -31
- package/package.json +1 -1
- package/scripts/test-recursive-mode-smoke.ts +5 -1
- package/src/errors.ts +12 -0
- package/src/goals-projection.ts +48 -7
- package/src/index.ts +9 -0
- package/src/recursive_ask.tool.ts +218 -17
- package/src/recursive_init.tool.ts +16 -1
- package/src/run-start.ts +123 -0
- package/src/runtime.ts +157 -14
|
@@ -1,7 +1,24 @@
|
|
|
1
1
|
import type { RecursiveRuntime } from './runtime.ts';
|
|
2
|
+
import { RUN_START_GATE_ID } from './run-start.ts';
|
|
2
3
|
/** The identifiers the workflow uses for its three human gates. */
|
|
3
4
|
export declare const ASK_GATE_IDS: readonly ["tdd-mode", "qa-signoff", "gate-block"];
|
|
4
5
|
export type AskGateId = (typeof ASK_GATE_IDS)[number];
|
|
6
|
+
/**
|
|
7
|
+
* PHASE 0 — THE FOURTH GATE, AND WHY IT IS NOT IN `ASK_GATE_IDS`.
|
|
8
|
+
*
|
|
9
|
+
* The three above are the WORKFLOW's gates, and their membership is asserted as exactly those three.
|
|
10
|
+
* Starting a run is a different kind of decision — it decides whether there is a run at all, and
|
|
11
|
+
* approving it ARMS A GOAL the harness will keep driving — so its data lives in `run-start.ts` with its
|
|
12
|
+
* own contract and its own options. Widening the workflow's gate list must not silently widen what may
|
|
13
|
+
* start a run.
|
|
14
|
+
*/
|
|
15
|
+
export type AskAnyGateId = AskGateId | typeof RUN_START_GATE_ID;
|
|
16
|
+
/** Is this gate id the run-start gate? */
|
|
17
|
+
export declare function isRunStartGate(gateId: string): boolean;
|
|
18
|
+
/** Every gate id `recursive_ask` accepts, workflow gates first. */
|
|
19
|
+
export declare function askGateIds(): string[];
|
|
20
|
+
/** The artifact a gate's answer belongs in — the run-start gate's is fixed to the Phase 0 requirements. */
|
|
21
|
+
export declare function askGateArtifact(gateId: AskAnyGateId): string;
|
|
5
22
|
/** The plugin's own documented bounds — see the module comment on why these are not a claimed mirror. */
|
|
6
23
|
export declare const MAX_HEADER_CHARS = 12;
|
|
7
24
|
export declare const MAX_LABEL_CHARS = 30;
|
|
@@ -41,6 +58,21 @@ export declare class AskValidationError extends Error {
|
|
|
41
58
|
export declare function validateAskQuestion(question: AskQuestion): AskQuestion;
|
|
42
59
|
/** Build the question for a gate, validated. */
|
|
43
60
|
export declare function buildAskQuestion(gateId: AskGateId): AskQuestion;
|
|
61
|
+
/**
|
|
62
|
+
* PHASE 0 — build the question for ANY accepted gate, including the run-start gate.
|
|
63
|
+
*
|
|
64
|
+
* A separate entry point rather than a widened `buildAskQuestion` so the three workflow gates keep the
|
|
65
|
+
* exact signature and behaviour their callers (and `runtime.phaseRules`) already rely on.
|
|
66
|
+
*/
|
|
67
|
+
export declare function buildAskQuestionFor(gateId: AskAnyGateId): AskQuestion;
|
|
68
|
+
/**
|
|
69
|
+
* PHASE 0 — validate an answer to ANY accepted gate.
|
|
70
|
+
*
|
|
71
|
+
* The run-start gate accepts only the labels IT offered, exactly like the other three, and the check is
|
|
72
|
+
* the same `ASK_GATES`-shaped test against its own options. An answer of `maybe` is refused rather than
|
|
73
|
+
* recorded, because a recorded non-answer is the failure mode this whole change exists to prevent.
|
|
74
|
+
*/
|
|
75
|
+
export declare function validateAskAnswerFor(gateId: AskAnyGateId, answer: string): string;
|
|
44
76
|
/**
|
|
45
77
|
* Validate an answer against its gate.
|
|
46
78
|
*
|
|
@@ -84,5 +116,33 @@ export declare function pendingGateFor(artifactFile: string, artifactText: strin
|
|
|
84
116
|
* ⚠ THE WRITE-BACK IS A MARKER LINE, REPLACED IN PLACE when the artifact already carries one. A
|
|
85
117
|
* second `TDD Mode:` line would leave two answers to one question and make "what was decided?"
|
|
86
118
|
* depend on which a reader found first.
|
|
119
|
+
*
|
|
120
|
+
* ⚠ PHASE 0 — `run-start` IS RECORDED BY THE PLUGIN, NEVER BY A BARE MARKER WRITE. See
|
|
121
|
+
* `recordRunStartAnswer`: it is the only path that can arm a run goal, it prefers the blocking human
|
|
122
|
+
* channel, and it fails closed when no person can be reached.
|
|
87
123
|
*/
|
|
88
124
|
export declare function createRecursiveAskTool(recursive: RecursiveRuntime): import("@deepseek-ai/dsh-tools").ToolDefinition;
|
|
125
|
+
/**
|
|
126
|
+
* PHASE 0 — record the answer to the run-start gate, and start the run only if it says so.
