pi-crew 0.1.45 → 0.1.49

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 (178) hide show
  1. package/CHANGELOG.md +97 -0
  2. package/README.md +5 -5
  3. package/agents/analyst.md +11 -11
  4. package/agents/critic.md +11 -11
  5. package/agents/executor.md +11 -11
  6. package/agents/explorer.md +11 -11
  7. package/agents/planner.md +11 -11
  8. package/agents/reviewer.md +11 -11
  9. package/agents/security-reviewer.md +11 -11
  10. package/agents/test-engineer.md +11 -11
  11. package/agents/verifier.md +11 -11
  12. package/agents/writer.md +11 -11
  13. package/docs/next-upgrade-roadmap.md +808 -0
  14. package/docs/research/AGENT-EXECUTION-ARCHITECTURE.md +261 -0
  15. package/docs/research/AGENT-LIFECYCLE-COMPARISON.md +111 -0
  16. package/docs/research/AUDIT_OH_MY_PI.md +261 -0
  17. package/docs/research/AUDIT_PI_CREW.md +457 -0
  18. package/docs/research/CAVEMAN-DEEP-RESEARCH.md +281 -0
  19. package/docs/research/COMPARISON_OH_MY_PI_VS_PI_CREW.md +264 -0
  20. package/docs/research/DEEP-RESEARCH-PI-POWERBAR.md +343 -0
  21. package/docs/research/DEEP_RESEARCH_SUBAGENT_ARCHITECTURE.md +480 -0
  22. package/docs/research/GAP_CLOSURE_IMPLEMENTATION_PLAN.md +354 -0
  23. package/docs/research/IMPLEMENTATION_PLAN.md +385 -0
  24. package/docs/research/LIVE-SESSION-PRODUCTION-READY-PLAN.md +502 -0
  25. package/docs/research/OH-MY-PI-DEEP-RESEARCH-v14.7.6.md +266 -0
  26. package/docs/research/REMAINING-GAPS-PLAN.md +363 -0
  27. package/docs/research/SESSION-SUMMARY-2026-05-08.md +146 -0
  28. package/docs/research/UI-RESPONSIVENESS-AUDIT.md +173 -0
  29. package/docs/research-awesome-agent-skills-distillation.md +100 -0
  30. package/docs/research-oh-my-pi-distillation.md +369 -0
  31. package/docs/source-runtime-refactor-map.md +24 -0
  32. package/docs/usage.md +3 -3
  33. package/install.mjs +52 -8
  34. package/package.json +99 -98
  35. package/schema.json +10 -1
  36. package/skills/async-worker-recovery/SKILL.md +42 -0
  37. package/skills/context-artifact-hygiene/SKILL.md +52 -0
  38. package/skills/delegation-patterns/SKILL.md +54 -0
  39. package/skills/mailbox-interactive/SKILL.md +40 -0
  40. package/skills/model-routing-context/SKILL.md +39 -0
  41. package/skills/multi-perspective-review/SKILL.md +58 -0
  42. package/skills/observability-reliability/SKILL.md +41 -0
  43. package/skills/orchestration/SKILL.md +157 -0
  44. package/skills/ownership-session-security/SKILL.md +41 -0
  45. package/skills/pi-extension-lifecycle/SKILL.md +39 -0
  46. package/skills/requirements-to-task-packet/SKILL.md +63 -0
  47. package/skills/resource-discovery-config/SKILL.md +41 -0
  48. package/skills/runtime-state-reader/SKILL.md +44 -0
  49. package/skills/secure-agent-orchestration-review/SKILL.md +45 -0
  50. package/skills/state-mutation-locking/SKILL.md +42 -0
  51. package/skills/systematic-debugging/SKILL.md +67 -0
  52. package/skills/ui-render-performance/SKILL.md +39 -0
  53. package/skills/verification-before-done/SKILL.md +57 -0
  54. package/skills/worktree-isolation/SKILL.md +39 -0
  55. package/src/agents/agent-config.ts +6 -0
  56. package/src/agents/agent-search.ts +98 -0
  57. package/src/agents/agent-serializer.ts +38 -34
  58. package/src/agents/discover-agents.ts +29 -15
  59. package/src/config/config.ts +72 -24
  60. package/src/config/defaults.ts +25 -0
  61. package/src/extension/autonomous-policy.ts +26 -33
  62. package/src/extension/help.ts +1 -0
  63. package/src/extension/management.ts +5 -0
  64. package/src/extension/project-init.ts +62 -2
  65. package/src/extension/register.ts +69 -22
  66. package/src/extension/registration/commands.ts +64 -25
  67. package/src/extension/registration/compaction-guard.ts +1 -1
  68. package/src/extension/registration/subagent-helpers.ts +8 -0
  69. package/src/extension/registration/subagent-tools.ts +149 -148
  70. package/src/extension/registration/team-tool.ts +14 -10
  71. package/src/extension/run-index.ts +35 -21
  72. package/src/extension/run-maintenance.ts +30 -5
  73. package/src/extension/team-tool/api.ts +47 -9
  74. package/src/extension/team-tool/cancel.ts +109 -5
  75. package/src/extension/team-tool/context.ts +8 -0
  76. package/src/extension/team-tool/intent-policy.ts +42 -0
  77. package/src/extension/team-tool/lifecycle-actions.ts +120 -79
  78. package/src/extension/team-tool/parallel-dispatch.ts +156 -0
  79. package/src/extension/team-tool/respond.ts +46 -18
  80. package/src/extension/team-tool/run.ts +55 -12
  81. package/src/extension/team-tool/status.ts +13 -2
