pi-crew 0.9.64 → 0.9.66
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +31 -0
- package/README.md +46 -1
- package/dist/index.mjs +423 -330
- package/package.json +3 -2
- package/scripts/pty_probe.py +10 -8
- package/skills/real-test-pi-crew/SKILL.md +6 -6
- package/src/config/config.ts +19 -3
- package/src/config/types.ts +2 -0
- package/src/extension/team-tool/cancel.ts +34 -0
- package/src/extension/team-tool/dispatch/manage.ts +7 -4
- package/src/extension/team-tool/explain.ts +3 -1
- package/src/extension/team-tool/lifecycle-actions.ts +4 -1
- package/src/extension/team-tool-types.ts +2 -0
- package/src/observability/event-to-metric.ts +29 -0
- package/src/observability/metrics-primitives.ts +41 -3
- package/src/runtime/README.md +1 -1
- package/src/runtime/broker/crew-broker.ts +0 -16
- package/src/runtime/effectiveness.ts +23 -1
- package/src/runtime/merge-gate.ts +202 -0
- package/src/runtime/model/model-fallback.ts +11 -0
- package/src/runtime/model/provider-extensions.ts +31 -12
- package/src/runtime/output/output-validator.ts +34 -6
- package/src/runtime/output/progress-tracker.ts +3 -33
- package/src/runtime/scheduling/scheduler.ts +67 -19
- package/src/runtime/scratchpad/engine.ts +40 -2
- package/src/runtime/scratchpad/snapshot-hmac.ts +161 -0
- package/src/runtime/team-runner.ts +128 -203
- package/src/schema/team-tool-schema.ts +2 -0
- package/src/teams/discover-teams.ts +2 -0
- package/src/teams/team-config.ts +7 -0
- package/src/teams/team-serializer.ts +1 -0
- package/src/ui/mascot.ts +1 -14
- package/teams/default.team.md +1 -0
- package/teams/fast-fix.team.md +1 -0
- package/src/observability/event-bus.ts +0 -86
- package/src/plugins/plugin-define.ts +0 -6
- package/src/plugins/plugin-registry.ts +0 -32
- package/src/plugins/plugins/index.ts +0 -3
- package/src/plugins/plugins/nextjs.ts +0 -19
- package/src/plugins/plugins/vite.ts +0 -10
- package/src/plugins/plugins/vitest.ts +0 -9
- package/src/runtime/child-pi/child-pi-pool.ts +0 -68
- package/src/runtime/iteration-hooks.ts +0 -305
|
@@ -0,0 +1,202 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Monotonic merge gate for parallel task updates.
|
|
3
|
+
*
|
|
4
|
+
* Extracted from team-runner.ts (2026-08-10, improvement-plan Tier 2
|
|
5
|
+
* "team-runner split" — self-contained portion). The merge gate protects
|
|
6
|
+
* terminal task states from being regressed by stale parallel worker
|
|
7
|
+
* snapshots: every terminal->non-terminal transition is rejected, plus a
|
|
8
|
+
* small set of bespoke policies (P2 completed integrity, P3
|
|
9
|
+
* waiting->running stale-snapshot regression) that are stricter than the
|
|
10
|
+
* lifecycle table on the parallel-merge path.
|
|
11
|
+
*
|
|
12
|
+
* Exhaustively tested in test/unit/team-runner-should-merge-table.test.ts.
|
|
13
|
+
*/
|
|
14
|
+
import { TEAM_TASK_STATUSES, TEAM_TERMINAL_TASK_STATUSES, type TeamTaskStatus } from "../state/contracts.ts";
|
|
15
|
+
import type { TeamTaskState } from "../state/types.ts";
|
|
16
|
+
import { refreshTaskGraphQueues } from "./scheduling/task-graph-scheduler.ts";
|
|
17
|
+
|
|
18
|
+
export function isNonTerminalTaskStatus(status: TeamTaskState["status"]): boolean {
|
|
19
|
+
return status === "queued" || status === "running" || status === "waiting";
|
|
20
|
+
}
|
|
21
|
+
|
|
22
|
+
export function safeFinishedAt(task: TeamTaskState): number {
|
|
23
|
+
if (!task.finishedAt) return -Infinity;
|
|
24
|
+
const ms = new Date(task.finishedAt).getTime();
|
|
25
|
+
return Number.isNaN(ms) ? Infinity : ms;
|
|
26
|
+
}
|
|
27
|
+
|
|
28
|
+
/**
|
|
29
|
+
* Returns true when the current task has a malformed finishedAt (NaN/Infinity)
|
|
30
|
+
* and the updated task has a valid finite finishedAt. Malformed finishedAt
|
|
31
|
+
* should be replaced rather than persisting corruption.
|
|
32
|
+
*/
|
|
33
|
+
export function isMalformedFinishedAtReplacement(currentTime: number, updatedTime: number): boolean {
|
|
34
|
+
return !Number.isFinite(currentTime) && Number.isFinite(updatedTime);
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
/**
|
|
38
|
+
* RT-16: status-level gate for shouldMergeTaskUpdate. Returns the stable
|
|
39
|
+
* "from->to" key used by REJECTED_STATUS_MERGE_TRANSITIONS.
