pi-pier 0.3.0 → 0.4.1

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.
Files changed (103) hide show
  1. package/README.md +16 -12
  2. package/package.json +3 -3
  3. package/schemas/role-manifest.schema.json +7 -1
  4. package/skills/pi-herdr-workflow/SKILL.md +4 -5
  5. package/src/ansi-text.ts +62 -11
  6. package/src/ask-dialog.ts +308 -0
  7. package/src/ask-user.ts +113 -112
  8. package/src/bootstrap.ts +10 -13
  9. package/src/compact-coordinator.ts +279 -261
  10. package/src/compact-economics-core.ts +77 -145
  11. package/src/config-catalog-core.ts +77 -197
  12. package/src/config-command.ts +271 -64
  13. package/src/cordis-services.ts +64 -0
  14. package/src/dashboard-command.ts +77 -33
  15. package/src/delegation-core.ts +115 -0
  16. package/src/efficiency-config-core.ts +286 -486
  17. package/src/efficiency-store.ts +16 -29
  18. package/src/gc-core.ts +10 -36
  19. package/src/guards.ts +6 -0
  20. package/src/herdr-client.ts +202 -233
  21. package/src/history-store.ts +38 -23
  22. package/src/index-gates.ts +118 -0
  23. package/src/index-locks.ts +2 -7
  24. package/src/index-master.ts +31 -29
  25. package/src/index-notices.ts +28 -15
  26. package/src/index-pipe.ts +31 -12
  27. package/src/index-roles.ts +306 -0
  28. package/src/index-runtime.ts +16 -6
  29. package/src/index-worker.ts +8 -18
  30. package/src/index.ts +561 -570
  31. package/src/jev-client.ts +12 -17
  32. package/src/jev-core.ts +138 -90
  33. package/src/ledger.ts +13 -21
  34. package/src/lock-core.ts +28 -56
  35. package/src/manifest-compose.ts +35 -20
  36. package/src/observation-core.ts +52 -100
  37. package/src/pane-title.ts +19 -38
  38. package/src/panel-open.ts +47 -0
  39. package/src/pi-surface.ts +34 -22
  40. package/src/pier-options.ts +53 -50
  41. package/src/pipe-channel.ts +44 -59
  42. package/src/platform-paths.ts +19 -52
  43. package/src/{core → plugins}/grid-shape.ts +8 -10
  44. package/src/{core → plugins}/heat-plan.ts +20 -22
  45. package/src/plugins/observation.ts +475 -0
  46. package/src/plugins/subagent.ts +766 -0
  47. package/src/{core → plugins}/terminal.ts +198 -139
  48. package/src/{core → plugins}/todo.ts +73 -134
  49. package/src/progress-core.ts +16 -53
  50. package/src/reconcile-core.ts +24 -33
  51. package/src/reducer-core.ts +26 -64
  52. package/src/reducer-invoker.ts +266 -239
  53. package/src/renderers.ts +108 -84
  54. package/src/role-loader.ts +33 -28
  55. package/src/role-manifest.ts +25 -31
  56. package/src/role-state.ts +119 -0
  57. package/src/roles/worker-default.json +3 -3
  58. package/src/routing-telemetry.ts +136 -0
  59. package/src/runtime-policy.ts +25 -40
  60. package/src/session-tail.ts +65 -41
  61. package/src/settle-wake-core.ts +61 -22
  62. package/src/slim-frame.ts +25 -122
  63. package/src/stale-core.ts +17 -24
  64. package/src/storage-layout.ts +14 -43
  65. package/src/subagent-core.ts +423 -120
  66. package/src/subagent-gc.ts +109 -67
  67. package/src/subagent-poller.ts +383 -24
  68. package/src/subagent-session.ts +516 -0
  69. package/src/subagent-spawn.ts +577 -52
  70. package/src/swallow.ts +7 -17
  71. package/src/terminal-core.ts +187 -134
  72. package/src/todo-core.ts +57 -109
  73. package/src/todo-read-hook.ts +46 -39
  74. package/src/todo-reminder-core.ts +26 -51
  75. package/src/todo-window.ts +11 -12
  76. package/src/todos-service.ts +15 -35
  77. package/src/tool-error.ts +6 -13
  78. package/src/tool-gate.ts +38 -21
  79. package/src/tree-view.ts +162 -0
  80. package/src/vocab.ts +52 -25
  81. package/src/worker-marker.ts +76 -0
  82. package/src/ask-multi.ts +0 -260
  83. package/src/config-guide.ts +0 -361
  84. package/src/core/observation.ts +0 -555
  85. package/src/core/subagent.ts +0 -633
  86. package/src/efficiency-config-io.ts +0 -53
  87. package/src/focus-poller.ts +0 -244
  88. package/src/git-adapter.ts +0 -113
  89. package/src/notice-buffer.ts +0 -23
  90. package/src/subagent-execute.ts +0 -328
  91. package/src/subagent-git-io.ts +0 -80
  92. package/src/subagent-launch.ts +0 -114
  93. package/src/subagent-list-action.ts +0 -69
  94. package/src/subagent-output-action.ts +0 -104
  95. package/src/subagent-output-adapter.ts +0 -27
  96. package/src/subagent-output-core.ts +0 -263
  97. package/src/subagent-poll-loop.ts +0 -305
  98. package/src/subagent-port.ts +0 -27
  99. package/src/subagent-ready.ts +0 -89
  100. package/src/subagent-registry.ts +0 -70
  101. package/src/subagent-resolution.ts +0 -41
  102. package/src/subagent-scope.ts +0 -41
  103. package/src/subagent-session-io.ts +0 -230
@@ -1,58 +1,284 @@
1
1
  /**
2
- * Task-tab pane spawn, launch-line construction, and trust-flag decisions.