|
|
127
|
+
*
|
|
128
|
+
* ⚠ THIS GATE NEVER ACCEPTS A RELAYED ANSWER WHILE A HUMAN CHANNEL IS MOUNTED. That is the rule that makes
|
|
129
|
+
* an approval a human act rather than an inference: when `ctx.userQuestions` is present, the question is
|
|
130
|
+
* PUT TO THE PERSON and nothing else can settle it — not the caller's own `answer` argument, and not a
|
|
131
|
+
* fabrication, because `ask()` resolves only with a real selection. A person's decline is likewise final
|
|
132
|
+
* for that call and cannot be overridden by a model that asked for `Start run` in the same breath.
|
|
133
|
+
*
|
|
134
|
+
* ⚠ AND WHEN NO CHANNEL IS MOUNTED, THE RELAYED ANSWER IS THE ONLY POSSIBLE SOURCE, so it is used — that
|
|
135
|
+
* is the same contract the other three gates have always had, and refusing it would leave a composition
|
|
136
|
+
* without the channel unable to start any run at all. The question is surfaced first by the ASK branch
|
|
137
|
+
* (the card data the host renders), and the model's `answer` is that person's selection coming back.
|
|
138
|
+
*
|
|
139
|
+
* ⚠ WHAT THE GATE THEREFORE DOES *NOT* CLAIM, stated rather than implied: in a composition with no
|
|
140
|
+
* `userQuestions` channel, a plugin cannot verify that a person was really asked, so a model could in
|
|
141
|
+
* principle relay a label nobody gave. That is a property of the relay, not of this gate — and it is the
|
|
142
|
+
* reason the channel is consulted in preference whenever it exists. See the header of `run-start.ts`.
|
|
143
|
+
*/
|
|
144
|
+
export declare function recordRunStartAnswer(recursive: RecursiveRuntime, root: string, runId: string, answer: string | undefined, exec: {
|
|
145
|
+
agent?: unknown;
|
|
146
|
+
signal?: unknown;
|
|
147
|
+
callId?: unknown;
|
|
148
|
+
}): Promise<Record<string, unknown>>;
|
|
@@ -1,2 +1,17 @@
|
|
|
1
1
|
import type { RecursiveRuntime } from './runtime.ts';
|
|
2
|
+
/**
|
|
3
|
+
* PHASE 0 — SCAFFOLDING IS NOT STARTING, AND THE TOOL SAYS SO AT THE MOMENT IT MATTERS.
|
|
4
|
+
*
|
|
5
|
+
* A spec may legitimately exist before a run does: this tool writes the run directory and every phase
|
|
6
|
+
* document, and it still does. What it must NOT do is start the run, because starting is creating and
|
|
7
|
+
* arming the goal the harness drives autonomous rounds from. That is the owner's rule — *"phase 0
|
|
8
|
+
* requires explicit approval to start a run and goal"* — so the description below names the gate and the
|
|
9
|
+
* result carries `runStartApproval`, which is the pointer a caller needs: the run is inert until
|
|
10
|
+
* `recursive_ask` answers `run-start`.
|
|
11
|
+
*
|
|
12
|
+
* `runStartApproval` is read from the run's own Phase 0 artifact on every call, so it is the TRUE state
|
|
13
|
+
* rather than "this call created something": re-initialising an APPROVED run reports `approved: true`
|
|
14
|
+
* (and the run keeps its goal), which is what a caller re-scaffolding a run it already started needs to
|
|
15
|
+
* see.
|
|
16
|
+
*/
|
|
2
17
|
export declare function createRecursiveInitTool(recursive: RecursiveRuntime): import("@deepseek-ai/dsh-tools").ToolDefinition;
|
|
@@ -0,0 +1,55 @@
|
|
|
1
|
+
/** The gate id `recursive_ask` answers for a run start. Deliberately NOT in ASK_GATE_IDS. */
|
|
2
|
+
export declare const RUN_START_GATE_ID = "run-start";
|
|
3
|
+
/** The Phase 0 artifact the approval is recorded in. */
|
|
4
|
+
export declare const RUN_START_ARTIFACT = "00-requirements.md";
|
|
5
|
+
/** The artifact field the approval reads back from. */
|
|
6
|
+
export declare const RUN_START_MARKER = "Run Start";
|
|
7
|
+
/** The approving label. The ONLY label that starts a run. */
|
|
8
|
+
export declare const RUN_START_APPROVE = "Start run";
|
|
9
|
+
/** The withholding label: the spec stays a spec. */
|
|
10
|
+
export declare const RUN_START_HOLD = "Hold";
|
|
11
|
+
/**
|
|
12
|
+
* WHY THIS GATE IS NOT IN `ASK_GATE_IDS`. Those three are the WORKFLOW's gates — phase-3 test
|
|
13
|
+
* evidence, phase-5 sign-off, resolving a gate block — and their membership is asserted as exactly
|
|
14
|
+
* three. Starting a run is a different kind of decision: it is the one that decides whether there is
|
|
15
|
+
* a run at all. It lives here, with its own contract, so widening the workflow's gate list cannot
|
|
16
|
+
* quietly widen what may start a run.