|
|
40
|
+
*/
|
|
41
|
+
export function statusMergeKey(from: TeamTaskStatus, to: TeamTaskStatus): string {
|
|
42
|
+
return `${from}->${to}`;
|
|
43
|
+
}
|
|
44
|
+
|
|
45
|
+
/**
|
|
46
|
+
* RT-16 — derived merge-gate transition table.
|
|
47
|
+
*
|
|
48
|
+
* The set of old->new status pairs that shouldMergeTaskUpdate must REJECT based
|
|
49
|
+
* solely on the status transition (before any field-level comparison). It
|
|
50
|
+
* replaces the former 13 hand-written status guards that hand-duplicated the
|
|
51
|
+
* lifecycle table. Built once from the single source-of-truth table
|
|
52
|
+
* (TEAM_TASK_STATUSES + TEAM_TERMINAL_TASK_STATUSES — the terminal half of
|
|
53
|
+
* TEAM_TASK_STATUS_TRANSITIONS) plus two merge-specific policies that are
|
|
54
|
+
* STRICTER than the lifecycle table on the parallel-merge path:
|
|
55
|
+
*
|
|
56
|
+
* P1 Terminal preservation — every terminal->non-terminal pair is rejected.
|
|
57
|
+
* The lifecycle table permits retries (e.g. completed->queued), but a stale
|
|
58
|
+
* worker snapshot must never resurrect a settled task.
|
|
59
|
+
* P2 Completed integrity — five terminal->terminal flips that touch the
|
|
60
|
+
* "completed" success terminal are rejected: completed->failed,
|
|
61
|
+
* completed->needs_attention, failed->completed, cancelled->completed,
|
|
62
|
+
* needs_attention->completed. (completed may still move to
|
|
63
|
+
* cancelled/skipped; that is intentionally allowed, so these are NOT simply
|
|
64
|
+
* "every illegal terminal->terminal flip".)
|
|
65
|
+
* P3 waiting->running regression — the single stale-snapshot case.
|
|
66
|
+
*
|
|
67
|
+
* The decision for every old->new pair is byte-for-byte identical to the former
|
|
68
|
+
* 7 status guards (verified exhaustively in
|
|
69
|
+
* test/unit/team-runner-should-merge-table.test.ts).
|
|
70
|
+
*/
|
|
71
|
+
export const REJECTED_STATUS_MERGE_TRANSITIONS: ReadonlySet<string> = (() => {
|
|
72
|
+
const rejected = new Set<string>();
|
|
73
|
+
// P1 — terminal preservation: reject every terminal->non-terminal pair.
|
|
74
|
+
for (const from of TEAM_TASK_STATUSES) {
|
|
75
|
+
if (!TEAM_TERMINAL_TASK_STATUSES.has(from)) continue;
|
|
76
|
+
for (const to of TEAM_TASK_STATUSES) {
|
|
77
|
+
if (!TEAM_TERMINAL_TASK_STATUSES.has(to)) rejected.add(statusMergeKey(from, to));
|
|
78
|
+
}
|
|
79
|
+
}
|
|
80
|
+
// P3 — waiting->running stale-snapshot regression.
|
|
81
|
+
rejected.add(statusMergeKey("waiting", "running"));
|
|
82
|
+
// P2 — completed integrity flips (bespoke terminal->terminal policy).
|
|
83
|
+
const completedIntegrityFlips: ReadonlyArray<[TeamTaskStatus, TeamTaskStatus]> = [
|
|
84
|
+
["completed", "failed"],
|
|
85
|
+
["completed", "needs_attention"],
|
|
86
|
+
["failed", "completed"],
|
|
87
|
+
["cancelled", "completed"],
|
|
88
|
+
["needs_attention", "completed"],
|
|
89
|
+
];
|
|
90
|
+
for (const [from, to] of completedIntegrityFlips) rejected.add(statusMergeKey(from, to));
|
|
91
|
+
return rejected;
|
|
92
|
+
})();
|
|
93
|
+
|
|
94
|
+
export function shouldMergeTaskUpdate(current: TeamTaskState, updated: TeamTaskState): boolean {
|
|
95
|
+
// RT-16: status-level gate — reject stale/dangerous transitions via the
|
|
96
|
+
// derived transition table (REJECTED_STATUS_MERGE_TRANSITIONS) instead of
|
|
97
|
+
// hand-written guards. Parallel workers receive the same input snapshot; a
|
|
98
|
+
// later result may still carry stale copies. The table encodes three
|
|
99
|
+
// merge-specific policies stricter than the lifecycle table: terminal
|
|
100
|
+
// preservation (no terminal->non-terminal resurrection), completed integrity
|
|
101
|
+
// (no flipping the "completed" success terminal to/from failed or
|
|
102
|
+
// needs_attention), and the waiting->running stale-snapshot regression.
|
|
103
|
+
if (REJECTED_STATUS_MERGE_TRANSITIONS.has(statusMergeKey(current.status, updated.status))) return false;
|
|
104
|
+
// Guard: when current is "running" but has resultArtifact (another worker already
|
|
105
|
+
// completed it), a stale updated with status="running" and no resultArtifact
|
|
106
|
+
// must not overwrite the actual completed state.
|
|
107
|
+
if (current.status === updated.status && updated.status === "running" && current.resultArtifact && !updated.resultArtifact)
|
|
108
|
+
return false;
|
|
109
|
+
// Guard: when current is "completed" and has resultArtifact but updated is also
|
|
110
|
+
// "completed" without resultArtifact, block the stale update from overwriting
|
|
111
|
+
// a task that successfully produced output.