3
- *
4
- * Why: spawn mixed herdr layout I/O with the subagent tool body. Placement
5
- * races (D26 mutex) and D86 worktree grouping belong in one adapter.
2
+ * Subagent spawn domain: git/worktree I/O, tab placement, launch line, readiness wait, and the `spawn`
3
+ * action; placement is serialized by a mutex (concurrent spawns raced on tab lookup).
6
4
  */
5
+ import { execFile as nodeExecFile } from 'node:child_process';
6
+ import { randomUUID } from 'node:crypto';
7
+ import { mkdirSync, stat as statCb } from 'node:fs';
8
+ import { basename, dirname, join, resolve as pathResolve } from 'node:path';
9
+ import { promisify } from 'node:util';
7
10
  import type { HerdrClientLike } from './herdr-client.ts';
8
- import { pipeNameCandidates, pipeRequest } from './pipe-channel.ts';
11
+ import { delegationEnv } from './delegation-core.ts';
12
+ import { bareSessionId } from './session-tail.ts';
13
+ import { pipeNameCandidates, pipeNameFor, pipeRequest, pipeRequestTo } from './pipe-channel.ts';
14
+ import { composeForRole } from './manifest-compose.ts';
15
+ import { platformPaths } from './platform-paths.ts';
16
+ import { planSpawnProfileRow, type RoutingTelemetryRecord } from './routing-telemetry.ts';
9
17
  import { runtimePolicy } from './runtime-policy.ts';
18
+ import { toolError } from './tool-error.ts';
19
+ import type { HistoryEntry } from './history-store.ts';
10
20
  import {
11
- READY_LIVENESS_SAMPLE_MS,
12
- READY_TAIL_CHARS,
13
- planReadyAttempt,
14
- readyFailureText,
15
- type ReadyFailure,
16
- } from './subagent-ready.ts';
17
- import {
21
+ FOREGROUND_POLL_MS,
18
22
  Semaphore,
23
+ allocateSubSlug,
24
+ buildAliveNotice,
25
+ buildBlockedGateNotice,
26
+ buildIsolatePreamble,
19
27
  buildLaunchLine,
20
28
  buildLaunchParts,
21
29
  classifyWorktreeZone,
30
+ formatSubagentResult,
31
+ formatWorktreeStat,
32
+ isAlive,
22
33
  isPathUnder,
34
+ makeProgressUpdate,
35
+ planForegroundTick,
36
+ planIsolateWorktree,
37
+ planLaunchValidation,
38
+ planShellReady,
23
39
  planTabPlacement,
40
+ sleep,
41
+ subRef,
42
+ type SubEntry,
24
43
  type TabPlacementPlan,
25
44
  type WorktreeZone,
26
45
  } from './subagent-core.ts';
27
- import { parseShapeTree, pickGridSplit, type PaneCell } from './core/grid-shape.ts';
28
- import { fromShapeTree, planSpawnSplitRatio, type LayoutNode } from './core/heat-plan.ts';
29
- import type { GitIo } from './subagent-git-io.ts';
46
+ import { parseShapeTree, pickGridSplit, type PaneCell } from './plugins/grid-shape.ts';
47
+ import { fromShapeTree, planSpawnSplitRatio, type LayoutNode } from './plugins/heat-plan.ts';
48
+ import type { Poller } from './subagent-poller.ts';
49
+ import type { SessionIo } from './subagent-session.ts';
50
+
51
+ const statAsync = promisify(statCb);
52
+
53
+ /** Bound on waiting for a fresh pane's shell to reach its prompt. On timeout the launch line is sent
54
+ * anyway: `waitSubReady` still diagnoses a failed boot, and an unsupported probe must not stall spawn. */
55
+ export const SHELL_READY_TIMEOUT_MS = 2_000;
56
+ const SHELL_READY_POLL_MS = 100;
57
+
58
+ /* ── git adapter ────────────────────────────────────────────────── */
59
+
60
+ export interface GitResult {
61
+ readonly stdout: string;
62
+ readonly stderr: string;
63
+ }
64
+
65
+ export interface GitError extends Error {
66
+ readonly code?: string | number;
67
+ readonly stdout?: string;
68
+ readonly stderr?: string;
69
+ }
70
+
71
+ export type GitExecFile = (
72
+ file: string,
73
+ args: readonly string[],
74
+ options: { timeout: number; encoding: 'utf8'; maxBuffer: number },
75
+ ) => Promise<{ stdout: string; stderr: string }>;
76
+
77
+ /** Output ceiling for one git call: a large checkout blows past execFile's 1MB default (ENOBUFS). */
78
+ const GIT_MAX_BUFFER = 64 * 1024 * 1024;
79
+
80
+ export interface GitAdapter {
81
+ listWorktrees(cwd: string): Promise<GitResult>;
82
+ status(cwd: string): Promise<GitResult>;
83
+ /** Throws a GitError on failure/timeout; callers wanting "null on error" catch at the call site. */
84
+ run(cwd: string, args: readonly string[]): Promise<GitResult>;
85
+ }
86
+
87
+ const execFile = promisify(nodeExecFile);
88
+
89
+ export class NodeGitAdapter implements GitAdapter {
90
+ private readonly gitExecutable: string;
91
+ private readonly timeoutMs: number;
92
+ private readonly exec: GitExecFile;
93
+
94
+ constructor(
95
+ gitExecutable: string = 'git',
96
+ timeoutMs: number = runtimePolicy.gitTimeoutMs,
97
+ exec: GitExecFile = execFile as unknown as GitExecFile,
98
+ ) {
99
+ this.gitExecutable = gitExecutable;
100
+ this.timeoutMs = timeoutMs;
101
+ this.exec = exec;
102
+ }
103
+
104
+ listWorktrees(cwd: string): Promise<GitResult> {
105
+ return this.run(cwd, ['worktree', 'list', '--porcelain']);
106
+ }
107
+
108
+ status(cwd: string): Promise<GitResult> {
109