  82. package/src/extension/team-tool-types.ts +3 -0
  83. package/src/extension/team-tool.ts +45 -14
  84. package/src/hooks/registry.ts +61 -0
  85. package/src/hooks/types.ts +41 -0
  86. package/src/observability/event-to-metric.ts +8 -1
  87. package/src/runtime/agent-control.ts +169 -63
  88. package/src/runtime/async-runner.ts +3 -1
  89. package/src/runtime/background-runner.ts +78 -53
  90. package/src/runtime/cancellation-token.ts +89 -0
  91. package/src/runtime/cancellation.ts +61 -0
  92. package/src/runtime/capability-inventory.ts +116 -0
  93. package/src/runtime/child-pi.ts +458 -444
  94. package/src/runtime/code-summary.ts +247 -0
  95. package/src/runtime/crash-recovery.ts +182 -0
  96. package/src/runtime/crew-agent-records.ts +70 -10
  97. package/src/runtime/crew-agent-runtime.ts +1 -0
  98. package/src/runtime/custom-tools/irc-tool.ts +201 -0
  99. package/src/runtime/custom-tools/submit-result-tool.ts +90 -0
  100. package/src/runtime/deadletter.ts +1 -0
  101. package/src/runtime/delivery-coordinator.ts +48 -25
  102. package/src/runtime/effectiveness.ts +81 -0
  103. package/src/runtime/event-stream-bridge.ts +90 -0
  104. package/src/runtime/live-agent-control.ts +2 -1
  105. package/src/runtime/live-agent-manager.ts +179 -85
  106. package/src/runtime/live-control-realtime.ts +1 -1
  107. package/src/runtime/live-extension-bridge.ts +150 -0
  108. package/src/runtime/live-irc.ts +92 -0
  109. package/src/runtime/live-session-health.ts +100 -0
  110. package/src/runtime/live-session-runtime.ts +599 -305
  111. package/src/runtime/manifest-cache.ts +17 -2
  112. package/src/runtime/mcp-proxy.ts +113 -0
  113. package/src/runtime/model-fallback.ts +6 -4
  114. package/src/runtime/notebook-helpers.ts +90 -0
  115. package/src/runtime/orphan-sentinel.ts +7 -0
  116. package/src/runtime/output-validator.ts +187 -0
  117. package/src/runtime/parallel-utils.ts +57 -0
  118. package/src/runtime/parent-guard.ts +80 -0
  119. package/src/runtime/pi-args.ts +18 -3
  120. package/src/runtime/process-status.ts +5 -1
  121. package/src/runtime/prose-compressor.ts +164 -0
  122. package/src/runtime/result-extractor.ts +121 -0
  123. package/src/runtime/retry-executor.ts +81 -64
  124. package/src/runtime/runtime-resolver.ts +23 -10
  125. package/src/runtime/semaphore.ts +131 -0
  126. package/src/runtime/sensitive-paths.ts +92 -0
  127. package/src/runtime/skill-instructions.ts +222 -0
  128. package/src/runtime/stale-reconciler.ts +4 -14
  129. package/src/runtime/stream-preview.ts +177 -0
  130. package/src/runtime/subagent-manager.ts +6 -2
  131. package/src/runtime/subprocess-tool-registry.ts +67 -0
  132. package/src/runtime/task-output-context.ts +177 -127
  133. package/src/runtime/task-runner/capabilities.ts +78 -0
  134. package/src/runtime/task-runner/live-executor.ts +107 -101
  135. package/src/runtime/task-runner/prompt-builder.ts +72 -8
  136. package/src/runtime/task-runner/prompt-pipeline.ts +64 -0
  137. package/src/runtime/task-runner/run-projection.ts +104 -0
  138. package/src/runtime/task-runner.ts +115 -5
  139. package/src/runtime/team-runner.ts +134 -19
  140. package/src/runtime/workspace-tree.ts +298 -0
  141. package/src/runtime/yield-handler.ts +189 -0
  142. package/src/schema/config-schema.ts +7 -0
  143. package/src/schema/team-tool-schema.ts +14 -4
  144. package/src/skills/discover-skills.ts +67 -0
  145. package/src/state/active-run-registry.ts +167 -0
  146. package/src/state/artifact-store.ts +4 -1
  147. package/src/state/atomic-write.ts +50 -1
  148. package/src/state/blob-store.ts +117 -0
  149. package/src/state/contracts.ts +2 -1
  150. package/src/state/event-log-rotation.ts +158 -0
  151. package/src/state/event-log.ts +52 -2
  152. package/src/state/mailbox.ts +129 -9
  153. package/src/state/state-store.ts +32 -5
  154. package/src/state/types.ts +64 -2
  155. package/src/teams/team-config.ts +1 -0
  156. package/src/ui/agent-management-overlay.ts +144 -0
  157. package/src/ui/crew-widget.ts +15 -5
  158. package/src/ui/dashboard-panes/cancellation-pane.ts +43 -0
  159. package/src/ui/dashboard-panes/capability-pane.ts +60 -0
  160. package/src/ui/dashboard-panes/mailbox-pane.ts +35 -11
  161. package/src/ui/dashboard-panes/progress-pane.ts +2 -0
  162. package/src/ui/live-run-sidebar.ts +4 -0
  163. package/src/ui/powerbar-publisher.ts +77 -15
  164. package/src/ui/render-coalescer.ts +51 -0
  165. package/src/ui/run-dashboard.ts +4 -0
  166. package/src/ui/run-event-bus.ts +209 -0
  167. package/src/ui/run-snapshot-cache.ts +78 -18