|
|
112
|
+
if (current.status === updated.status && current.status === "completed" && current.resultArtifact && !updated.resultArtifact)
|
|
113
|
+
return false;
|
|
114
|
+
// Prevent a stale completed task from overwriting a fresher one.
|
|
115
|
+
// Restructure to handle undefined current.finishedAt as a special case:
|
|
116
|
+
// - undefined current + valid updated: allow the update
|
|
117
|
+
// - valid current + undefined updated: block the update (don't lose completion time)
|
|
118
|
+
// - both undefined: finishedAt guard does not apply, fall through to heartbeat check
|
|
119
|
+
// - both valid: compare timestamps as before
|
|
120
|
+
if (current.finishedAt !== undefined && updated.finishedAt !== undefined) {
|
|
121
|
+
const currentTime = safeFinishedAt(current);
|
|
122
|
+
const updatedTime = safeFinishedAt(updated);
|
|
123
|
+
// Malformed finishedAt (NaN) is treated as Infinity — invalid state should be
|
|
124
|
+
// replaced rather than persisting corruption. Log warning for visibility.
|
|
125
|
+
if (!Number.isFinite(currentTime)) {
|
|
126
|
+
console.warn(`[merge-gate] Task ${current.id} has malformed finishedAt: ${current.finishedAt}`);
|
|
127
|
+
}
|
|
128
|
+
if (isMalformedFinishedAtReplacement(currentTime, updatedTime)) {
|
|
129
|
+
return true;
|
|
130
|
+
}
|
|
131
|
+
if (updatedTime < currentTime) return false;
|
|
132
|
+
}
|
|
133
|
+
// Block if updated is trying to establish a terminal status without a finishedAt
|
|
134
|
+
// timestamp. Heartbeat-only updates (status='running', no finishedAt) are
|
|
135
|
+
// allowed if heartbeat has changed (checked separately in hasMeaningfulUpdate).
|
|
136
|
+
if (!updated.finishedAt && !isNonTerminalTaskStatus(updated.status)) return false;
|
|
137
|
+
// Explicitly enumerate all fields that constitute a meaningful update so that
|
|
138
|
+
// adding a new important field requires updating this list (rather than silently
|
|
139
|
+
// losing data if a field is forgotten in the boolean OR chain below).
|
|
140
|
+
const hasMeaningfulUpdate =
|
|
141
|
+
updated.status !== current.status ||
|
|
142
|
+
updated.finishedAt !== current.finishedAt ||
|
|
143
|
+
updated.startedAt !== current.startedAt ||
|
|
144
|
+
Boolean(updated.resultArtifact) !== Boolean(current.resultArtifact) ||
|
|
145
|
+
(Boolean(updated.resultArtifact) && updated.resultArtifact !== current.resultArtifact) ||
|
|
146
|
+
Boolean(updated.error) ||
|
|
147
|
+
Boolean(updated.modelAttempts?.length) ||
|
|
148
|
+
Boolean(updated.usage) ||
|
|
149
|
+
Boolean(updated.attempts?.length) ||
|
|
150
|
+
updated.heartbeat?.lastSeenAt !== current.heartbeat?.lastSeenAt ||
|
|
151
|
+
updated.jsonEvents !== current.jsonEvents ||
|
|
152
|
+
updated.agentProgress?.lastActivityAt !== current.agentProgress?.lastActivityAt;
|
|
153
|
+
return hasMeaningfulUpdate;
|
|
154
|
+
}
|
|
155
|
+
/** Exposed for the exhaustive status-merge table test (RT-16). */
|
|
156
|
+
export const __test__shouldMergeTaskUpdate = shouldMergeTaskUpdate;
|
|
157
|
+
|
|
158
|
+
// H4 fix: rename to descriptive name. Kept __test__ as alias for backward
|
|
159
|
+
// compat test imports.
|
|
160
|
+
// FIX (perf P10): replace O(N×M) .find() + .map() inside nested loops with a
|
|
161
|
+
// single-pass Map-based merge. Build an index of `merged` once, then for each
|
|
162
|
+
// incoming updated task do O(1) lookup; the final pass reassembles `merged`
|
|
163
|
+
// preserving original order. For a 20-task run × 5-batch merger with
|
|
164
|
+
// ~10 updates per result, this reduces from O(50×20) = 1000 ops to O(120).
|
|
165
|
+
// Behavior is unchanged: skipped updates (shouldMergeTaskUpdate=false) still
|
|
166
|
+
// leave the existing task in place.
|
|
167
|
+
export function mergeTaskUpdatesPreservingTerminal(base: TeamTaskState[], results: Array<{ tasks: TeamTaskState[] }>): TeamTaskState[] {
|
|
168
|
+
// Index current merged state by id for O(1) lookup during the merge pass.
|
|
169
|
+
const indexById = new Map<string, TeamTaskState>();
|
|
170
|
+
for (const task of base) indexById.set(task.id, task);
|
|
171
|
+
|
|
172
|
+
let skipped = 0;
|
|
173
|
+
for (const result of results) {
|
|
174
|
+
for (const updated of result.tasks) {
|
|
175
|
+
const current = indexById.get(updated.id);
|
|
176
|
+
if (!current) continue;
|
|
177
|
+
if (!shouldMergeTaskUpdate(current, updated)) {
|
|
178
|
+
// Log skipped merges for visibility into rejected parallel updates.