+ return this.run(cwd, ['status', '--short']);
110
+ }
111
+
112
+ async run(cwd: string, args: readonly string[]): Promise<GitResult> {
113
+ try {
114
+ const { stdout, stderr } = await this.exec(
115
+ this.gitExecutable,
116
+ ['-C', cwd, ...args],
117
+ { timeout: this.timeoutMs, encoding: 'utf8', maxBuffer: GIT_MAX_BUFFER },
118
+ );
119
+ return { stdout, stderr };
120
+ } catch (err) {
121
+ const message = err instanceof Error ? err.message : String(err);
122
+ throw Object.assign(err instanceof Error ? err : new Error(), {
123
+ message: `Git ${args[0] ?? 'command'} failed: ${message}`,
124
+ }) as GitError;
125
+ }
126
+ }
127
+ }
128
+
129
+ export const defaultGitAdapter = new NodeGitAdapter();
130
+
131
+ /* ── git I/O used by spawn, isolate, GC and settlement stat lines ── */
132
+
133
+ const WORKTREES_CACHE_MS = 5000;
134
+
135
+ export interface GitIo {
136
+ listWorktrees(cwd: string): Promise<string[]>;
137
+ runGit(cwd: string, args: string[]): Promise<string | null>;
138
+ worktreeStatLine(entry: SubEntry): Promise<string | null>;
139
+ invalidateWorktreesCache(): void;
140
+ }
141
+
142
+ export function createGitIo(git: GitAdapter = defaultGitAdapter): GitIo {
143
+ let worktreesCache: { at: number; list: string[] } | null = null;
144
+
145
+ async function listWorktrees(cwd: string): Promise<string[]> {
146
+ if (worktreesCache && Date.now() - worktreesCache.at < WORKTREES_CACHE_MS) return worktreesCache.list;
147
+ let list: string[] = [];
148
+ try {
149
+ const { stdout } = await git.listWorktrees(cwd);
150
+ for (const line of String(stdout).split('\n')) {
151
+ const m = /^worktree (.+)$/.exec(line.trim());
152
+ if (m) list.push(m[1]!);
153
+ }
154
+ } catch {
155
+ list = [];
156
+ }
157
+ worktreesCache = { at: Date.now(), list };
158
+ return list;
159
+ }
160
+
161
+ async function runGit(cwd: string, args: string[]): Promise<string | null> {
162
+ try {
163
+ const { stdout } = await git.run(cwd, args);
164
+ return stdout == null ? null : String(stdout);
165
+ } catch {
166
+ return null;
167
+ }
168
+ }
169
+
170
+ function lastStatLine(out: string | null): string | null {
171
+ if (!out) return null;
172
+ const lines = out.replace(/\r/g, '').split('\n').map((l) => l.trim()).filter(Boolean);
173
+ return lines.length ? lines[lines.length - 1]! : null;
174
+ }
175
+
176
+ async function worktreeStatLine(entry: SubEntry): Promise<string | null> {
177
+ const porcelain = await runGit(entry.cwd, ['status', '--porcelain']);
178
+ if (porcelain === null) return null;
179
+ const dirtyCount = porcelain.split('\n').filter((l) => l.trim() !== '').length;
180
+ if (entry.isolate) {
181
+ const statOut = await runGit(entry.cwd, ['diff', '--stat', `${entry.isolate.baseSha}...HEAD`]);
182
+ const commitsOut = await runGit(entry.cwd, ['rev-list', '--count', `${entry.isolate.baseSha}..HEAD`]);
183
+ const commits = commitsOut != null && /^\d+$/.test(commitsOut.trim()) ? Number(commitsOut.trim()) : null;
184
+ return formatWorktreeStat({ branch: entry.isolate.branch, commits, statLine: lastStatLine(statOut), dirtyCount });
185
+ }
186
+ return formatWorktreeStat({
187
+ branch: null,
188
+ commits: null,
189
+ statLine: lastStatLine(await runGit(entry.cwd, ['diff', '--stat', 'HEAD'])),
190
+ dirtyCount,
191
+ });
192
+ }
193
+
194
+ return {
195
+ listWorktrees,
196
+ runGit,
197
+ worktreeStatLine,
198
+ invalidateWorktreesCache: () => { worktreesCache = null; },
199
+ };
200
+ }
201
+
202
+ /* ── readiness wait (pure planners) ─────────────────────────────── */
203
+
204
+ const READY_BASE_INTERVAL_MS = 500;
205
+ const READY_MAX_INTERVAL_MS = 4000;
206
+ /** Liveness/agent sampling is far more expensive than a ping, so it runs on its own cadence. */
207
+ const READY_LIVENESS_SAMPLE_MS = 2500;
208
+ /** Enough of the pane tail to carry a stack trace head into the failure text. */
209
+ const READY_TAIL_CHARS = 1200;
210
+
211
+ interface ReadyFailure {
212
+ readonly paneId: string;
213
+ readonly reason: 'pane-gone' | 'timeout';
214
+ readonly elapsedMs: number;
215
+ readonly timeoutMs: number;
216
+ readonly lastStatus?: string | null;
217
+ readonly tail?: string | null;
218
+ readonly hint?: string | null;
219
+ }
220
+
221
+ type ReadyAttemptPlan =
222
+ | { readonly kind: 'ready' }
223
+ | { readonly kind: 'retry'; readonly delayMs: number }
224
+ | { readonly kind: 'give-up'; readonly reason: 'pane-gone' | 'timeout' };
30
225
 
31
- export interface SpawnEnv {
226
+ /** Decide one readiness attempt. `alive` is tri-state on purpose: an unavailable liveness probe
227
+ * (`null`) must not be read as death, or a slow herdr socket fails spawns that are merely slow. */
228
+ export function planReadyAttempt(opts: {
229
+ elapsedMs: number;
230
+ attempt: number;
231
+ timeoutMs: number;
232
+ alive: boolean | null;
233
+ ready: boolean;
234
+ }): ReadyAttemptPlan {
235
