@davesheffer/hunch 1.32.3 → 1.32.4

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -61,7 +61,7 @@ hunch integrations check --harness codex --require context,edit-blocking
61
61
 
62
62
  Capabilities are reported as **verified**, **advisory-only**, **unsupported** or **untested**. `--require` fails unless every named capability is verified. `mcp` is verified by a fresh-server probe; hook capabilities become verified only from lifecycle events actually delivered to Hunch's hook on the expected version within the last 30 days (machine-local evidence, the same trust level as the served ledger), so a repository whose agent has actually run shows it, and one that only has configuration does not.
63
63
 
64
- The Codex integration currently supplies MCP and instructions, with no native lifecycle adapter. The opt-in `--probe` verifies a fresh MCP process, not whether an existing host session or model actually followed the memory.
64
+ Codex CLI 0.153+ gets a native lifecycle adapter (`.codex/hooks.json`: session orientation, prompt task IDs from `turn_id`, `apply_patch` pre-edit grounding and strict denial, Stop cards); project-layer hooks load only for a trusted project and must be trusted once in Codex with `/hooks`. The opt-in `--probe` verifies a fresh MCP process, not whether an existing host session or model actually followed the memory.
65
65
 
66
66
  Use `hunch integrations check` in CI to prevent pin drift; add `--require` for capabilities your workflow cannot operate without.
67
67
 
package/dist/cli/index.js CHANGED
@@ -31,6 +31,7 @@ import { registerUpdateCommand } from "./update.js";
31
31
  import { registerReviewMemoryCommands } from "./reviewMemory.js";
32
32
  import { detectInitiator, normalizeInitiator } from "../synthesis/initiator.js";
33
33
  import { inspectIntegrations, formatIntegrationHealth, integrationHealthFails, integrationSessionWarning } from "../integrations/health.js";
34
+ import { publishedStatus } from "../integrations/registry.js";
34
35
  import { HunchStore } from "../store/hunchStore.js";
35
36
  import { JsonStore } from "../store/jsonStore.js";
36
37
  import { selectEmbedder } from "../store/embedder.js";
@@ -4451,15 +4452,21 @@ program
4451
4452
  .option("--provider <provider>", "hook event dialect: claude | vscode | cursor | windsurf | antigravity", "claude")
4452
4453
  .action(async (opts) => {
4453
4454
  // A hook MUST NEVER break the agent: on ANY error or unrecognized input we
4454
- // emit nothing and exit 0 (the action defers to Claude Code's normal flow).
4455
+ // exit 0 and never deny (the action defers to the host's normal flow).
4456
+ // Unrecognized input stays silent; an error after the event was recognized
4457
+ // says "grounding unavailable" in one context line — see the catch below.
4455
4458
  let store = null;
4459
+ // Hoisted so the fail-open catch below can still say which grounding was lost.
4460
+ let provider = null;
4461
+ let eventName = null;
4456
4462
  try {
4457
- const provider = hookProvider(opts.provider);
4463
+ provider = hookProvider(opts.provider);
4458
4464
  if (!provider)
4459
4465
  return;
4460
4466
  const evt = normalizeHookEvent(JSON.parse(await readStdin()), provider);
4461
4467
  if (!evt)
4462
4468
  return;
4469
+ eventName = evt.hook_event_name;
4463
4470
  const root = findRoot();
4464
4471
  // The host delivered this event: runtime evidence for `hunch integrations check`,
4465
4472
  // recorded before any policy decision so firmness never hides delivery itself.
@@ -4769,15 +4776,16 @@ program
4769
4776
  }
4770
4777
  // Pre-edit grounding must resolve the same advertised graph as every CLI
4771
4778
  // and MCP consumer. Any unavailable/mismatched team route falls through to
4772
- // the outer fail-open catch and emits nothing, preserving the hook's
4773
- // non-blocking invariant without false-passing against public/stale memory.
4779
+ // the outer fail-open catch: one "grounding unavailable" line, never a deny,
4780
+ // never grounding from public/stale memory (dec_77d99014e0).
4774
4781
  const opened = openTeamStore(root, { requireFreshTeamMemory: firmness === "strict" });
4775
4782
  store = opened.store;
4776
4783
  if (firmness === "strict" && opened.teamPullStatus
4777
4784
  && opened.teamPullStatus !== "updated" && opened.teamPullStatus !== "current") {
4778
4785
  // A strict deny is only trustworthy when it includes the latest team
4779
4786
  // rules. Offline/busy/unconfigured team memory is unavailable, so the
4780
- // non-blocking hook emits nothing instead of denying from stale state.
4787
+ // non-blocking hook says so instead of denying from stale state.
4788
+ emitContext(provider, "PreToolUse", `Hunch grounding unavailable for this edit; it proceeds ungrounded (team memory is ${opened.teamPullStatus}; strict mode never denies from stale rules). Run \`hunch doctor\`.`);
4781
4789
  return;
4782
4790
  }
4783
4791
  // strict: refuse an edit that hits a BLOCKING invariant (direct OR via blast
@@ -4900,8 +4908,18 @@ program
4900
4908
  }
4901
4909
  emitContext(provider, "PreToolUse", text + reportNotice, recalled ?? undefined);
4902
4910
  }
4903
- catch {
4904
- // swallow — never block an edit on a hook failure
4911
+ catch (e) {
4912
+ // Never block an edit on a hook failure — and never go silent either: an
4913
+ // ungrounded edit that looks grounded gets diagnosed as model flakiness.
