@stigmer/runner 3.1.2 → 3.1.4
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/dist/.build-fingerprint +1 -1
- package/dist/activities/execute-cursor/attachment-resolver.d.ts +1 -1
- package/dist/activities/execute-cursor/attachment-resolver.js +1 -1
- package/dist/activities/execute-cursor/capture-flow.d.ts +29 -0
- package/dist/activities/execute-cursor/capture-flow.js +59 -0
- package/dist/activities/execute-cursor/capture-flow.js.map +1 -1
- package/dist/activities/execute-cursor/command-provenance.d.ts +11 -25
- package/dist/activities/execute-cursor/command-provenance.js +25 -115
- package/dist/activities/execute-cursor/command-provenance.js.map +1 -1
- package/dist/activities/execute-cursor/index.js +358 -557
- package/dist/activities/execute-cursor/index.js.map +1 -1
- package/dist/activities/execute-cursor/todo-tracker.d.ts +6 -1
- package/dist/activities/execute-cursor/todo-tracker.js +15 -43
- package/dist/activities/execute-cursor/todo-tracker.js.map +1 -1
- package/dist/activities/execute-cursor/turn-boundary.d.ts +108 -0
- package/dist/activities/execute-cursor/turn-boundary.js +163 -0
- package/dist/activities/execute-cursor/turn-boundary.js.map +1 -0
- package/dist/activities/execute-cursor/turn-stream.d.ts +141 -0
- package/dist/activities/execute-cursor/turn-stream.js +249 -0
- package/dist/activities/execute-cursor/turn-stream.js.map +1 -0
- package/dist/activities/execute-deep-agent/command-provenance.d.ts +61 -0
- package/dist/activities/execute-deep-agent/command-provenance.js +72 -0
- package/dist/activities/execute-deep-agent/command-provenance.js.map +1 -0
- package/dist/activities/execute-deep-agent/index.js +73 -4
- package/dist/activities/execute-deep-agent/index.js.map +1 -1
- package/dist/activities/execute-deep-agent/status-builder.js +8 -1
- package/dist/activities/execute-deep-agent/status-builder.js.map +1 -1
- package/dist/activities/execute-deep-agent/streaming-v3.js +6 -1
- package/dist/activities/execute-deep-agent/streaming-v3.js.map +1 -1
- package/dist/activities/execute-deep-agent/streaming.d.ts +7 -0
- package/dist/activities/execute-deep-agent/streaming.js +5 -1
- package/dist/activities/execute-deep-agent/streaming.js.map +1 -1
- package/dist/activities/execute-deep-agent/v3-status-builder.js +12 -1
- package/dist/activities/execute-deep-agent/v3-status-builder.js.map +1 -1
- package/dist/otel.js +10 -0
- package/dist/otel.js.map +1 -1
- package/dist/shared/filereview/capture.d.ts +18 -1
- package/dist/shared/filereview/capture.js +105 -28
- package/dist/shared/filereview/capture.js.map +1 -1
- package/dist/shared/filereview/cas-progress.d.ts +63 -0
- package/dist/shared/filereview/cas-progress.js +128 -0
- package/dist/shared/filereview/cas-progress.js.map +1 -0
- package/dist/shared/filereview/cas-substrate.d.ts +38 -0
- package/dist/shared/filereview/cas-substrate.js +48 -15
- package/dist/shared/filereview/cas-substrate.js.map +1 -1
- package/dist/shared/filereview/command-provenance.d.ts +93 -0
- package/dist/shared/filereview/command-provenance.js +132 -0
- package/dist/shared/filereview/command-provenance.js.map +1 -0
- package/dist/shared/filereview/events.d.ts +16 -0
- package/dist/shared/filereview/events.js +37 -0
- package/dist/shared/filereview/events.js.map +1 -1
- package/dist/shared/filereview/git-substrate.d.ts +48 -0
- package/dist/shared/filereview/git-substrate.js +78 -0
- package/dist/shared/filereview/git-substrate.js.map +1 -1
- package/dist/shared/filereview/index.d.ts +5 -2
- package/dist/shared/filereview/index.js +5 -2
- package/dist/shared/filereview/index.js.map +1 -1
- package/dist/shared/filereview/line-counts.d.ts +48 -0
- package/dist/shared/filereview/line-counts.js +72 -0
- package/dist/shared/filereview/line-counts.js.map +1 -0
- package/dist/shared/filereview/progress.d.ts +167 -0
- package/dist/shared/filereview/progress.js +196 -0
- package/dist/shared/filereview/progress.js.map +1 -0
- package/dist/shared/implement-plan-prompt.d.ts +6 -4
- package/dist/shared/implement-plan-prompt.js +38 -6
- package/dist/shared/implement-plan-prompt.js.map +1 -1
- package/dist/shared/plan-artifact.d.ts +59 -11
- package/dist/shared/plan-artifact.js +144 -16
- package/dist/shared/plan-artifact.js.map +1 -1
- package/dist/shared/plan-mode-prompt.d.ts +15 -4
- package/dist/shared/plan-mode-prompt.js +28 -5
- package/dist/shared/plan-mode-prompt.js.map +1 -1
- package/dist/shared/todos.d.ts +56 -0
- package/dist/shared/todos.js +98 -0
- package/dist/shared/todos.js.map +1 -0
- package/dist/shared/tool-row.d.ts +16 -0
- package/dist/shared/tool-row.js +31 -0
- package/dist/shared/tool-row.js.map +1 -1
- package/package.json +3 -2
- package/src/__tests__/otel-turn-span.test.ts +61 -0
- package/src/activities/execute-cursor/__tests__/build-prompt.test.ts +7 -0
- package/src/activities/execute-cursor/__tests__/delta-enricher.test.ts +68 -0
- package/src/activities/execute-cursor/__tests__/progress-substrate.test.ts +169 -0
- package/src/activities/execute-cursor/__tests__/turn-boundary.test.ts +280 -0
- package/src/activities/execute-cursor/__tests__/turn-stream.test.ts +349 -0
- package/src/activities/execute-cursor/attachment-resolver.ts +1 -1
- package/src/activities/execute-cursor/capture-flow.ts +78 -0
- package/src/activities/execute-cursor/command-provenance.ts +25 -120
- package/src/activities/execute-cursor/index.ts +419 -602
- package/src/activities/execute-cursor/todo-tracker.ts +17 -59
- package/src/activities/execute-cursor/turn-boundary.ts +274 -0
- package/src/activities/execute-cursor/turn-stream.ts +418 -0
- package/src/activities/execute-deep-agent/__tests__/command-provenance.test.ts +252 -0
- package/src/activities/execute-deep-agent/__tests__/prompt-builder.test.ts +11 -0
- package/src/activities/execute-deep-agent/__tests__/status-builder.test.ts +78 -0
- package/src/activities/execute-deep-agent/__tests__/v3-status-builder.test.ts +105 -1
- package/src/activities/execute-deep-agent/command-provenance.ts +102 -0
- package/src/activities/execute-deep-agent/index.ts +91 -4
- package/src/activities/execute-deep-agent/status-builder.ts +9 -0
- package/src/activities/execute-deep-agent/streaming-v3.ts +6 -1
- package/src/activities/execute-deep-agent/streaming.ts +12 -1
- package/src/activities/execute-deep-agent/v3-status-builder.ts +13 -0
- package/src/otel.ts +8 -0
- package/src/shared/__tests__/implement-plan-prompt.test.ts +40 -2
- package/src/shared/__tests__/plan-artifact.test.ts +142 -9
- package/src/shared/__tests__/todos.test.ts +216 -0
- package/src/shared/filereview/__tests__/capture.test.ts +182 -1
- package/src/shared/filereview/__tests__/cas-progress.test.ts +228 -0
- package/src/shared/filereview/__tests__/cas-substrate.test.ts +100 -0
- package/src/shared/filereview/__tests__/command-provenance.test.ts +252 -0
- package/src/shared/filereview/__tests__/events.test.ts +105 -0
- package/src/shared/filereview/__tests__/line-counts.test.ts +73 -0
- package/src/shared/filereview/__tests__/progress.test.ts +338 -0
- package/src/shared/filereview/capture.ts +134 -30
- package/src/shared/filereview/cas-progress.ts +170 -0
- package/src/shared/filereview/cas-substrate.ts +79 -18
- package/src/shared/filereview/command-provenance.ts +180 -0
- package/src/shared/filereview/events.ts +50 -0
- package/src/shared/filereview/git-substrate.ts +114 -0
- package/src/shared/filereview/index.ts +36 -1
- package/src/shared/filereview/line-counts.ts +83 -0
- package/src/shared/filereview/progress.ts +310 -0
- package/src/shared/implement-plan-prompt.ts +39 -8
- package/src/shared/plan-artifact.ts +152 -16
- package/src/shared/plan-mode-prompt.ts +28 -5
- package/src/shared/todos.ts +126 -0
- package/src/shared/tool-row.ts +34 -0
|
@@ -0,0 +1,163 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* The Cursor harness's turn boundary — the single post-run pipeline that turns
|
|
3
|
+
* a finished agent run into the durable review surfaces:
|
|
4
|
+
*
|
|
5
|
+
* 1. read the denial ledger the preToolUse hook appended this turn;
|
|
6
|
+
* 2. derive the approved-command provenance (DD-28 auto-keep facts);
|
|
7
|
+
* 3. capture the turn's net file change set to the file_review ledger
|
|
8
|
+
* (CANDIDATE_CAPTURED) and stamp the flowed edit rows;
|
|
9
|
+
* 4. reconcile denied tool calls to WAITING_APPROVAL gate rows and redact the
|
|
10
|
+
* model's provisional post-denial narration.
