@dogfood-lab/findings 1.4.0 → 1.5.0

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/cli.js CHANGED
@@ -279,6 +279,14 @@ Synthesis artifacts (patterns / recommendations / doctrine):
279
279
  doctrine <accept|reject> <id> [--actor X] [--reason Y]
280
280
  doctrine queue
281
281
 
282
+ Advice (read-only, derived from accepted artifacts):
283
+ advise --surface <surface> [--execution-mode <mode>] [--repo <org/repo>] [--json]
284
+ Bootstrap guidance for a new repo/surface.
285
+ --json emits the structured advice bundle
286
+ as pure JSON (pipeable; no human text on
287
+ stdout) for shipcheck / repo-knowledge.
288
+ sync-export [--json] Export accepted artifacts for downstream sync.
289
+
282
290
  Note: patterns carry a literal "invalidated" status; recommendations and
283
291
  doctrine do not, so invalidate is supported for patterns only. The review /
284
292
  reopen verbs target the intermediate "reviewed" state, which the artifact
@@ -1095,6 +1103,14 @@ Filters (for list):
1095
1103
  const bundle = generateAdviceBundle(ROOT, { surface, executionMode, repo });
1096
1104
  const a = bundle.advice;
1097
1105
 
1106
+ // --json keeps stdout pure JSON so downstream tools (shipcheck,
1107
+ // repo-knowledge) can pipe the structured bundle without scraping the
1108
+ // human formatter. Matches the sync-export verb's flags.json idiom below.
1109
+ if (flags.json) {
1110
+ console.log(JSON.stringify(bundle, null, 2));
1111
+ process.exit(0);
1112
+ }
1113
+
1098
1114
  console.log(`Advice for: ${[surface, executionMode, repo].filter(Boolean).join(', ') || 'general'}\n`);
1099
1115
 
1100
1116
  if (a.starter_checks.length > 0) {
package/derive/ids.js CHANGED
@@ -5,17 +5,33 @@
5
5
  * No timestamp noise in IDs.
6
6
  */
7
7
 
8
+ import { createHash } from 'node:crypto';
9
+
10
+ /**
11
+ * Delimiter for the boundary hash: a NUL code unit cannot appear in a repo
12
+ * slug (`[a-zA-Z0-9_.-]`) nor in a derived lesson slug, so joining components
13
+ * with it means no component value can forge a false boundary between them.
14
+ */
15
+ const BOUNDARY_DELIM = '\u0000';
16
+
8
17
  /**
9
18
  * Generate a stable finding ID from derivation context.
10
- * Format: dfind-<repo-slug>-<lesson-slug>
19
+ * Format: dfind-<repo-slug>-<lesson-slug>-<boundary-hash>
20
+ *
21
+ * The trailing boundary hash disambiguates inputs whose component boundary is
22
+ * itself an underscore (findings-A-002): `sanitize` folds `_` to `-`, so
23
+ * (repoSlug='a_b', lessonSlug='c') and (repoSlug='a', lessonSlug='b_c') both
24
+ * flatten the human-readable middle to `a-b-c` and previously produced the
25
+ * same id. The hash is computed over the components joined by BOUNDARY_DELIM,
26
+ * so the two tuples hash differently and never collapse to one finding_id.
11
27
  *
12
28
  * @param {string} repoSlug - e.g. "repo-crawler-mcp"
13
29
  * @param {string} lessonSlug - e.g. "surface-misclassification"
14
30
  * @returns {string}
15
31
  */
16
32
  export function generateFindingId(repoSlug, lessonSlug) {
17
- const normalized = `dfind-${sanitize(repoSlug)}-${sanitize(lessonSlug)}`;
18
- return normalized;
33
+ const boundary = boundaryHash([repoSlug, lessonSlug]);
34
+ return `dfind-${sanitize(repoSlug)}-${sanitize(lessonSlug)}-${boundary}`;
19
35
  }
20
36
 