+ if (opts.ready) return { kind: 'ready' };
236
+ if (opts.alive === false) return { kind: 'give-up', reason: 'pane-gone' };
237
+ if (opts.elapsedMs >= opts.timeoutMs) return { kind: 'give-up', reason: 'timeout' };
238
+ return { kind: 'retry', delayMs: readyBackoffMs(opts.attempt) };
239
+ }
240
+
241
+ /** Exponential backoff with a cap: 500ms, 1s, 2s, 4s, 4s, … */
242
+ export function readyBackoffMs(attempt: number, baseMs = READY_BASE_INTERVAL_MS, capMs = READY_MAX_INTERVAL_MS): number {
243
+ const n = Math.max(0, Math.floor(attempt));
244
+ return Math.min(capMs, baseMs * 2 ** n);
245
+ }
246
+
247
+ /** Failure text for the model; the three cases stay distinguishable: pane died (the tail carries the
248
+ * reason), pane alive but never opened its pipe, pane alive and just slower than the timeout. */
249
+ export function readyFailureText(failure: ReadyFailure): string {
250
+ const seconds = Math.round(failure.elapsedMs / 1000);
251
+ const head = failure.reason === 'pane-gone'
252
+ ? `subagent pane ${failure.paneId} exited before its pipe became ready (waited ${seconds}s)`
253
+ : `subagent pane ${failure.paneId} pipe not ready within ${seconds}s (limit ${Math.round(failure.timeoutMs / 1000)}s)`;
254
+ const why = failure.reason === 'pane-gone'
255
+ ? 'The child process is gone — its last output below usually carries the reason.'
256
+ : failure.lastStatus === 'working'
257
+ ? 'The pane is alive and working, so the prompt may simply be slow to boot; retrying the same call is safe.'
258
+ : 'The pane is alive but never registered its pipe — the child pi process may still be starting, or it failed to load the pier extension.';
259
+ const tail = failure.tail ? `\nlast output of ${failure.paneId}:\n${failure.tail}` : '';
260
+ const hint = failure.hint ? `\n${failure.hint}` : '';
261
+ return `${head}\n${why}${hint}${tail}`;
262
+ }
263
+
264
+ /* ── spawner ────────────────────────────────────────────────────── */
265
+
266
+ interface SpawnEnv {
32
267
  paneId: string;
33
268
  tabId: string;
34
269
  workspaceId: string;
35
270
  }
36
271
 
37
- export interface SpawnRuntime {
38
- nodePath: string;
39
- cliPath: string;
40
- extPath: string;
41
- }
42
-
43
- export interface SpawnerHost {
272
+ interface SpawnerHost {
44
273
  client: HerdrClientLike;
45
274
  env: SpawnEnv | null;
46
- runtime: SpawnRuntime;
275
+ runtime: { nodePath: string; cliPath: string; extPath: string };
47
276
  git: GitIo;
48
277
  /** Test override. Production reads `runtimePolicy.readinessTimeoutMs`. */
49
278
  readinessTimeoutMs?: number;
50
279
  }
51
280
 
52
- /** A14: readiness outcome — failures explain themselves instead of collapsing into `false`. */
53
- export type ReadyOutcome =
54
- | { ok: true }
55
- | { ok: false; failure: ReadyFailure; message: string };
281
+ type ReadyOutcome = { ok: true } | { ok: false; message: string };
56
282
 
57
283
  export interface Spawner {
58
284
  spawnPaneInTaskTab(
@@ -64,7 +290,6 @@ export interface Spawner {
64
290
  launchLine(resumeFile?: string | null, roleModel?: string | null, approve?: boolean): string;
65
291
  approveFor(cwd: string, masterCwd: string): Promise<boolean>;
66
292
  waitSubReady(cwd: string, paneId: string): Promise<ReadyOutcome>;
67
- liveTabs(): Promise<Array<{ tabName: string; tabId: string }>>;
68
293
  findExistingPane(sessionFile: string | null): Promise<{ paneId: string; tabId: string } | null>;
69
294
  }
70
295
 
@@ -87,17 +312,26 @@ export function createSpawner(h: SpawnerHost): Spawner {
87
312
  return buildLaunchLine(buildLaunchParts(h.runtime, { resumeFile, roleModel, approve }));
88
313
  }
89
314
 
315
+ /** herdr 0.9.4 refuses over-long commands on a shell that is still starting; `pane.send_text`
316
+ * has no such guard. Poll the foreground process until the shell owns the PTY (or give up). */
317
+ async function waitShellReady(paneId: string): Promise<void> {
318
+ const deadline = Date.now() + SHELL_READY_TIMEOUT_MS;
319
+ while (Date.now() < deadline) {
320
+ const info = await h.client.processInfo(paneId).catch(() => null);
321
+ if (info === null) return; // older server or pane already gone: keep the old best-effort path
322
+ if (planShellReady(info.foreground) === 'ready') return;
323
+ await sleep(SHELL_READY_POLL_MS);
324
+ }
325
+ }
326
+
90
327
  async function approveFor(cwd: string, masterCwd: string): Promise<boolean> {
91
328
  if (isPathUnder(cwd, masterCwd)) return true;
92
329
  const zone = classifyWorktreeZone({ cwd, masterCwd, worktrees: await h.git.listWorktrees(masterCwd) });
93
330
  return zone.zone === 'worktree';
94
331
  }
95
332
 
96
- /**
97
- * A14: wait for the child's pipe with backoff, and give up early when the pane
98
- * itself is gone from pane.list. Screen tail is sampled while the pane exists
99
- * so a crash that leaves a shell still explains itself on timeout.