|
|
179
|
+
// In distributed systems with parallel workers, rejected merges may
|
|
180
|
+
// indicate bugs (wrong status, timestamp corruption) if they accumulate.
|
|
181
|
+
console.debug("[merge-gate] Skipping stale merge for task", updated.id, {
|
|
182
|
+
currentStatus: current.status,
|
|
183
|
+
updatedStatus: updated.status,
|
|
184
|
+
currentFinishedAt: current.finishedAt,
|
|
185
|
+
updatedFinishedAt: updated.finishedAt,
|
|
186
|
+
});
|
|
187
|
+
skipped += 1;
|
|
188
|
+
continue;
|
|
189
|
+
}
|
|
190
|
+
indexById.set(updated.id, updated);
|
|
191
|
+
}
|
|
192
|
+
}
|
|
193
|
+
// Reassemble in original `base` order so downstream snapshots stay stable.
|
|
194
|
+
const merged = base.map((task) => indexById.get(task.id) ?? task);
|
|
195
|
+
// `skipped` is intentional visibility — currently no caller reads it but
|
|
196
|
+
// we'd rather leave the count available for future instrumentation than
|
|
197
|
+
// remove the cumulative silent-rejection signal it provides.
|
|
198
|
+
void skipped;
|
|
199
|
+
return refreshTaskGraphQueues(merged);
|
|
200
|
+
}
|
|
201
|
+
/** @deprecated Use mergeTaskUpdatesPreservingTerminal. Kept for backward test import compat. */
|
|
202
|
+
export const __test__mergeTaskUpdates = mergeTaskUpdatesPreservingTerminal;
|
|
@@ -365,6 +365,17 @@ const RETRYABLE_MODEL_FAILURE_PATTERNS = [
|
|
|
365
365
|
/safety/i,
|
|
366
366
|
/is[_ ]?overloaded/i,
|
|
367
367
|
/\b408\b/,
|
|
368
|
+
//
|
|
369
|
+
// EPIPE / broken-pipe. In the child-pi worker path this typically means
|
|
370
|
+
// the child `pi` process exited (crash or early exit) while the parent
|
|
371
|
+
// was still writing to its stdin — spawning a fresh child on the next
|
|
372
|
+
// model in the fallback chain usually recovers. In the network path it
|
|
373
|
+
// is a transient pipe close. Both are retryable on a different model.
|
|
374
|
+
// See docs/failure-mode-inventory.md EPIPE gap; NON_RETRYABLE patterns
|
|
375
|
+
// (auth/billing) are checked first, so an auth error mentioning EPIPE
|
|
376
|
+
// stays non-retryable.
|
|
377
|
+
/epipe/i,
|
|
378
|
+
/broken pipe/i,
|
|
368
379
|
];
|
|
369
380
|
|
|
370
381
|
// These patterns indicate auth/key/billing issues that will never succeed on retry.
|
|
@@ -22,7 +22,7 @@
|
|
|
22
22
|
*/
|
|
23
23
|
import * as fs from "node:fs";
|
|
24
24
|
import * as path from "node:path";
|
|
25
|
-
import { userPiRoot } from "../../utils/paths.ts";
|
|
25
|
+
import { packageRoot, userPiRoot } from "../../utils/paths.ts";
|
|
26
26
|
|
|
27
27
|
export interface DiscoveredProviderExtension {
|
|
28
28
|
/** Package specifier as written in settings.json packages (e.g. "npm:pi-commandcode-provider"). */
|
|
@@ -95,12 +95,17 @@ function resolvePackageEntry(pkgDir: string): string | undefined {
|
|
|
95
95
|
|
|
96
96
|
/**
|
|
97
97
|
* Discover provider extension entry points from Pi's installed package registry.
|
|
98
|
-
* Reads `~/.pi/agent/settings.json` → `packages`
|
|
99
|
-
*
|
|
100
|
-
*
|
|
98
|
+
* Reads `~/.pi/agent/settings.json` → `packages` and resolves each spec:
|
|
99
|
+
* - `npm:<name>` → `~/.pi/agent/npm/node_modules/<name>/`
|
|
100
|
+
* - local path spec → resolved relative to the settings.json dir (the way
|
|
101
|
+
* `pi install <local-path>` records them), e.g. "../../src/foo"
|
|
102
|
+
* - git:/file: specs → skipped (not resolvable on disk)
|
|
101
103
|
*
|
|
102
|
-
*
|
|
103
|
-
*
|
|
104
|
+
* Both npm: and local-path specs are SANCTIONED channels — the user wrote them
|
|
105
|
+
* into settings.json (directly or via `pi install`), so they are trusted at the
|
|
106
|
+
* same level. This is distinct from project-sourced AGENT extensions
|
|
107
|
+
* (`.crew/agents/*.md` `extensions:` frontmatter), which are repo-adjacent
|
|
108
|
+
* untrusted data and stay gated by SEC-1 in discover-agents.ts.