|
|
11
|
+
*
|
|
12
|
+
* Extracted from the activity entry point (index.ts Phase 12) so it is directly
|
|
13
|
+
* unit-testable AND re-enterable: the poisoned-handle / transport-timeout
|
|
14
|
+
* recoveries re-run the agent with a fresh handle AFTER the primary boundary
|
|
15
|
+
* already ran, so their edits must flow through this exact pipeline again or
|
|
16
|
+
* they silently escape review (production case aex_01kws27q1e2esvkqjpvectttxf,
|
|
17
|
+
* where a Build-from-plan retry created a file with no review gate).
|
|
18
|
+
*
|
|
19
|
+
* Re-entry is safe by construction: a retry is only reachable when the primary
|
|
20
|
+
* boundary captured nothing (a captured change pauses the turn before
|
|
21
|
+
* run.wait() is ever consulted), `stampFlowedFileEditRows` skips already-stamped
|
|
22
|
+
* rows, and the denial ledger is per-turn append-only.
|
|
23
|
+
*
|
|
24
|
+
* The caller owns everything around the boundary: the stream epilogue
|
|
25
|
+
* (accumulator/enricher finalize), the WAITING_FOR_APPROVAL phase flip +
|
|
26
|
+
* persist, and the terminal result mapping.
|
|
27
|
+
*/
|
|
28
|
+
import { collectSubAgentToolCallIds } from "../../shared/tool-row.js";
|
|
29
|
+
import { LocalWorkspaceBackend } from "../../shared/workspace/local-backend.js";
|
|
30
|
+
import { primaryToken, readDenialLedger } from "./approval-state.js";
|
|
31
|
+
import { deriveTurnCommandProvenance } from "./command-provenance.js";
|
|
32
|
+
import { captureTurnToLedger } from "./capture-flow.js";
|
|
33
|
+
import { clearProvisionalPostDenialNarration, reconcileDeniedToolCalls, } from "./message-translator.js";
|
|
34
|
+
// How long the boundary waits for the first-denial-stop's run.cancel() to
|
|
35
|
+
// settle before reading the final denial ledger and capturing the turn's tree.
|
|
36
|
+
// Long enough for the SDK's normal teardown, short enough that a wedged cancel
|
|
37
|
+
// cannot noticeably delay the approval pause the user is already waiting on.
|
|
38
|
+
const FIRST_DENIAL_CANCEL_TIMEOUT_MS = 5_000;
|
|
39
|
+
/**
|
|
40
|
+
* Run the turn boundary: author this turn's change set to the file_review
|
|
41
|
+
* ledger and overlay the hook's denials as approval gates. Mutates
|
|
42
|
+
* `opts.status` in place and reports whether the turn must pause.
|
|
43
|
+
*
|
|
44
|
+
* The hook records each denial to the ledger; we mark the corresponding tool
|
|
45
|
+
* calls WAITING_APPROVAL. The backend projects pending_approvals from that
|
|
46
|
+
* tool-call status (PendingApprovalComputer), so — exactly like the native
|
|
47
|
+
* harness — the approval surface is driven entirely by tool-call status. We
|
|
48
|
+
* deliberately do NOT set status.pendingApprovals here: any value would be
|
|
49
|
+
* discarded by the backend's recompute on the next updateStatus.
|
|
50
|
+
*/
|
|
51
|
+
export async function runTurnBoundary(opts) {
|
|
52
|
+
const { status, executionId, changeSetId, hitlDir, captureMode, baselineTree, primaryWorkspaceDir, gitWorkspace, turnStartMessageIndex, approvalGrants, globalBypass, seededSubAgents, artifactStorage, mergedPolicies, denialCancelSettled, } = opts;
|
|
53
|
+
// The timebox keeps a wedged cancel from hanging the pause; the reconcile
|
|
54
|
+
// trims below remain the backstop for that degraded case.
|
|
55
|
+
if (denialCancelSettled) {
|
|
56
|
+
await Promise.race([
|
|
57
|
+
denialCancelSettled,
|
|
58
|
+
new Promise((resolve) => {
|
|
59
|
+
const timer = setTimeout(resolve, FIRST_DENIAL_CANCEL_TIMEOUT_MS);
|
|
60
|
+
timer.unref();
|
|
61
|
+
}),
|
|
62
|
+
]);
|
|
63
|
+
}
|
|
64
|
+
const deniedLedger = await readDenialLedger(hitlDir ?? "");
|
|
65
|
+
// Capture mode: author the net change set to the file_review ledger as the
|
|
66
|
+
// CANDIDATE_CAPTURED event (projected server-side to a file_change_set
|
|
67
|
+
// AWAITING_REVIEW — the single review surface). The runner-owned gate files
|
|
68
|
+
// are excluded from the capture. The agent's edits are LEFT applied on the
|
|
69
|
+
// working tree (Cursor parity — the user reviews the real change; nothing is
|
|
70
|
+
// committed and the next turn is blocked until approval, and a reject snaps
|
|
71
|
+
// each file back on resume). Runs BEFORE the denial reconcile so a denied
|
|
72
|
+
// (gitignored) write stays on the deny-gate path while every flowed edit is
|
|
73
|
+
// captured to the ledger.
|
|
74
|
+
let capturedChangeCount = 0;
|
|
75
|
+
if (captureMode && baselineTree !== undefined && primaryWorkspaceDir) {
|
|
76
|
+
const deniedTokens = new Set(deniedLedger.map((e) => e.token));
|
|
77
|
+
// Approved-command turn facts (DD-28): when every mutation-capable call
|
|
78
|
+
// this turn was a consented shell command, attach the provenance so the
|
|
79
|
+
// backend can verify the cited consent rows and auto-keep the set instead
|
|
80
|
+
// of arming a second gate. Fail-closed: any non-qualifying turn attaches
|
|
81
|
+
// nothing and reviews manually exactly as before.
|
|
82
|
+
const commandProvenance = deriveTurnCommandProvenance({
|
|
83
|
+
messages: status.messages,
|
|
84
|
+
turnStartIndex: turnStartMessageIndex,
|
|
85
|
+
deniedTokens,
|
|
86
|
+
grantTokenToConsentId: new Map((approvalGrants ?? []).map((g) => [
|
|
87
|
+
primaryToken(g.key, g.salient, g.contentDigest),
|
|
88
|
+
g.sourceToolCallId,
|
|
89
|
+
])),
|
|
90
|
+
globalBypass,
|
|
91
|
+
});
|
|
92
|
+
if (commandProvenance) {
|
|
93
|
+
console.log(`ExecuteCursor capture: turn qualifies for approved-command auto-keep ` +
|
|
94
|
+
`(consent rows: ${commandProvenance.consentToolCallIds.join(",") || "(auto_approve_all)"}); ` +
|
|
95
|
+
`attaching provenance to candidate (execution=${executionId})`);
|
|
96
|
+
}
|
|
97
|
+
const captured = await captureTurnToLedger({
|
|
98
|
+
status,
|
|
99
|
+
gitRoot: primaryWorkspaceDir,
|
|
100
|
+
executionId,
|
|
101
|
+
changeSetId,
|
|
102
|
+
baselineTree,
|
|
103
|
+
messages: status.messages,
|
|
104
|
+
deniedTokens,
|
|
105
|
+
commandProvenance,
|
|
106
|
+
// Scope sub-agent row stamping to this turn: the seeded prior sub-agents
|
|
107
|
+
// (cloned in on resume) are the "before this turn" rows to skip.
|
|
108
|
+
priorSubAgentToolCallIds: collectSubAgentToolCallIds(seededSubAgents),
|
|
109
|
+
// The CAS half: read the sidecar the hook staged this turn and compose it
|
|
110
|
+
// into the change set. hitlDir + storage are present when captureIgnored
|
|
111
|
+
// was on (a git tree's gitignored writes, or ALL writes in a non-git
|
|
112
|
+
// workspace). In a git tree this composes with the git diff (HYBRID); in a
|
|
113
|
+
// non-git workspace it IS the whole change set (CAS-only).
|
|
114
|
+
hitlDir,
|
|
115
|
+
storage: artifactStorage,
|
|
116
|
+
gitWorkspace,
|
|
117
|
+
});
|
|
118
|
+
capturedChangeCount = captured.length;
|
|
119
|
+
if (capturedChangeCount > 0) {
|
|
120
|
+
console.log(`ExecuteCursor capture: ${capturedChangeCount} file change(s) authored to the ` +
|
|
121
|
+
`file_review ledger (change_set=${changeSetId}), working tree left applied ` +
|
|
122
|
+
`for review (execution=${executionId})`);
|
|
123
|
+
}
|
|
124
|
+
}
|
|
125
|
+
// The gate reads each denied file's pre-edit `before` from the workspace the
|
|
126
|
+
// runner is co-located with (local FS for OSS; the sandbox in cloud), so a
|
|
127
|
+
// whole-file rewrite gate renders a true before/after diff. The tool was
|
|
128
|
+
// DENIED, so disk still holds the old content. User files are never platform
|
|
129
|
+
// paths, so no platformDir routing is needed here.
|
|
130
|
+
const gateWorkspaceBackend = new LocalWorkspaceBackend(primaryWorkspaceDir);
|
|
131
|
+
const deniedToolCalls = await reconcileDeniedToolCalls(status.messages, deniedLedger, mergedPolicies, gateWorkspaceBackend);
|
|
132
|
+
// Observability: a synthesized placeholder (id `approval:*`) means a denial
|
|
133
|
+
// correlated to NO streamed tool call in either the exact or the normalized
|
|
134
|
+
// pass. After the normalized-path fallback this should be ~0; a non-zero rate
|
|
135
|
+
// is the early-warning signal of a NEW identity drift (the gate would then
|
|
136
|
+
// show "No preview available" with no diff). Logged, not thrown — the
|
|
137
|
+
// synthesized gate still safely surfaces the approval.