|
|
17
|
+
*/
|
|
18
|
+
export declare const RUN_START_GATE: {
|
|
19
|
+
readonly id: "run-start";
|
|
20
|
+
readonly header: "Start run";
|
|
21
|
+
readonly question: "Approve phase 0 and start this run? Approving creates an armed goal the harness will keep driving.";
|
|
22
|
+
readonly options: readonly [{
|
|
23
|
+
readonly label: "Start run";
|
|
24
|
+
readonly description: "Record the approval and arm the run goal.";
|
|
25
|
+
}, {
|
|
26
|
+
readonly label: "Hold";
|
|
27
|
+
readonly description: "Leave the spec inert: no run goal, no autonomous rounds.";
|
|
28
|
+
}];
|
|
29
|
+
readonly marker: "Run Start";
|
|
30
|
+
};
|
|
31
|
+
/** The durable line an approval writes. */
|
|
32
|
+
export declare function runStartApprovalLine(): string;
|
|
33
|
+
/** Is a `run-start` answer the approving one? */
|
|
34
|
+
export declare function isRunStartApproval(answer: string): boolean;
|
|
35
|
+
/** Where the Phase 0 requirements artifact lives for a run rooted at `root`. */
|
|
36
|
+
export declare function runStartArtifactPath(root: string, runId: string): string;
|
|
37
|
+
/** The artifact text, or null when the file is absent (a read failure is not an approval). */
|
|
38
|
+
export declare function readRunStartArtifact(root: string, runId: string): string | null;
|
|
39
|
+
/**
|
|
40
|
+
* The approval state of a run, read from its Phase 0 artifact.
|
|
41
|
+
*
|
|
42
|
+
* ⚠ MATCHED ON THE VALUE, NOT ON THE LINE'S PRESENCE. `getMdFieldValue` returns the field's VALUE, so
|
|
43
|
+
* a recorded `- Run Start: Hold` is refused here — a check for "is there a Run Start line?" would read
|
|
44
|
+
* a refusal as consent, which is the one mistake this whole module exists to prevent.
|
|
45
|
+
*/
|
|
46
|
+
export declare function readRunStartApproval(root: string, runId: string): {
|
|
47
|
+
approved: boolean;
|
|
48
|
+
artifact: string;
|
|
49
|
+
reason: string;
|
|
50
|
+
};
|
|
51
|
+
/**
|
|
52
|
+
* The ONE refusal reason the projection returns before approval, exported so every caller branches on
|
|
53
|
+
* the same string instead of re-typing it (a re-typed reason is a caller that silently stops matching).
|
|
54
|
+
*/
|
|
55
|
+
export declare const RUN_START_NOT_APPROVED = "run not started: phase 0 approval has not been granted";
|
package/lib/runtime.d.ts
CHANGED
|
@@ -15,7 +15,7 @@ import { type SubagentsRuntimeLike, type SubagentStartRequestLike, type Subagent
|
|
|
15
15
|
import { type RecursivePhaseState } from './lifecycle.ts';
|
|
16
16
|
import { type EnforcementConfig, type ToolGuardDecision, type ToolExecLike } from './enforcement.ts';
|
|
17
17
|
import { type CreateWorktreeResult, type PromoteBranchResult } from './worktree.ts';
|
|
18
|
-
import { syncRunGoal, type GoalServiceLike } from './goals-projection.ts';
|
|
18
|
+
import { syncRunGoal, type GoalServiceLike, type SyncResult } from './goals-projection.ts';
|
|
19
19
|
import { type TeamRuntimeLike, type AuditToPassResult, type TeamCallerHandle, type TeamTaskViewLike, type AuditRoundOutcome } from './teams-loop.ts';
|
|
20
20
|
import type { ContinuableChildId, ContinuableMessageId } from './delegation.ts';
|
|
21
21
|
declare module '@deepseek-ai/cordis' {
|
|
@@ -39,6 +39,41 @@ export interface LintArtifactResult {
|
|
|
39
39
|
warnings: string[];
|
|
40
40
|
passed: boolean;
|
|
41
41
|
}
|
|
42
|
+
/**
|
|
43
|
+
* PHASE 0 — the structural seam for the host's human-question channel (`ctx.userQuestions`).
|
|
44
|
+
*
|
|
45
|
+
* Declared here as a minimal seam for the same reason as every other harness touchpoint in this plugin:
|
|
46
|
+
* the live `UserQuestionService` satisfies it structurally, so the plugin never imports the host package,
|
|
47
|
+
* and a test can drive the run-start gate with a fake that behaves like the real one. The error case is
|
|
48
|
+
* part of the contract, not an afterthought: the real `ask()` REJECTS (NO_PROVIDER / CALLER_NOT_LIVE /
|
|
49
|
+
* ASK_ABORTED) instead of resolving with something that could be mistaken for an answer, which is what
|
|
50
|
+
* lets `recursive_ask` fail closed rather than invent an approval.