|
|
104
109
|
*/
|
|
105
110
|
export function discoverProviderExtensions(settingsPath?: string): DiscoveredProviderExtension[] {
|
|
106
111
|
const root = userPiRoot();
|
|
@@ -125,13 +130,27 @@ export function discoverProviderExtensions(settingsPath?: string): DiscoveredPro
|
|
|
125
130
|
const npmBase = path.join(baseDir, "npm", "node_modules");
|
|
126
131
|
for (const spec of settings.packages ?? []) {
|
|
127
132
|
if (typeof spec !== "string") continue;
|
|
128
|
-
|
|
129
|
-
|
|
130
|
-
|
|
131
|
-
|
|
132
|
-
|
|
133
|
-
|
|
133
|
+
let pkgDir: string;
|
|
134
|
+
if (spec.startsWith("npm:")) {
|
|
135
|
+
// Scoped packages: "@scope/name" → "@scope/name"; plain: "name".
|
|
136
|
+
pkgDir = path.join(npmBase, spec.slice(4));
|
|
137
|
+
} else if (spec.startsWith("./") || spec.startsWith("../") || path.isAbsolute(spec)) {
|
|
138
|
+
// Local path spec — resolve relative to the settings.json dir, matching
|
|
139
|
+
// how `pi install <local-path>` records it. Same trust level as npm:
|
|
140
|
+
// (user wrote it into settings.json). Not to be confused with project
|
|
141
|
+
// AGENT extensions (.crew/agents/* frontmatter) — those stay SEC-1 gated.
|
|
142
|
+
pkgDir = path.resolve(baseDir, spec);
|
|
143
|
+
} else {
|
|
144
|
+
// git:/file:/http: specs etc. — not resolvable on disk, skip.
|
|
145
|
+
continue;
|
|
146
|
+
}
|
|
134
147
|
if (!fs.existsSync(pkgDir)) continue;
|
|
148
|
+
// Skip self: pi-crew's own package is a settings package (the orchestrator
|
|
149
|
+
// extension the parent loads), but a child WORKER must not re-load it — it
|
|
150
|
+
// would register the team tool / observability / MCP wiring intended for
|
|
151
|
+
// the orchestrator process, not a worker. Provider + adapter extensions
|
|
152
|
+
// (pi-other-provider, pi-mcp-adapter, pi-rlm, ...) stay.
|
|
153
|
+
if (path.resolve(pkgDir) === path.resolve(packageRoot())) continue;
|
|
135
154
|
const entryPath = resolvePackageEntry(pkgDir);
|
|
136
155
|
if (entryPath) out.push({ spec, entryPath });
|
|
137
156
|
}
|
|
@@ -9,13 +9,41 @@
|
|
|
9
9
|
* (headings, code blocks, URLs) after compression.
|
|
10
10
|
*/
|
|
11
11
|
|
|
12
|
-
/**
|
|
12
|
+
/**
|
|
13
|
+
* Why relax: real worker LLMs emit markdown handoffs (`## Handoff`,
|
|
14
|
+
* `### Summary`, `## Follow-ups`, `- bullet`, `**bold**`) instead of the
|
|
15
|
+
* caveman formats (`file:line — text`, `PASS:`/`FAIL:`, emoji findings)
|
|
16
|
+
* these patterns originally required. Each role pattern below ORs its
|
|
17
|
+
* strict contract with a markdown-structured alternation so structured
|
|
18
|
+
* handoffs validate while empty/garbage output still fails.
|
|
19
|
+
*
|
|
20
|
+
* What changed: added the MARKDOWN_STRUCTURED alternation to every role.
|
|
21
|
+
* What is preserved: the strict patterns still match verbatim, and the
|
|
22
|
+
* structural-preservation checks (code blocks, URLs, headings) in
|
|
23
|
+
* validateWorkerOutput are UNCHANGED.
|
|
24
|
+
*/
|
|
25
|
+
|
|
26
|
+
/** Accepts atx headings (`## X`), bold (`**x**`), bullets (`- x` / `* x`), and numbered lists (`1. x`) */
|
|
27
|
+
const MARKDOWN_STRUCTURED = /^(?:#{1,6}\s|\*\*|[-*]\s|\d+\.\s)/m;
|
|
28
|
+
|
|
29
|
+
/** Strict per-role contract patterns (kept verbatim; `.source` is embedded in the alternations below) */
|
|
30
|
+
const STRICT_ROLE_PATTERNS: Record<string, RegExp> = {
|
|
31
|
+
explorer: /^(\S+:\d+|Defs:|Refs:|Callers:|Tests:|Sites:|No match\.|totals:)/m,
|
|
32
|
+
executor: /^(\S+:\d+(-\d+)? — .{1,80}\.|verified:|too-big\.|needs-confirm\.|ambiguous\.|regressed\.)/m,
|
|
33
|
+
reviewer: /^([^:\s]+:\d+:\s+\p{Emoji_Presentation}|No issues\.|totals:)/mu,
|
|
34
|
+
"security-reviewer": /^([^:\s]+:\d+:\s+\p{Emoji_Presentation}|No issues\.|totals:)/mu,
|
|
35
|
+
verifier: /^(PASS:|FAIL:)/m,
|
|
36
|
+
};
|
|
37
|
+
|
|
38
|
+
/** Role-specific output format patterns — constructed fresh per call to avoid /g lastIndex leak.