|
|
138
|
+
const synthesizedGateCount = deniedToolCalls.filter((tc) => tc.id.startsWith("approval:")).length;
|
|
139
|
+
if (synthesizedGateCount > 0) {
|
|
140
|
+
console.warn(`ExecuteCursor reconcile synthesized ${synthesizedGateCount} placeholder gate(s) ` +
|
|
141
|
+
`with no correlated stream call (execution=${executionId}); ` +
|
|
142
|
+
`possible hook/stream identity drift — gate(s) will lack a diff`);
|
|
143
|
+
}
|
|
144
|
+
if (deniedToolCalls.length > 0) {
|
|
145
|
+
// Deterministic clean-pause: a turn that pauses for approval must read as
|
|
146
|
+
// the same shape the native harness produces — pre-tool text + the gated
|
|
147
|
+
// tool calls — never the model's provisional reaction to Cursor's deny
|
|
148
|
+
// (e.g. "blocked by a hook; enable it in your Cursor settings"). We blank
|
|
149
|
+
// that reaction in place (keeping the message count, so the finalize stays
|
|
150
|
+
// append-only) rather than removing it. See
|
|
151
|
+
// clearProvisionalPostDenialNarration for the full rationale.
|
|
152
|
+
const redactedNarration = clearProvisionalPostDenialNarration(status.messages, deniedToolCalls);
|
|
153
|
+
if (redactedNarration.length > 0) {
|
|
154
|
+
console.log(`ExecuteCursor redacted ${redactedNarration.length} provisional post-denial narration message(s) before pausing for approval`);
|
|
155
|
+
}
|
|
156
|
+
}
|
|
157
|
+
return {
|
|
158
|
+
waiting: deniedToolCalls.length > 0 || capturedChangeCount > 0,
|
|
159
|
+
capturedChangeCount,
|
|
160
|
+
deniedToolCallCount: deniedToolCalls.length,
|
|
161
|
+
};
|
|
162
|
+
}
|
|
163
|
+
//# sourceMappingURL=turn-boundary.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"turn-boundary.js","sourceRoot":"","sources":["../../../src/activities/execute-cursor/turn-boundary.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;GA0BG;AAIH,OAAO,EAAE,0BAA0B,EAAE,MAAM,0BAA0B,CAAC;AACtE,OAAO,EAAE,qBAAqB,EAAE,MAAM,yCAAyC,CAAC;AAGhF,OAAO,EAAE,YAAY,EAAE,gBAAgB,EAAsB,MAAM,qBAAqB,CAAC;AACzF,OAAO,EAAE,2BAA2B,EAAE,MAAM,yBAAyB,CAAC;AACtE,OAAO,EAAE,mBAAmB,EAAE,MAAM,mBAAmB,CAAC;AACxD,OAAO,EACL,mCAAmC,EACnC,wBAAwB,GACzB,MAAM,yBAAyB,CAAC;AAEjC,0EAA0E;AAC1E,+EAA+E;AAC/E,+EAA+E;AAC/E,6EAA6E;AAC7E,MAAM,8BAA8B,GAAG,KAAK,CAAC;AAmE7C;;;;;;;;;;;GAWG;AACH,MAAM,CAAC,KAAK,UAAU,eAAe,CAAC,IAAyB;IAC7D,MAAM,EACJ,MAAM,EACN,WAAW,EACX,WAAW,EACX,OAAO,EACP,WAAW,EACX,YAAY,EACZ,mBAAmB,EACnB,YAAY,EACZ,qBAAqB,EACrB,cAAc,EACd,YAAY,EACZ,eAAe,EACf,eAAe,EACf,cAAc,EACd,mBAAmB,GACpB,GAAG,IAAI,CAAC;IAET,0EAA0E;IAC1E,0DAA0D;IAC1D,IAAI,mBAAmB,EAAE,CAAC;QACxB,MAAM,OAAO,CAAC,IAAI,CAAC;YACjB,mBAAmB;YACnB,IAAI,OAAO,CAAO,CAAC,OAAO,EAAE,EAAE;gBAC5B,MAAM,KAAK,GAAG,UAAU,CAAC,OAAO,EAAE,8BAA8B,CAAC,CAAC;gBAClE,KAAK,CAAC,KAAK,EAAE,CAAC;YAChB,CAAC,CAAC;SACH,CAAC,CAAC;IACL,CAAC;IACD,MAAM,YAAY,GAAG,MAAM,gBAAgB,CAAC,OAAO,IAAI,EAAE,CAAC,CAAC;IAE3D,2EAA2E;IAC3E,uEAAuE;IACvE,4EAA4E;IAC5E,2EAA2E;IAC3E,6EAA6E;IAC7E,4EAA4E;IAC5E,0EAA0E;IAC1E,4EAA4E;IAC5E,0BAA0B;IAC1B,IAAI,mBAAmB,GAAG,CAAC,CAAC;IAC5B,IAAI,WAAW,IAAI,YAAY,KAAK,SAAS,IAAI,mBAAmB,EAAE,CAAC;QACrE,MAAM,YAAY,GAAG,IAAI,GAAG,CAAC,YAAY,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC;QAC/D,wEAAwE;QACxE,wEAAwE;QACxE,0EAA0E;QAC1E,yEAAyE;QACzE,kDAAkD;QAClD,MAAM,iBAAiB,GAAG,2BAA2B,CAAC;YACpD,QAAQ,EAAE,MAAM,CAAC,QAAQ;YACzB,cAAc,EAAE,qBAAqB;YACrC,YAAY;YACZ,qBAAqB,EAAE,IAAI,GAAG,CAC5B,CAAC,cAAc,IAAI,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC;gBAChC,YAAY,CAAC,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,aAAa,CAAC;gBAC/C,CAAC,CAAC,gBAAgB;aACnB,CAAC,CACH;YACD,YAAY;SACb,CAAC,CAAC;QACH,IAAI,iBAAiB,EAAE,CAAC;YACtB,OAAO,CAAC,GAAG,CACT,uEAAuE;gBACvE,kBAAkB,iBAAiB,CAAC,kBAAkB,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,oBAAoB,KAAK;gBAC7F,gDAAgD,WAAW,GAAG,CAC/D,CAAC;QACJ,CAAC;QACD,MAAM,QAAQ,GAAG,MAAM,mBAAmB,CAAC;YACzC,MAAM;YACN,OAAO,EAAE,mBAAmB;YAC5B,WAAW;YACX,WAAW;YACX,YAAY;YACZ,QAAQ,EAAE,MAAM,CAAC,QAAQ;YACzB,YAAY;YACZ,iBAAiB;YACjB,yEAAyE;YACzE,iEAAiE;YACjE,wBAAwB,EAAE,0BAA0B,CAAC,eAAe,CAAC;YACrE,0EAA0E;YAC1E,yEAAyE;YACzE,qEAAqE;YACrE,2EAA2E;YAC3E,2DAA2D;YAC3D,OAAO;YACP,OAAO,EAAE,eAAe;YACxB,YAAY;SACb,CAAC,CAAC;QACH,mBAAmB,GAAG,QAAQ,CAAC,MAAM,CAAC;QACtC,IAAI,mBAAmB,GAAG,CAAC,EAAE,CAAC;YAC5B,OAAO,CAAC,GAAG,CACT,0BAA0B,mBAAmB,kCAAkC;gBAC/E,kCAAkC,WAAW,+BAA+B;gBAC5E,yBAAyB,WAAW,GAAG,CACxC,CAAC;QACJ,CAAC;IACH,CAAC;IAED,6EAA6E;IAC7E,2EAA2E;IAC3E,yEAAyE;IACzE,6EAA6E;IAC7E,mDAAmD;IACnD,MAAM,oBAAoB,GAAG,IAAI,qBAAqB,CAAC,mBAAmB,CAAC,CAAC;IAC5E,MAAM,eAAe,GAAG,MAAM,wBAAwB,CACpD,MAAM,CAAC,QAAQ,EACf,YAAY,EACZ,cAAc,EACd,oBAAoB,CACrB,CAAC;IACF,4EAA4E;IAC5E,4EAA4E;IAC5E,8EAA8E;IAC9E,2EAA2E;IAC3E,sEAAsE;IACtE,uDAAuD;IACvD,MAAM,oBAAoB,GAAG,eAAe,CAAC,MAAM,CAAC,CAAC,EAAE,EAAE,EAAE,CACzD,EAAE,CAAC,EAAE,CAAC,UAAU,CAAC,WAAW,CAAC,CAC9B,CAAC,MAAM,CAAC;IACT,IAAI,oBAAoB,GAAG,CAAC,EAAE,CAAC;QAC7B,OAAO,CAAC,IAAI,CACV,uCAAuC,oBAAoB,uBAAuB;YAChF,6CAA6C,WAAW,KAAK;YAC7D,gEAAgE,CACnE,CAAC;IACJ,CAAC;IACD,IAAI,eAAe,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC/B,0EAA0E;QAC1E,yEAAyE;QACzE,uEAAuE;QACvE,0EAA0E;QAC1E,2EAA2E;QAC3E,4CAA4C;QAC5C,8DAA8D;QAC9D,MAAM,iBAAiB,GAAG,mCAAmC,CAAC,MAAM,CAAC,QAAQ,EAAE,eAAe,CAAC,CAAC;QAChG,IAAI,iBAAiB,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACjC,OAAO,CAAC,GAAG,CACT,0BAA0B,iBAAiB,CAAC,MAAM,2EAA2E,CAC9H,CAAC;QACJ,CAAC;IACH,CAAC;IAED,OAAO;QACL,OAAO,EAAE,eAAe,CAAC,MAAM,GAAG,CAAC,IAAI,mBAAmB,GAAG,CAAC;QAC9D,mBAAmB;QACnB,mBAAmB,EAAE,eAAe,CAAC,MAAM;KAC5C,CAAC;AACJ,CAAC"}
|
|
@@ -0,0 +1,141 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* The Cursor harness's stream phase — the single per-event consumption loop that
|
|
3
|
+
* turns a live Cursor SDK run into the streamed transcript, mid-run progress, and
|
|
4
|
+
* the flags that decide how the turn ends.