4914
+ // One context line, exit 0 (the launcher itself failing stays out of reach).
4915
+ const reason = e instanceof Error ? e.message.split("\n")[0] : "unknown error";
4916
+ if (provider && (eventName === "PreToolUse" || eventName === "SessionStart" || eventName === "UserPromptSubmit")) {
4917
+ const what = eventName === "PreToolUse" ? "for this edit; it proceeds ungrounded" : "for this session";
4918
+ try {
4919
+ emitContext(provider, eventName, `Hunch grounding unavailable ${what} (${reason}). Run \`hunch doctor\`.`);
4920
+ }
4921
+ catch { /* stdout gone */ }
4922
+ }
4905
4923
  }
4906
4924
  finally {
4907
4925
  store?.close();
@@ -6626,6 +6644,21 @@ program
6626
6644
  .action(async () => {
6627
6645
  const integrations = inspectIntegrations(findRoot());
6628
6646
  console.log(formatIntegrationHealth(integrations));
6647
+ // A pin npm cannot serve kills every npx launcher (hooks and MCP) before Hunch
6648
+ // runs, and the hosts report nothing. Name it here; bounded, offline-safe.
6649
+ const published = new Map();
6650
+ for (const v of new Set([integrations.expectedVersion, ...integrations.pins.map(p => p.version)]))
6651
+ published.set(v, publishedStatus(v));
6652
+ const expectedUnpublished = published.get(integrations.expectedVersion) === "unpublished";
6653
+ for (const pin of integrations.pins) {
6654
+ if (published.get(pin.version) !== "unpublished")
6655
+ continue;
6656
+ console.log(`ERROR ${pin.file}: pins Hunch ${pin.version}, which npm cannot serve (ETARGET) — every hook run and MCP launch from this file fails before Hunch starts, silently. Run \`hunch integrations repair-pins\`${expectedUnpublished ? " once the release publishes" : ""}.`);
6657
+ process.exitCode = 1;
6658
+ }
6659
+ if (expectedUnpublished && integrations.pins.every(p => p.version !== integrations.expectedVersion)) {
6660
+ console.log(`note: package.json says ${integrations.expectedVersion}, which is not on npm yet; machine-local pins stay on the last published release until it is (then run \`hunch integrations repair-pins\`).`);
6661
+ }
6629
6662
  // A shared-memory or CLI-only checkout may intentionally have no local
6630
6663
  // assistant config. The explicit integrations check still fails that case.
6631
6664
  if (integrations.harnesses.length > 0 && integrationHealthFails(integrations))
@@ -1,7 +1,8 @@
1
1
  import { existsSync, mkdirSync, readFileSync } from "node:fs";
2
2
  import { findRoot } from "../core/paths.js";
3
3
  import { writeFileAtomic } from "../core/io.js";
4
- import { finishReportTask, forgetReportTask, listReportTasks, pruneReportHistory, readTaskReport, readLessonHistory, startReportTask } from "../core/taskReport.js";
4
+ import { finishReportTask, forgetReportTask, listReportTasks, listTaskSummaries, pruneReportHistory, readTaskReport, readLessonHistory, renderTaskStatusLine, startReportTask, summarizeTaskReport, taskReportStats } from "../core/taskReport.js";
5
+ import { promptTaskId } from "../core/taskReportHook.js";
5
6
  import { DEFAULT_CHECK_TIMEOUT_MS, MAX_CHECK_TIMEOUT_MS, reportSourceSnapshot, runReportCheck, runReportConformance } from "../core/taskReportEvidence.js";
6
7
  import { renderTaskReport, writeTaskReportHtml } from "../core/taskReportRender.js";
7
8
  import { assertReportPath } from "../core/taskReportPaths.js";
@@ -40,6 +41,72 @@ export function registerTaskReportCommands(program, openStore) {
40
41
  finishReportTask(root, id, opts.interrupted ? "interrupted" : "completed");
41
42
  console.log(renderTaskReport(readTaskReport(root, id, reportSourceSnapshot(root).hash)));
42
43
  });
44
+ task.command("list").description("Recent tasks observed in this repository with what Hunch delivered, saved, guarded, and checked")
45
+ .option("--limit <n>", "how many recent tasks (max 30)", "30")
46
+ .option("--json", "machine-readable summaries (consumed by the VS Code Contribution view)")
47
+ .action((opts) => {
48
+ const root = findRoot();
49
+ const summaries = listTaskSummaries(root, Number(opts.limit) || 30, reportSourceSnapshot(root).hash);
50
+ if (opts.json) {
51
+ console.log(JSON.stringify(summaries, null, 2));
52
+ return;
53
+ }
54
+ if (!summaries.length) {
55
+ console.log("No task activity observed yet.");
56
+ return;
57
+ }
58
+ for (const s of summaries)
59
+ console.log(`${s.task.started_at.slice(0, 16).replace("T", " ")} ${s.task.task_id} ${s.task.state.padEnd(11)} ${renderTaskStatusLine(s) || "nothing observed"}`);