  168. package/src/ui/snapshot-types.ts +10 -0
  169. package/src/ui/transcript-entries.ts +258 -0
  170. package/src/utils/ids.ts +5 -0
  171. package/src/utils/incremental-reader.ts +104 -0
  172. package/src/utils/paths.ts +4 -2
  173. package/src/utils/scan-cache.ts +137 -0
  174. package/src/utils/sse-parser.ts +134 -0
  175. package/src/utils/task-name-generator.ts +337 -0
  176. package/src/utils/visual.ts +33 -2
  177. package/src/workflows/workflow-config.ts +1 -0
  178. package/src/worktree/cleanup.ts +2 -1
@@ -2,6 +2,7 @@ import * as fs from "node:fs";
2
2
  import * as path from "node:path";
3
3
  import { closeWatcher, watchWithErrorHandler } from "../utils/fs-watch.ts";
4
4
  import { findRepoRoot, projectCrewRoot, userCrewRoot } from "../utils/paths.ts";
5
+ import { activeRunEntries } from "../state/active-run-registry.ts";
5
6
  import { isSafePathId, resolveContainedRelativePath, resolveRealContainedPath } from "../utils/safe-paths.ts";
6
7
  import type { TeamRunManifest } from "../state/types.ts";
7
8
  import { DEFAULT_CACHE, DEFAULT_PATHS } from "../config/defaults.ts";
@@ -106,8 +107,10 @@ function parseManifestIfChanged(root: string, runId: string, filePath: string, p
106
107
  }
107
108
 
108
109
  function listRunRoots(cwd: string): string[] {
110
+ const roots = new Set<string>();
109
111
  const base = findRepoRoot(cwd) ? projectCrewRoot(cwd) : userCrewRoot();
110
- return [path.join(base, DEFAULT_PATHS.state.runsSubdir)];
112
+ roots.add(path.join(base, DEFAULT_PATHS.state.runsSubdir));
113
+ return [...roots];
111
114
  }
112
115
 
113
116
  function collectRoots(root: string): ParsedEntry[] {
@@ -156,6 +159,15 @@ export function createManifestCache(cwd: string, options: ManifestCacheOptions =
156
159
  function loadManifest(runId: string, rootsToCheck: string[]): CachedManifest | undefined {
157
160
  let cached = manifestIndex.get(runId);
158
161
  if (!isSafePathId(runId)) return undefined;
162
+ const activeEntry = activeRunEntries().find((entry) => entry.runId === runId);
163
+ if (activeEntry) {
164
+ const activeRoot = path.dirname(activeEntry.stateRoot);
165
+ const parsed = parseManifestIfChanged(activeRoot, runId, activeEntry.manifestPath, cached);
166
+ if (parsed) {
167
+ manifestIndex.set(runId, parsed);
168
+ return parsed;
169
+ }
170
+ }
159
171
  for (const root of rootsToCheck) {
160
172
  const manifestPath = manifestPathForRun(root, runId);
161
173
  if (!manifestPath) continue;
@@ -180,7 +192,10 @@ export function createManifestCache(cwd: string, options: ManifestCacheOptions =
180
192
  if (cached && cached.expireAtMs > now) {
181
193
  return cached.runs;
182
194
  }
183
- const parsedEntries = roots.flatMap((root) => collectRoots(root));
195
+ const parsedEntries = [
196
+ ...roots.flatMap((root) => collectRoots(root)),
197
+ ...activeRunEntries().map((entry) => ({ runId: entry.runId, path: entry.manifestPath })),
198
+ ];
184
199
  const unique = new Map<string, CachedManifest | undefined>();
185
200
  for (const entry of parsedEntries) {
186
201
  if (entry.runId.length === 0) continue;
@@ -0,0 +1,113 @@
1
+ /**
2
+ * G2: MCP Proxy for live-session workers.
3
+ *
4
+ * When the parent process has MCP servers configured, live-session workers
5
+ * can reuse those connections instead of establishing their own. This module
6
+ * discovers MCP tools available in the parent environment and creates proxy
7
+ * tool definitions that forward calls through the parent's connections.
8
+ *
9
+ * Strategy:
10
+ * 1. If the Pi SDK session has MCP tools after bindExtensions → use them directly
11
+ * 2. If not → create proxy custom tools that wrap MCP calls
12
+ * 3. If no MCP config exists → disable MCP in the session
13
+ *
14
+ * The Pi SDK's `createAgentSession` accepts a `customTools` array for injecting
15
+ * proxy tools. The session also accepts `enableMCP: false` to skip MCP discovery
16
+ * when proxying from the parent.
17
+ */
18
+
19
+ import { defineTool, type ToolDefinition } from "@mariozechner/pi-coding-agent";
20
+ import { Type, type Static, type TSchema } from "@sinclair/typebox";
21
+
22
+ export interface McpProxyConfig {
23
+ /** Whether to enable MCP in the child session. */
24
+ enableMcp: boolean;
25
+ /** Proxy tools to inject via customTools (replaces MCP connection). */
26
+ proxyTools: Array<ToolDefinition<TSchema, unknown>>;
27
+ /** Names of MCP tools available (for metadata/tracking). */
28
+ proxyToolNames: string[];
29
+ }
30
+
31
+ /**
32
+ * Build MCP proxy configuration for a live-session worker.