|
|
5
|
+
*
|
|
6
|
+
* Extracted from the activity entry point (index.ts) so that BOTH the primary
|
|
7
|
+
* turn and the two recovery retries (poisoned-handle / transport-timeout) drive
|
|
8
|
+
* the exact same code. The retries previously re-implemented a stripped "bare"
|
|
9
|
+
* loop that dropped live persist, DD-32/DD-33 mid-run progress, sub-agent
|
|
10
|
+
* tracking, the first-denial early stop, and correct pause/stall handling — so a
|
|
11
|
+
* retry froze the UI and mis-handled a mid-retry pause. One loop removes that
|
|
12
|
+
* drift by construction.
|
|
13
|
+
*
|
|
14
|
+
* This mirrors the deep-agent harness's `streamExecution` seam
|
|
15
|
+
* (execute-deep-agent/streaming.ts): injected `heartbeat`/`isCancelled` so it
|
|
16
|
+
* runs without Temporal, a structural `StreamableRun` so it runs without the live
|
|
17
|
+
* SDK, and a result (the terminal reason) instead of internal terminal mapping.
|
|
18
|
+
* The one deliberate difference: this consumer REPORTS why the stream ended and
|
|
19
|
+
* leaves the terminal MAPPING to the activity, because Cursor's mapping is
|
|
20
|
+
* entangled with worker-shutdown-vs-pause disambiguation, the turn boundary,
|
|
21
|
+
* structured output, and recovery retries — all of which must stay in index.ts.
|
|
22
|
+
*/
|
|
23
|
+
import type { AgentExecutionStatus } from "@stigmer/protos/ai/stigmer/agentic/agentexecution/v1/api_pb";
|
|
24
|
+
import { ExecutionControlSignal } from "@stigmer/protos/ai/stigmer/agentic/agentexecution/v1/enum_pb";
|
|
25
|
+
import type { SDKMessage, InteractionUpdate } from "@cursor/sdk";
|
|
26
|
+
import { StallTimeoutError, type StallWatchdog } from "../../shared/stall-watchdog.js";
|
|
27
|
+
import { type ProgressCaptureState, type ProgressSubstrate } from "../../shared/filereview/progress.js";
|
|
28
|
+
import type { MessageAccumulator } from "./message-translator.js";
|
|
29
|
+
import type { DeltaEnricher } from "./delta-enricher.js";
|
|
30
|
+
import type { TodoTracker } from "./todo-tracker.js";
|
|
31
|
+
import type { StreamingUpdateScheduler } from "../../shared/streaming-scheduler.js";
|
|
32
|
+
import type { UsageAccumulator } from "./usage-accumulator.js";
|
|
33
|
+
import type { createCursorEventRecorder } from "./cursor-event-recorder.js";
|
|
34
|
+
/**
|
|
35
|
+
* The subset of the Cursor SDK `Run` the stream phase consumes. Kept structural
|
|
36
|
+
* so the loop is unit-testable with a mock run — no live SDK, no Temporal.
|
|
37
|
+
*/
|
|
38
|
+
export interface StreamableRun {
|
|
39
|
+
stream(): AsyncIterable<SDKMessage>;
|
|
40
|
+
supports?(feature: string): boolean;
|
|
41
|
+
cancel(): Promise<void>;
|
|
42
|
+
}
|
|
43
|
+
export type CursorTurnEventRecorder = ReturnType<typeof createCursorEventRecorder>;
|
|
44
|
+
/**
|
|
45
|
+
* Why the turn's stream ended. The activity maps each reason to a terminal
|
|
46
|
+
* outcome (see resolvePreBoundaryTerminal in index.ts): `completed` and
|
|
47
|
+
* `first-denial` proceed to the turn boundary; the rest are pre-boundary
|
|
48
|
+
* terminals that return or throw.
|
|
49
|
+
*/
|
|
50
|
+
export type TurnStreamReason = "completed" | "paused" | "stalled" | "first-denial" | "platform-stop";
|
|
51
|
+
/**
|
|
52
|
+
* Single owner for every flag the turn's stream produces. Before this, these
|
|
53
|
+
* eleven fields lived as scattered `let`s co-written by FOUR concurrent
|
|
54
|
+
* producers — the SDK `onDelta`, the fs denial-watcher, the stall-watchdog, and
|
|
55
|
+
* the loop — the classic "who writes this field?" hazard. Consolidating them into
|
|
56
|
+
* one object (created once per turn, shared by every producer, the epilogue, and
|
|
57
|
+
* the outer catch/finally) makes ownership explicit and lets the primary turn and
|
|
58
|
+
* both recovery retries drive the identical stream code.
|
|
59
|
+
*
|
|
60
|
+
* Field ownership is annotated per field: exactly one primary writer, plus any
|
|
61
|
+
* documented producer/consumer handshakes.
|
|
62
|
+
*/
|
|
63
|
+
export interface TurnStreamState {
|
|
64
|
+
/** Written by: onDelta (heartbeat cancel) + the loop's cancellation check. Read by: the loop, epilogue, outer catch. */
|
|
65
|
+
pauseDetected: boolean;
|
|
66
|
+
/** Written by: the stall-watchdog callback. Read by: the loop + epilogue. */
|
|
67
|
+
stallDetected: boolean;
|
|
68
|
+
stallError: StallTimeoutError | undefined;
|
|
69
|
+
/** Written by: the loop (on the first ledger entry). Read by: the loop, epilogue, settleRetryTurn. */
|
|
70
|
+
firstDenialDetected: boolean;
|
|
71
|
+
/** Written by: the fs denial-watcher (→ true) + the loop (→ false after a read). Read by: the loop. */
|
|
72
|
+
denialLedgerDirty: boolean;
|
|
73
|
+
/** Written by: the loop (the first-denial run.cancel()). Read by: the turn boundary. */
|
|
74
|
+
denialCancelSettled: Promise<void> | undefined;
|
|
75
|
+
/** Written by: the loop (a platform STOP signal from persist). Read by: the loop + epilogue. */
|
|
76
|
+
platformStopSignaled: boolean;
|
|
77
|
+
/** Written by: the loop (a stream ERROR status event) + the retry setup (reset). Read by: error classification. */
|
|
78
|
+
streamErrorMessage: string | undefined;
|
|
79
|
+
/** Written by: the loop (each tool_call). Read by: the stall-watchdog (message enrichment). */
|
|
80
|
+
lastToolName: string | undefined;
|
|
81
|
+
/** Written by: the loop. Read by: the event recorder, the persist cadence, and logs. */
|
|
82
|
+
eventCount: number;
|
|
83
|
+
/** Written + read by: onDelta (log the first turn's context attribution exactly once). */
|
|
84
|
+
firstTurnAttributionLogged: boolean;
|
|
85
|
+
/** The stall watchdog for the CURRENT stream, armed by consumeCursorTurnStream; held so onDelta can reset it and the activity's finally can stop it. */
|
|
86
|
+
stallWatchdog: StallWatchdog | undefined;
|
|
87
|
+
}
|
|
88
|
+
export declare function newTurnStreamState(): TurnStreamState;
|
|
89
|
+
/**
|
|
90
|
+
* The subset of collaborators the shared onDelta needs. Broken out from
|
|
91
|
+
* CursorTurnStreamDeps because the primary send happens BEFORE the accumulator
|
|
92
|
+
* exists — onDelta only touches usage, the enricher, the heartbeat, and state.
|
|
93
|
+
*/
|
|
94
|
+
export interface TurnOnDeltaDeps {
|
|
95
|
+
readonly usageAccumulator: UsageAccumulator;
|
|
96
|
+
readonly deltaEnricher: DeltaEnricher;
|
|
97
|
+
/** Temporal heartbeat. Injected so the callback is testable without Temporal. */
|
|
98
|
+
readonly heartbeat: () => void;
|
|
99
|
+
readonly promptEstimatedTokens: number;
|
|
100
|
+
readonly executionId: string;
|
|
101
|
+
readonly state: TurnStreamState;
|
|
102
|
+
}
|
|
103
|
+
export interface CursorTurnStreamDeps extends TurnOnDeltaDeps {
|
|
104
|
+
readonly status: AgentExecutionStatus;
|
|
105
|
+
readonly accumulator: MessageAccumulator;
|
|
106
|
+
readonly todoTracker: TodoTracker;
|
|
107
|
+
readonly eventRecorder: CursorTurnEventRecorder | undefined;
|
|
108
|
+
readonly scheduler: StreamingUpdateScheduler;
|
|
109
|
+
readonly progressSubstrate: ProgressSubstrate | undefined;
|
|
110
|
+
readonly progressState: ProgressCaptureState;
|
|
111
|
+
readonly changeSetId: string;
|
|
112
|
+
/** Session HITL dir holding the denial ledger; undefined → no gate installed → no first-denial stop. */
|
|
113
|
+
readonly hitlDir: string | undefined;
|
|
114
|
+
readonly stallTimeoutMs: number;
|
|
115
|
+
/** The single size-bounded persist chokepoint. Returns the platform control signal. */
|
|
116
|
+
readonly persist: (status: AgentExecutionStatus) => Promise<ExecutionControlSignal>;
|
|
117
|
+
/** Temporal cancellation check. Injected so the loop is testable without Temporal. */
|
|
118
|
+
readonly isCancelled: () => boolean;
|
|
119
|
+
}
|
|
120
|
+
/**
|
|
121
|
+
* Build the shared onDelta callback. The Cursor SDK's fine-grained delta channel
|
|
122
|
+
* carries token usage, live shell output, and precise tool-call timings; it also
|
|
123
|
+
* fires far more often than discrete stream events, so it is where the stall
|
|
124
|
+
* timer is reset during a long model generation. A heartbeat CancelledFailure
|
|
125
|
+
* here flags the pause so the loop and epilogue treat it as a user pause (the
|
|
126
|
+
* bare retry loops swallowed this — the cause of the mid-retry mislabel).
