@feigi/fleet-ctl 3.24.2 → 3.24.4
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/package.json +1 -1
- package/scripts/board.html +22 -22
- package/scripts/board.mjs +143 -145
- package/scripts/compute-board.mjs +39 -38
- package/scripts/derive-testcmd.sh +1 -1
- package/scripts/drop-merged-label.sh +10 -11
- package/scripts/fleet-bootstrap +8 -8
- package/scripts/fleet-provenance +8 -8
- package/scripts/fleet-run +3 -3
- package/scripts/frontmatter-allowlist.json +6 -6
- package/scripts/instruments.sh +30 -31
- package/scripts/json.sh +12 -13
- package/scripts/net.sh +23 -21
- package/scripts/prove-merge.sh +7 -7
- package/scripts/review-core.mjs +57 -58
- package/scripts/review-eval.mjs +14 -15
- package/scripts/verify-sha.sh +4 -4
package/scripts/review-core.mjs
CHANGED
|
@@ -1,5 +1,4 @@
|
|
|
1
|
-
// review-core.mjs — the review body for PR review
|
|
2
|
-
// on #1296). Read that ruling before touching this file.
|
|
1
|
+
// review-core.mjs — the review body for PR review.
|
|
3
2
|
//
|
|
4
3
|
// This file holds every host-independent declaration PR review needs: the
|
|
5
4
|
// JSON schemas, `usableDiff`, `readRules`, `resolveTestCmd`,
|
|
@@ -22,15 +21,15 @@
|
|
|
22
21
|
// `null`, which every guard here — `if (snap) {...}`, `unrunCrashed`,
|
|
23
22
|
// `verdictFor` — is written against).
|
|
24
23
|
//
|
|
25
|
-
// Two imports. arg.mjs's isDigits()
|
|
24
|
+
// Two imports. arg.mjs's isDigits(): this file is an ordinary
|
|
26
25
|
// module, so it consumes the repo's digits rule directly rather than
|
|
27
|
-
// declaring its own copy. node:path's `posix`,
|
|
26
|
+
// declaring its own copy. node:path's `posix`, to normalise a
|
|
28
27
|
// scratch path before `runnerScratchRefusal` reads its last component.
|
|
29
28
|
//
|
|
30
|
-
// `.mjs`, not `.js
|
|
29
|
+
// `.mjs`, not `.js`: this file was once `review-core.js`, and
|
|
31
30
|
// nothing that ships declares a `type`, so Node below 20.19.0/22.7.0 —
|
|
32
31
|
// inside the declared consumer floor — read it as CommonJS and
|
|
33
|
-
// review-eval.mjs's import of it failed.
|
|
32
|
+
// review-eval.mjs's import of it failed. A suite sweep keeps a
|
|
34
33
|
// shipped `.mjs` from importing a relative module under any other extension.
|
|
35
34
|
|
|
36
35
|
import { posix } from "node:path";
|
|
@@ -91,7 +90,7 @@ export const SNAPSHOT_SCHEMA = {
|
|
|
91
90
|
path: { type: "string" },
|
|
92
91
|
head: { type: "string" },
|
|
93
92
|
pathVerified: { type: "boolean" },
|
|
94
|
-
//
|
|
93
|
+
// `pathVerified` says the tree is THERE; this says the tree is
|
|
95
94
|
// MEASURABLE — that the snapshot is a git repository holding the reviewed
|
|
96
95
|
// commit's tree. Required for `pathVerified`'s reason: an omitted boolean
|
|
97
96
|
// must not read as a verified environment, because the whole defect was a
|
|
@@ -111,7 +110,7 @@ export const SNAPSHOT_SCHEMA = {
|
|
|
111
110
|
},
|
|
112
111
|
};
|
|
113
112
|
|
|
114
|
-
//
|
|
113
|
+
// What the review's one shared test run reports back. `command` and the
|
|
115
114
|
// log path are NOT asked for: the caller handed both out, so it already knows
|
|
116
115
|
// them, and an agent's echo of either could only disagree. Every count is
|
|
117
116
|
// optional and stays absent when the log does not state it — an absent
|
|
@@ -133,7 +132,7 @@ export const TEST_RUN_SCHEMA = {
|
|
|
133
132
|
};
|
|
134
133
|
|
|
135
134
|
// --- Dimension catalog --------------------------------------------------
|
|
136
|
-
//
|
|
135
|
+
// No dimension carries a `model`/`effort` field —
|
|
137
136
|
// dispatch tier lives ONLY in each fleet-owned agent definition's own
|
|
138
137
|
// frontmatter. `agentType` is a bare frontmatter `name:` (omp's native
|
|
139
138
|
// `agent()` lookup is an exact match on one).