60
+ });
61
+ task.command("stats").description("Adherence over a window: how many prompts Hunch reached (delivery), checked, saved, or guarded — from the ledger, never from agent claims")
62
+ .option("--days <days>", "window in days", "7")
63
+ .option("--json", "machine-readable")
64
+ .action((opts) => {
65
+ const stats = taskReportStats(findRoot(), Number(opts.days) || 7);
66
+ if (opts.json) {
67
+ console.log(JSON.stringify(stats, null, 2));
68
+ return;
69
+ }
70
+ const pct = (n) => stats.tasks ? `${Math.round((n / stats.tasks) * 100)}%` : "–";
71
+ console.log(`Hunch adherence, last ${Number(opts.days) || 7} day(s): ${stats.tasks} task(s), ${stats.completed} completed`);
72
+ console.log(` reached by memory (delivery) ${stats.with_delivery} ${pct(stats.with_delivery)}`);
73
+ console.log(` independent check recorded ${stats.with_check} ${pct(stats.with_check)}`);
74
+ console.log(` application claimed by agent ${stats.with_claim} ${pct(stats.with_claim)}`);
75
+ console.log(` memory saved ${stats.with_save} ${pct(stats.with_save)}`);
76
+ console.log(` edit denied ${stats.with_refusal} ${pct(stats.with_refusal)}`);
77
+ console.log(` nothing observed ${stats.empty} ${pct(stats.empty)}`);
78
+ });
79
+ task.command("status").description("One line for a terminal status line: the current prompt's task when Claude Code's status-line JSON arrives on stdin, otherwise the most recent task here")
80
+ .option("--json", "machine-readable summary")
81
+ .action(async (opts) => {
82
+ const input = process.stdin.isTTY ? "" : await readStdinText();
83
+ let root = findRoot();
84
+ let taskId = null;
85
+ try {
86
+ const host = input.trim() ? JSON.parse(input) : {};
87
+ const dir = host.workspace?.current_dir ?? host.cwd;
88
+ if (dir)
89
+ root = findRoot(dir);
90
+ if (host.session_id && host.prompt_id)
91
+ taskId = promptTaskId(root, host.session_id, host.prompt_id);
92
+ }
93
+ catch { /* a malformed host payload falls back to the most recent task */ }
94
+ let summary = null;
95
+ try {
96
+ const snapshot = reportSourceSnapshot(root).hash;
97
+ summary = taskId ? summarizeTaskReport(root, taskId, snapshot) : listTaskSummaries(root, 1, snapshot)[0] ?? null;
98
+ }
99
+ catch {
100
+ summary = null;
101
+ }
102
+ if (opts.json) {
103
+ console.log(JSON.stringify(summary));
104
+ return;
105
+ }
106
+ const line = renderTaskStatusLine(summary);
107
+ if (line)
108
+ console.log(line);
109
+ });
43
110
  task.command("forget <id>").description("Delete this closed task's local observations and generated report; retains project memory")
44
111
  .action((id) => { forgetReportTask(findRoot(), id); console.log("Task report history removed; project memory retained."); });
45
112
  task.command("prune").description("Delete local report history for tasks closed more than the retention period ago")
@@ -125,4 +192,16 @@ export function registerTaskReportCommands(program, openStore) {
125
192
  console.log(opts.json ? JSON.stringify(report, null, 2) : renderTaskReport(report));
126
193
  });
127
194
  }
195
+ function readStdinText() {
196
+ return new Promise(resolve => {
197
+ let data = "";
198
+ const done = () => resolve(data);
199
+ process.stdin.setEncoding("utf8");
200
+ process.stdin.on("data", chunk => { data += chunk; });
201
+ process.stdin.on("end", done);
202
+ process.stdin.on("error", done);
203
+ // A host that opened stdin but never writes must not hang the status line.
204
+ setTimeout(done, 1500).unref();
205
+ });
206
+ }
128
207
  //# sourceMappingURL=taskReport.js.map
@@ -5,7 +5,7 @@
5
5
  * fields or tools. Keep that variability here so the policy engine receives
6
6
  * the same small, fail-open shape regardless of the assistant that emitted it.
7
7
  */
8
- export declare const HOOK_PROVIDERS: readonly ["claude", "vscode", "windsurf", "antigravity", "cursor"];
8
+ export declare const HOOK_PROVIDERS: readonly ["claude", "codex", "vscode", "windsurf", "antigravity", "cursor"];
9
9
  export type HookProvider = (typeof HOOK_PROVIDERS)[number];
10
10
  export type HunchHookEvent = "PreToolUse" | "PostToolUse" | "PostToolUseFailure" | "UserPromptSubmit" | "SessionStart" | "SubagentStart" | "PreCompact" | "Stop";
11
11
  export interface HunchToolInput {
@@ -5,7 +5,7 @@
5
5
  * fields or tools. Keep that variability here so the policy engine receives
6
6
  * the same small, fail-open shape regardless of the assistant that emitted it.