|
|
39
|
+
* Each factory: strict contract OR markdown-structured alternation. */
|
|
13
40
|
const ROLE_PATTERN_DEFS: Record<string, () => RegExp> = {
|
|
14
|
-
explorer: () =>
|
|
15
|
-
executor: () =>
|
|
16
|
-
reviewer: () =>
|
|
17
|
-
"security-reviewer": () =>
|
|
18
|
-
|
|
41
|
+
explorer: () => new RegExp(`(?:${STRICT_ROLE_PATTERNS.explorer.source})|(?:${MARKDOWN_STRUCTURED.source})`, "m"),
|
|
42
|
+
executor: () => new RegExp(`(?:${STRICT_ROLE_PATTERNS.executor.source})|(?:${MARKDOWN_STRUCTURED.source})`, "m"),
|
|
43
|
+
reviewer: () => new RegExp(`(?:${STRICT_ROLE_PATTERNS.reviewer.source})|(?:${MARKDOWN_STRUCTURED.source})`, "mu"),
|
|
44
|
+
"security-reviewer": () =>
|
|
45
|
+
new RegExp(`(?:${STRICT_ROLE_PATTERNS["security-reviewer"].source})|(?:${MARKDOWN_STRUCTURED.source})`, "mu"),
|
|
46
|
+
verifier: () => new RegExp(`(?:${STRICT_ROLE_PATTERNS.verifier.source})|(?:${MARKDOWN_STRUCTURED.source})`, "m"),
|
|
19
47
|
};
|
|
20
48
|
|
|
21
49
|
/** Fresh RegExp factories for structural preservation checks (avoids /g lastIndex leak) */
|
|
@@ -1,5 +1,4 @@
|
|
|
1
1
|
import type { AgentSessionEvent } from "@earendil-works/pi-coding-agent";
|
|
2
|
-
import { crewEventBus } from "../../observability/event-bus.ts";
|
|
3
2
|
|
|
4
3
|
export interface AgentProgress {
|
|
5
4
|
toolCalls: number;
|
|
@@ -25,7 +24,7 @@ export class ProgressTracker {
|
|
|
25
24
|
subscribe: (listener: (event: AgentSessionEvent) => void) => () => void;
|
|
26
25
|
},
|
|
27
26
|
agentId: string,
|
|
28
|
-
|
|
27
|
+
_runId: string,
|
|
29
28
|
): AgentProgress {
|
|
30
29
|
if (this.sessions.has(agentId)) {
|
|
31
30
|
return this.sessions.get(agentId)!.progress;
|
|
@@ -42,26 +41,19 @@ export class ProgressTracker {
|
|
|
42
41
|
};
|
|
43
42
|
|
|
44
43
|
const unsubscribe = session.subscribe((event: AgentSessionEvent) => {
|
|
45
|
-
this.handleEvent(event, progress
|
|
44
|
+
this.handleEvent(event, progress);
|
|
46
45
|
});
|
|
47
46
|
|
|
48
47
|
this.sessions.set(agentId, { unsubscribe, progress });
|
|
49
48
|
return progress;
|
|
50
49
|
}
|
|
51
50
|
|
|
52
|
-
private handleEvent(event: AgentSessionEvent, progress: AgentProgress
|
|
51
|
+
private handleEvent(event: AgentSessionEvent, progress: AgentProgress): void {
|
|
53
52
|
switch (event.type) {
|
|
54
53
|
case "tool_execution_start":
|
|
55
54
|
progress.toolCalls++;
|
|
56
55
|
progress.currentTool = event.toolName;
|
|
57
56
|
progress.toolStartTime = Date.now();
|
|
58
|
-
crewEventBus.emit({
|
|
59
|
-
type: "agent:progress",
|
|
60
|
-
runId,
|
|
61
|
-
agentId,
|
|
62
|
-
payload: { ...progress },
|
|
63
|
-
timestamp: Date.now(),
|
|
64
|
-
});
|
|
65
57
|
break;
|
|
66
58
|
|
|
67
59
|
case "tool_execution_end":
|
|
@@ -69,21 +61,7 @@ export class ProgressTracker {
|
|
|
69
61
|
progress.toolStartTime = null;
|
|
70
62
|
if (event.isError) {
|
|
71
63
|
progress.errors.push(String(event.result ?? "Unknown error"));
|
|
72
|
-
crewEventBus.emit({
|
|
73
|
-
type: "agent:error",
|
|
74
|
-
runId,
|
|
75
|
-
agentId,
|
|
76
|
-
payload: String(event.result ?? "Unknown error"),
|
|
77
|
-
timestamp: Date.now(),
|
|
78
|
-
});
|
|
79
64
|
}
|
|
80
|
-
crewEventBus.emit({
|
|
81
|
-
type: "agent:progress",
|
|
82
|
-
runId,
|
|
83
|
-
agentId,
|
|
84
|
-
payload: { ...progress },
|
|
85
|
-
timestamp: Date.now(),
|
|
86
|
-
});
|
|
87
65
|
break;
|
|
88
66
|
|
|
89
67
|
case "turn_start":
|
|
@@ -92,13 +70,6 @@ export class ProgressTracker {
|
|
|
92
70
|
|
|
93
71
|
case "agent_end":
|
|
94
72
|
progress.status = "completed";
|
|
95
|
-
crewEventBus.emit({
|
|
96
|
-
type: "agent:complete",
|
|
97
|
-
runId,
|
|
98
|
-
agentId,
|
|
99
|
-
payload: { ...progress },
|
|
100
|
-
timestamp: Date.now(),
|
|
101
|
-
});
|
|
102
73
|
break;
|
|
103
74
|
|
|
104
75
|
case "agent_start":
|
|
@@ -120,5 +91,4 @@ export class ProgressTracker {
|
|
|
120
91
|
}
|
|
121
92
|
}
|
|
122
93
|
|
|
123
|
-
// Export singleton instance
|
|
124
94
|
export const globalProgressTracker = new ProgressTracker();
|
|
@@ -229,24 +229,72 @@ function parseIntervalMs(s: string): number | undefined {