|
|
@@ -174,14 +173,14 @@ export const DEFAULT_DIMENSIONS = [
|
|
|
174
173
|
},
|
|
175
174
|
];
|
|
176
175
|
|
|
177
|
-
//
|
|
176
|
+
// The omp `fleet-review-runner` agent definition's `spawns:`
|
|
178
177
|
// frontmatter must allow exactly the agent types dispatched below (the
|
|
179
|
-
// snapshot agent, the shared test run's agent
|
|
178
|
+
// snapshot agent, the shared test run's agent, every
|
|
180
179
|
// DEFAULT_DIMENSIONS entry, and the verifier) — omp denies every spawn by
|
|
181
180
|
// default, and the runner declared no `spawns` at all until this fix.
|
|
182
181
|
// Exporting the three literal-only types (snapshot/test-run/verifier have no
|
|
183
182
|
// DEFAULT_DIMENSIONS entry of their own) plus the derived full set lets
|
|
184
|
-
//
|
|
183
|
+
// a test pin the frontmatter against what this file
|
|
185
184
|
// actually dispatches, instead of a hand-copied list that can drift.
|
|
186
185
|
export const SNAPSHOT_AGENT_TYPE = "fleet-review-snapshot";
|
|
187
186
|
export const TEST_RUN_AGENT_TYPE = "fleet-review-test-run";
|
|
@@ -197,15 +196,17 @@ export const SPAWNED_AGENT_TYPES = [
|
|
|
197
196
|
// computeStats owns both thresholds).
|
|
198
197
|
export const SIZE_TIER_PROFILES = new Set(["single-file", "small"]);
|
|
199
198
|
// The refuter-budget floor: `correctness`/`silent-failure` miss silently and
|
|
200
|
-
// permanently; `comments` sits beside them
|
|
199
|
+
// permanently; `comments` sits beside them because a comment-only edit to a
|
|
200
|
+
// code file classifies as `src`, never `docs`, and would otherwise get no
|
|
201
|
+
// comment review at all.
|
|
201
202
|
// Keyed on recoverability of a MISS, never on tier — `verifiersFor` takes a
|
|
202
203
|
// severity and nothing else.
|
|
203
204
|
export const SIZE_TIER_DIMS = new Set(["correctness", "silent-failure", "comments"]);
|
|
204
205
|
|
|
205
206
|
// The refuter budget per severity, derived once per run from the caller's
|
|
206
207
|
// own `A.verifiers`/`A.verifiersBySeverity` overrides (or the defaults) and
|
|
207
|
-
// bound to a `(severity) => count` closure —
|
|
208
|
-
//
|
|
208
|
+
// bound to a `(severity) => count` closure — the suite imports this directly
|
|
209
|
+
// rather than re-deriving it from a lifted copy.
|
|
209
210
|
export function verifiersFor(A) {
|
|
210
211
|
const verifiers = A.verifiers || 2;
|
|
211
212
|
const by = A.verifiersBySeverity || { critical: verifiers, important: verifiers, suggestion: 0 };
|
|
@@ -214,8 +215,8 @@ export function verifiersFor(A) {
|
|
|
214
215
|
|
|
215
216
|
// --- Pure functions -------------------------------------------------------
|
|
216
217
|
// Every function below was ported byte-identical (function body) from this
|
|
217
|
-
// repo's
|
|
218
|
-
//
|
|
218
|
+
// repo's pre-cutover review workflow script's own declaration, before that
|
|
219
|
+
// file was retired — see this repo's git history for the full
|
|
219
220
|
// historical rationale on each; it is not repeated here to avoid a second
|
|
220
221
|
// copy of PROSE disconnecting the way this repo's own "recurring pin defect"
|
|
221
222
|
// comment warns a second copy of CODE does.