7
7
  */
8
- export const HOOK_PROVIDERS = ["claude", "vscode", "windsurf", "antigravity", "cursor"];
8
+ export const HOOK_PROVIDERS = ["claude", "codex", "vscode", "windsurf", "antigravity", "cursor"];
9
9
  function obj(value) {
10
10
  return value && typeof value === "object" && !Array.isArray(value) ? value : null;
11
11
  }
@@ -49,10 +49,24 @@ function edits(value) {
49
49
  .map((item) => ({ new_string: stringAt(item, "new_string", "newString", "ReplacementContent", "replacementContent") }));
50
50
  return normalized.length ? normalized : undefined;
51
51
  }
52
+ /** Codex edits files through `apply_patch`, whose input is the patch text itself
53
+ * (`*** Update File: path`). The first touched path becomes the edit target so the
54
+ * per-file pre-edit gate applies; the whole patch stands in for the new content. */
55
+ const PATCH_FILE = /^\*\*\* (?:Update|Add|Delete) File: (.+?)\s*$/m;
56
+ function applyPatchInput(raw) {
57
+ const patch = [raw.input, raw.patch, raw.content].find((v) => typeof v === "string" && /\*\*\* Begin Patch/.test(v));
58
+ if (!patch)
59
+ return undefined;
60
+ const file = PATCH_FILE.exec(patch)?.[1];
61
+ return file ? { file_path: file, content: patch } : undefined;
62
+ }
52
63
  function normalizeToolInput(value) {
53
64
  const raw = obj(value);
54
65
  if (!raw)
55
66
  return undefined;
67
+ const patched = applyPatchInput(raw);
68
+ if (patched)
69
+ return patched;
56
70
  const replacementChunks = Array.isArray(raw.ReplacementChunks) ? raw.ReplacementChunks : raw.replacementChunks;
57
71
  const chunkEdits = Array.isArray(replacementChunks)
58
72
  ? replacementChunks.map((chunk) => obj(chunk)).filter((chunk) => !!chunk)
@@ -63,7 +77,9 @@ function normalizeToolInput(value) {
63
77
  new_string: stringAt(raw, "new_string", "newString", "ReplacementContent", "replacementContent", "TargetContent", "targetContent"),
64
78
  content: stringAt(raw, "content", "contents", "CodeContent", "codeContent"),
65
79
  edits: edits(raw.edits) ?? edits(raw.files) ?? chunkEdits,
66
- command: stringAt(raw, "command", "commandLine", "CommandLine", "cmd"),
80
+ // Codex's shell tools carry argv arrays (["bash", "-lc", ""]); policies read one string.
81
+ command: stringAt(raw, "command", "commandLine", "CommandLine", "cmd")
82
+ ?? (Array.isArray(raw.command) && raw.command.every(p => typeof p === "string") ? raw.command.join(" ") : undefined),
67
83
  skill: stringAt(raw, "skill", "skillName", "name"),
68
84
  };
69
85
  return Object.values(out).some((v) => v !== undefined) ? out : undefined;
@@ -194,7 +210,10 @@ export function normalizeHookEvent(raw, provider) {
194
210
  return {
195
211
  hook_event_name: event,
196
212
  session_id: stringAt(input, "session_id", "sessionId", "conversation_id", "conversationId"),
197
- ...(provider === "claude" ? Object.fromEntries(["prompt_id", "cwd", "agent_id"].filter(key => input[key] !== undefined).map(key => [key, typeof input[key] === "string" ? input[key] : ""])) : {}),
213
+ // Codex delivers the same lifecycle payload with `turn_id` where Claude Code
214
+ // says `prompt_id`; both are native per-prompt identities, never synthesized.