|
|
127
|
+
*/
|
|
128
|
+
export declare function makeCursorTurnOnDelta(deps: TurnOnDeltaDeps): (event: {
|
|
129
|
+
update: InteractionUpdate;
|
|
130
|
+
}) => void;
|
|
131
|
+
/**
|
|
132
|
+
* Consume a Cursor SDK run to completion (or to a clean early stop), driving the
|
|
133
|
+
* transcript, todos, sub-agent tracking, live persist, and DD-32/DD-33 mid-run
|
|
134
|
+
* progress, and reporting WHY the stream ended.
|
|
135
|
+
*
|
|
136
|
+
* Arms + stops its own stall watchdog (stored on `state.stallWatchdog` so the
|
|
137
|
+
* shared onDelta can reset it and the activity's finally can stop it as a
|
|
138
|
+
* backstop). Mutates `deps.status` + `deps.state` in place. The caller owns the
|
|
139
|
+
* post-stream finalize and the terminal mapping.
|
|
140
|
+
*/
|
|
141
|
+
export declare function consumeCursorTurnStream(run: StreamableRun, deps: CursorTurnStreamDeps): Promise<TurnStreamReason>;
|
|
@@ -0,0 +1,249 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* The Cursor harness's stream phase — the single per-event consumption loop that
|
|
3
|
+
* turns a live Cursor SDK run into the streamed transcript, mid-run progress, and
|
|
4
|
+
* the flags that decide how the turn ends.
|
|
5
|
+
*
|
|
6
|
+
* Extracted from the activity entry point (index.ts) so that BOTH the primary
|
|
7
|
+
* turn and the two recovery retries (poisoned-handle / transport-timeout) drive
|
|
8
|
+
* the exact same code. The retries previously re-implemented a stripped "bare"
|
|
9
|
+
* loop that dropped live persist, DD-32/DD-33 mid-run progress, sub-agent
|
|
10
|
+
* tracking, the first-denial early stop, and correct pause/stall handling — so a
|
|
11
|
+
* retry froze the UI and mis-handled a mid-retry pause. One loop removes that
|
|
12
|
+
* drift by construction.
|
|
13
|
+
*
|
|
14
|
+
* This mirrors the deep-agent harness's `streamExecution` seam
|
|
15
|
+
* (execute-deep-agent/streaming.ts): injected `heartbeat`/`isCancelled` so it
|
|
16
|
+
* runs without Temporal, a structural `StreamableRun` so it runs without the live
|
|
17
|
+
* SDK, and a result (the terminal reason) instead of internal terminal mapping.
|
|
18
|
+
* The one deliberate difference: this consumer REPORTS why the stream ended and
|
|
19
|
+
* leaves the terminal MAPPING to the activity, because Cursor's mapping is
|
|
20
|
+
* entangled with worker-shutdown-vs-pause disambiguation, the turn boundary,
|
|
21
|
+
* structured output, and recovery retries — all of which must stay in index.ts.
|
|
22
|
+
*/
|
|
23
|
+
import { create } from "@bufbuild/protobuf";
|
|
24
|
+
import { CancelledFailure } from "@temporalio/activity";
|
|
25
|
+
import { ExecutionControlSignal } from "@stigmer/protos/ai/stigmer/agentic/agentexecution/v1/enum_pb";
|
|
26
|
+
import { StreamingUsageSummarySchema } from "@stigmer/protos/ai/stigmer/agentic/agentexecution/v1/usage_pb";
|
|
27
|
+
import { startStallWatchdog, StallTimeoutError, } from "../../shared/stall-watchdog.js";
|
|
28
|
+
import { shouldPersistStreamingStatus } from "./persist-decision.js";
|
|
29
|
+
import { readDenialLedger } from "./approval-state.js";
|
|
30
|
+
import { captureFileChangeProgress, } from "../../shared/filereview/progress.js";
|
|
31
|
+
export function newTurnStreamState() {
|
|
32
|
+
return {
|
|
33
|
+
pauseDetected: false,
|
|
34
|
+
stallDetected: false,
|
|
35
|
+
stallError: undefined,
|
|
36
|
+
firstDenialDetected: false,
|
|
37
|
+
denialLedgerDirty: false,
|
|
38
|
+
denialCancelSettled: undefined,
|
|
39
|
+
platformStopSignaled: false,
|
|
40
|
+
streamErrorMessage: undefined,
|
|
41
|
+
lastToolName: undefined,
|
|
42
|
+
eventCount: 0,
|
|
43
|
+
firstTurnAttributionLogged: false,
|
|
44
|
+
stallWatchdog: undefined,
|
|
45
|
+
};
|
|
46
|
+
}
|
|
47
|
+
/**
|
|
48
|
+
* Build the shared onDelta callback. The Cursor SDK's fine-grained delta channel
|
|
49
|
+
* carries token usage, live shell output, and precise tool-call timings; it also
|
|
50
|
+
* fires far more often than discrete stream events, so it is where the stall
|
|
51
|
+
* timer is reset during a long model generation. A heartbeat CancelledFailure
|
|
52
|
+
* here flags the pause so the loop and epilogue treat it as a user pause (the
|
|
53
|
+
* bare retry loops swallowed this — the cause of the mid-retry mislabel).
|
|
54
|
+
*/
|
|
55
|
+
export function makeCursorTurnOnDelta(deps) {
|
|
56
|
+
const { usageAccumulator, deltaEnricher, heartbeat, promptEstimatedTokens, executionId, state } = deps;
|
|
57
|
+
return ({ update }) => {
|
|
58
|
+
// Reset the stall timer on the delta channel too: a long model generation
|
|
59
|
+
// emits token deltas but few discrete stream events, so resetting only in the
|
|
60
|
+
// stream loop would false-positive a stall.
|
|
61
|
+
state.stallWatchdog?.recordActivity();
|
|
62
|
+
if (update.type === "turn-ended" && update.usage) {
|
|
63
|
+
usageAccumulator.addTurn(update.usage);
|
|
64
|
+
if (!state.firstTurnAttributionLogged) {
|
|
65
|
+
state.firstTurnAttributionLogged = true;
|
|
66
|
+
const sdkInputTokens = update.usage.inputTokens ?? 0;
|
|
67
|
+
const cursorOverhead = Math.max(0, sdkInputTokens - promptEstimatedTokens);
|
|
68
|
+
console.log(`ExecuteCursor context attribution (first turn): execution=${executionId}, ` +
|
|
69
|
+
`sdkInputTokens=${sdkInputTokens}, stigmerPreamble=${promptEstimatedTokens}, ` +
|
|
70
|
+
`cursorOverhead=${cursorOverhead} (estimated)`);
|
|
71
|
+
}
|
|
72
|
+
}
|
|
73
|
+
deltaEnricher.processDelta(update);
|
|
74
|
+
try {
|
|
75
|
+
heartbeat();
|
|
76
|
+
}
|
|
77
|
+
catch (hbErr) {
|
|
78
|
+
if (hbErr instanceof CancelledFailure) {
|
|
79
|
+
state.pauseDetected = true;
|
|
80
|
+
return;
|
|
81
|
+
}
|
|
82
|
+
throw hbErr;
|
|
83
|
+
}
|
|
84
|
+
};
|
|
85
|
+
}
|
|
86
|
+
/**
|
|
87
|
+
* Consume a Cursor SDK run to completion (or to a clean early stop), driving the
|
|
88
|
+
* transcript, todos, sub-agent tracking, live persist, and DD-32/DD-33 mid-run
|
|
89
|
+
* progress, and reporting WHY the stream ended.
|
|
90
|
+
*
|
|
91
|
+
* Arms + stops its own stall watchdog (stored on `state.stallWatchdog` so the
|
|
92
|
+
* shared onDelta can reset it and the activity's finally can stop it as a
|
|
93
|
+
* backstop). Mutates `deps.status` + `deps.state` in place. The caller owns the
|
|
94
|
+
* post-stream finalize and the terminal mapping.
|
|
95
|
+
*/
|
|
96
|
+
export async function consumeCursorTurnStream(run, deps) {
|
|
97
|
+
const { status, accumulator, todoTracker, deltaEnricher, eventRecorder, scheduler, usageAccumulator, progressSubstrate, progressState, changeSetId, hitlDir, executionId, stallTimeoutMs, persist, heartbeat, isCancelled, state, } = deps;
|
|
98
|
+
// Arm the stall watchdog now that the run exists. The activity's periodic
|
|
99
|
+
// heartbeat proves the process is alive, not that the agent is progressing: if
|
|
100
|
+
// the stream wedges (a tool call or model connection that never returns), no
|
|
101
|
+
// event/delta arrives and the turn would hang forever. On stall we end the run
|
|
102
|
+
// cleanly via the SDK's run.cancel() (guarded by supports("cancel")), which
|
|
103
|
+
// unblocks the for-await; the caller then reports EXECUTION_FAILED with a
|
|
104
|
+
// recognizable, actionable message.