|
|
51
|
+
*/
|
|
52
|
+
export interface UserQuestionsLike {
|
|
53
|
+
ask(request: {
|
|
54
|
+
questions: Array<{
|
|
55
|
+
id: string;
|
|
56
|
+
header?: string;
|
|
57
|
+
question: string;
|
|
58
|
+
options?: Array<{
|
|
59
|
+
label: string;
|
|
60
|
+
description?: string;
|
|
61
|
+
}>;
|
|
62
|
+
}>;
|
|
63
|
+
agent?: unknown;
|
|
64
|
+
signal?: AbortSignal;
|
|
65
|
+
/** Links the card to the tool call that asked, the way plan-mode's exit does. */
|
|
66
|
+
wait?: {
|
|
67
|
+
callId?: unknown;
|
|
68
|
+
};
|
|
69
|
+
}): Promise<{
|
|
70
|
+
answers: Array<{
|
|
71
|
+
id: string;
|
|
72
|
+
selected: string[];
|
|
73
|
+
custom?: string;
|
|
74
|
+
}>;
|
|
75
|
+
}>;
|
|
76
|
+
}
|
|
42
77
|
/**
|
|
43
78
|
* T15 (G): the folded status PLUS the rolling guard-decision evidence. Declared
|
|
44
79
|
* as an intersection rather than by editing RecursiveStatusResult/foldRun — the
|
|
@@ -79,8 +114,20 @@ export declare class RecursiveRuntime extends Service {
|
|
|
79
114
|
subagents?: SubagentsRuntimeLike | null;
|
|
80
115
|
workflow?: WorkflowEngineLike | null;
|
|
81
116
|
});
|
|
82
|
-
/**
|
|
117
|
+
/**
|
|
118
|
+
* T10: the native jobs registry, when the composition mounts one. */
|
|
83
119
|
private readonly jobs;
|
|
120
|
+
/**
|
|
121
|
+
* PHASE 0 — attach the goals service after construction.
|
|
122
|
+
*
|
|
123
|
+
* The composition resolves `goals` with ONE `ctx.get` at apply time and passes it to the constructor,
|
|
124
|
+
* which is fine for a service that is already mounted. This seam exists for the two cases that pattern
|
|
125
|
+
* cannot cover: a composition that mounts `goals` later (the same late-attach reason `attachSubagents`
|
|
126
|
+
* and `attachLlmInventory` exist), and a test that needs the REAL runtime wired to a structural fake —
|
|
127
|
+
* a fake passed through the plugin's Config is dropped, because the Config schema is the settings
|
|
128
|
+
* namespace and strips keys it does not declare.
|
|
129
|
+
*/
|
|
130
|
+
attachGoals(service: GoalServiceLike | null): void;
|
|
84
131
|
/**
|
|
85
132
|
* T23 — write a gate's answer into an artifact as a marker line.
|
|
86
133
|
*
|
|
@@ -160,7 +207,7 @@ export declare class RecursiveRuntime extends Service {
|
|
|
160
207
|
knownProviderNames(): string[];
|
|
161
208
|
private readonly repoRoot;
|
|
162
209
|
private readonly workspaceRegistry;
|
|
163
|
-
private
|
|
210
|
+
private goalsService;
|
|
164
211
|
/**
|
|
165
212
|
* T27 — the hook registry, EXPOSED so a sibling plugin can participate in a run
|
|
166
213
|
* without patching this one:
|
|
@@ -216,10 +263,49 @@ export declare class RecursiveRuntime extends Service {
|
|
|
216
263
|
history?: string;
|
|
217
264
|
lock?: LockArtifactResult;
|
|
218
265
|
}>;
|
|
266
|
+
/**
|
|
267
|
+
* PHASE 0 — read a run's start approval from its own Phase 0 artifact.
|
|
268
|
+
*
|
|
269
|
+
* The approval is a DURABLE line in `.recursive/run/<runId>/00-requirements.md`, not a value held in
|
|
270
|
+
* memory, for the reason every other gate here is durable: a decision that only exists in a session
|
|
271
|
+
* cannot be cited, and cannot survive the session it was made in. Read-only; asking changes nothing.
|
|
272
|
+
*/
|
|
273
|
+
readRunStartApproval(root: string, runId: string): {
|
|
274
|
+
approved: boolean;
|
|
275
|
+
artifact: string;
|
|
276
|
+
reason: string;
|
|
277
|
+
};
|
|
278
|
+
/**
|
|
279
|
+
* PHASE 0 — the harness's blocking human-question channel (`ctx.userQuestions`), when this composition
|
|
280
|
+
* mounts one.
|
|
281
|
+
*
|
|
282
|
+
* ⚠ WHY THE PLUGIN REACHES FOR THIS AT ALL. The other three gates answer through a question card and a
|
|
283
|
+
* relayed label, which is fine for a decision the workflow acts on later. STARTING A RUN is different:
|
|
284
|
+
* the first human turn is the only place the harness can say "arming this goal means autonomous rounds"
|
|
285
|
+
* BEFORE arming it. This channel is the same one plan-mode's exit uses; `ask()` resolves only with a
|
|
286
|
+
* real answer from a real person, and it THROWS when there is no answerer or no live root agent. So the
|
|
287
|
+
* absence of this service cannot be papered over: `recursive_ask` refuses the run-start gate and names
|
|
288
|
+
* the missing channel (RM5502).
|
|
289
|
+
*/
|
|
290
|
+
private userQuestions;
|
|
291
|
+
/** Late-bind the human-question channel when the composition mounts it. */
|
|
292
|
+
attachUserQuestions(service: UserQuestionsLike | null): void;
|
|
293
|
+
/** The human-question channel this composition mounted, or null. */
|
|
294
|
+
get userQuestionsChannel(): UserQuestionsLike | null;
|
|
219
295
|
/**
|
|
220
296
|
* T1 (goals projection): project the run into the native goals service so it is
|
|
221
297
|
* a first-class durable, resumable, blockable object. Best-effort — the run's
|
|
222
298
|
* filesystem state is the source of truth; a goal is the durable projection.