21
37
  /**
@@ -46,3 +62,15 @@ function sanitize(s) {
46
62
  .replace(/-+/g, '-')
47
63
  .replace(/^-|-$/g, '');
48
64
  }
65
+
66
+ /**
67
+ * Short stable hash of an ordered component tuple, joined by BOUNDARY_DELIM so
68
+ * no component value can forge a false boundary. 8 lowercase-hex chars are
69
+ * ample for the finding/pattern id space (a per-repo, per-lesson namespace).
70
+ *
71
+ * @param {string[]} components
72
+ * @returns {string}
73
+ */
74
+ export function boundaryHash(components) {
75
+ return createHash('sha256').update(components.join(BOUNDARY_DELIM)).digest('hex').slice(0, 8);
76
+ }
@@ -34,6 +34,7 @@ import { mkdirSync, existsSync } from 'node:fs';
34
34
  import { resolve, dirname } from 'node:path';
35
35
  import yaml from 'js-yaml';
36
36
 
37
+ import { isUnsafeSegment } from '@dogfood-lab/ingest/lib/unsafe-segment.js';
37
38
  import { atomicWriteFileSync } from '../lib/atomic-write.js';
38
39
  import { validateFinding } from '../validate.js';
39
40
 
@@ -75,6 +76,26 @@ export class FindingIdCollisionError extends Error {
75
76
  }
76
77
  }
77
78
 
79
+ /**
80
+ * Structured error thrown when a finding's `repo` splits into an org/repo
81
+ * segment that is a path-traversal vector (`..` or a path separator).
82
+ *
83
+ * findings-A-001 — write-side traversal guard. The READ side
84
+ * (`loadRecordsForRepoWithSkips`) already rejects such segments via
85
+ * `isUnsafeSegment`; the WRITE side did not, so a schema-valid
86
+ * `repo: '../policies'` (the dogfood-finding `repo` pattern admits `.`/`..`)
87
+ * resolved one level under rootDir and wrote outside `findings/`.
88
+ */
89
+ export class FindingUnsafeRepoError extends Error {
90
+ constructor(repo, findingId) {
91
+ super(`unsafe repo path segment in finding${findingId ? ` (${findingId})` : ''}: '${repo}' contains a path-traversal or separator and was refused before any write (findings-A-001).`);
92
+ this.name = 'FindingUnsafeRepoError';
93
+ this.code = 'FINDING_UNSAFE_REPO';
94
+ this.repo = repo;
95
+ this.findingId = findingId;
96
+ }
97
+ }
98
+
78
99
  /**
79
100
  * Process-level memory of ids that have already been written via
80
101
  * `writeFinding` (singleton) OR the singleton-call inside `writeFindings`
@@ -127,6 +148,13 @@ export function writeFinding(rootDir, finding) {
127
148
  const [org, repo] = (finding.repo || '').split('/');
128
149
  if (!org || !repo) throw new Error(`Invalid repo in finding: ${finding.repo}`);
129
150
 
151
+ // findings-A-001 — write-side path-traversal guard. Mirrors the READ side
152
+ // (load-records.js loadRecordsForRepoWithSkips) so a schema-valid
153
+ // `repo: '../policies'` cannot escape `findings/` into a sibling data dir.
154
+ if (isUnsafeSegment(org) || isUnsafeSegment(repo)) {
155
+ throw new FindingUnsafeRepoError(finding.repo, finding.finding_id);
156
+ }
157
+
130
158
  // L3-001 (Wave A2 amend2): same-process same-id refusal. Programmatic
131
159
  // callers that loop over assembleFinding output now fail-closed at the
132
160
  // singleton path, not silently at the atomicWriteFileSync.
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@dogfood-lab/findings",
3
- "version": "1.4.0",
3
+ "version": "1.5.0",
4
4
  "type": "module",
5
5
  "description": "Finding contract spine for testing-os. Validates, reads, lists, and queries evidence-bound findings — the fourth contract alongside record, scenario, and policy.",
6
6
  "main": "index.js",
package/reader.js CHANGED
@@ -4,7 +4,7 @@
4
4
  */
5
5
 