|
|
105
|
+
state.stallWatchdog = startStallWatchdog(stallTimeoutMs, (idleMs) => {
|
|
106
|
+
state.stallDetected = true;
|
|
107
|
+
state.stallError = new StallTimeoutError(idleMs, state.lastToolName ? `last tool: ${state.lastToolName}` : undefined);
|
|
108
|
+
console.warn(`ExecuteCursor stall detected: execution=${executionId}, idleMs=${idleMs}, lastTool=${state.lastToolName ?? "none"}`);
|
|
109
|
+
if (run.supports?.("cancel")) {
|
|
110
|
+
void run.cancel().catch((cancelErr) => {
|
|
111
|
+
console.warn(`ExecuteCursor run.cancel() after stall failed (non-fatal): execution=${executionId}, ` +
|
|
112
|
+
`error=${cancelErr instanceof Error ? cancelErr.message : cancelErr}`);
|
|
113
|
+
});
|
|
114
|
+
}
|
|
115
|
+
});
|
|
116
|
+
try {
|
|
117
|
+
for await (const event of run.stream()) {
|
|
118
|
+
if (state.pauseDetected || isCancelled()) {
|
|
119
|
+
state.pauseDetected = true;
|
|
120
|
+
break;
|
|
121
|
+
}
|
|
122
|
+
if (state.stallDetected)
|
|
123
|
+
break;
|
|
124
|
+
// Progress: reset the stall timer on every stream event.
|
|
125
|
+
state.stallWatchdog.recordActivity();
|
|
126
|
+
if (event.type === "tool_call" && typeof event.name === "string") {
|
|
127
|
+
state.lastToolName = event.name;
|
|
128
|
+
}
|
|
129
|
+
eventRecorder?.record(event, state.eventCount);
|
|
130
|
+
accumulator.processEvent(event);
|
|
131
|
+
todoTracker.processEvent(event);
|
|
132
|
+
if (event.type === "tool_call" && event.name === "task") {
|
|
133
|
+
accumulator.trackSubAgentExecution(event);
|
|
134
|
+
}
|
|
135
|
+
// First-denial stop (HITL clean pause). In CAPTURE mode this fires only for
|
|
136
|
+
// an IRREVERSIBLE tool the hook still gates (shell, MCP, or a gitignored
|
|
137
|
+
// write/delete) — file edits flow freely and are captured at the turn
|
|
138
|
+
// boundary, so they never enter the ledger. In the deny-gate FALLBACK
|
|
139
|
+
// (non-git workspace) it fires for every gated file edit too. Either way:
|
|
140
|
+
// the preToolUse hook appends to the denial ledger the instant it gates a
|
|
141
|
+
// tool — before Cursor surfaces the failure to the model — and the fs
|
|
142
|
+
// watcher flips denialLedgerDirty the moment that write lands. Confirming
|
|
143
|
+
// the flag with a read on the very next event (of ANY type — thinking
|
|
144
|
+
// deltas arrive within milliseconds) ends the turn before the model's
|
|
145
|
+
// reaction can persist: waiting for the next tool_call event let the full
|
|
146
|
+
// post-denial reaction (thinking, narration, a workaround shell) stream and
|
|
147
|
+
// persist live (production case aex_01kwj07f7g23c3wp9sn8496z5g). The
|
|
148
|
+
// tool_call-event read stays as the backstop for platforms where fs.watch
|
|
149
|
+
// is unreliable; the current event was already accumulated above, so the
|
|
150
|
+
// anchor's own row is always present for the turn-boundary gate overlay.
|
|
151
|
+
if (!state.firstDenialDetected && hitlDir && (state.denialLedgerDirty || event.type === "tool_call")) {
|
|
152
|
+
state.denialLedgerDirty = false;
|
|
153
|
+
const denials = await readDenialLedger(hitlDir);
|
|
154
|
+
if (denials.length > 0) {
|
|
155
|
+
state.firstDenialDetected = true;
|
|
156
|
+
console.log(`ExecuteCursor first denial detected (${denials.length} ledger ` +
|
|
157
|
+
`entr${denials.length === 1 ? "y" : "ies"}); stopping turn to pause ` +
|
|
158
|
+
`cleanly for approval: execution=${executionId}`);
|
|
159
|
+
if (run.supports?.("cancel")) {
|
|
160
|
+
// Kept (not fire-and-forget): awaited timeboxed by the turn boundary
|
|
161
|
+
// so the ledger read and tree capture see a stopped agent.
|
|
162
|
+
state.denialCancelSettled = run.cancel().then(() => { }, (cancelErr) => {
|
|
163
|
+
console.warn(`ExecuteCursor run.cancel() after first denial failed (non-fatal): ` +
|
|
164
|
+
`execution=${executionId}, ` +
|
|
165
|
+
`error=${cancelErr instanceof Error ? cancelErr.message : cancelErr}`);
|
|
166
|
+
});
|
|
167
|
+
}
|
|
168
|
+
break;
|
|
169
|
+
}
|
|
170
|
+
}
|
|
171
|
+
deltaEnricher.applyEnrichments(status.messages);
|
|
172
|
+
state.eventCount++;
|
|
173
|
+
if (event.type === "status") {
|
|
174
|
+
console.log(`ExecuteCursor stream status: execution=${executionId}, status=${JSON.stringify(event)}`);
|
|
175
|
+
const statusEvent = event;
|
|
176
|
+
if (statusEvent.status === "ERROR" && statusEvent.message) {
|
|
177
|
+
state.streamErrorMessage = statusEvent.message;
|
|
178
|
+
}
|
|
179
|
+
}
|
|
180
|
+
const shouldPersist = shouldPersistStreamingStatus({
|
|
181
|
+
deltaEnricherDirty: deltaEnricher.isDirty,
|
|
182
|
+
todosDirty: todoTracker.isDirty,
|
|
183
|
+
contentDirty: accumulator.isDirty,
|
|
184
|
+
}, scheduler, state.eventCount);
|
|
185
|
+
if (usageAccumulator.hasTurns) {
|
|
186
|
+
status.streamingUsage = create(StreamingUsageSummarySchema, usageAccumulator.snapshot());
|
|
187
|
+
}
|
|
188
|
+
if (shouldPersist) {
|
|
189
|
+
// Sync sub-agent executions into status before every persist so the live
|
|
190
|
+
// UI reflects delegation (including the IN_PROGRESS state) while the
|
|
191
|
+
// parent is still running — matching the native harness.
|
|
192
|
+
status.subAgentExecutions = accumulator.subAgentExecutions;
|
|
193
|
+
// Mid-run live capture (DD-32 / DD-33): attach the "N files changed so
|
|
194
|
+
// far" snapshot onto status.file_change_progress, throttled internally by
|
|
195
|
+
// the floor. Never authoritative — the turn-boundary candidate remains
|
|
196
|
+
// the reviewed diff.
|
|
197
|
+
if (progressSubstrate) {
|
|
198
|
+
await captureFileChangeProgress({
|
|
199
|
+
status,
|
|
200
|
+
changeSetId,
|
|
201
|
+
substrate: progressSubstrate,
|
|
202
|
+
state: progressState,
|
|
203
|
+
});
|
|
204
|
+
}
|
|
205
|
+
const signal = await persist(status);
|
|
206
|
+
deltaEnricher.markPersisted();
|
|
207
|
+
todoTracker.markPersisted();
|
|
208
|
+
accumulator.markPersisted();
|
|
209
|
+
scheduler.markUpdateSent(state.eventCount);
|
|
210
|
+
heartbeat();
|
|
211
|
+
if (signal === ExecutionControlSignal.STOP) {
|
|
212
|
+
state.platformStopSignaled = true;
|
|
213
|
+
console.warn(`ExecuteCursor platform stop signal received: execution=${executionId}`);
|
|
214
|
+
}
|
|
215
|
+
}
|
|
216
|
+
if (state.platformStopSignaled) {
|
|
217
|
+
console.log(`ExecuteCursor stopping stream due to platform stop signal: execution=${executionId}`);
|
|
218
|
+
break;
|
|
219
|
+
}
|
|
220
|
+
}
|
|
221
|
+
}
|
|
222
|
+
catch (streamErr) {
|
|
223
|
+
// run.cancel() — from the stall watchdog or the first-denial stop — can make
|
|
224
|
+
// the stream iterator reject as it tears down; that is the expected teardown
|
|
225
|
+
// for both, so swallow it and fall through. Anything else is a genuine stream
|
|
226
|
+
// failure — rethrow it to the activity's error handler.
|
|
227
|
+
if (!state.stallDetected && !state.firstDenialDetected)
|
|
228
|
+
throw streamErr;
|
|
229
|
+
console.warn(`ExecuteCursor stream ended via cancel: execution=${executionId}, ` +
|
|
230
|
+
`stall=${state.stallDetected}, firstDenial=${state.firstDenialDetected}`);
|
|
231
|
+
}
|
|
232
|
+
finally {
|
|
233
|
+
state.stallWatchdog.stop();
|
|
234
|
+
}
|
|
235
|
+
// Report why the stream ended, in the same precedence the activity's epilogue
|
|
236
|
+
// applies: a stall (an EXECUTION_FAILED terminal) outranks everything; a
|
|
237
|
+
// platform stop and a first denial are distinct proceed/return outcomes; a
|
|
238
|
+
// pause is the fallback for a cancellation with no other cause.