33
+ *
34
+ * @param options.parentMcpTools — MCP tool names from the parent session (if available)
35
+ * @param options.shareMcp — Whether to share MCP connections (default: true)
36
+ */
37
+ export function buildMcpProxyConfig(options: {
38
+ parentMcpTools?: string[];
39
+ shareMcp?: boolean;
40
+ }): McpProxyConfig {
41
+ if (options.shareMcp === false) {
42
+ return { enableMcp: true, proxyTools: [], proxyToolNames: [] };
43
+ }
44
+
45
+ const parentTools = options.parentMcpTools ?? [];
46
+ if (parentTools.length === 0) {
47
+ // No MCP tools in parent — let session discover on its own
48
+ return { enableMcp: true, proxyTools: [], proxyToolNames: [] };
49
+ }
50
+
51
+ // MCP tools exist in parent — try to create proxy tools.
52
+ // If proxy tools are not available (stub), keep enableMcp: true
53
+ // so the child session can self-discover MCP instead of losing all access.
54
+ const proxyTools = createMcpProxyTools(parentTools);
55
+ if (proxyTools.length === 0) {
56
+ // No proxy tools available — let child discover MCP on its own
57
+ return { enableMcp: true, proxyTools: [], proxyToolNames: parentTools };
58
+ }
59
+ return {
60
+ enableMcp: false,
61
+ proxyTools,
62
+ proxyToolNames: parentTools,
63
+ };
64
+ }
65
+
66
+ /**
67
+ * Create lightweight proxy tools that represent MCP tools from the parent.
68
+ *
69
+ * These tools tell the model that the MCP tools are available, but actual
70
+ * execution is forwarded through the parent's MCP connections. Since we
71
+ * can't directly access the parent's MCP manager from a child session,
72
+ * the tools return a message indicating the model should use them normally.
73
+ *
74
+ * In a future iteration, these can be wired to the actual MCP connections
75
+ * via an inter-process bridge.
76
+ */
77
+ function createMcpProxyTools(
78
+ toolNames: string[],
79
+ ): Array<ToolDefinition<TSchema, unknown>> {
80
+ // For now, we don't create individual proxy tools because we can't
81
+ // forward MCP calls without the parent's MCP manager reference.
82
+ //
83
+ // Instead, we let the child session discover MCP on its own (enableMcp: true)
84
+ // or share the parent's MCP config directory.
85
+ //
86
+ // This will be enhanced when we add inter-process MCP call forwarding.
87
+ return [];
88
+ }
89
+
90
+ /**
91
+ * Discover MCP tool names from a live session's active tools.
92
+ * MCP tools typically have names containing "__" (e.g., "mcp__filesystem__read_file").
93
+ */
94
+ export function discoverMcpToolNames(activeToolNames: string[]): string[] {
95
+ return activeToolNames.filter(
96
+ (name) =>
97
+ name.startsWith("mcp__") ||
98
+ name.startsWith("mcp-") ||
99
+ (name.includes("__") && !name.startsWith("submit_result")),
100
+ );
101
+ }
102
+
103
+ /**
104
+ * Build MCP proxy config from a real Pi SDK session's active tools.
105
+ * This is the preferred way — inspect what the parent session has available.
106
+ */
107
+ export function buildMcpProxyFromSession(
108
+ activeToolNames: string[],
109
+ options?: { shareMcp?: boolean },
110
+ ): McpProxyConfig {
111
+ const mcpTools = discoverMcpToolNames(activeToolNames);
112
+ return buildMcpProxyConfig({ parentMcpTools: mcpTools, shareMcp: options?.shareMcp });
113
+ }
@@ -117,9 +117,10 @@ export function configuredModelInfosFromPiConfig(cwd?: string): AvailableModelIn
117
117
  const globalSettings = readJsonObject(path.join(agentDir, "settings.json")) as PiSettingsLike | undefined;
118
118
  const projectSettings = cwd ? readJsonObject(path.join(cwd, ".pi", "settings.json")) as PiSettingsLike | undefined : undefined;
119
119
  const effectiveSettings = { ...(globalSettings ?? {}), ...(projectSettings ?? {}) };
120
+ const defaultModel = settingsModelInfo(effectiveSettings);
120
121
  return uniqueModelInfos([
122
+ ...(defaultModel ? [defaultModel] : []),
121
123
  ...modelsJsonInfos(readJsonObject(path.join(agentDir, "models.json")) as PiModelsJsonLike | undefined),
122
- ...(settingsModelInfo(effectiveSettings) ? [settingsModelInfo(effectiveSettings)!] : []),
123
124
  ]);
124
125
  }
125
126
 
@@ -236,6 +237,7 @@ export interface ConfiguredModelRouting {
236
237
  export function buildConfiguredModelRouting(input: {
237
238
  overrideModel?: string;
238
239
  stepModel?: string;
240
+ teamRoleModel?: string;
239
241
  agentModel?: string;
240
242
  fallbackModels?: string[];
241
243
  parentModel?: unknown;
@@ -250,11 +252,11 @@ export function buildConfiguredModelRouting(input: {
250
252
  // B3: Parent model inheritance — when agent has no model specified,
251
253
  // inherit from parent session model before falling back to defaults.