6
6
  import { readdirSync, existsSync, statSync } from 'node:fs';
7
- import { resolve, join, basename, extname } from 'node:path';
7
+ import { resolve, join, basename, extname, relative } from 'node:path';
8
8
  import { parseFinding, validateFinding } from './validate.js';
9
9
 
10
10
  /**
@@ -129,9 +129,21 @@ export function loadFindings(rootDir, opts = {}) {
129
129
  * @returns {{ path: string, data: object, valid: boolean, errors: Array } | null}
130
130
  */
131
131
  export function findById(rootDir, findingId) {
132
+ // B-002 — collect torn finding files encountered during the scan. The
133
+ // finding_id lives INSIDE the file, so a torn file cannot be matched to the
134
+ // requested id; if the id turns out absent from every clean file, any torn
135
+ // file may be the one hiding it. Surfacing them on stderr (below) mirrors the
136
+ // derive CLI's "N finding(s) skipped (torn/unreadable)" honesty so a torn
137
+ // YAML never silently masquerades as a not-found id.
138
+ const torn = [];
139
+
132
140
  // Search real findings
133
141
  for (const filePath of discoverFindings(rootDir)) {
134
- const { data } = parseFinding(filePath);
142
+ const { data, error } = parseFinding(filePath);
143
+ if (error) {
144
+ torn.push({ path: filePath, error });
145
+ continue;
146
+ }
135
147
  if (data && data.finding_id === findingId) {
136
148
  const result = validateFinding(data);
137
149
  return { path: filePath, data, ...result };
@@ -140,13 +152,26 @@ export function findById(rootDir, findingId) {
140
152
 
141
153
  // Search valid fixtures
142
154
  for (const filePath of discoverFixtures(rootDir, 'valid')) {
143
- const { data } = parseFinding(filePath);
155
+ const { data, error } = parseFinding(filePath);
156
+ if (error) {
157
+ torn.push({ path: filePath, error });
158
+ continue;
159
+ }
144
160
  if (data && data.finding_id === findingId) {
145
161
  const result = validateFinding(data);
146
162
  return { path: filePath, data, ...result };
147
163
  }
148
164
  }
149
165
 
166
+ if (torn.length > 0) {
167
+ // eslint-disable-next-line no-console
168
+ console.error(`${torn.length} finding(s) skipped (torn/unreadable):`);
169
+ for (const t of torn) {
170
+ // eslint-disable-next-line no-console
171
+ console.error(` ${relative(rootDir, t.path)} — ${t.error}`);
172
+ }
173
+ }
174
+
150
175
  return null;
151
176
  }
152
177
 
@@ -1,27 +1,41 @@
1
1
  /**
2
2
  * Append-only review event log.
3
3
  *
4
- * Events are stored as YAML arrays in reviews/<YYYY>/<date>-finding-review-log.yaml
4
+ * Events are stored ONE PER FILE under reviews/<YYYY>/<YYYY-MM-DD>/<event-id>.yaml
5
+ * (the same sharded-immutable-file pattern the records/ store uses). This
6
+ * replaced an earlier "daily YAML array, read-modify-rename under a file lock"
7
+ * design: a Phase-9 50-fork race detector proved that NO lock FILE is reliably
8
+ * exclusive on NTFS under heavy concurrent create churn — two appenders could
9
+ * both win the lock, both read N events, both rename, and one clobber the
10
+ * other's event (~1/3 of saturated runs, every child still exiting 0). A
11
+ * shared mutable array file cannot be made safe under concurrency on that fs.
12
+ * One immutable file per event removes the shared mutable state entirely: each
13
+ * append writes a uniquely-named file, so concurrent appends CANNOT collide or
14
+ * lose an event, and no lock is needed for correctness. `getAllEventsWithSkips`
15
+ * already globs reviews/ recursively and merges per-file events.
5
16
  */
6
17
 