|
|
299
|
+
*
|
|
300
|
+
* ⚠ PHASE 0 — AND IT ARMS NOTHING UNLESS THE RUN WAS STARTED. `create` returns an ARMED goal, and an
|
|
301
|
+
* armed goal is the harness driving autonomous rounds, so this is the one place where "project the
|
|
302
|
+
* state" can quietly equal "start the run". The approval is therefore REQUIRED from the caller and
|
|
303
|
+
* has no default here: a caller that has not resolved the run's approval cannot arm a goal by
|
|
304
|
+
* forgetting to pass one, and `syncRunGoal` refuses every branch that would create one without it.
|
|
305
|
+
*
|
|
306
|
+
* Unapproved is the EXPECTED state for a scaffolded run, so the refusal comes back as a plain
|
|
307
|
+
* `{ ok: false }` carrying {@link RUN_START_NOT_APPROVED}: callers must treat that as normal work,
|
|
308
|
+
* never as a warning (see `armRunGoalIfApproved`, the one caller that arms).
|
|
223
309
|
*/
|
|
224
310
|
projectRunToGoal(agent: {
|
|
225
311
|
session?: {
|
|
@@ -227,15 +313,21 @@ export declare class RecursiveRuntime extends Service {
|
|
|
227
313
|
cwd?: string;
|
|
228
314
|
};
|
|
229
315
|
};
|
|
230
|
-
} | null | undefined, runId: string, state?: Parameters<typeof syncRunGoal>[3]):
|
|
231
|
-
|
|
232
|
-
|
|
233
|
-
|
|
234
|
-
|
|
235
|
-
|
|
236
|
-
|
|
237
|
-
|
|
238
|
-
|
|
316
|
+
} | null | undefined, runId: string, state?: Parameters<typeof syncRunGoal>[3], approved?: boolean): SyncResult;
|
|
317
|
+
/**
|
|
318
|
+
* PHASE 0 — the approved-run arm step: read the run's approval from `root` and project the goal only
|
|
319
|
+
* if it is there. This is the phase-progress path (`syncRunGoal` reached on ordinary work), so the
|
|
320
|
+
* unapproved case is deliberately silent: `{ ok: false, reason: RUN_START_NOT_APPROVED }` with no
|
|
321
|
+
* goal, no write and no throw. Read on EVERY call rather than cached, because the approval can arrive
|
|
322
|
+
* mid-session and a cached "not yet" would leave an approved run unable to arm until a plugin reload.
|
|
323
|
+
*/
|
|
324
|
+
armRunGoalIfApproved(agent: {
|
|
325
|
+
session?: {
|
|
326
|
+
header?: {
|
|
327
|
+
cwd?: string;
|
|
328
|
+
};
|
|
329
|
+
};
|
|
330
|
+
} | null | undefined, root: string, runId: string, state?: Parameters<typeof syncRunGoal>[3]): SyncResult;
|
|
239
331
|
/** T1: block the run's goal on a gate-block (durable + UI-visible). */
|
|
240
332
|
blockRunToGoal(agent: {
|
|
241
333
|
session?: {
|
|
@@ -246,31 +338,20 @@ export declare class RecursiveRuntime extends Service {
|
|
|
246
338
|
} | null | undefined, runId: string, reason: {
|
|
247
339
|
code: string;
|
|
248
340
|
message: string;
|
|
249
|
-
}):
|
|
250
|
-
|
|
251
|
-
|
|
252
|
-
|
|
253
|
-
|
|
254
|
-
|
|
255
|
-
|
|
256
|
-
reason: string;
|
|
257
|
-
};
|
|
258
|
-
/** T1: re-arm the run's goal on a reopen (blocked/paused -> active). */
|
|
341
|
+
}): SyncResult;
|
|
342
|
+
/**
|
|
343
|
+
* T1: re-arm the run's goal on a reopen (blocked/paused -> active). Never starts an unstarted run —
|
|
344
|
+
* REOPEN IS NOT A BACK DOOR TO STARTING A RUN. `approved` is required for the same reason as in
|
|
345
|
+
* `projectRunToGoal`: the phase-0 gate cannot be defaulted open. An approved run's approval outlives
|
|
346
|
+
* a reopen because it is a durable line in the run's own Phase 0 artifact, not a held value.
|
|
347
|
+
*/
|
|
259
348
|
resumeRunToGoal(agent: {
|
|
260
349
|
session?: {
|
|
261
350
|
header?: {
|
|
262
351
|
cwd?: string;
|
|
263
352
|
};
|
|
264
353
|
};
|
|
265
|
-
} | null | undefined, runId: string):
|
|
266
|
-
ok: true;
|
|
267
|
-
phase: import("./goals-projection.ts").GoalPhase;
|
|
268
|
-
ref?: import("./goals-projection.ts").GoalRefLike;
|
|
269
|
-
created?: boolean;
|
|
270
|
-
} | {
|
|
271
|
-
ok: boolean;
|
|
272
|
-
reason: string;
|
|
273
|
-
};
|
|
354
|
+
} | null | undefined, runId: string, approved?: boolean): SyncResult;
|
|
274
355
|
/**
|
|
275
356
|
* Workspace-scoped control-plane root (R1 binding invariant).