252
254
  const effectiveAgentModel = input.agentModel?.trim() ? input.agentModel : parentModel;
253
- const requested = [input.overrideModel, input.stepModel, effectiveAgentModel].find((model): model is string => Boolean(model?.trim()));
255
+ const requested = [input.overrideModel, input.stepModel, input.teamRoleModel, effectiveAgentModel].find((model): model is string => Boolean(model?.trim()));
254
256
  if (availableModels && availableModels.length === 0) return { requested, candidates: [], reason: "no configured Pi models available" };
255
257
  const rawModels = availableModels
256
- ? [input.overrideModel, input.stepModel, effectiveAgentModel, ...(input.fallbackModels ?? []), ...availableModels.map((model) => model.fullId)]
257
- : [input.overrideModel, parentModel];
258
+ ? [input.overrideModel, input.stepModel, input.teamRoleModel, effectiveAgentModel, ...(input.fallbackModels ?? []), ...availableModels.map((model) => model.fullId)]
259
+ : [input.overrideModel, input.stepModel, input.teamRoleModel, effectiveAgentModel, ...(input.fallbackModels ?? []), parentModel];
258
260
  const configuredModels = rawModels
259
261
  .filter((model): model is string => Boolean(model?.trim()))
260
262
  .filter((model) => isAvailableModel(model.trim(), availableModels));
@@ -0,0 +1,90 @@
1
+ export interface NotebookCell {
2
+ index: number;
3
+ cellType: "code" | "markdown" | "raw";
4
+ source: string;
5
+ outputs?: unknown[];
6
+ metadata?: Record<string, unknown>;
7
+ }
8
+
9
+ export interface Notebook {
10
+ cells: NotebookCell[];
11
+ metadata?: Record<string, unknown>;
12
+ nbformat: number;
13
+ }
14
+
15
+ /** Check if a file path is a notebook */
16
+ export function isNotebookPath(filePath: string): boolean {
17
+ return filePath.toLowerCase().endsWith(".ipynb");
18
+ }
19
+
20
+ function normalizeSource(source: unknown): string {
21
+ if (typeof source === "string") return source;
22
+ if (Array.isArray(source)) return source.join("");
23
+ return "";
24
+ }
25
+
26
+ /** Parse a .ipynb JSON file into a Notebook structure */
27
+ export function parseNotebook(content: string): Notebook {
28
+ let raw: unknown;
29
+ try {
30
+ raw = JSON.parse(content);
31
+ } catch {
32
+ return { cells: [], nbformat: 4 };
33
+ }
34
+ if (typeof raw !== "object" || raw === null) return { cells: [], nbformat: 4 };
35
+ const obj = raw as Record<string, unknown>;
36
+ const nbformat = typeof obj.nbformat === "number" ? obj.nbformat : 4;
37
+ const rawCells = Array.isArray(obj.cells) ? obj.cells : [];
38
+ const cells: NotebookCell[] = rawCells
39
+ .map((c: unknown, i: number): NotebookCell | null => {
40
+ if (typeof c !== "object" || c === null) return null;
41
+ const cell = c as Record<string, unknown>;
42
+ const cellType = cell.cell_type;
43
+ if (cellType !== "code" && cellType !== "markdown" && cellType !== "raw") return null;
44
+ return {
45
+ index: i,
46
+ cellType,
47
+ source: normalizeSource(cell.source),
48
+ outputs: Array.isArray(cell.outputs) ? cell.outputs : undefined,
49
+ metadata:
50
+ cell.metadata && typeof cell.metadata === "object" && !Array.isArray(cell.metadata)
51
+ ? (cell.metadata as Record<string, unknown>)
52
+ : undefined,
53
+ };
54
+ })
55
+ .filter((c: NotebookCell | null): c is NotebookCell => c !== null);
56
+ const metadata =
57
+ obj.metadata && typeof obj.metadata === "object" && !Array.isArray(obj.metadata)
58
+ ? (obj.metadata as Record<string, unknown>)
59
+ : undefined;
60
+ return { cells, metadata, nbformat };
61
+ }
62
+
63
+ /** Get a specific cell by index */
64
+ export function getCell(notebook: Notebook, index: number): NotebookCell | undefined {
65
+ return notebook.cells.find((c) => c.index === index);
66
+ }
67
+
68
+ /** Update a cell's source content, returning a new Notebook */
69
+ export function updateCell(notebook: Notebook, index: number, source: string): Notebook {
70
+ const cells = notebook.cells.map((c) =>
71
+ c.index === index ? { ...c, source } : c,
72
+ );
73
+ return { ...notebook, cells };
74
+ }
75
+
76
+ /** Serialize a Notebook back to .ipynb JSON string */
77
+ export function serializeNotebook(notebook: Notebook): string {
78
+ const raw = {
79
+ nbformat: notebook.nbformat,
80
+ nbformat_minor: 5,
81
+ metadata: notebook.metadata ?? {},
82
+ cells: notebook.cells.map((c) => ({
83
+ cell_type: c.cellType,
84
+ source: c.source,
85
+ metadata: c.metadata ?? {},
86
+ ...(c.cellType === "code" ? { outputs: c.outputs ?? [], execution_count: null } : {}),
87
+ })),
88
+ };
89
+ return JSON.stringify(raw, null, 2) + "\n";
90
+ }
@@ -0,0 +1,7 @@
1
+ /**
2
+ * Orphan sentinel — DEPRECATED, replaced by parent-guard.ts.
3
+ *
4
+ * Keeping this file as a no-op placeholder until all references are removed.
5
+ * Workers now self-monitor via startParentGuard() instead of relying on
6
+ * an external sentinel process.
7
+ */
@@ -0,0 +1,187 @@
1
+ /**
2
+ * Phase 4 (caveman): Output format validation for live-session workers.
3
+ *
4
+ * Validates that worker output conforms to the structured output contract
5
+ * for the given role. If validation fails, returns structured error info
6
+ * that can be used for retry or fallback.
7
+ *
8
+ * Inspired by caveman's validate.py — check structural preservation
9
+ * (headings, code blocks, URLs) after compression.