7
- import { readFileSync, writeFileSync, mkdirSync, existsSync } from 'node:fs';
8
- import { resolve, dirname } from 'node:path';
18
+ import { mkdirSync, existsSync, openSync, writeSync, closeSync } from 'node:fs';
19
+ import { resolve } from 'node:path';
9
20
  import { randomBytes } from 'node:crypto';
10
21
  import yaml from 'js-yaml';
11
22
 
12
- import { withFileLock } from '../lib/file-lock.js';
13
- import { renameWithRetry } from '../lib/rename-with-retry.js';
14
23
  import { loadYamlDir } from '../lib/safe-yaml-load.js';
15
24
 
16
25
  let _eventCounter = 0;
17
26
 
18
27
  /**
19
- * Generate a unique event ID.
28
+ * Generate a unique event ID. Includes a random suffix so the id is unique
29
+ * ACROSS processes (the `_eventCounter` is per-process, so two forks would
30
+ * otherwise mint the same `rev-<ts>-<seq>`); this id also names the per-event
31
+ * file, so cross-process uniqueness keeps two concurrent appends from picking
32
+ * the same filename.
20
33
  */
21
34
  export function generateEventId() {
22
35
  const ts = Date.now().toString(36);
23
36
  const seq = (++_eventCounter).toString(36).padStart(4, '0');
24
- return `rev-${ts}-${seq}`;
37
+ const rand = randomBytes(4).toString('hex');
38
+ return `rev-${ts}-${seq}-${rand}`;
25
39
  }
26
40
 
27
41
  /**
@@ -71,82 +85,52 @@ export function getLogPath(rootDir, date = new Date()) {
71
85
  }
72
86
 
73
87
  /**
74
- * Append an event to the review log.
88
+ * Directory holding one date-sharded set of per-event files:
89
+ * reviews/<YYYY>/<YYYY-MM-DD>/. Each appended event is its own file inside.
90
+ */
91
+ export function getEventDir(rootDir, date = new Date()) {
92
+ const year = String(date.getFullYear());
93
+ const month = String(date.getMonth() + 1).padStart(2, '0');
94
+ const day = String(date.getDate()).padStart(2, '0');
95
+ return resolve(rootDir, 'reviews', year, `${year}-${month}-${day}`);
96
+ }
97
+
98
+ /**
99
+ * Append a review event by writing it to its OWN immutable file
100
+ * (reviews/<YYYY>/<YYYY-MM-DD>/<event-id>.yaml). See the module header for why
101
+ * this replaced the daily-array-under-a-lock design: no shared mutable file
102
+ * means concurrent appends physically cannot collide or lose an event, so no
103
+ * lock is needed for correctness — the Phase-9 50-fork race detector that lost
104
+ * an event ~1/3 of saturated runs is structurally impossible here.
75
105
  *
76
- * Atomicity: writes the new event list to a unique temp file then renames it
77
- * over the canonical log file. `rename` is atomic on POSIX and Windows, so a
78
- * concurrent reader sees either the old contents or the new contents — never
79
- * a half-written file. This also makes the operation crash-safe: a Ctrl+C
80
- * between read and write leaves the original log intact.
106
+ * Crash-safety: a single `openSync(path, 'wx')` + write of one event is atomic
107
+ * at the file granularity a crash mid-write leaves a complete event file or
108
+ * none, never a torn shared array. The filename is the (cross-process-unique)
109
+ * event id, so two concurrent appends never target the same path.
81
110
  *
82
- * Concurrency: serialized at the choke point via `withFileLock` on the daily
83
- * log file (F-PIPELINE-011 / W3-PIPE-001 Pattern #4 choke-point fix). The
84
- * read-then-write window is closed by holding a directory-mutex (`<logPath>.lock`)
85
- * across the read → push → rename sequence. Two concurrent `appendEvent` calls
86
- * to the SAME daily log serialize against each other; calls to DIFFERENT daily
87
- * logs (e.g. across a midnight boundary) do not contend. The lock is reclaimed
88
- * if the holder process dies — see `lib/file-lock.js` for the full design
89
- * rationale (why a lock dir, why not `O_APPEND`, stale recovery semantics,
90
- * single-machine scope).
111
+ * @param {string} rootDir
112
+ * @param {object} event - a `createEvent()` result.
113
+ * @returns {string} the path of the event file written.
91
114
  */
92
115
  export function appendEvent(rootDir, event) {
93
- const logPath = getLogPath(rootDir);
94
- const dir = dirname(logPath);
116
+ const dir = getEventDir(rootDir);
95
117
  mkdirSync(dir, { recursive: true });
96
118
 