|
|
276
357
|
* Resolves the session agent's canonical cwd -> workspace path via the
|
|
@@ -571,6 +652,31 @@ export declare class RecursiveRuntime extends Service {
|
|
|
571
652
|
existing: string[];
|
|
572
653
|
worktree?: CreateWorktreeResult;
|
|
573
654
|
}>;
|
|
655
|
+
/**
|
|
656
|
+
* PHASE 0 — THE APPROVAL ACT: record the human's `Start run` decision and arm the run's goal.
|
|
657
|
+
*
|
|
658
|
+
* ⚠ THE ONLY PATH THAT STARTS A RUN. It exists as one method rather than as "write a line, then
|
|
659
|
+
* project the goal" at the tool, because those two steps must not be separable: an approval recorded
|
|
660
|
+
* without the arm (or an arm without the record) is exactly the half-state that made this defect hard
|
|
661
|
+
* to see. `tests/run-start-approval.spec.ts` drives both halves through this one call.
|
|
662
|
+
*
|
|
663
|
+
* The approval line goes into the run's own Phase 0 artifact, so it is durable, citable, and survives
|
|
664
|
+
* the session — and so a reader of the run can answer "was this run started, and by what?" without the
|
|
665
|
+
* transcript. `answer` is validated against the gate's own labels before it reaches here.
|
|
666
|
+
*/
|
|
667
|
+
approveRunStart(root: string, runId: string, agent?: {
|
|
668
|
+
session?: {
|
|
669
|
+
header?: {
|
|
670
|
+
cwd?: string;
|
|
671
|
+
};
|
|
672
|
+
};
|
|
673
|
+
} | null, answer?: string): {
|
|
674
|
+
ok: boolean;
|
|
675
|
+
reason: string;
|
|
676
|
+
path: string;
|
|
677
|
+
replaced: boolean;
|
|
678
|
+
goal: SyncResult;
|
|
679
|
+
};
|
|
574
680
|
/**
|
|
575
681
|
* Create a linked worktree for a run under the given workspace root. The
|
|
576
682
|
* worktree branch defaults to `recursive/<runId>` and is cut from the given
|
package/package.json
CHANGED
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@try-works/dsh-recursive-mode",
|
|
3
3
|
"description": "recursive-mode workflow as a DeepSeek Harness bundle: RecursiveRuntime service + 13 recursive_* tools (recursive_status, recursive_init, recursive_lock, recursive_lint, recursive_closeout, recursive_scratch, recursive_worktree, recursive_phase, recursive_audit_team, recursive_review, recursive_delegate, recursive_ask, recursive_preview)",
|
|
4
|
-
"version": "0.4.
|
|
4
|
+
"version": "0.4.5",
|
|
5
5
|
"type": "module",
|
|
6
6
|
"main": "lib/index.js",
|
|
7
7
|
"types": "lib/index.d.ts",
|
|
@@ -185,7 +185,11 @@ async function main() {
|
|
|
185
185
|
complete: (_agent, ref) => { goalCurrent = { ...(goalCurrent ?? { id: ref.id, revision: ref.revision }), phase: 'complete', revision: ref.revision + 1 }; return goalCurrent },
|
|
186
186
|
clear: (_agent, ref) => { goalCurrent = undefined; return { id: ref.id, revision: ref.revision + 1 } },
|
|
187
187
|
}
|
|
188
|
-
|
|
188
|
+
// PHASE 0: scaffolding a run must not start it. Without an approval the projection creates NOTHING —
|
|
189
|
+
// this is the defect the gate closes, so the smoke script proves the refusal as well as the arm.
|
|
190
|
+
const unapproved = syncRunGoal(goalService, {}, '10-smoke', 'active')
|
|
191
|
+
check('PHASE 0 no goal without approval', unapproved.ok === false && goalCurrent === undefined)
|
|
192
|
+
const goalSync = syncRunGoal(goalService, {}, '10-smoke', 'active', true)
|
|
189
193
|
check('T1 run goal armed', goalSync.ok === true && goalCurrent?.objective === 'recursive-run:10-smoke · active')
|
|
190
194
|
const goalBlock = blockRunGoal(goalService, {}, '10-smoke', { code: 'prerequisite-blockers', message: 'monotonic lock-order' })
|
|
191
195
|
check('T1 gate-block blocks the run goal', goalBlock.ok === true && goalCurrent?.phase === 'blocked')
|
package/src/errors.ts
CHANGED
|
@@ -147,6 +147,18 @@ export const TOOL_ERRORS = {
|
|
|
147
147
|
|
|
148
148
|
/* 5xxx — the runtime refused an operation it understands. */
|
|
149
149
|
|
|
150
|
+
RUN_START_NO_CHANNEL: {
|
|
151
|
+
code: 'RM5502',
|
|
152
|
+
klass: 'runtime',
|
|
153
|
+
problem: 'the run-start gate needs an answer, and this composition mounts no user-questions channel to ask one directly',
|
|
154
|
+
next: 'call recursive_ask with gate: run-start and no answer to surface the question, then retry with answer: ' + '"Start run"',
|
|
155
|
+
},
|
|
156
|
+
RUN_START_UNANSWERED: {
|
|
157
|
+
code: 'RM5503',
|
|
158
|
+
klass: 'runtime',
|
|
159
|
+
problem: 'the user-questions channel mounted in this composition refused the run-start question, so no person was asked',
|
|
160
|
+
next: 'use recursive_ask without an answer to surface the question, and retry it with answer: ' + '"Start run" once the user has approved the run start',
|
|
161
|
+
},
|
|
150
162
|
RUNTIME_REFUSED: {
|
|
151
163
|
code: 'RM5501',
|
|
152
164
|
klass: 'runtime',
|
package/src/goals-projection.ts
CHANGED
|
@@ -15,6 +15,7 @@
|
|
|
15
15
|
* completed goal may be replaced, per the service contract).