|
|
@@ -270,12 +271,12 @@ reads carried 434 KB.`;
|
|
|
270
271
|
|
|
271
272
|
// The step a missing Test entrypoint is derived by, named in the refusal so the
|
|
272
273
|
// caller — a controller or a reviewer — knows what to run rather than what went
|
|
273
|
-
// wrong
|
|
274
|
+
// wrong. derive-testcmd.sh's refusals about the cache name the same
|
|
274
275
|
// step; its exit-3 refusals name an environment fault instead, which the caller
|
|
275
276
|
// fixes rather than deriving. This one covers the case where the snapshot agent
|
|
276
277
|
// reported no reason at all.
|
|
277
278
|
const DERIVATION_STEP =
|
|
278
|
-
"run the Recipe derivation step (run-team phase 0, before the first claim
|
|
279
|
+
"run the Recipe derivation step (run-team phase 0, before the first claim) to derive, prove and write the Recipe cache";
|
|
279
280
|
|
|
280
281
|
export function resolveTestCmd(explicit, snap) {
|
|
281
282
|
if (explicit) return explicit;
|
|
@@ -323,12 +324,12 @@ export function selectDimensions(all, stats) {
|
|
|
323
324
|
// documents `args.dimensions` as accepting "keys or dimension objects" — but
|
|
324
325
|
// until now only objects worked: a key array passed straight through and
|
|
325
326
|
// every dereference below (`d.key`, `d.prompt`, `d.agentType` — three) came
|
|
326
|
-
// back `undefined`, with no throw and no warning
|
|
327
|
+
// back `undefined`, with no throw and no warning. Resolve strings
|
|
327
328
|
// against the workflow's own catalog, and check every object has the three
|
|
328
329
|
// fields it REQUIRES AND that each of those is a string — presence alone let
|
|
329
330
|
// a non-string field reach the specialist dispatch machinery downstream
|
|
330
331
|
// instead of failing at this validated boundary. `model` is no longer a
|
|
331
|
-
// fourth optional field
|
|
332
|
+
// fourth optional field: an override entry that still sends one is
|
|
332
333
|
// refused outright, loudly, rather than silently accepted and ignored — see
|
|
333
334
|
// the check below.
|
|
334
335
|
// Anything unresolvable stops the run and names what was not recognised — a
|
|
@@ -394,7 +395,7 @@ export function snapshotMissing(snap, runRootPrefix) {
|
|
|
394
395
|
return null;
|
|
395
396
|
}
|
|
396
397
|
|
|
397
|
-
//
|
|
398
|
+
// What a suite run inside the snapshot is evidence ABOUT, in one
|
|
398
399
|
// paragraph every specialist prompt and the returned payload both carry.