97
- // FAILS-then-PASSES proof gate (W3-PIPE-001):
98
- // Set DISABLE_APPEND_LOCK=1 in the env to bypass the lock for the explicit
99
- // purpose of demonstrating the race-detection test fails without the fix.
100
- // Wave-30 receipt documents the proof: with the lock, the multi-process
101
- // test passes 50/50 forks across 3 iterations, 20 consecutive test runs.
102
- // With the lock disabled, the test reliably fails (rename collisions on
103
- // unprotected concurrent rebuilds, dropped events).
104
- if (process.env.DISABLE_APPEND_LOCK) {
105
- let events = [];
106
- if (existsSync(logPath)) {
107
- const raw = readFileSync(logPath, 'utf-8');
108
- events = yaml.load(raw) || [];
109
- if (!Array.isArray(events)) events = [events];
110
- }
111
- events.push(event);
112
- const tmpSuffix = randomBytes(4).toString('hex');
113
- const tmpPath = `${logPath}.${tmpSuffix}.tmp`;
114
- writeFileSync(tmpPath, yaml.dump(events, { lineWidth: 120, noRefs: true }), 'utf-8');
115
- // Windows EPERM/EBUSY on rename can fire transiently when AV or
116
- // Search Indexer holds a handle to the freshly written temp. Retry.
117
- renameWithRetry(tmpPath, logPath);
118
- return logPath;
119
+ const eventId = event.review_event_id || generateEventId();
120
+ const eventPath = resolve(dir, `${eventId}.yaml`);
121
+ const body = yaml.dump(event, { lineWidth: 120, noRefs: true });
122
+
123
+ // 'wx' = O_EXCL exclusive-create: asserts the unique id has not collided
124
+ // (it carries a random suffix, so it won't) rather than relying on it. No
125
+ // temp+rename is needed there is no shared file to atomically replace; one
126
+ // event per file is already all-or-nothing.
127
+ const fd = openSync(eventPath, 'wx');
128
+ try {
129
+ writeSync(fd, body);
130
+ } finally {
131
+ closeSync(fd);
119
132
  }
120
-
121
- return withFileLock(logPath, () => {
122
- let events = [];
123
- // Read-or-empty without an `existsSync` precheck: the readFileSync call
124
- // either returns the bytes or throws ENOENT. Avoiding `existsSync` here
125
- // closes a Windows-specific TOCTOU window where the dirent cache could
126
- // report `existsSync(logPath) === false` immediately after a sibling
127
- // process renamed a fresh file into place — which would cause us to
128
- // start with `events = []` and silently OVERWRITE the sibling's events.
129
- // The lock alone wasn't enough; the existsSync gate was the bug.
130
- try {
131
- const raw = readFileSync(logPath, 'utf-8');
132
- const parsed = yaml.load(raw);
133
- if (parsed) events = Array.isArray(parsed) ? parsed : [parsed];
134
- } catch (err) {
135
- if (!err || err.code !== 'ENOENT') throw err;
136
- }
137
-
138
- events.push(event);
139
-
140
- // Atomic write: temp file → rename. Same pattern persist.js + rebuild-indexes.js use.
141
- const tmpSuffix = randomBytes(4).toString('hex');
142
- const tmpPath = `${logPath}.${tmpSuffix}.tmp`;
143
- writeFileSync(tmpPath, yaml.dump(events, { lineWidth: 120, noRefs: true }), 'utf-8');
144
- // renameWithRetry: tolerate the Windows EPERM/EBUSY transient handle race
145
- // even though we hold the per-file lock — antivirus/Search Indexer can
146
- // still grab a handle on the temp during the rename window.
147
- renameWithRetry(tmpPath, logPath);
148
- return logPath;
149
- });
133
+ return eventPath;
150
134
  }
151
135
 