100
- */
333
+ /** Wait for the child's pipe with backoff; give up early when the pane is gone. The screen tail
334
+ * is sampled while the pane exists so a crash that leaves a shell still explains itself. */
101
335
  async function waitSubReady(cwd: string, paneId: string): Promise<ReadyOutcome> {
102
336
  const names = pipeNameCandidates(cwd, paneId);
103
337
  const startedAt = Date.now();
@@ -110,7 +344,6 @@ export function createSpawner(h: SpawnerHost): Spawner {
110
344
  while (true) {
111
345
  const now = Date.now();
112
346
  const elapsedMs = now - startedAt;
113
- // Liveness/agent sampling is much more expensive than a ping, so it runs on its own cadence.
114
347
  if (now - lastProbeAt >= READY_LIVENESS_SAMPLE_MS || attempt === 0) {
115
348
  lastProbeAt = now;
116
349
  const probe = await probeChildPane(paneId);
@@ -136,7 +369,7 @@ export function createSpawner(h: SpawnerHost): Spawner {
136
369
  if (parts.length > 0) explainHint = `Herdr detection diagnosis: ${parts.join('; ')}`;
137
370
  }
138
371
  } catch {
139
- /* Best effort diagnostics; failures never mask the root error. */
372
+ /* diagnostics never mask the root error */
140
373
  }
141
374
  }
142
375
  const defaultHint = plan.reason === 'timeout' && lastStatus === 'working'
@@ -152,10 +385,10 @@ export function createSpawner(h: SpawnerHost): Spawner {
152
385
  tail,
153
386
  hint,
154
387
  };
155
- return { ok: false, failure, message: readyFailureText(failure) };
388
+ return { ok: false, message: readyFailureText(failure) };
156
389
  }
157
390
  attempt += 1;
158
- await new Promise((r) => setTimeout(r, plan.delayMs));
391
+ await sleep(plan.delayMs);
159
392
  }
160
393
  }
161
394
 
@@ -165,26 +398,22 @@ export function createSpawner(h: SpawnerHost): Spawner {
165
398
  try {
166
399
  const res = await pipeRequest(name, { type: 'ping', id: `ping-${Date.now()}` }, 3000);
167
400
  if (res.type === 'ok') return true;
168
- } catch {
169
- /* not ready on this name */
170
- }
401
+ } catch { /* not ready on this name */ }
171
402
  }
172
403
  return false;
173
404
  }
174
405
 
175
- /** Pane liveness + last visible output; socket failures degrade to "unknown" rather than "dead". */
406
+ /** Pane liveness + last visible output; socket failures degrade to "unknown", never "dead". */
176
407
  async function probeChildPane(paneId: string): Promise<{ alive: boolean | null; status: string | null; tail: string | null }> {
177
408
  let alive: boolean | null = null;
178
409
  let status: string | null = null;
179
410
  try {
180
- // pane.list includes unknown shells; agent.list does not (herdr 0.9.1 live probe).
181
- // A fresh split is absent from agent.list until pi is identified — that is boot, not death.
182
- const panes = await h.client.listPanes();
183
- const pane = panes.find((p) => p.paneId === paneId);
411
+ // pane.list includes unknown shells; agent.list does not. A fresh split is absent from
412
+ // agent.list until pi is identified — that is boot, not death.
413
+ const pane = (await h.client.listPanes()).find((p) => p.paneId === paneId);
184
414
  alive = pane != null;
185
415
  status = pane?.agentStatus ?? null;
186
416
  } catch {
187
- /* keep null: an unreachable socket must not be read as a dead child */
188
417
  }
189
418
  let tail: string | null = null;
190
419
  try {
@@ -205,8 +434,7 @@ export function createSpawner(h: SpawnerHost): Spawner {
205
434
  const release = await tabMutex.acquire();
206
435
  try {
207
436
  const allPanes = await h.client.listPanes();
208
- const mainTabId = allPanes.find((p) => p.paneId === h.env?.paneId)?.tabId
209
- ?? (h.env?.tabId ? h.env.tabId : null);
437
+ const mainTabId = allPanes.find((p) => p.paneId === h.env?.paneId)?.tabId ?? (h.env?.tabId ? h.env.tabId : null);
210
438
  let plan: TabPlacementPlan = planTabPlacement({
211
439
  desiredTab: placement.desiredTab,
212
440
  description: placement.description,
@@ -265,6 +493,7 @@ export function createSpawner(h: SpawnerHost): Spawner {
265
493
  focus: false,
266
494
  ...(ratio != null ? { ratio } : {}),
267
495
  });
496
+ await waitShellReady(paneId);
268
497
  await h.client.sendPaneText(paneId, launch);
269
498
  return { tabId: plan.tabId, paneId, tabName: plan.tabName };
270
499
  }
@@ -276,6 +505,7 @@ export function createSpawner(h: SpawnerHost): Spawner {
276
505
  cwd,
277
506
  env: envOver,
278
507
  });
508
+ await waitShellReady(created.paneId);
279
509
  await h.client.sendPaneText(created.paneId, launch);
280
510
  return { tabId: created.tabId, paneId: created.paneId, tabName: plan.tabName };
281
511
  } finally {
@@ -283,19 +513,314 @@ export function createSpawner(h: SpawnerHost): Spawner {
283
513
  }
284
514
  }
285
515
 
516
+ /** D94: reuse a pane that already runs the same session instead of competing on one transcript. */
286
517