10
+ */
11
+
12
+ /** Role-specific output format patterns — constructed fresh per call to avoid /g lastIndex leak */
13
+ const ROLE_PATTERN_DEFS: Record<string, () => RegExp> = {
14
+ explorer: () => /^(\S+:\d+|Defs:|Refs:|Callers:|Tests:|Sites:|No match\.|totals:)/m,
15
+ executor: () => /^(\S+:\d+(-\d+)? — .{1,80}\.|verified:|too-big\.|needs-confirm\.|ambiguous\.|regressed\.)/m,
16
+ reviewer: () => /^([^:\s]+:\d+:\s+\p{Emoji_Presentation}|No issues\.|totals:)/mu,
17
+ "security-reviewer": () => /^([^:\s]+:\d+:\s+\p{Emoji_Presentation}|No issues\.|totals:)/mu,
18
+ verifier: () => /^(PASS:|FAIL:)/m,
19
+ };
20
+
21
+ /** Fresh RegExp factories for structural preservation checks (avoids /g lastIndex leak) */
22
+ const makeUrlRe = () => /\bhttps?:\/\/[^\s<>)\]"',;]+/gi;
23
+ const makeFencedCodeRe = () => /```[\s\S]*?```/g;
24
+ const makeInlineCodeRe = () => /`[^`\n]+`/g;
25
+ const makeHeadingRe = () => /^#{1,6}\s+.+/gm;
26
+
27
+ export interface OutputValidationResult {
28
+ /** Whether the output passes validation */
29
+ valid: boolean;
30
+ /** Whether the output follows the role's contract format */
31
+ formatMatch: boolean;
32
+ /** Whether structural elements (code, URLs, headings) are preserved */
33
+ structurePreserved: boolean;
34
+ /** Specific issues found */
35
+ issues: string[];
36
+ }
37
+
38
+ /**
39
+ * Validate worker output against role-specific contract + structural preservation.
40
+ */
41
+ export function validateWorkerOutput(role: string, output: string): OutputValidationResult {
42
+ const issues: string[] = [];
43
+
44
+ // Empty output always fails
45
+ if (!output || !output.trim()) {
46
+ return { valid: false, formatMatch: false, structurePreserved: false, issues: ["Empty output"] };
47
+ }
48
+
49
+ // Check role-specific format
50
+ const patternFactory = ROLE_PATTERN_DEFS[role];
51
+ const pattern = patternFactory ? patternFactory() : undefined;
52
+ const formatMatch = !pattern || pattern.test(output);
53
+ if (!formatMatch) {
54
+ issues.push(`Output does not match expected ${role} contract format`);
55
+ }
56
+
57
+ // Check structural preservation (code blocks, URLs, headings)
58
+ let structurePreserved = true;
59
+ const trimmedOutput = output.trim();
60
+
61
+ // Detect if output was truncated mid-code-block
62
+ const opens = (trimmedOutput.match(/```/g) ?? []).length;
63
+ if (opens % 2 !== 0) {
64
+ structurePreserved = false;
65
+ issues.push("Unclosed code block — output may be truncated");
66
+ }
67
+
68
+ // Check for malformed URLs
69
+ const urls = trimmedOutput.match(makeUrlRe()) ?? [];
70
+ for (const url of urls) {
71
+ if (url.endsWith(".") || url.endsWith(",")) {
72
+ structurePreserved = false;
73
+ issues.push(`URL with trailing punctuation: ${url.slice(-20)}`);
74
+ }
75
+ }
76
+
77
+ return {
78
+ valid: formatMatch && structurePreserved,
79
+ formatMatch,
80
+ structurePreserved,
81
+ issues,
82
+ };
83
+ }
84
+
85
+ /**
86
+ * Extract structured findings from reviewer output.
87
+ * Returns array of { file, line, severity, message } objects.
88
+ */
89
+ export function parseReviewerFindings(output: string): Array<{ file: string; line: number; severity: string; message: string }> {
90
+ const findings: Array<{ file: string; line: number; severity: string; message: string }> = [];
91
+ const lines = output.split("\n");
92
+
93
+ const SEVERITY_MAP: Record<string, string> = {
94
+ "🔴": "bug",
95
+ "🟡": "risk",
96
+ "🔵": "nit",
97
+ "❓": "question",
98
+ };
99
+
100
+ for (const line of lines) {
101
+ // Match: path/to/file.ts:42: 🔴 bug: problem. fix.
102
+ const match = line.match(/^([^:\s]+):(\d+):\s+(\p{Emoji_Presentation}) (\w+):\s+(.+)/u);
103
+ if (match) {
104
+ findings.push({
105
+ file: match[1],
106
+ line: Number(match[2]),
107
+ severity: SEVERITY_MAP[match[3]] ?? match[3],
108
+ message: match[5].trim(),
109
+ });
110
+ }
111
+ }
112
+
113
+ return findings;
114
+ }
115
+
116
+ /**
117
+ * Extract explorer results from structured output.
118
+ * Returns array of { file, line, symbol, note } objects.
119
+ */
120
+ export function parseExplorerResults(output: string): Array<{ file: string; line: number; symbol: string; note: string }> {
121
+ const results: Array<{ file: string; line: number; symbol: string; note: string }> = [];
122
+ const lines = output.split("\n");
123
+
124
+ for (const line of lines) {
125
+ // Match: path/to/file.ts:42 — `symbol` — note
126
+ const match = line.match(/^[- ]*(\S+):(\d+)\s*[—–-]\s*`([^`]+)`\s*[—–-]\s*(.+)/);
127
+ if (match) {
128
+ results.push({
129
+ file: match[1],
130
+ line: Number(match[2]),
131
+ symbol: match[3],
132
+ note: match[4].trim(),
133
+ });
134
+ }
135
+ }
136
+
137
+ return results;
138
+ }
139
+
140
+ /**
141
+ * Validate that compressed prose preserves structural elements from original.