152
136
  /**
@@ -32,10 +32,12 @@
32
32
  * partial coverage, not a silent no-op.
33
33
  */
34
34
 
35
+ import { relative } from 'node:path';
36
+
35
37
  import { validateTransition, ACTION_TARGET_STATUS, REASON_REQUIRED, REQUIRES_ACCEPTED, REQUIRES_CLOSED } from './transitions.js';
36
38
  import { createEvent, appendEvent } from './event-log.js';
37
39
  import {
38
- resetSeenArtifactWrites,
40
+ resetSeenArtifactWrite,
39
41
  writePattern,
40
42
  writeRecommendation,
41
43
  writeDoctrine,
@@ -44,6 +46,34 @@ import {
44
46
  loadDoctrinesWithSkips
45
47
  } from '../synthesis/write-artifacts.js';
46
48
 
49
+ /**
50
+ * Surface torn/unreadable artifact YAML on stderr — the lookup-path analogue of
51
+ * the derive CLI's "N pattern(s) skipped (torn/unreadable)" reporting
52
+ * (cli.js) and `findRecordFile`'s stderr note (derive/load-records.js).
53
+ *
54
+ * B-001 — `findArtifactById` / `getArtifactReviewQueue` consumed only the
55
+ * loader's `entries`, discarding `skipped[]`. A torn artifact YAML for the very
56
+ * id under accept/show/queue therefore vanished behind a bare "<type> not
57
+ * found", giving the operator no signal that a file was unreadable. The id
58
+ * lives INSIDE the file, so a torn file can't be matched to a requested id;
59
+ * the honest signal is to name the torn files whenever any are present so the
60
+ * operator can distinguish "absent" from "present but unparseable". Returns the
61
+ * skipped list so callers can decide whether the not-found was hint-worthy.
62
+ *
63
+ * @param {string} rootDir
64
+ * @param {string} type - artifact kind, for the message prefix
65
+ * @param {Array<{ path: string, error: string }>} skipped
66
+ */
67
+ function reportArtifactSkips(rootDir, type, skipped) {
68
+ if (!skipped || skipped.length === 0) return;
69
+ // eslint-disable-next-line no-console
70
+ console.error(`${skipped.length} ${type}(s) skipped (torn/unreadable):`);
71
+ for (const s of skipped) {
72
+ // eslint-disable-next-line no-console
73
+ console.error(` ${relative(rootDir, s.path)} — ${s.error}`);
74
+ }
75
+ }
76
+
47
77
  /**
48
78
  * Per-artifact-type configuration: the on-disk directory, the id field name,
49
79
  * the loader (with-skips) and the writer. Mirrors how the finding engine pairs
@@ -87,12 +117,15 @@ const ARTIFACT_TYPES = {
87
117
  export function findArtifactById(rootDir, type, id) {
88
118
  const cfg = ARTIFACT_TYPES[type];
89
119
  if (!cfg) return null;
90
- const { entries } = cfg.loadWithSkips(rootDir);
120
+ const { entries, skipped } = cfg.loadWithSkips(rootDir);
91
121
  for (const entry of entries) {
92
122
  if (entry.data && entry.data[cfg.idKey] === id) {
93
123
  return { data: entry.data, path: entry.path, type };
94
124
  }
95
125
  }
126
+ // B-001 — id absent from the clean entries. A torn file may be hiding it;
127
+ // name the torn files rather than letting a bare not-found mislead.
128
+ reportArtifactSkips(rootDir, type, skipped);
96
129
  return null;
97
130
  }
98
131
 