|
|
399
400
|
//
|
|
400
401
|
// `repoVerified` false is deliberately NOT a refusal (`snapshotMissing` says
|
|
@@ -434,7 +435,7 @@ validation of the tree: its counts are snapshot-measured, and a failure in it
|
|
|
434
435
|
cannot be told apart from a regression.`;
|
|
435
436
|
}
|
|
436
437
|
|
|
437
|
-
//
|
|
438
|
+
// Every run-quality verdict below reads the review's ONE shared test
|
|
438
439
|
// run — `{command, logPath, exitCode?, tests?, pass?, fail?, cancelled?,
|
|
439
440
|
// skipped?, error?}`, `command`/`logPath` the caller's own and the rest the
|
|
440
441
|
// test-run agent's report — never a specialist's own run: no specialist runs
|
|
@@ -474,7 +475,7 @@ function failingDesc(run) {
|
|
|
474
475
|
return run.cancelled ? `${run.fail || 0} failing and ${run.cancelled} cancelled tests` : `${run.fail} failing tests`;
|
|
475
476
|
}
|
|
476
477
|
|
|
477
|
-
// Failing tests are owned by the review as a whole
|
|
478
|
+
// Failing tests are owned by the review as a whole: `findings` is
|
|
478
479
|
// every finding that reached the payload, and ONE from any dimension
|
|
479
480
|
// satisfies the check for all — no dimension is asked to duplicate a
|
|
480
481
|
// sibling's, and none is marked unrun because a sibling filed it. A refuted
|
|
@@ -509,7 +510,7 @@ function countsOf(run) {
|
|
|
509
510
|
.join(", ");
|
|
510
511
|
}
|
|
511
512
|
|
|
512
|
-
//
|
|
513
|
+
// The Tests paragraph of every specialist prompt: the shared run's
|
|
513
514
|
// command, counts, exit status and log path, in place of the instruction to
|
|
514
515
|
// run the full suite that each specialist used to follow — one review was up
|
|
515
516
|
// to six full sweeps of one immutable snapshot. `owner` is the one dimension
|
|
@@ -547,7 +548,7 @@ bears on your own lens.`,
|
|
|
547
548
|
return lines.join("\n");
|
|
548
549
|
}
|
|
549
550
|
|
|
550
|
-
//
|
|
551
|
+
// The CWD-AUDIT line the Review dispatch below asks a specialist
|
|
551
552
|
// to fold into `scope_searched` is a convention, not schema — FINDINGS_SCHEMA
|
|
552
553
|
// accepts any string there, so a specialist that satisfies the schema while
|
|
553
554
|
// never emitting the line, or misspelling it, produces a fully valid,
|
|
@@ -574,7 +575,7 @@ export function verdictFor(dispatched, votes) {
|
|
|
574
575
|
return { verdict, votes: live, refutersDispatched: dispatched };
|
|
575
576
|
}
|
|
576
577
|
|
|
577
|
-
//
|
|
578
|
+
// One in-run re-dispatch for a crashed dispatch, before the result is
|
|
578
579
|
// assembled — the specialist call and the refuter PAIR are the two units it
|
|
579
580
|
// wraps. `crashed(value)` says whether a settled first attempt counts as a
|
|
580
581
|
// crash; a THROWN first attempt is handed to it as `null`, because a
|
|
@@ -589,21 +590,21 @@ export function retryCrashed(dispatch, crashed) {
|
|
|
589
590
|
.then(again, () => again(null));
|
|
590
591
|
}
|
|
591
592
|
|
|
592
|
-
// omp has no cached-replay mechanism
|
|
593
|
+
// omp has no cached-replay mechanism: a fresh
|
|
593
594
|
// review re-dispatches every agent() live rather than only the crashed
|
|
594
595
|
// legs, and the one re-dispatch this run gets was already spent in-run
|
|
595
|
-
//
|
|
596
|
+
// — a crashed finding is reported and deferred, never resumed.
|
|
596
597
|
export function resumeFor(unverified) {
|
|
597
598
|
const crashed = unverified.filter((f) => f.refutersDispatched > 0);
|
|
598
599
|
if (!crashed.length) return { crashed, resume: null };
|
|
599
600
|
const claim =
|
|
600
601
|
"Findings in `unverified` with `refutersDispatched` above zero and no surviving vote had every refuter die, and die again on the in-run retry — nothing looked at them. ";
|
|
601
602
|
const verb =
|
|
602
|
-
"Defer them as crashed — reported, not acted on: omp's eval has no cached-replay mechanism
|
|
603
|
+
"Defer them as crashed — reported, not acted on: omp's eval has no cached-replay mechanism, so a fresh review re-dispatches every agent() live rather than only the crashed legs, and the in-run retry was this review's one re-dispatch. Re-run nothing for them.";
|
|
603
604
|
return { crashed, resume: claim + verb };
|
|
604
605
|
}
|
|
605
606
|
|
|
606
|
-
//
|
|
607
|
+
// The digest: every field a controller acts on, and nothing bulky.