142
+ * Returns list of specific issues (empty = valid).
143
+ */
144
+ export function validateCompressionPreservation(original: string, compressed: string): string[] {
145
+ const issues: string[] = [];
146
+
147
+ // Check code blocks preserved
148
+ const origBlocks = original.match(makeFencedCodeRe()) ?? [];
149
+ const compBlocks = compressed.match(makeFencedCodeRe()) ?? [];
150
+ if (origBlocks.length !== compBlocks.length) {
151
+ issues.push(`Code block count: ${origBlocks.length} → ${compBlocks.length}`);
152
+ }
153
+ for (let i = 0; i < Math.min(origBlocks.length, compBlocks.length); i++) {
154
+ if (origBlocks[i] !== compBlocks[i]) {
155
+ issues.push(`Code block ${i + 1} content changed`);
156
+ }
157
+ }
158
+
159
+ // Check URLs preserved
160
+ const origUrls = new Set(original.match(makeUrlRe()) ?? []);
161
+ const compUrls = new Set(compressed.match(makeUrlRe()) ?? []);
162
+ for (const url of origUrls) {
163
+ if (!compUrls.has(url)) {
164
+ issues.push(`URL lost: ${url.slice(0, 60)}...`);
165
+ }
166
+ }
167
+
168
+ // Check inline code preserved
169
+ const origInline = original.match(makeInlineCodeRe()) ?? [];
170
+ const compInline = compressed.match(makeInlineCodeRe()) ?? [];
171
+ const origInlineSet = new Set(origInline);
172
+ const compInlineSet = new Set(compInline);
173
+ for (const code of origInlineSet) {
174
+ if (!compInlineSet.has(code)) {
175
+ issues.push(`Inline code lost: ${code}`);
176
+ }
177
+ }
178
+
179
+ // Check headings preserved
180
+ const origHeadings = original.match(makeHeadingRe()) ?? [];
181
+ const compHeadings = compressed.match(makeHeadingRe()) ?? [];
182
+ if (origHeadings.length !== compHeadings.length) {
183
+ issues.push(`Heading count: ${origHeadings.length} → ${compHeadings.length}`);
184
+ }
185
+
186
+ return issues;
187
+ }
@@ -58,6 +58,63 @@ export async function mapConcurrent<T, R>(items: T[], limit: number, fn: (item:
58
58
  return results;
59
59
  }
60
60
 
61
+ /**
62
+ * Phase 6: mapConcurrent with AbortSignal and fail-fast support.
63
+ * On abort: returns partial results (may contain undefined entries).
64
+ * On error: throws immediately (fail-fast) and cancels remaining work.
65
+ */
66
+ export async function mapConcurrentWithSignal<T, R>(
67
+ items: T[],
68
+ limit: number,
69
+ fn: (item: T, i: number, signal: AbortSignal) => Promise<R>,
70
+ signal?: AbortSignal,
71
+ ): Promise<{ results: (R | undefined)[]; aborted: boolean }> {
72
+ const safeLimit = Math.max(1, Math.floor(limit) || 1);
73
+ const results: (R | undefined)[] = new Array(items.length);
74
+ let next = 0;
75
+ let aborted = false;
76
+
77
+ const abortController = new AbortController();
78
+ const workerSignal = signal
79
+ ? AbortSignal.any([signal, abortController.signal])
80
+ : abortController.signal;
81
+
82
+ let rejectFirst: (error: unknown) => void;
83
+ const firstErrorPromise = new Promise<never>((_, reject) => {
84
+ rejectFirst = reject;
85
+ });
86
+
87
+ const worker = async (): Promise<void> => {
88
+ while (!workerSignal.aborted) {
89
+ const i = next++;
90
+ if (i >= items.length) return;
91
+ try {
92
+ results[i] = await fn(items[i], i, workerSignal);
93
+ } catch (error) {
94
+ if (!workerSignal.aborted) {
95
+ abortController.abort();
96
+ rejectFirst(error);
97
+ throw error;
98
+ }
99
+ }
100
+ }
101
+ };
102
+
103
+ const workers = Array.from({ length: Math.min(safeLimit, items.length) }, () => worker());
104
+
105
+ try {
106
+ await Promise.race([Promise.all(workers), firstErrorPromise]);
107
+ } catch (error) {
108
+ if (signal?.aborted) {
109
+ aborted = true;
110
+ return { results, aborted };
111
+ }
112
+ throw error;
113
+ }
114
+
115
+ return { results, aborted: signal?.aborted ?? false };
116
+ }
117
+
61
118
  export interface ParallelTaskResult {
62
119
  agent: string;
63
120
  taskIndex?: number;
@@ -0,0 +1,80 @@
1
+ /**
2
+ * Parent liveness guard for pi-crew worker processes.
3
+ *
4
+ * Workers call `startParentGuard(parentPid)` at startup. A lightweight
5
+ * interval checks if the parent PID is still alive. When the parent dies
6
+ * (SIGKILL, crash, power loss, terminal close), the worker self-terminates
7
+ * immediately — no sentinel process needed.
8
+ *
9
+ * Note: `process.kill(pid, 0)` works on both Unix and Windows in Node.js
10
+ * for checking process existence. On Windows, it may throw for processes
11
+ * owned by other users (permission error), but correctly detects dead PIDs.