  async function findExistingPane(sessionFile: string | null): Promise<{ paneId: string; tabId: string } | null> {
287
518
  if (!sessionFile) return null;
288
519
  try {
289
520
  const agents = await h.client.listAgents();
290
521
  const match = agents.find((a) => a.session === sessionFile && a.status !== 'unknown');
291
522
  if (!match) return null;
292
- const panes = await h.client.listPanes();
293
- const pane = panes.find((p) => p.paneId === match.paneId);
523
+ const pane = (await h.client.listPanes()).find((p) => p.paneId === match.paneId);
294
524
  return pane ? { paneId: match.paneId, tabId: pane.tabId } : null;
295
525
  } catch {
296
526
  return null;
297
527
  }
298
528
  }
299
529
 
300
- return { spawnPaneInTaskTab, launchLine, approveFor, waitSubReady, liveTabs, findExistingPane };
530
+ return { spawnPaneInTaskTab, launchLine, approveFor, waitSubReady, findExistingPane };
531
+ }
532
+
533
+ /* ── spawn action ───────────────────────────────────────────────── */
534
+
535
+ interface SpawnActionHost {
536
+ client: HerdrClientLike;
537
+ env: SpawnEnv | null;
538
+ subSemaphore: Semaphore;
539
+ subs: Map<string, SubEntry>;
540
+ persistSubs(): void;
541
+ writeHistory(entry: SubEntry, patch?: Partial<HistoryEntry>, via?: string): void;
542
+ pendingIsolateBranches: Set<string>;
543
+ lastMachineInjectAt: Map<string, number>;
544
+ lastRequestIdByPane: Map<string, number | string>;
545
+ git: GitIo;
546
+ session: SessionIo;
547
+ spawn: Spawner;
548
+ poller: Poller;
549
+ /** Master session id for the delegation identity handed to the worker. */
550
+ getSessionId: () => string;
551
+ /** Phase 0 routing telemetry (RFC rfc-jev-role-routing §8): best-effort spawn profile; absent in tests. */
552
+ logRouting?: (row: RoutingTelemetryRecord) => void;
553
+ }
554
+
555
+ export function createSpawnAction(h: SpawnActionHost) {
556
+ const { client, env, git, subs } = h;
557
+ const { resolveSessionFile, collectFinalText, readAskFlag, probeAlive, subSessionState } = h.session;
558
+ const { spawnPaneInTaskTab, launchLine, waitSubReady } = h.spawn;
559
+ const { startPoller } = h.poller;
560
+
561
+ return async function executeSubagentSpawn(
562
+ params: Record<string, unknown> | undefined,
563
+ toolCtx: unknown,
564
+ onUpdate?: (update: unknown) => void,
565
+ ): Promise<unknown> {
566
+ const launch = planLaunchValidation(params, client.available);
567
+ if (launch.kind === 'error') return toolError(launch.text);
568
+ const { spec, background, isolate, cwdParam, roleKind: kind, suggested, manifestRole, tab } = launch;
569
+ const masterCwd = (toolCtx as { cwd?: string })?.cwd ?? process.cwd();
570
+ const taskId = randomUUID();
571
+ // Validate the role before any isolate worktree exists: this early return has no cleanup path.
572
+ let roleManifestEnv: Record<string, string> = {};
573
+ let roleModel: string | null = null;
574
+ // Delegation identity (RFC rfc-delegation-identity §2.1): the worker learns who delegated it.
575
+ let delegationEnvPatch: Record<string, string> = {};
576
+ try {
577
+ const { role, manifest } = composeForRole(manifestRole ?? 'worker-default', suggested, { loadRoleOpts: { baseDir: masterCwd } });
578
+ roleManifestEnv = {
579
+ PI_HERDR_ROLE_MANIFEST: JSON.stringify({
580
+ role: role.role,
581
+ version: role.version,
582
+ tools: manifest.tools,
583
+ permissions: manifest.permissions,
584
+ unknownTools: manifest.unknownTools,
585
+ services: role.services ?? {},
586
+ // Without guidelines parseRuntimeManifest yields none and the pier-role prompt section silently no-ops.
587
+ ...(role.guidelines?.length ? { guidelines: role.guidelines } : {}),
588
+ }),
589
+ // Role base: /pier-role and the pipe switch resolve .pi-herdr/roles/ against the MASTER's checkout.
590
+ PI_HERDR_ROLE_BASE: masterCwd,
591
+ };
592
+ delegationEnvPatch = delegationEnv({
593
+ parentPaneId: env?.paneId ?? '',
594
+ parentSessionId: bareSessionId(h.getSessionId()),
595
+ // Phase 2: the master's own pipe so the worker's report_progress can dial home (D49).
596
+ parentPipe: pipeNameFor(masterCwd, env?.paneId ?? ''),
597
+ taskId,
598
+ description: spec.description,
599
+ role: role.role,
600
+ });
601
+ if (typeof role.model === 'string' && role.model.trim()) roleModel = role.model.trim();
602
+ h.logRouting?.(
603
+ planSpawnProfileRow({
604
+ now: Date.now(),
605
+ roleExplicit: typeof manifestRole === 'string' && manifestRole.trim() !== '',
606
+ role: role.role,
607
+ allowedTools: suggested,
608
+ manifestTools: manifest.tools,
609
+ task: spec.prompt,
610
+ }),
611
+ );
612
+ } catch (err) {
613
+ const msg = err instanceof Error ? err.message : String(err);
614
+ return toolError(`Error: role "${manifestRole}" manifest invalid — ${msg}`);
615
+ }
616
+ // D98: create the isolate through pier's execFile git path, not herdr's socket worktree.create
617
+ // (bootstrap races, missing root-pane env, linked_worktree_source rejection).