|
|
607
608
|
// The result object LEADS with exactly these keys, in this order
|
|
608
609
|
// (`runReview`'s return below) — historically so the pre-cutover harness's
|
|
609
610
|
// ~8 KB inline `<result>` cut landed in the finding arrays, never in the
|
|
@@ -621,18 +622,17 @@ export function digestOf(result) {
|
|
|
621
622
|
return Object.fromEntries(DIGEST_KEYS.map((k) => [k, result[k]]));
|
|
622
623
|
}
|
|
623
624
|
|
|
624
|
-
//
|
|
625
|
+
// The digits refusal review-eval.mjs's `runReviewToFile` runs before
|
|
625
626
|
// its first attempt, so a dispatch mistake is refused once instead of being
|
|
626
627
|
// retried and sent to the fallback reviewer. It lives here, not there, because
|
|
627
628
|
// review-eval.mjs is a library and arg.mjs's `scripts/*.mjs` importers are its
|
|
628
|
-
// CLI roster, each one probed to refuse a stray flag
|
|
629
|
-
//
|
|
630
|
-
// rule on the review path. Null when `pr` is a PR number.
|
|
629
|
+
// CLI roster, each one probed to refuse a stray flag; this module already
|
|
630
|
+
// consumes the digits rule on the review path. Null when `pr` is a PR number.
|
|
631
631
|
export function runnerPrRefusal(pr) {
|
|
632
632
|
return isDigits(pr) ? null : `args.pr must be a PR number, got ${JSON.stringify(pr)}`;
|
|
633
633
|
}
|
|
634
634
|
|
|
635
|
-
//
|
|
635
|
+
// `runReview` owns the review side's `pr<N>/` partition — it appends
|
|
636
636
|
// `/pr${pr}` to the scratch it is given — so a scratch whose last component
|
|
637
637
|
// is already `pr` plus digits is a caller that partitioned too, and its run
|
|
638
638
|
// roots would nest at `<root>/pr<N>/pr<N>/run-*`, below the snapshot prompt's
|
|
@@ -662,7 +662,7 @@ export function runnerScratchRefusal(scratch) {
|
|
|
662
662
|
// contract every guard below is written against), `phase(title)`,
|
|
663
663
|
// `log(message)`, and OPTIONALLY `pipeline`/`parallel` (see
|
|
664
664
|
// defaultPipeline/defaultParallel below for the omp shim's implementation
|
|
665
|
-
// of both
|
|
665
|
+
// of both).
|
|
666
666
|
function defaultParallel(fns) {
|
|
667
667
|
return Promise.all(fns.map((fn) => fn()));
|
|
668
668
|
}
|
|
@@ -670,7 +670,7 @@ function defaultParallel(fns) {
|
|
|
670
670
|
// Per-item independence, INCLUDING the null short-circuit: a stage-1 throw
|
|
671
671
|
// or a stage-1 falsy result must land in the SAME null slot a stage-2 throw
|
|
672
672
|
// would, so `unrunCrashed`'s index-aligned read of the pipeline result sees
|
|
673
|
-
// one uniform shape for every crash cause
|
|
673
|
+
// one uniform shape for every crash cause.