|
|
16
16
|
*/
|
|
17
17
|
import type { RunState } from './lifecycle.ts'
|
|
18
|
+
import { RUN_START_NOT_APPROVED } from './run-start.ts'
|
|
18
19
|
|
|
19
20
|
/** Native goal phase (mirrors @deepseek-ai/dsh-goal GoalPhase). */
|
|
20
21
|
export type GoalPhase = 'active' | 'paused' | 'blocked' | 'complete'
|
|
@@ -97,8 +98,31 @@ function mutatePhase(service: GoalServiceLike, agent: AgentHandle, ref: GoalRefL
|
|
|
97
98
|
* Sync a run's durable goal to the requested phase. Safe: never touches a goal
|
|
98
99
|
* whose objective is not this run's marker, and never re-creates over a
|
|
99
100
|
* non-complete foreign goal.
|
|
101
|
+
*
|
|
102
|
+
* ⚠ `approved` IS THE PHASE-0 GATE, and it defaults to the SAFE direction. A goal is not a label:
|
|
103
|
+
* `create` returns an ARMED view and the harness starts driving autonomous goal rounds for the
|
|
104
|
+
* session, so creating one is starting the run. The owner's rule is that phase 0 requires explicit
|
|
105
|
+
* approval, which means the projection must be unable to arm anything on its own — hence a default of
|
|
106
|
+
* `false` and an explicit refusal in EVERY branch that would call `create`, including the two
|
|
107
|
+
* replace-a-completed-goal branches (an unapproved run cannot have reached `complete`, but "cannot
|
|
108
|
+
* happen" is what the single unguarded branch relied on too).
|
|
109
|
+
*
|
|
110
|
+
* ⚠ AND IT IS REACHED ON ORDINARY WORK, so the unapproved path is QUIET AND IDEMPOTENT: no goal is
|
|
111
|
+
* created, nothing is written, no error is thrown, and the run's artifacts are untouched. The caller
|
|
112
|
+
* reads {@link RUN_START_NOT_APPROVED} to tell "this run has not been started yet" apart from a real
|
|
113
|
+
* failure, so a normal phase step never surfaces a warning.
|
|
114
|
+
*
|
|
115
|
+
* `approved` is passed IN rather than read here because this module is pure: it takes the goal service
|
|
116
|
+
* seam and nothing else, and the plugin's own filesystem reads live in the runtime (see
|
|
117
|
+
* `RecursiveRuntime.readRunStartApproval`).
|
|
100
118
|
*/
|
|
101
|
-
export function syncRunGoal(
|
|
119
|
+
export function syncRunGoal(
|
|
120
|
+
service: GoalServiceLike | undefined | null,
|
|
121
|
+
agent: AgentHandle,
|
|
122
|
+
runId: string,
|
|
123
|
+
runState: RunState,
|
|
124
|
+
approved = false,
|
|
125
|
+
): SyncResult {
|
|
102
126
|
if (!service) return { ok: false, reason: 'no goals service' }
|
|
103
127
|
const target = RUN_TO_GOAL_PHASE[runState]
|
|
104
128
|
const current = service.get(agent)
|
|
@@ -110,6 +134,7 @@ export function syncRunGoal(service: GoalServiceLike | undefined | null, agent:
|
|
|
110
134
|
if (phase === target) return { ok: true, phase: target, ref }
|
|
111
135
|
// A completed goal is final: the contract allows it to be REPLACED, not resumed.
|
|
112
136
|
if (phase === 'complete') {
|
|
137
|
+
if (!approved) return { ok: false, reason: RUN_START_NOT_APPROVED }
|
|
113
138
|
const created = service.create(agent, { objective: goalObjective(runId, runState) })
|
|
114
139
|
return { ok: true, phase: target, ref: refOf(created), created: true }
|
|
115
140
|
}
|
|
@@ -121,29 +146,45 @@ export function syncRunGoal(service: GoalServiceLike | undefined | null, agent:
|
|
|
121
146
|
// cleared or resumed instead. Never clobber a foreign goal.
|
|
122
147
|
if (current) {
|
|
123
148
|
if (current.phase === 'complete') {
|
|
149
|
+
if (!approved) return { ok: false, reason: RUN_START_NOT_APPROVED }
|
|
124
150
|
const created = service.create(agent, { objective: goalObjective(runId, runState) })
|
|
125
151
|
return { ok: true, phase: target, ref: refOf(created), created: true }
|
|
126
152
|
}
|
|
127
153
|
return { ok: false, reason: 'a non-matching active goal exists (foreign goal not touched)' }
|
|
128
154
|
}
|
|
129
155
|
|
|
130
|
-
// 3. No current goal
|
|
156
|
+
// 3. No current goal. This is where the defect lived: scaffolding a run armed it. A run with no
|
|
157
|
+
// phase-0 approval stays goal-less — the spec exists, the run does not.