@@ -217,11 +250,13 @@ export function reviewArtifact(rootDir, params) {
217
250
  });
218
251
 
219
252
  // Persist via the synthesis writer — which RE-VALIDATES the promoted artifact
220
- // against its JSON Schema (fail-closed). `resetSeenArtifactWrites` is the
221
- // documented opt-in for a legitimate re-write of an id already touched in this
222
- // process (the synthesis collision guard otherwise refuses the second write).
253
+ // against its JSON Schema (fail-closed). The reset is the documented opt-in
254
+ // for a legitimate re-write of an id already touched in this process (the
255
+ // synthesis collision guard otherwise refuses the second write). B-003 —
256
+ // scope it to THIS id so a future batch reviewer doesn't disarm the guard for
257
+ // the other ids it has written this process.
223
258
  try {
224
- resetSeenArtifactWrites(rootDir);
259
+ resetSeenArtifactWrite(rootDir, type, id);
225
260
  cfg.write(rootDir, artifact);
226
261
  } catch (err) {
227
262
  return { success: false, error: err.message, code: err.code };
@@ -249,7 +284,10 @@ export function getArtifactReviewQueue(rootDir, type) {
249
284
  for (const t of types) {
250
285
  const cfg = ARTIFACT_TYPES[t];
251
286
  if (!cfg) continue;
252
- const { entries } = cfg.loadWithSkips(rootDir);
287
+ const { entries, skipped } = cfg.loadWithSkips(rootDir);
288
+ // B-001 — a torn artifact silently shrinks the review queue; name it so the
289
+ // operator knows the queue is partial rather than complete.
290
+ reportArtifactSkips(rootDir, t, skipped);
253
291
  for (const entry of entries) {
254
292
  const data = entry.data;
255
293
  if (!data) continue;
@@ -34,6 +34,7 @@ import { resolve } from 'node:path';
34
34
  import { existsSync, readFileSync } from 'node:fs';
35
35
  import yaml from 'js-yaml';
36
36
 
37
+ import { isUnsafeSegment } from '@dogfood-lab/ingest/lib/unsafe-segment.js';
37
38
  import { atomicWriteFileSync } from '../lib/atomic-write.js';
38
39
  import { findArtifactById } from '../review/review-artifacts.js';
39
40
  import { createEvent, appendEvent } from '../review/event-log.js';
@@ -89,6 +90,22 @@ export function applyRecommendation(rootDir, params) {
89
90
  const action = rec.action || {};
90
91
  const surfaces = rec.applies_to?.product_surfaces || [];
91
92
 
93
+ // findings-A-001 — path-traversal guard on the operator-supplied --policy
94
+ // <org/repo>. `policyPathFor` resolves `policies/repos/<org>/<repo>.yaml`; an
95
+ // org/repo carrying `..` or a separator would escape the policies tree on
96
+ // BOTH the dry-run (path leaked in preview) and write (file touched) paths,
97
+ // so reject here before either branch resolves a path.
98
+ if (params.policyRepo) {
99
+ const [pOrg, pRepo] = String(params.policyRepo).split('/');
100
+ if (!pOrg || !pRepo || isUnsafeSegment(pOrg) || isUnsafeSegment(pRepo)) {
101
+ return structuredError(
102
+ 'RECOMMENDATION_UNSAFE_POLICY',
103
+ `policy repo "${params.policyRepo}" is not a safe org/repo path segment`,
104
+ 'Pass --policy <org/repo> with no ".." or path separators inside the org or repo name.'
105
+ );
106
+ }
107
+ }
108
+
92
109
  // Build the resolution context shared by dry-run and write.
93
110
  const isStructured = STRUCTURED_LIST_ACTIONS.has(action.type);
94
111
  const surface = surfaces.length === 1 ? surfaces[0] : null;
@@ -8,6 +8,7 @@
8
8
  */
9
9
 