215
+ ...(provider === "codex" && input.prompt_id === undefined && input.turn_id !== undefined ? { prompt_id: typeof input.turn_id === "string" ? input.turn_id : "" } : {}),
216
+ ...(provider === "claude" || provider === "codex" ? Object.fromEntries(["prompt_id", "cwd", "agent_id"].filter(key => input[key] !== undefined).map(key => [key, typeof input[key] === "string" ? input[key] : ""])) : {}),
198
217
  tool_name: hunchToolName(stringAt(input, "tool_name", "toolName"), toolInput ?? {}),
199
218
  tool_input: toolInput,
200
219
  ...(toolOutcome ? { tool_outcome: toolOutcome } : {}),
@@ -11,6 +11,9 @@ export declare const FIRMNESS_LEVELS: readonly Firmness[];
11
11
  export declare const DEFAULT_FIRMNESS: Firmness;
12
12
  export interface HunchConfig {
13
13
  firmness: Firmness;
14
+ /** MCP tool groups beyond the everyday set: `all`, `core`, or `core,nuryel`
15
+ * (see src/mcp/toolset.ts). Undefined = decide from the root's contents. */
16
+ mcp_tools?: string;
14
17
  }
15
18
  export declare function isFirmness(v: unknown): v is Firmness;
16
19
  /** Read `.hunch/config.json`. A missing/unparseable file, or an unknown firmness
@@ -19,7 +19,10 @@ export function readConfig(paths) {
19
19
  return defaults();
20
20
  try {
21
21
  const raw = JSON.parse(readFileSync(paths.config, "utf8"));
22
- return { firmness: isFirmness(raw.firmness) ? raw.firmness : DEFAULT_FIRMNESS };
22
+ return {
23
+ firmness: isFirmness(raw.firmness) ? raw.firmness : DEFAULT_FIRMNESS,
24
+ ...(typeof raw.mcp_tools === "string" && raw.mcp_tools.trim() ? { mcp_tools: raw.mcp_tools.trim() } : {}),
25
+ };
23
26
  }
24
27
  catch {
25
28
  return defaults();
@@ -222,6 +222,55 @@ export declare function recordReportConformance(root: string, taskId: string, co
222
222
  export declare function recordReportCheck(root: string, taskId: string, check: ReportCheck): string;
223
223
  export declare function beginReportCheck(root: string, taskId: string, label: string): string;
224
224
  export declare function finishReportTask(root: string, taskId: string, state?: "completed" | "interrupted"): ReportTask;
225
+ /** A report with no observation of any kind. Presentation surfaces may stay
226
+ * silent for it; the task row itself is retained so "never touched Hunch" is
227
+ * still countable (hunch report / the VS Code view / task list). */
228
+ export declare function isEmptyTaskReport(report: Pick<TaskReport, "deliveries" | "claims" | "checks" | "conformance" | "saves" | "refusals">): boolean;
229
+ export interface TaskSummary {
230
+ task: ReportTask;
231
+ deliveries: number;
232
+ lessons: number;
233
+ claims: number;
234
+ saves: number;
235
+ refusals: number;
236
+ /** Standing of the last recorded check, or null when none ran. */
237
+ check: {
238
+ label: string;
239
+ state: "passed" | "failed" | "timed out" | "cancelled";
240
+ current: boolean;
241
+ } | null;
242
+ /** Any delivered rule evaluated as violated on the changed files. */
243
+ violated: boolean;
244
+ coverage: TaskReport["coverage"];
245
+ empty: boolean;
246
+ /** Generated evidence view, when one has been written for this task. */
247
+ report_html: string | null;
248
+ /** Set when the observation ledger could not be read for this task. */
249
+ error: string | null;
250
+ }
251
+ /** One bounded summary per recent task for status lines and host views; the
252
+ * card and evidence view remain the authoritative renderings. */
253
+ export declare function summarizeTaskReport(root: string, taskId: string, currentSnapshot?: string | null): TaskSummary;
254
+ export declare function listTaskSummaries(root: string, limit?: number, currentSnapshot?: string | null): TaskSummary[];
255
+ /** One line for a terminal status line. Empty string when nothing was observed
256
+ * for the task, so a bare prompt shows no Hunch noise at all. */
257
+ export declare function renderTaskStatusLine(summary: TaskSummary | null): string;
258
+ export interface TaskReportStats {
259
+ since: string;
260
+ tasks: number;
261
+ completed: number;
262
+ with_delivery: number;
263
+ with_check: number;
264
+ with_claim: number;
265
+ with_save: number;
266
+ with_refusal: number;
267
+ empty: number;
268
+ /** with_delivery / tasks, the adherence number worth watching; null when no tasks. */
269
+ delivery_rate: number | null;
270
+ }
271
+ /** Adherence over a window: how many prompts Hunch actually reached. Counts
272
+ * come from the ledger, never from agent claims; a claim is counted as a claim. */
273
+ export declare function taskReportStats(root: string, days?: number): TaskReportStats;
225
274
  export declare function listReportTasks(root: string): ReportTask[];
226
275
  export declare function reportActivity(root: string): string;
227
276
  /** Exact task deletion is a user-invoked operation, never a memory deletion. */
@@ -368,6 +368,105 @@ export function finishReportTask(root, taskId, state = "completed") {
368
368
  return finished;
369
369
  }));
370
370
  }
371
+ /** A report with no observation of any kind. Presentation surfaces may stay
372
+ * silent for it; the task row itself is retained so "never touched Hunch" is
373
+ * still countable (hunch report / the VS Code view / task list). */
374
+ export function isEmptyTaskReport(report) {
375
+ return !report.deliveries.length && !report.claims.length && !report.checks.length && !report.conformance.length && !report.saves.length && !report.refusals.length;
376
+ }
377