|
|
158
|
+
if (!approved) return { ok: false, reason: RUN_START_NOT_APPROVED }
|
|
131
159
|
const created = service.create(agent, { objective: goalObjective(runId, runState) })
|
|
132
160
|
return { ok: true, phase: target, ref: refOf(created), created: true }
|
|
133
161
|
}
|
|
134
162
|
|
|
135
|
-
/**
|
|
163
|
+
/**
|
|
164
|
+
* Block the current run goal (used on a gate-block). Never touches a foreign goal.
|
|
165
|
+
*
|
|
166
|
+
* ⚠ A RUN THAT WAS NEVER STARTED HAS NO GOAL TO BLOCK, so this reports the unapproved state in the
|
|
167
|
+
* same words as {@link syncRunGoal} rather than "no current goal to block": the caller's question is
|
|
168
|
+
* "why is there no goal", and the answer must not depend on which entry point happened to ask.
|
|
169
|
+
*/
|
|
136
170
|
export function blockRunGoal(service: GoalServiceLike | undefined | null, agent: AgentHandle, runId: string, reason: { code: string; message: string }): SyncResult {
|
|
137
171
|
if (!service) return { ok: false, reason: 'no goals service' }
|
|
138
172
|
const current = service.get(agent)
|
|
139
|
-
if (!current) return { ok: false, reason:
|
|
173
|
+
if (!current) return { ok: false, reason: RUN_START_NOT_APPROVED }
|
|
140
174
|
if (!isRunGoal(current, runId)) return { ok: false, reason: 'current goal is not for this run (foreign goal not touched)' }
|
|
141
175
|
const ref = refOf(current)
|
|
142
176
|
const ok = !!service.block(agent, ref, reason)
|
|
143
177
|
return ok ? { ok: true, phase: 'blocked', ref } : { ok: false, reason: 'goal block failed' }
|
|
144
178
|
}
|
|
145
179
|
|
|
146
|
-
/**
|
|
147
|
-
|
|
148
|
-
|
|
180
|
+
/**
|
|
181
|
+
* Bridge a run's blocked goal back to active (used on a reopen).
|
|
182
|
+
*
|
|
183
|
+
* ⚠ REOPEN IS NOT A BACK DOOR TO STARTING A RUN. It routes through {@link syncRunGoal}, so a reopen of
|
|
184
|
+
* an unapproved run cannot create the goal that init deliberately withheld. An APPROVED run is
|
|
185
|
+
* unaffected: its approval outlives the reopen, because the approval is a durable line in the run's
|
|
186
|
+
* own Phase 0 artifact rather than a value held in memory (verified in `tests/run-start-approval.spec.ts`).
|
|
187
|
+
*/
|
|
188
|
+
export function resumeRunGoal(service: GoalServiceLike | undefined | null, agent: AgentHandle, runId: string, approved = false): SyncResult {
|
|
189
|
+
return syncRunGoal(service, agent, runId, 'active', approved)
|
|
149
190
|
}
|
package/src/index.ts
CHANGED
|
@@ -4,6 +4,7 @@ import type { ContextFormed } from '@deepseek-ai/dsh-llm'
|
|
|
4
4
|
import { existsSync } from 'node:fs'
|
|
5
5
|
import { join } from 'node:path'
|
|
6
6
|
import { RecursiveRuntime } from './runtime.ts'
|
|
7
|
+
import type { UserQuestionsLike } from './runtime.ts'
|
|
7
8
|
import type { JobsRegistryLike } from './jobs-runner.ts'
|
|
8
9
|
import { planGateForExit } from './plan-gate.ts'
|
|
9
10
|
import { registerPhaseSkills, type SkillRegistryLike } from './skills-phase.ts'
|
|
@@ -237,6 +238,14 @@ export function apply(ctx: Context, config?: RecursiveModeConfig) {
|
|
|
237
238
|
ctx.inject(['llm'], (llmCtx: Context) => {
|
|
238
239
|
recursive.attachLlmInventory((llmCtx.get('llm') as LlmInventoryLike | undefined) ?? null)
|
|
239
240
|
})
|
|
241
|
+
// ⚠ PHASE 0 — THE HUMAN-QUESTION CHANNEL, resolved the same late-attaching way for the same reason:
|
|
242
|
+
// a composition may mount `userQuestions` after this plugin applies, and the run-start gate asks the
|
|
243
|
+
// person DIRECTLY through it before it arms a goal — which is what keeps a relayed `Start run` from
|
|
244
|
+
// being an approval while a person can actually be asked (see run-start.ts).
|
|
245
|
+
recursive.attachUserQuestions((ctx.get('userQuestions') as UserQuestionsLike | undefined) ?? null)
|
|
246
|
+
ctx.inject(['userQuestions'], (questionsCtx: Context) => {
|
|
247
|
+
recursive.attachUserQuestions((questionsCtx.get('userQuestions') as UserQuestionsLike | undefined) ?? null)
|
|
248
|
+
})
|
|
240
249
|
|
|
241
250
|
// T7 — THE SETTINGS NAMESPACE, APPLIED ON EVERY APPLY. The settings service edits the
|
|
242
251
|
// Loader entry's config and the Loader RE-APPLIES this plugin, so a toggle in the UI
|