12
+ *
13
+ * Usage in worker entry points:
14
+ * ```ts
15
+ * const parentPid = Number(process.env.PI_CREW_PARENT_PID);
16
+ * if (parentPid > 0) startParentGuard(parentPid);
17
+ * ```
18
+ */
19
+
20
+ const POLL_INTERVAL_MS = 3_000;
21
+
22
+ let guardInterval: ReturnType<typeof setInterval> | undefined;
23
+
24
+ function isPidAlive(pid: number): boolean {
25
+ try {
26
+ process.kill(pid, 0);
27
+ return true;
28
+ } catch {
29
+ return false;
30
+ }
31
+ }
32
+
33
+ function selfTerminate(parentPid: number): never {
34
+ // Best-effort: try to log why we're dying
35
+ try {
36
+ if (typeof process.stderr?.write === "function") {
37
+ process.stderr.write(`[pi-crew] Parent process ${parentPid} is dead — self-terminating worker ${process.pid}\n`);
38
+ }
39
+ } catch {
40
+ // Ignore
41
+ }
42
+ process.exit(124); // 124 = "parent died" exit code
43
+ }
44
+
45
+ /**
46
+ * Start a lightweight poll that checks if the parent process is still alive.
47
+ * If the parent dies, this worker exits immediately with code 124.
48
+ *
49
+ * The interval is `unref()`'d so it doesn't keep the event loop alive
50
+ * on its own — but if the worker has other work (LLM calls, tool execution),
51
+ * the check continues running in the background.
52
+ */
53
+ export function startParentGuard(parentPid: number): void {
54
+ if (!parentPid || !Number.isFinite(parentPid) || parentPid <= 0) return;
55
+
56
+ // Immediate check — if parent is already dead, don't even start
57
+ if (!isPidAlive(parentPid)) {
58
+ selfTerminate(parentPid);
59
+ }
60
+
61
+ guardInterval = setInterval(() => {
62
+ if (!isPidAlive(parentPid)) {
63
+ if (guardInterval) clearInterval(guardInterval);
64
+ selfTerminate(parentPid);
65
+ }
66
+ }, POLL_INTERVAL_MS);
67
+
68
+ guardInterval.unref();
69
+ }
70
+
71
+ /**
72
+ * Stop the parent guard. Called when the worker finishes normally
73
+ * and doesn't need to watch the parent anymore.
74
+ */
75
+ export function stopParentGuard(): void {
76
+ if (guardInterval) {
77
+ clearInterval(guardInterval);
78
+ guardInterval = undefined;
79
+ }
80
+ }
@@ -15,6 +15,7 @@ export interface BuildPiWorkerArgsInput {
15
15
  model?: string;
16
16
  sessionEnabled?: boolean;
17
17
  maxDepth?: number;
18
+ skillPaths?: string[];
18
19
  env?: NodeJS.ProcessEnv;
19
20
  }
20
21
 
@@ -24,10 +25,16 @@ export interface BuildPiWorkerArgsResult {
24
25
  tempDir?: string;
25
26
  }
26
27
 
28
+ function isValidThinkingLevel(value: string | undefined): value is string {
29
+ return value !== undefined && THINKING_LEVELS.includes(value);
30
+ }
31
+
27
32
  export function applyThinkingSuffix(model: string | undefined, thinking: string | undefined): string | undefined {
28
33
  if (!model || !thinking || thinking === "off") return model;
29
34
  const colonIdx = model.lastIndexOf(":");
30
- if (colonIdx !== -1 && THINKING_LEVELS.includes(model.substring(colonIdx + 1))) return model;
35
+ if (colonIdx !== -1 && isValidThinkingLevel(model.substring(colonIdx + 1))) return model;
36
+ // Invalid config values fall back to Pi's default thinking behavior.
37
+ if (!isValidThinkingLevel(thinking)) return model;
31
38
  return `${model}:${thinking}`;
32
39
  }
33
40
 
@@ -54,8 +61,15 @@ export function buildPiWorkerArgs(input: BuildPiWorkerArgsInput): BuildPiWorkerA
54
61
  const args = ["--mode", "json", "-p"];
55
62
  if (input.sessionEnabled === false) args.push("--no-session");
56
63
 
57
- const model = applyThinkingSuffix(input.model ?? input.agent.model, input.agent.thinking);
58
- if (model) args.push("--model", model);
64
+ const resolvedModel = input.model ?? input.agent.model;
65
+ if (resolvedModel) {
66
+ const modelWithThinking = applyThinkingSuffix(resolvedModel, input.agent.thinking);
67
+ if (modelWithThinking) args.push("--model", modelWithThinking);
68
+ }
69
+ // When no model resolved, pass thinking separately so Pi can apply it to the inherited parent model.
70
+ if (!resolvedModel && input.agent.thinking && input.agent.thinking !== "off" && isValidThinkingLevel(input.agent.thinking)) {
71
+ args.push("--thinking", input.agent.thinking);
72
+ }
59
73
 
60
74
  if (input.agent.tools?.length) args.push("--tools", input.agent.tools.join(","));
61
75
  if (input.agent.extensions !== undefined) {
@@ -65,6 +79,7 @@ export function buildPiWorkerArgs(input: BuildPiWorkerArgsInput): BuildPiWorkerA
65
79
  args.push("--extension", PROMPT_RUNTIME_EXTENSION_PATH);
66
80
  }
67
81
  if (!input.agent.inheritSkills) args.push("--no-skills");
82
+ for (const skillPath of input.skillPaths ?? []) args.push("--skill", skillPath);
68
83
 
69
84
  let tempDir: string | undefined;
70
85
  if (input.agent.systemPrompt) {