+ /** One bounded summary per recent task for status lines and host views; the
378
+ * card and evidence view remain the authoritative renderings. */
379
+ export function summarizeTaskReport(root, taskId, currentSnapshot = null) {
380
+ const html = join(root, ".hunch-cache", "reports", `${taskId}.html`);
381
+ try {
382
+ const report = readTaskReport(root, taskId, currentSnapshot);
383
+ const last = report.checks.at(-1);
384
+ const standing = new Map(report.conformance.map(r => [`${r.kind}:${r.record_id}:${r.content_hash}`, r]));
385
+ return {
386
+ task: report.task,
387
+ deliveries: report.deliveries.length,
388
+ lessons: new Set(report.deliveries.flatMap(d => d.records).map(r => `${r.kind}:${r.record_id}`)).size,
389
+ claims: report.claims.length,
390
+ saves: report.saves.length,
391
+ refusals: report.refusals.length,
392
+ check: last ? { label: last.label, state: last.cancelled ? "cancelled" : last.timed_out ? "timed out" : last.exit_code === 0 ? "passed" : "failed", current: last.current } : null,
393
+ violated: [...standing.values()].some(r => r.outcome === "violated"),
394
+ coverage: report.coverage,
395
+ empty: isEmptyTaskReport(report),
396
+ report_html: existsSync(html) ? html : null,
397
+ error: null,
398
+ };
399
+ }
400
+ catch (e) {
401
+ return taskDb(root, db => {
402
+ const row = db.prepare("SELECT body FROM report_tasks WHERE task_id = ?").get(taskId);
403
+ if (!row)
404
+ throw e;
405
+ return { task: TaskSchema.parse(JSON.parse(row.body)), deliveries: 0, lessons: 0, claims: 0, saves: 0, refusals: 0, check: null, violated: false, coverage: "no-delivery-observed", empty: true, report_html: existsSync(html) ? html : null, error: e.message };
406
+ });
407
+ }
408
+ }
409
+ export function listTaskSummaries(root, limit = 30, currentSnapshot = null) {
410
+ if (!existsSync(join(root, ".hunch-cache", "served.db")))
411
+ return [];
412
+ return listReportTasks(root).slice(0, Math.max(1, Math.min(limit, 30))).map(task => summarizeTaskReport(root, task.task_id, currentSnapshot));
413
+ }
414
+ /** One line for a terminal status line. Empty string when nothing was observed
415
+ * for the task, so a bare prompt shows no Hunch noise at all. */
416
+ export function renderTaskStatusLine(summary) {
417
+ if (!summary || summary.empty)
418
+ return "";
419
+ const parts = [`Hunch`];
420
+ parts.push(summary.lessons ? `${summary.lessons} lesson${summary.lessons === 1 ? "" : "s"} recalled` : summary.deliveries ? "memory delivered" : "no delivery");
421
+ if (summary.violated)
422
+ parts.push("rule violated");
423
+ else if (summary.claims)
424
+ parts.push(`${summary.claims} applied`);
425
+ if (summary.saves)
426
+ parts.push(`${summary.saves} saved`);
427
+ if (summary.refusals)
428
+ parts.push("edit denied");
429
+ if (summary.check)
430
+ parts.push(`${summary.check.label}: ${summary.check.state}${summary.check.current ? "" : " (source changed)"}`);
431
+ else
432
+ parts.push("no check recorded");
433
+ return parts.join(" · ");
434
+ }
435
+ /** Adherence over a window: how many prompts Hunch actually reached. Counts
436
+ * come from the ledger, never from agent claims; a claim is counted as a claim. */
437
+ export function taskReportStats(root, days = 7) {
438
+ const since = new Date(Date.now() - Math.max(1, days) * 86_400_000).toISOString();
439
+ const empty = { since, tasks: 0, completed: 0, with_delivery: 0, with_check: 0, with_claim: 0, with_save: 0, with_refusal: 0, empty: 0, delivery_rate: null };
440
+ if (!existsSync(join(root, ".hunch-cache", "served.db")))
441
+ return empty;
442
+ return taskDb(root, db => {
443
+ const tasks = db.prepare("SELECT body FROM report_tasks WHERE scope = ? AND json_extract(body, '$.started_at') >= ?").all(scopeOf(root), since)
444
+ .map(r => TaskSchema.parse(JSON.parse(r.body)));
445
+ if (!tasks.length)
446
+ return empty;
447
+ const kinds = (taskId) => new Set(db.prepare("SELECT DISTINCT kind FROM report_events WHERE task_id = ?").all(taskId).map(r => r.kind));
448
+ const stats = { ...empty, tasks: tasks.length };
449
+ for (const task of tasks) {
450
+ const k = kinds(task.task_id);
451
+ if (task.state === "completed")
452
+ stats.completed++;
453
+ if (k.has("delivery"))
454
+ stats.with_delivery++;
455
+ if (k.has("check") || k.has("check-start"))
456
+ stats.with_check++;
457
+ if (k.has("claim"))
458
+ stats.with_claim++;
459
+ if (k.has("save"))
460
+ stats.with_save++;
461
+ if (k.has("refusal"))
462
+ stats.with_refusal++;
463
+ if (!k.size)
464
+ stats.empty++;
465
+ }
466
+ stats.delivery_rate = stats.with_delivery / stats.tasks;
467
+ return stats;
468
+ });
469
+ }
371
470
  export function listReportTasks(root) {
372
471
  return taskDb(root, db => db.prepare("SELECT body FROM report_tasks WHERE scope = ? ORDER BY rowid DESC LIMIT 30").all(scopeOf(root)).map(r => TaskSchema.parse(JSON.parse(r.body))));
373
472
  }
@@ -1,10 +1,16 @@
1
1
  import type { HookProvider, HunchHookInput } from "./agenthook.js";
2
+ /** The exact task identity a Claude Code prompt maps to. The status line receives
3
+ * the same session_id/prompt_id on stdin, so it can name the prompt's task too. */
4
+ export declare function promptTaskId(root: string, sessionId: string, promptId: string, agentId?: string | null, provider?: HookProvider): string;
2
5
  export declare function hookReportTaskId(root: string, provider: HookProvider, event: HunchHookInput): string | null;
3
6
  /** Every prompt receives its exact ID, even when ambient reminders were deduped.