618
+ let isolateMeta: SubEntry['isolate'] | null = null;
619
+ let cwd = masterCwd;
620
+ if (isolate) {
621
+ const baseSha = await git.runGit(masterCwd, ['rev-parse', 'HEAD']);
622
+ if (baseSha == null) return toolError('Error: isolate requires a git repository with at least one commit');
623
+ const sha = baseSha.trim();
624
+ const pierBranches = new Set(
625
+ ((await git.runGit(masterCwd, ['for-each-ref', '--format=%(refname:short)', 'refs/heads/pier/'])) ?? '')
626
+ .split('\n').filter(Boolean),
627
+ );
628
+ const repoName = basename((await git.listWorktrees(masterCwd))[0] ?? masterCwd);
629
+ const plan = planIsolateWorktree({
630
+ description: spec.description,
631
+ taskHex: taskId.slice(0, 6),
632
+ existingPierBranches: pierBranches,
633
+ });
634
+ const wtPath = join(platformPaths.worktreeBaseDir, repoName, plan.worktreeDirName);
635
+ try {
636
+ // git worktree add does not create its parent directory, so a first run would fail.
637
+ mkdirSync(dirname(wtPath), { recursive: true });
638
+ } catch { /* existing directory or permissions failure is handled by worktree add */ }
639
+ // Keep the branch pending between worktree add and subs.set: the readiness wait lets the
640
+ // ticker scan, which could otherwise reclaim a new clean =HEAD-ancestor worktree as an orphan.
641
+ h.pendingIsolateBranches.add(plan.branch);
642
+ const added = await git.runGit(masterCwd, ['worktree', 'add', '-b', plan.branch, wtPath, sha]);
643
+ if (added === null) {
644
+ h.pendingIsolateBranches.delete(plan.branch);
645
+ return toolError(`Error: failed to create worktree ${wtPath} (branch ${plan.branch}) — run \`git worktree prune\` and retry if it reports stale entries`);
646
+ }
647
+ git.invalidateWorktreesCache(); // so the zone classifier sees the new worktree immediately
648
+ cwd = wtPath;
649
+ isolateMeta = { worktreePath: wtPath, branch: plan.branch, baseSha: sha, releasedAt: null, retainNotified: false };
650
+ spec.prompt = `${buildIsolatePreamble({ worktreePath: wtPath, branch: plan.branch, baseShort: sha.slice(0, 7) })}\n\n${spec.prompt}`;
651
+ } else if (cwdParam) {
652
+ cwd = pathResolve(masterCwd, cwdParam);
653
+ try {
654
+ if (!(await statAsync(cwd)).isDirectory()) throw new Error('not a directory');
655
+ } catch {
656
+ return toolError(`Error: \`cwd\` is not an existing directory: ${cwdParam}`);
657
+ }
658
+ }
659
+ // D86 R1: group by the git worktree containing cwd (main checkout → main tab; other → its tab).
660
+ const zone = classifyWorktreeZone({ cwd, masterCwd, worktrees: await git.listWorktrees(masterCwd) });
661
+ // D86 trust: pass -a only for the master's checkout/worktrees; external dirs stay behind pi's Trust dialog.
662
+ const approve = isPathUnder(cwd, masterCwd) || zone.zone === 'worktree';
663
+
664
+ const release = await h.subSemaphore.acquire();
665
+ let paneId = '';
666
+ try {
667
+ const spawnedAt = Date.now();
668
+ const spawned = await spawnPaneInTaskTab(
669
+ { desiredTab: tab, description: spec.description, zone },
670
+ cwd,
671
+ // pier owns worker state reporting (report_agent with todo/blocked badges); pi 1.1+'s own
672
+ // OSC 7501 program-status emission would become a second, competing source once a terminal
673
+ // or dashboard starts reading it. One reporter per pane.
674
+ { PI_HERDR_SUBAGENT: '1', PI_PROGRAM_STATUS: '0', ...roleManifestEnv, ...delegationEnvPatch },
675
+ launchLine(null, roleModel, approve),
676
+ );
677
+ paneId = spawned.paneId;
678
+ // Phase 2: allocate once per spawn; revives/history keep the row's slug (stability beats beauty).
679
+ const slug = allocateSubSlug(
680
+ spec.description,
681
+ taskId,
682
+ new Set([...subs.values()].map((s) => s.slug ?? '').filter(Boolean)),
683
+ );
684
+ const entry: SubEntry = {
685
+ taskId,
686
+ kind,
687
+ paneId,
688
+ slug,
689
+ tabId: spawned.tabId,
690
+ tabName: spawned.tabName,
691
+ cwd,
692
+ description: spec.description,
693
+ background,
694
+ status: 'running',
695
+ sessionFile: null,
696
+ launchCommand: [launchLine(null, roleModel, approve)],
697
+ createdAt: Date.now(),
698
+ revivedFrom: null,
699
+ ...(isolateMeta ? { isolate: isolateMeta } : {}),
700
+ };
701
+ const ready = await waitSubReady(cwd, paneId);
702
+ if (!ready.ok) throw new Error(ready.message);
703
+ // herdr's per-pane report lags right after spawn, so the mtime fallback would otherwise adopt
704
+ // the newest PRE-EXISTING session in cwd; only a file written after the pane existed counts.
705
+ entry.sessionFile = await resolveSessionFile(paneId, cwd, undefined, { minMtimeMs: spawnedAt - 2_000 });
706
+ subs.set(paneId, entry);
707
+ h.persistSubs();
708
+ onUpdate?.(makeProgressUpdate(`subagent ready in pane ${paneId}; injecting prompt via pipe…`));
709
+ // M11: inject through the extension pipe, reserving the PTY keyboard channel for the human;
710
+ // injection and injectTs must stay together or the spawn leaves a ghost running ledger entry.