|
|
|
229
229
|
return ms;
|
|
230
230
|
}
|
|
231
231
|
|
|
232
|
-
|
|
233
|
-
|
|
234
|
-
|
|
235
|
-
|
|
232
|
+
/** Named-token maps for cron DOW (SUN=0..SAT=6) and month (JAN=1..DEC=12). */
|
|
233
|
+
const CRON_DOW_NAMES: Record<string, number> = { SUN: 0, MON: 1, TUE: 2, WED: 3, THU: 4, FRI: 5, SAT: 6 };
|
|
234
|
+
const CRON_MONTH_NAMES: Record<string, number> = {
|
|
235
|
+
JAN: 1,
|
|
236
|
+
FEB: 2,
|
|
237
|
+
MAR: 3,
|
|
238
|
+
APR: 4,
|
|
239
|
+
MAY: 5,
|
|
240
|
+
JUN: 6,
|
|
241
|
+
JUL: 7,
|
|
242
|
+
AUG: 8,
|
|
243
|
+
SEP: 9,
|
|
244
|
+
OCT: 10,
|
|
245
|
+
NOV: 11,
|
|
246
|
+
DEC: 12,
|
|
247
|
+
};
|
|
236
248
|
|
|
237
|
-
|
|
238
|
-
|
|
239
|
-
|
|
240
|
-
|
|
241
|
-
|
|
242
|
-
|
|
243
|
-
|
|
249
|
+
/**
|
|
250
|
+
* Match a single cron field value against a cron field expression.
|
|
251
|
+
* Supports the standard cron grammar: wildcard, single N, range a-b, list a,b,c,
|
|
252
|
+
* step syntax (wildcard-step, range-step, from-step), and named tokens
|
|
253
|
+
* (MON, JAN) when a `names` map is passed. Fixes the prior matcher that
|
|
254
|
+
* rejected step values and named DOW (parseInt returned NaN).
|
|
255
|
+
*/
|
|
256
|
+
function cronFieldMatches(value: number, field: string, min: number, max: number, names?: Record<string, number>): boolean {
|
|
257
|
+
let normalized = field.trim().toUpperCase();
|
|
258
|
+
if (names) {
|
|
259
|
+
for (const name of Object.keys(names).sort((a, b) => b.length - a.length)) {
|
|
260
|
+
normalized = normalized.split(name).join(String(names[name]));
|
|
261
|
+
}
|
|
262
|
+
}
|
|
263
|
+
// Cron permits 7 for Sunday in the DOW field — normalize to 0.
|
|
264
|
+
if (min === 0 && max === 6) normalized = normalized.replace(/\b7\b/g, "0");
|
|
265
|
+
const matched = new Set<number>();
|
|
266
|
+
for (const rawPart of normalized.split(",")) {
|
|
267
|
+
const part = rawPart.trim();
|
|
268
|
+
if (part === "") return false;
|
|
269
|
+
const stepMatch = part.match(/^(.*)\/(\d+)$/);
|
|
270
|
+
const step = stepMatch ? Number.parseInt(stepMatch[2], 10) : 1;
|
|
271
|
+
if (!Number.isFinite(step) || step < 1) return false;
|
|
272
|
+
const rangeStr = stepMatch ? stepMatch[1] : part;
|
|
273
|
+
let lo: number;
|
|
274
|
+
let hi: number;
|
|
275
|
+
if (rangeStr === "*") {
|
|
276
|
+
lo = min;
|
|
277
|
+
hi = max;
|
|
278
|
+
} else if (/^\d+$/.test(rangeStr)) {
|
|
279
|
+
lo = Number.parseInt(rangeStr, 10);
|
|
280
|
+
hi = stepMatch ? max : lo; // bare `N` = single value; `N/S` = N..max step S
|
|
281
|
+
} else {
|
|
282
|
+
const rm = rangeStr.match(/^(\d+)-(\d+)$/);
|
|
283
|
+
if (!rm) return false;
|
|
284
|
+
lo = Number.parseInt(rm[1], 10);
|
|
285
|
+
hi = Number.parseInt(rm[2], 10);
|
|
244
286
|
}
|
|
245
|
-
|
|
246
|
-
|
|
287
|
+
for (let v = lo; v <= hi; v += step) {
|
|
288
|
+
if (v >= min && v <= max) matched.add(v);
|
|
247
289
|
}
|
|
248
|
-
return false;
|
|
249
290
|
}
|
|
291
|
+
return matched.has(value);
|
|
292
|
+
}
|
|
293
|
+
|
|
294
|
+
function nextCronDate(spec: string, from: Date): Date | { error: string } | null {
|
|
295
|
+
const parts = spec.split(/\s+/);
|
|
296
|
+
if (parts.length < 5) return { error: "Invalid cron expression" };
|
|
297
|
+
const [minStr, hourStr, domStr, monthStr, dowStr] = parts;
|
|
250
298
|
|
|
251
299
|
let cursor = new Date(from.getTime());
|
|
252
300
|
cursor.setSeconds(0, 0);
|
|
@@ -261,11 +309,11 @@ function nextCronDate(spec: string, from: Date): Date | { error: string } | null
|
|
|
261
309
|
const dow = cursor.getUTCDay();
|
|
262
310
|
|
|
263
311
|
if (
|
|
264
|
-
|
|
265
|
-
|
|
266
|
-
|
|
267
|
-
|
|
268
|
-
|
|
312
|
+
cronFieldMatches(min, minStr, 0, 59) &&
|
|
313
|
+
cronFieldMatches(hour, hourStr, 0, 23) &&
|
|
314
|
+
cronFieldMatches(dom, domStr, 1, 31) &&
|
|
315
|
+
cronFieldMatches(month, monthStr, 1, 12, CRON_MONTH_NAMES) &&
|
|
316
|
+
cronFieldMatches(dow, dowStr, 0, 6, CRON_DOW_NAMES)
|
|
269
317
|
) {
|
|
270
318
|
return cursor;
|
|
271
319
|
}
|
|
@@ -17,7 +17,7 @@
|
|
|
17
17
|
* grace, truncateWithMarker, the childClosed race guard — is ported 1:1.