4
7
  * No raw prompt, host session identifier, or transcript is retained. */
5
8
  export declare function startHookReport(root: string, provider: HookProvider, event: HunchHookInput): string | null;
6
9
  /** A presentation notice never denies Stop or injects another model turn. Stop
7
- * can precede another hook's continuation, so it does not close an open task. */
10
+ * can precede another hook's continuation, so it does not close an open task.
11
+ * A prompt with no observation at all prints nothing: the empty task row stays
12
+ * in the ledger (hunch task list, the VS Code Contribution view) so "never
13
+ * touched Hunch" remains countable without a five-line notice per prompt. */
8
14
  export declare function stopHookReport(root: string, provider: HookProvider, event: HunchHookInput): {
9
15
  systemMessage: string;
10
16
  } | null;
@@ -3,11 +3,19 @@
3
3
  import { realpathSync } from "node:fs";
4
4
  import { pathToFileURL } from "node:url";
5
5
  import { findRoot } from "./paths.js";
6
- import { readTaskReport, recordReportRefusal, reportHash, reportPresentationEnabled, startReportTask } from "./taskReport.js";
6
+ import { isEmptyTaskReport, readTaskReport, recordReportRefusal, reportHash, reportPresentationEnabled, startReportTask } from "./taskReport.js";
7
7
  import { reportSourceSnapshot } from "./taskReportEvidence.js";
8
8
  import { renderTaskReport, writeTaskReportHtml } from "./taskReportRender.js";
9
+ /** The exact task identity a Claude Code prompt maps to. The status line receives
10
+ * the same session_id/prompt_id on stdin, so it can name the prompt's task too. */
11
+ export function promptTaskId(root, sessionId, promptId, agentId = null, provider = "claude") {
12
+ return `htask_${reportHash([realpathSync(root), provider, sessionId, promptId, agentId]).slice(7, 31)}`;
13
+ }
14
+ /** Hosts whose hooks deliver a native per-prompt identity (Claude Code's
15
+ * prompt_id, Codex's turn_id). Others get no task from a hook. */
16
+ const NATIVE_PROMPT_HOSTS = new Set(["claude", "codex"]);
9
17
  function identity(root, provider, event) {
10
- if (provider !== "claude" || !event.cwd || realpathSync(findRoot(event.cwd)) !== realpathSync(root))
18
+ if (!NATIVE_PROMPT_HOSTS.has(provider) || !event.cwd || realpathSync(findRoot(event.cwd)) !== realpathSync(root))
11
19
  return null;
12
20
  for (const value of [event.session_id, event.prompt_id, event.agent_id]) {
13
21
  if (value !== undefined && (!value.length || value.length > 1024 || /[\u0000-\u001f\u007f]/.test(value)))
@@ -17,7 +25,7 @@ function identity(root, provider, event) {
17
25
  return null;
18
26
  if (!event.prompt_id)
19
27
  return "legacy";
20
- return `htask_${reportHash([realpathSync(root), provider, event.session_id, event.prompt_id, event.agent_id ?? null]).slice(7, 31)}`;
28
+ return promptTaskId(root, event.session_id, event.prompt_id, event.agent_id ?? null, provider);
21
29
  }
22
30
  export function hookReportTaskId(root, provider, event) {
23
31
  try {
@@ -38,7 +46,10 @@ export function startHookReport(root, provider, event) {
38
46
  return `Hunch has opened this prompt's report: ${task.task_id}. Reuse this exact ID for this prompt. Call hunch_task(action: "start", task_id: "${task.task_id}", title: "Claude task") to obtain verification_argv; do not create another report. Pass this task_id to hunch_context and decision/correction/finding captures, and finish with hunch_task before responding. A host Stop notice will show the evidence even if no task-linked memory was observed.`;
39
47
  }
40
48
  /** A presentation notice never denies Stop or injects another model turn. Stop
41
- * can precede another hook's continuation, so it does not close an open task. */
49
+ * can precede another hook's continuation, so it does not close an open task.