|
|
239
|
+
if (state.stallDetected)
|
|
240
|
+
return "stalled";
|
|
241
|
+
if (state.platformStopSignaled)
|
|
242
|
+
return "platform-stop";
|
|
243
|
+
if (state.firstDenialDetected)
|
|
244
|
+
return "first-denial";
|
|
245
|
+
if (state.pauseDetected || isCancelled())
|
|
246
|
+
return "paused";
|
|
247
|
+
return "completed";
|
|
248
|
+
}
|
|
249
|
+
//# sourceMappingURL=turn-stream.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"turn-stream.js","sourceRoot":"","sources":["../../../src/activities/execute-cursor/turn-stream.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAEH,OAAO,EAAE,MAAM,EAAE,MAAM,oBAAoB,CAAC;AAC5C,OAAO,EAAE,gBAAgB,EAAE,MAAM,sBAAsB,CAAC;AAExD,OAAO,EAAE,sBAAsB,EAAE,MAAM,8DAA8D,CAAC;AACtG,OAAO,EAAE,2BAA2B,EAAE,MAAM,+DAA+D,CAAC;AAE5G,OAAO,EACL,kBAAkB,EAClB,iBAAiB,GAElB,MAAM,gCAAgC,CAAC;AACxC,OAAO,EAAE,4BAA4B,EAAE,MAAM,uBAAuB,CAAC;AACrE,OAAO,EAAE,gBAAgB,EAAE,MAAM,qBAAqB,CAAC;AACvD,OAAO,EACL,yBAAyB,GAG1B,MAAM,qCAAqC,CAAC;AAuE7C,MAAM,UAAU,kBAAkB;IAChC,OAAO;QACL,aAAa,EAAE,KAAK;QACpB,aAAa,EAAE,KAAK;QACpB,UAAU,EAAE,SAAS;QACrB,mBAAmB,EAAE,KAAK;QAC1B,iBAAiB,EAAE,KAAK;QACxB,mBAAmB,EAAE,SAAS;QAC9B,oBAAoB,EAAE,KAAK;QAC3B,kBAAkB,EAAE,SAAS;QAC7B,YAAY,EAAE,SAAS;QACvB,UAAU,EAAE,CAAC;QACb,0BAA0B,EAAE,KAAK;QACjC,aAAa,EAAE,SAAS;KACzB,CAAC;AACJ,CAAC;AAmCD;;;;;;;GAOG;AACH,MAAM,UAAU,qBAAqB,CACnC,IAAqB;IAErB,MAAM,EAAE,gBAAgB,EAAE,aAAa,EAAE,SAAS,EAAE,qBAAqB,EAAE,WAAW,EAAE,KAAK,EAAE,GAC7F,IAAI,CAAC;IACP,OAAO,CAAC,EAAE,MAAM,EAAE,EAAE,EAAE;QACpB,0EAA0E;QAC1E,8EAA8E;QAC9E,4CAA4C;QAC5C,KAAK,CAAC,aAAa,EAAE,cAAc,EAAE,CAAC;QACtC,IAAI,MAAM,CAAC,IAAI,KAAK,YAAY,IAAI,MAAM,CAAC,KAAK,EAAE,CAAC;YACjD,gBAAgB,CAAC,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;YAEvC,IAAI,CAAC,KAAK,CAAC,0BAA0B,EAAE,CAAC;gBACtC,KAAK,CAAC,0BAA0B,GAAG,IAAI,CAAC;gBACxC,MAAM,cAAc,GAAG,MAAM,CAAC,KAAK,CAAC,WAAW,IAAI,CAAC,CAAC;gBACrD,MAAM,cAAc,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,cAAc,GAAG,qBAAqB,CAAC,CAAC;gBAC3E,OAAO,CAAC,GAAG,CACT,6DAA6D,WAAW,IAAI;oBAC1E,kBAAkB,cAAc,qBAAqB,qBAAqB,IAAI;oBAC9E,kBAAkB,cAAc,cAAc,CACjD,CAAC;YACJ,CAAC;QACH,CAAC;QACD,aAAa,CAAC,YAAY,CAAC,MAAM,CAAC,CAAC;QACnC,IAAI,CAAC;YACH,SAAS,EAAE,CAAC;QACd,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,IAAI,KAAK,YAAY,gBAAgB,EAAE,CAAC;gBACtC,KAAK,CAAC,aAAa,GAAG,IAAI,CAAC;gBAC3B,OAAO;YACT,CAAC;YACD,MAAM,KAAK,CAAC;QACd,CAAC;IACH,CAAC,CAAC;AACJ,CAAC;AAED;;;;;;;;;GASG;AACH,MAAM,CAAC,KAAK,UAAU,uBAAuB,CAC3C,GAAkB,EAClB,IAA0B;IAE1B,MAAM,EACJ,MAAM,EACN,WAAW,EACX,WAAW,EACX,aAAa,EACb,aAAa,EACb,SAAS,EACT,gBAAgB,EAChB,iBAAiB,EACjB,aAAa,EACb,WAAW,EACX,OAAO,EACP,WAAW,EACX,cAAc,EACd,OAAO,EACP,SAAS,EACT,WAAW,EACX,KAAK,GACN,GAAG,IAAI,CAAC;IAET,0EAA0E;IAC1E,+EAA+E;IAC/E,6EAA6E;IAC7E,+EAA+E;IAC/E,4EAA4E;IAC5E,0EAA0E;IAC1E,oCAAoC;IACpC,KAAK,CAAC,aAAa,GAAG,kBAAkB,CAAC,cAAc,EAAE,CAAC,MAAM,EAAE,EAAE;QAClE,KAAK,CAAC,aAAa,GAAG,IAAI,CAAC;QAC3B,KAAK,CAAC,UAAU,GAAG,IAAI,iBAAiB,CACtC,MAAM,EACN,KAAK,CAAC,YAAY,CAAC,CAAC,CAAC,cAAc,KAAK,CAAC,YAAY,EAAE,CAAC,CAAC,CAAC,SAAS,CACpE,CAAC;QACF,OAAO,CAAC,IAAI,CACV,2CAA2C,WAAW,YAAY,MAAM,cAAc,KAAK,CAAC,YAAY,IAAI,MAAM,EAAE,CACrH,CAAC;QACF,IAAI,GAAG,CAAC,QAAQ,EAAE,CAAC,QAAQ,CAAC,EAAE,CAAC;YAC7B,KAAK,GAAG,CAAC,MAAM,EAAE,CAAC,KAAK,CAAC,CAAC,SAAS,EAAE,EAAE;gBACpC,OAAO,CAAC,IAAI,CACV,wEAAwE,WAAW,IAAI;oBACrF,SAAS,SAAS,YAAY,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC,OAAO,CAAC,CAAC,CAAC,SAAS,EAAE,CACxE,CAAC;YACJ,CAAC,CAAC,CAAC;QACL,CAAC;IACH,CAAC,CAAC,CAAC;IAEH,IAAI,CAAC;QACH,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,GAAG,CAAC,MAAM,EAAE,EAAE,CAAC;YACvC,IAAI,KAAK,CAAC,aAAa,IAAI,WAAW,EAAE,EAAE,CAAC;gBACzC,KAAK,CAAC,aAAa,GAAG,IAAI,CAAC;gBAC3B,MAAM;YACR,CAAC;YACD,IAAI,KAAK,CAAC,aAAa;gBAAE,MAAM;YAE/B,yDAAyD;YACzD,KAAK,CAAC,aAAa,CAAC,cAAc,EAAE,CAAC;YACrC,IAAI,KAAK,CAAC,IAAI,KAAK,WAAW,IAAI,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;gBACjE,KAAK,CAAC,YAAY,GAAG,KAAK,CAAC,IAAI,CAAC;YAClC,CAAC;YAED,aAAa,EAAE,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,UAAU,CAAC,CAAC;YAE/C,WAAW,CAAC,YAAY,CAAC,KAAK,CAAC,CAAC;YAChC,WAAW,CAAC,YAAY,CAAC,KAAK,CAAC,CAAC;YAEhC,IAAI,KAAK,CAAC,IAAI,KAAK,WAAW,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;gBACxD,WAAW,CAAC,sBAAsB,CAAC,KAAmD,CAAC,CAAC;YAC1F,CAAC;YAED,4EAA4E;YAC5E,yEAAyE;YACzE,sEAAsE;YACtE,sEAAsE;YACtE,0EAA0E;YAC1E,0EAA0E;YAC1E,sEAAsE;YACtE,0EAA0E;YAC1E,sEAAsE;YACtE,sEAAsE;YACtE,0EAA0E;YAC1E,4EAA4E;YAC5E,qEAAqE;YACrE,0EAA0E;YAC1E,yEAAyE;YACzE,yEAAyE;YACzE,IAAI,CAAC,KAAK,CAAC,mBAAmB,IAAI,OAAO,IAAI,CAAC,KAAK,CAAC,iBAAiB,IAAI,KAAK,CAAC,IAAI,KAAK,WAAW,CAAC,EAAE,CAAC;gBACrG,KAAK,CAAC,iBAAiB,GAAG,KAAK,CAAC;gBAChC,MAAM,OAAO,GAAG,MAAM,gBAAgB,CAAC,OAAO,CAAC,CAAC;gBAChD,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;oBACvB,KAAK,CAAC,mBAAmB,GAAG,IAAI,CAAC;oBACjC,OAAO,CAAC,GAAG,CACT,wCAAwC,OAAO,CAAC,MAAM,UAAU;wBAC9D,OAAO,OAAO,CAAC,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,KAAK,4BAA4B;wBACrE,mCAAmC,WAAW,EAAE,CACnD,CAAC;oBACF,IAAI,GAAG,CAAC,QAAQ,EAAE,CAAC,QAAQ,CAAC,EAAE,CAAC;wBAC7B,qEAAqE;wBACrE,2DAA2D;wBAC3D,KAAK,CAAC,mBAAmB,GAAG,GAAG,CAAC,MAAM,EAAE,CAAC,IAAI,CAC3C,GAAG,EAAE,GAAE,CAAC,EACR,CAAC,SAAkB,EAAE,EAAE;4BACrB,OAAO,CAAC,IAAI,CACV,oEAAoE;gCAClE,aAAa,WAAW,IAAI;gCAC5B,SAAS,SAAS,YAAY,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC,OAAO,CAAC,CAAC,CAAC,SAAS,EAAE,CACxE,CAAC;wBACJ,CAAC,CACF,CAAC;oBACJ,CAAC;oBACD,MAAM;gBACR,CAAC;YACH,CAAC;YAED,aAAa,CAAC,gBAAgB,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;YAChD,KAAK,CAAC,UAAU,EAAE,CAAC;YAEnB,IAAI,KAAK,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;gBAC5B,OAAO,CAAC,GAAG,CACT,0CAA0C,WAAW,YAAY,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,EAAE,CACzF,CAAC;gBACF,MAAM,WAAW,GAAG,KAA8C,CAAC;gBACnE,IAAI,WAAW,CAAC,MAAM,KAAK,OAAO,IAAI,WAAW,CAAC,OAAO,EAAE,CAAC;oBAC1D,KAAK,CAAC,kBAAkB,GAAG,WAAW,CAAC,OAAO,CAAC;gBACjD,CAAC;YACH,CAAC;YAED,MAAM,aAAa,GAAG,4BAA4B,CAChD;gBACE,kBAAkB,EAAE,aAAa,CAAC,OAAO;gBACzC,UAAU,EAAE,WAAW,CAAC,OAAO;gBAC/B,YAAY,EAAE,WAAW,CAAC,OAAO;aAClC,EACD,SAAS,EACT,KAAK,CAAC,UAAU,CACjB,CAAC;YACF,IAAI,gBAAgB,CAAC,QAAQ,EAAE,CAAC;gBAC9B,MAAM,CAAC,cAAc,GAAG,MAAM,CAAC,2BAA2B,EAAE,gBAAgB,CAAC,QAAQ,EAAE,CAAC,CAAC;YAC3F,CAAC;YACD,IAAI,aAAa,EAAE,CAAC;gBAClB,yEAAyE;gBACzE,qEAAqE;gBACrE,yDAAyD;gBACzD,MAAM,CAAC,kBAAkB,GAAG,WAAW,CAAC,kBAAkB,CAAC;gBAC3D,uEAAuE;gBACvE,0EAA0E;gBAC1E,uEAAuE;gBACvE,qBAAqB;gBACrB,IAAI,iBAAiB,EAAE,CAAC;oBACtB,MAAM,yBAAyB,CAAC;wBAC9B,MAAM;wBACN,WAAW;wBACX,SAAS,EAAE,iBAAiB;wBAC5B,KAAK,EAAE,aAAa;qBACrB,CAAC,CAAC;gBACL,CAAC;gBACD,MAAM,MAAM,GAAG,MAAM,OAAO,CAAC,MAAM,CAAC,CAAC;gBACrC,aAAa,CAAC,aAAa,EAAE,CAAC;gBAC9B,WAAW,CAAC,aAAa,EAAE,CAAC;gBAC5B,WAAW,CAAC,aAAa,EAAE,CAAC;gBAC5B,SAAS,CAAC,cAAc,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC;gBAC3C,SAAS,EAAE,CAAC;gBACZ,IAAI,MAAM,KAAK,sBAAsB,CAAC,IAAI,EAAE,CAAC;oBAC3C,KAAK,CAAC,oBAAoB,GAAG,IAAI,CAAC;oBAClC,OAAO,CAAC,IAAI,CAAC,0DAA0D,WAAW,EAAE,CAAC,CAAC;gBACxF,CAAC;YACH,CAAC;YAED,IAAI,KAAK,CAAC,oBAAoB,EAAE,CAAC;gBAC/B,OAAO,CAAC,GAAG,CACT,wEAAwE,WAAW,EAAE,CACtF,CAAC;gBACF,MAAM;YACR,CAAC;QACH,CAAC;IACH,CAAC;IAAC,OAAO,SAAS,EAAE,CAAC;QACnB,6EAA6E;QAC7E,6EAA6E;QAC7E,8EAA8E;QAC9E,wDAAwD;QACxD,IAAI,CAAC,KAAK,CAAC,aAAa,IAAI,CAAC,KAAK,CAAC,mBAAmB;YAAE,MAAM,SAAS,CAAC;QACxE,OAAO,CAAC,IAAI,CACV,oDAAoD,WAAW,IAAI;YACjE,SAAS,KAAK,CAAC,aAAa,iBAAiB,KAAK,CAAC,mBAAmB,EAAE,CAC3E,CAAC;IACJ,CAAC;YAAS,CAAC;QACT,KAAK,CAAC,aAAa,CAAC,IAAI,EAAE,CAAC;IAC7B,CAAC;IAED,8EAA8E;IAC9E,yEAAyE;IACzE,2EAA2E;IAC3E,gEAAgE;IAChE,IAAI,KAAK,CAAC,aAAa;QAAE,OAAO,SAAS,CAAC;IAC1C,IAAI,KAAK,CAAC,oBAAoB;QAAE,OAAO,eAAe,CAAC;IACvD,IAAI,KAAK,CAAC,mBAAmB;QAAE,OAAO,cAAc,CAAC;IACrD,IAAI,KAAK,CAAC,aAAa,IAAI,WAAW,EAAE;QAAE,OAAO,QAAQ,CAAC;IAC1D,OAAO,WAAW,CAAC;AACrB,CAAC"}
|
|
@@ -0,0 +1,61 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Approved-command turn provenance (DD-28) — the DEEP-AGENT (native) harness
|
|
3
|
+
* adapter over the shared qualification rule ({@link qualifyTurnCommandProvenance}).
|
|
4
|
+
*
|
|
5
|
+
* The deep-agent's consent model differs from Cursor's in exactly two ways, and
|
|
6
|
+
* this adapter is precisely those two differences:
|
|
7
|
+
*
|
|
8
|
+
* 1. IDENTITY turn scoping (not positional). Cursor scopes a turn by the message
|
|
9
|
+
* index its stream started at. The deep-agent cannot: an approved command
|
|
10
|
+
* executes IN PLACE at its seeded transcript position (it was proposed in a
|
|
11
|
+
* prior segment, seeded into this segment, and updated to COMPLETED by
|
|
12
|
+
* `StatusBuilder.findResumableSeededToolCall`). So "this turn's calls" are
|
|
13
|
+
* those whose id is absent from a pre-stream {@link collectSettledToolCallIds}
|
|
14
|
+
* snapshot, and "executed" is a COMPLETED status.
|
|
15
|
+
*
|
|
16
|
+
* 2. SAME-ROW direct consent (not a grant token). A gated deep-agent tool pauses
|
|
17
|
+
* on a LangGraph interrupt; SubmitApproval writes `approval_action` onto that
|
|
18
|
+
* same row, and the row keeps its id + `approval_action` when it executes on
|
|
19
|
+
* resume. So the executed command IS its own consent row — cite its own id.
|
|
20
|
+
* (The server's PreserveApprovalFields keeps that `approval_action` on the row
|
|
21
|
+
* across the status write regardless of its new status, so the backend's
|
|
22
|
+
* verification finds it.)
|
|
23
|
+
*
|
|
24
|
+
* A turn that delegated ANY sub-agent fails closed (DD-28 D1: "delegated zero
|
|
25
|
+
* sub-agents"). A sub-agent's writes fold into the parent turn's change set
|
|
26
|
+
* (DD-19) but are not attributable to a consented top-level command. The
|
|
27
|
+
* top-level `task` call (a SUBAGENT kind) already trips the shared rule's
|
|
28
|
+
* `!== SHELL` fail-closed; the explicit sub-agent-novelty guard below also covers
|
|
29
|
+
* the (currently-impossible under the replace-per-turn model) case of a sub-agent
|
|
30
|
+
* that mutates across a turn boundary without a fresh top-level `task` row.
|
|
31
|
+
*
|
|
32
|
+
* (Trust boundary + fail-closed contract: see the shared module.)
|
|
33
|
+
*/
|
|
34
|
+
import type { AgentExecutionStatus } from "@stigmer/protos/ai/stigmer/agentic/agentexecution/v1/api_pb";
|
|
35
|
+
import type { TurnCommandProvenance } from "@stigmer/protos/ai/stigmer/agentic/agentexecution/v1/filereview_pb";
|
|
36
|
+
export interface DeepAgentCommandProvenanceInputs {
|
|
37
|
+
/** The status at capture time — top-level messages + sub-agent executions. */
|
|
38
|
+
readonly status: AgentExecutionStatus;
|
|
39
|
+
/**
|
|
40
|
+
* The top-level tool-call ids that had already SETTLED before this turn's
|
|
41
|
+
* stream (from {@link collectSettledToolCallIds}). A top-level call whose id is
|
|
42
|
+
* absent is this-turn's; the shared rule then keeps only the COMPLETED ones.
|
|
43
|
+
*/
|
|
44
|
+
readonly priorSettledToolCallIds: ReadonlySet<string>;
|
|
45
|
+
/**
|
|
46
|
+
* The sub-agent tool-call ids that existed before this turn's stream (from
|
|
47
|
+
* {@link collectSubAgentToolCallIds}). Any growth means a sub-agent acted this
|
|
48
|
+
* turn → fail closed.
|
|
49
|
+
*/
|
|
50
|
+
readonly priorSubAgentToolCallIds: ReadonlySet<string>;
|
|
51
|
+
/** True when the pre-armed spec.auto_approve_all bypassed the gate. */
|
|
52
|
+
readonly globalBypass: boolean;
|
|
53
|
+
}
|
|
54
|
+
/**
|
|
55
|
+
* Derive the {@link TurnCommandProvenance} for a deep-agent turn, or undefined
|
|
56
|
+
* when it does not qualify. Scopes this turn's top-level calls by id-novelty,
|
|
57
|
+
* resolves consent from the row's own server-authored `approval_action`, fails
|
|
58
|
+
* closed on any sub-agent activity, and delegates the DD-28 rule to
|
|
59
|
+
* {@link qualifyTurnCommandProvenance}.
|
|
60
|
+
*/
|
|
61
|
+
export declare function deriveTurnCommandProvenance(inputs: DeepAgentCommandProvenanceInputs): TurnCommandProvenance | undefined;
|