|
|
18
18
|
*/
|
|
19
19
|
|
|
20
|
-
import { type ChildProcess, spawn } from "node:child_process";
|
|
20
|
+
import { type ChildProcess, spawn, spawnSync } from "node:child_process";
|
|
21
21
|
import { randomUUID } from "node:crypto";
|
|
22
22
|
import { closeSync, constants as fsConstants, fstatSync, mkdirSync, openSync, readFileSync, renameSync, writeFileSync } from "node:fs";
|
|
23
23
|
import { dirname } from "node:path";
|
|
@@ -198,6 +198,15 @@ export class EngineManager {
|
|
|
198
198
|
...(this.options.env ?? {}),
|
|
199
199
|
[NONCE_ENV]: this.nonce,
|
|
200
200
|
},
|
|
201
|
+
// Run the guest as its own session leader so descendants
|
|
202
|
+
// (cell-spawned subprocesses) share its process group. teardown
|
|
203
|
+
// then uses `process.kill(-pid, ...)` (POSIX) or `taskkill /T`
|
|
204
|
+
// (Windows) to clean up the whole tree — closing the
|
|
205
|
+
// "cell-subprocess orphan on session_shutdown" gap declared in
|
|
206
|
+
// docs/failure-mode-inventory.md (D.4 narrowed).
|
|
207
|
+
// stdio pipes are still owned by the parent; `detached` only
|
|
208
|
+
// affects process-group / kill-on-parent-exit semantics.
|
|
209
|
+
detached: true,
|
|
201
210
|
// fd 3 carries protocol traffic so stdout/stderr stay pure user output.
|
|
202
211
|
stdio: ["pipe", "pipe", "pipe", "pipe"],
|
|
203
212
|
});
|
|
@@ -303,7 +312,36 @@ export class EngineManager {
|
|
|
303
312
|
this.engineState = "shutdown";
|
|
304
313
|
liveEngines.delete(this);
|
|
305
314
|
this.failAllPending(new Error("Engine has been shut down"));
|
|
306
|
-
|
|
315
|
+
// Kill the whole process group / job tree so cell-spawned
|
|
316
|
+
// subprocesses (grandchildren of the host) do not orphan when the
|
|
317
|
+
// session shuts down. Best-effort: fall back to a single-pid kill
|
|
318
|
+
// if the group/tree kill fails (already dead, EPERM, etc.).
|
|
319
|
+
const child = this.child;
|
|
320
|
+
const pid = child?.pid;
|
|
321
|
+
if (pid !== undefined && pid > 0) {
|
|
322
|
+
try {
|
|
323
|
+
if (process.platform === "win32") {
|
|
324
|
+
// taskkill /T /F kills the entire process tree rooted at pid.
|
|
325
|
+
spawnSync("taskkill", ["/PID", String(pid), "/T", "/F"], { stdio: "ignore" });
|
|
326
|
+
} else {
|
|
327
|
+
// Negative pid = signal the entire process group.
|
|
328
|
+
// Requires the guest to be its own session leader (detached:true above).
|
|
329
|
+
process.kill(-pid, "SIGKILL");
|
|
330
|
+
}
|
|
331
|
+
} catch {
|
|
332
|
+
try {
|
|
333
|
+
child?.kill("SIGKILL");
|
|
334
|
+
} catch {
|
|
335
|
+
/* already gone — nothing to clean up */
|
|
336
|
+
}
|
|
337
|
+
}
|
|
338
|
+
} else if (child) {
|
|
339
|
+
try {
|
|
340
|
+
child.kill("SIGKILL");
|
|
341
|
+
} catch {
|
|
342
|
+
/* already gone */
|
|
343
|
+
}
|
|
344
|
+
}
|
|
307
345
|
this.child = undefined;
|
|
308
346
|
this.protocolReader?.close();
|
|
309
347
|
this.protocolReader = undefined;
|