50
+ * A prompt with no observation at all prints nothing: the empty task row stays
51
+ * in the ledger (hunch task list, the VS Code Contribution view) so "never
52
+ * touched Hunch" remains countable without a five-line notice per prompt. */
42
53
  export function stopHookReport(root, provider, event) {
43
54
  if (!reportPresentationEnabled(root))
44
55
  return null;
@@ -46,9 +57,11 @@ export function stopHookReport(root, provider, event) {
46
57
  if (!id)
47
58
  return null;
48
59
  if (id === "legacy")
49
- return { systemMessage: "Hunch hook active. This Claude version does not provide an exact prompt identifier, so contribution for this response is unverified. Explicit task reports remain available with hunch report." };
60
+ return { systemMessage: "Hunch hook active. This host version does not provide an exact prompt identifier, so contribution for this response is unverified. Explicit task reports remain available with hunch report." };
50
61
  try {
51
62
  const report = readTaskReport(root, id, reportSourceSnapshot(root).hash);
63
+ if (isEmptyTaskReport(report))
64
+ return null;
52
65
  let card = renderTaskReport(report);
53
66
  try {
54
67
  const file = writeTaskReportHtml(root, id);
@@ -47,7 +47,7 @@ export function renderHunchSection(store, root) {
47
47
  lines.push("**Consult Hunch via the `hunch_*` MCP tools — pick by MOMENT, not from memory:**");
48
48
  lines.push("");
49
49
  lines.push("**Orient (session/task start):**");
50
- lines.push("- For a new user task, call `hunch_task(action: \"start\", title: <short task title>)` once and retain its `task_id`. If a native prompt hook already supplied a task ID, reuse its exact start arguments instead of creating another task; each new native prompt has its own ID. Otherwise reuse the ID for follow-up work on the same task; never borrow another task's ID. This is task bookkeeping; `hunch_context` remains the first memory lookup. If reporting fails, continue the work and disclose the gap.");
50
+ lines.push("- For a new user task, call `hunch_task(action: \"start\", title: <short task title>)` once and retain its `task_id`. Claude Code's prompt hook supplies a task ID natively — reuse its exact start arguments instead of creating another task (each new prompt has its own ID). Codex supplies one the same way once its `.codex/hooks.json` is trusted (`/hooks`). Hosts without prompt hooks (Windsurf, Cursor) never receive one: start the task yourself. Reuse the ID for follow-up work on the same task; never borrow another task's ID. This is task bookkeeping; `hunch_context` remains the first memory lookup. If reporting fails, continue the work and disclose the gap.");
51
51
  lines.push("- When the user asks to **update Hunch**, run `hunch update` from this repository root. It updates to the latest release and repairs all configured harness pins. Use `hunch update --global` to also update a global CLI alongside a repository dependency; reconnect active MCP sessions afterward.");
52
52
  lines.push("- `hunch_context(target, task_id)` — the minimal relevant slice for what you're about to do; a task phrase falls back to the closest graph matches. **Call FIRST** for memory. Include the current task ID on each context call so its contribution is inspectable.");
53
53
  lines.push("- `hunch_structure(target?)` — the indexed shape of the repo/dir/file/symbol — orient from the graph, not grep rounds.");
@@ -10,9 +10,9 @@ export declare const HARNESSES: {
10
10
  };
11
11
  readonly codex: {
12
12
  readonly mcp: ".codex/config.toml";
13
- readonly hooks: "";
14
- readonly key: "";
15
- readonly events: readonly [];
13
+ readonly hooks: ".codex/hooks.json";
14
+ readonly key: "hooks";
15
+ readonly events: readonly ["SessionStart", "PreToolUse", "PostToolUse", "PreCompact"];
16
16
  };
17
17
  readonly cursor: {
18
18
  readonly mcp: ".cursor/mcp.json";
@@ -58,6 +58,11 @@ export interface IntegrationHealth {
58
58
  scope: "repository-config";
59
59
  issues: HealthIssue[];
60
60
  harnesses: HarnessHealth[];
61
+ /** Every exact Hunch pin found in repository launch config, once per file+version. */
62
+ pins: Array<{
63
+ file: string;
64
+ version: string;
65
+ }>;
61
66
  }
62
67
  export declare function readLauncher(root: string, harness: Harness): {
63
68
  command: string;
@@ -65,9 +70,18 @@ export declare function readLauncher(root: string, harness: Harness): {
65
70
  customEnvironment: boolean;
66
71
  };
67
72
  export declare function inspectIntegrations(root: string, selected?: Harness): IntegrationHealth;
73
+ /** Harness launch files git ignores: this machine's config, never the tag's. A
74
+ * release cut may keep these at the last published version (see
75
+ * tooling/sync-version-pins.mjs) so hooks and MCP never point at a version npm
76
+ * cannot serve. Unknown git state yields [] — callers then treat nothing as local. */
77
+ export declare function machineLocalIntegrationFiles(root: string): string[];
68
78
  /** Repair only exact published pins. Preserve formatting and all other values.
69
- * Preflight every affected file before writing any; reject malformed JSON/TOML. */
70
- export declare function repairIntegrationPins(root: string): string[];
79
+ * Preflight every affected file before writing any; reject malformed JSON/TOML.
80
+ * `skip` leaves a file untouched (used to keep machine-local pins on a version
81
+ * npm can actually serve while a release is still publishing). */
82
+ export declare function repairIntegrationPins(root: string, opts?: {
83
+ skip?: (file: string) => boolean;
84
+ }): string[];
71
85
  export declare function integrationHealthFails(report: IntegrationHealth, required?: readonly Capability[]): boolean;
72
86
  export declare function formatIntegrationHealth(report: IntegrationHealth): string;
73
87
  /** Bounded session warning; diagnostics must never break hook execution. */