711
+ const injectTs = Date.now();
712
+ const injected = await pipeRequestTo(cwd, paneId, {
713
+ type: 'prompt',
714
+ id: `prompt-${taskId}`,
715
+ text: spec.prompt,
716
+ // The pipe name MUST be scoped by the MASTER's cwd: the server binds pipeNameFor(master
717
+ // session cwd, own paneId); the worker's cwd names a pipe nobody listens on.
718
+ from: pipeNameFor(masterCwd, env?.paneId ?? ''),
719
+ push: background,
720
+ });
721
+ if (injected.type !== 'ok') {
722
+ throw new Error(`pipe prompt rejected: ${injected.type === 'error' ? injected.message : 'unknown response'}`);
723
+ }
724
+ h.lastMachineInjectAt.set(paneId, injectTs); // attribute working state during the observation window
725
+ const handToPoller = (requestId: string): void => {
726
+ entry.background = true;
727
+ h.lastRequestIdByPane.set(paneId, requestId);
728
+ void startPoller(paneId, cwd, injectTs, spec.description, requestId);
729
+ h.persistSubs();
730
+ h.writeHistory(entry, undefined, 'to-background');
731
+ };
732
+
733
+ if (background) {
734
+ handToPoller(`prompt-${taskId}`);
735
+ return {
736
+ content: [{ type: 'text', text: `started subagent ${subRef(entry)} (task ${taskId})` }],
737
+ details: { paneId, taskId, background: true, role: kind, slug },
738
+ };
739
+ }
740
+
741
+ // Foreground waiting: content gate + patience threshold before backgrounding an idle pane.
742
+ const patienceDeadline = Date.now() + runtimePolicy.foregroundPatienceMs;
743
+ let text: string | null = null;
744
+ let settledKind: 'settled' | 'timeout' = 'timeout';
745
+ while (Date.now() < Math.min(patienceDeadline, spawnedAt + runtimePolicy.subagentTimeoutMs)) {
746
+ const state = await client.waitAgent(paneId, ['idle', 'done', 'blocked'], 30_000);
747
+ const session = (state === 'idle' || state === 'done')
748
+ ? await subSessionState(paneId, cwd, injectTs)
749
+ : { text: null, pendingTool: false, activity: false };
750
+ const tick = planForegroundTick({ state, session });
751
+ if (tick.kind === 'blocked') {
752
+ const question = await readAskFlag(paneId);
753
+ handToPoller(`prompt-${taskId}`);
754
+ return {
755
+ content: [{ type: 'text', text: buildBlockedGateNotice({ paneId, description: spec.description, question }) }],
756
+ details: { paneId, taskId, background: true, blocked: true, role: kind },
757
+ };
758
+ }
759
+ if (tick.kind === 'settled') {
760
+ text = tick.text;
761
+ settledKind = 'settled';
762
+ break;
763
+ }
764
+ if (tick.kind === 'wait') {
765
+ await sleep(tick.delayMs);
766
+ continue;
767
+ }
768
+ if (tick.kind === 'collect-final') {
769
+ text = await collectFinalText(paneId, cwd, injectTs, 6);
770
+ if (text) { settledKind = 'settled'; break; }
771
+ await sleep(FOREGROUND_POLL_MS);
772
+ }
773
+ }
774
+ if (!text) {
775
+ const probe = await probeAlive(paneId, cwd);
776
+ if (isAlive(probe, Date.now())) {
777
+ handToPoller(`prompt-${taskId}`);
778
+ return {
779
+ content: [{ type: 'text', text: buildAliveNotice({ paneId, description: spec.description, scenario: 'moved-to-bg', probe }, Date.now()) }],
780
+ details: { paneId, taskId, background: true, movedToBackground: true, role: kind },
781
+ };
782
+ }
783
+ settledKind = 'timeout';
784
+ }
785
+ const outcome =
786
+ settledKind === 'timeout' ? { kind: 'timeout' as const, text: text ?? '' }
787
+ : text ? { kind: 'completed' as const, text }
788
+ : { kind: 'no-output' as const, text: '' };
789
+ entry.status = 'consumed';
790
+ entry.consumedAt = Date.now();
791
+ h.persistSubs();
792
+ h.writeHistory(entry, { outcome: outcome.kind === 'completed' ? text : null }, 'fg-settle');
793
+ return {
794
+ content: [{ type: 'text', text: formatSubagentResult(outcome, spec.description) }],
795
+ details: { paneId, taskId, background, role: kind },
796
+ };
797
+ } catch (err) {
798
+ // A mid-spawn failure must remove the ledger entry and close its pane, or a ghost entry never settles.
799
+ if (paneId) {
800
+ subs.delete(paneId);
801
+ h.persistSubs();
802
+ void client.closePane(paneId).catch(() => { /* GC covers board exceptions */ });
803
+ }
804
+ // D98: an isolate whose startup never became ready has no work to preserve.
805
+ if (isolateMeta) {
806
+ const { branch, worktreePath } = isolateMeta;
807
+ void git.runGit(masterCwd, ['worktree', 'remove', '--force', worktreePath])
808
+ .then(() => git.runGit(masterCwd, ['branch', '-D', branch]))
809
+ .catch(() => { /* best effort only */ });
810
+ }
811
+ return {
812
+ content: [{
813
+ type: 'text',
814
+ text: formatSubagentResult(
815
+ { kind: 'spawn-failed', text: String((err as Error)?.message ?? err) },
816
+ spec.description,
817
+ ),
818
+ }],
819
+ details: {},
820
+ };
821
+ } finally {
822
+ release();
823
+ if (isolateMeta) h.pendingIsolateBranches.delete(isolateMeta.branch);
824
+ }
825
+ };
301
826
  }