10
10
  import { loadFindings } from '../reader.js';
11
+ import { boundaryHash } from '../derive/ids.js';
11
12
 
12
13
  /**
13
14
  * Derive candidate patterns from accepted findings.
@@ -120,7 +121,7 @@ function isFalseRecurrence(findings) {
120
121
  /**
121
122
  * Build a pattern candidate from a cluster.
122
123
  */
123
- function buildPatternCandidate(cluster) {
124
+ export function buildPatternCandidate(cluster) {
124
125
  const { findings, issue_kind, root_cause_kind, remediation_kind } = cluster;
125
126
  const now = new Date().toISOString();
126
127
 
@@ -141,8 +142,16 @@ function buildPatternCandidate(cluster) {
141
142
  // otherwise two clusters that differ only by root_cause_kind collide on pattern_id and the
142
143
  // second writePattern() silently overwrites the first on disk. Surface is added for readability,
143
144
  // not for uniqueness.
145
+ //
146
+ // The human-readable slug runs `.replace(/_/g, '-')` AFTER concatenation, which folds the
147
+ // underscore inside a component into the same `-` that delimits components: issue_kind='a_b' +
148
+ // root_cause='c' and issue_kind='a' + root_cause='b_c' both flatten to `...-a-b-c`
149
+ // (findings-A-002). A trailing boundary hash over the UN-flattened cluster key keeps the two
150
+ // distinct, so legitimately different clusters no longer trip the L3-001 *_ID_COLLISION guard.
144
151
  const surfaceStr = surfaces.size === 1 ? [...surfaces][0] : 'multi-surface';
145
- const slug = `${surfaceStr}-${issue_kind}-${root_cause_kind}`.replace(/_/g, '-');
152
+ const readable = `${surfaceStr}-${issue_kind}-${root_cause_kind}`.replace(/_/g, '-');
153
+ const boundary = boundaryHash([surfaceStr, issue_kind, root_cause_kind]);
154
+ const slug = `${readable}-${boundary}`;
146
155
 
147
156
  // Determine strength
148
157
  const strength = repos.size >= 3 ? 'strong' : repos.size >= 2 ? 'emerging' : 'emerging';
@@ -59,6 +59,24 @@ export function resetSeenArtifactWrites(rootDir) {
59
59
  }
60
60
  }
61
61
 
62
+ /**
63
+ * Reset the collision memory for a SINGLE artifact id only.
64
+ *
65
+ * B-003 — `reviewArtifact` re-writes the one id it is promoting, which
66
+ * legitimately needs that id's guard cleared. Clearing the whole root (via
67
+ * `resetSeenArtifactWrites(rootDir)`) also disarms the silent-clobber guard for
68
+ * every OTHER kind/id — latent for a future batch caller reviewing several ids
69
+ * in one process. Scope the reset to `${rootDir}:${kind}:${id}` so only the id
70
+ * under review is cleared.
71
+ *
72
+ * @param {string} rootDir
73
+ * @param {'pattern'|'recommendation'|'doctrine'} kind
74
+ * @param {string} id
75
+ */
76
+ export function resetSeenArtifactWrite(rootDir, kind, id) {
77
+ seenArtifactWrites.delete(`${rootDir}:${kind}:${id}`);
78
+ }
79
+
62
80
  /**
63
81
  * Structured collision errors thrown by the singleton writers when called
64
82
  * twice in the same process with the same id. Mirror the batch helpers'