|
|
674
674
|
async function defaultPipeline(items, stage1, stage2) {
|
|
675
675
|
return Promise.all(
|
|
676
676
|
items.map(async (item) => {
|
|
@@ -710,7 +710,7 @@ export async function runReview(host, args) {
|
|
|
710
710
|
|
|
711
711
|
if (!pr || !worktree) throw new Error("review-pr: args.pr and args.worktree are required");
|
|
712
712
|
|
|
713
|
-
//
|
|
713
|
+
// The truthiness check above is what made the CLI fail-open REACHABLE:
|
|
714
714
|
// `pr` is interpolated into `gh pr diff ${pr}`, `gh pr view ${pr}` and
|
|
715
715
|
// `diff-stats.mjs --pr ${pr}` in the snapshot prompt below, and a branch name
|
|
716
716
|
// passes all three — `gh` resolves a non-numeric ref as a BRANCH, so the
|
|
@@ -771,7 +771,7 @@ both hashes — in \`repoError\`. The init is what makes the snapshot MEASURABLE
|
|
|
771
771
|
the review runs this repository's own suite in there, and a suite with tests
|
|
772
772
|
that need a working tree reports fewer passes and more failures in a bare
|
|
773
773
|
extraction than in a checkout at the same commit, with nothing in the payload
|
|
774
|
-
saying the measurement happened somewhere else
|
|
774
|
+
saying the measurement happened somewhere else. The tree-hash compare is
|
|
775
775
|
the verification: two commits whose trees hash the same hold byte-identical
|
|
776
776
|
content, so a match settles that the snapshot IS the reviewed tree — a stronger
|
|
777
777
|
check than reading a few files, and the reason the commit comes BEFORE the
|
|
@@ -785,7 +785,7 @@ every file in the tree rather than a couple of them. Do not modify ${worktree}.
|
|
|
785
785
|
Every path in the block above is absolute or \`-C\`-anchored on purpose: this
|
|
786
786
|
dispatch carries no working directory of its own either, so you start in the
|
|
787
787
|
controller's own checkout, and a relative path — a \`tar -x\` with no \`-C\`, a
|
|
788
|
-
bare \`git\` — reads or writes THERE
|
|
788
|
+
bare \`git\` — reads or writes THERE. Add nothing relative to it, and
|
|
789
789
|
chain a \`cd\` into "$RUN" or "$SNAP" for anything you run beyond it.
|
|
790
790
|
|
|
791
791
|
Then capture the PR's diff for the specialists, plus the three facts the caller
|
|
@@ -811,7 +811,7 @@ as before. A cross-repo PR (a fork) skips the branch-ref read entirely and goes
|
|
|
811
811
|
straight to 'refs/pull/${pr}/head' — the base repo's own copy of the PR head —
|
|
812
812
|
because a fork's branch NAME is not guaranteed unique against the base
|
|
813
813
|
repository, and a same-named hit on 'origin' would silently answer for the
|
|
814
|
-
wrong repository rather than for the fork
|
|
814
|
+
wrong repository rather than for the fork. The fallback to
|
|
815
815
|
'refs/pull/${pr}/head' also still fires for a same-repo PR whose branch ref
|
|
816
816
|
came back empty, unchanged from before. Omit \`refHead\` when every read this
|
|
817
817
|
PR was entitled to came back empty or was skipped: an absent value is neither
|
|
@@ -886,7 +886,7 @@ a false repoVerified.`,
|
|
|
886
886
|
);
|
|
887
887
|
log(`agents dispatched ${dimensions.map((d) => `${d.key}=${d.agentType}`).join(" ")}`);
|
|
888
888
|
|
|
889
|
-
//
|
|
889
|
+
// The review's ONE run of the test command, before any specialist is
|
|
890
890
|
// dispatched: every dimension used to run the full suite itself, so one
|
|
891
891
|
// review was up to six full sweeps of one immutable snapshot. NOT wrapped in
|
|
892
892
|
// `retryCrashed`: a crashed dispatch may already have launched the command,
|
|
@@ -946,7 +946,7 @@ learns the run produced none.`,
|
|
|
946
946
|
const failureOwner = dimensions[0].key;
|
|
947
947
|
|
|
948
948
|
const dimensionsUnrun = [];
|
|
949
|
-
//
|
|
949
|
+
// Per-dimension record of the specialist's own CWD-AUDIT line (see
|
|
950
950
|
// `cwdAuditFrom` above) — `{dimension, state, line}`, `state` one of
|
|
951
951
|
// "clean"/"dirty"/"unrepo"/"missing". Populated for every dispatched
|
|
952
952
|
// review that returned at all (a crashed dispatch has nothing to audit,
|
|
@@ -955,10 +955,10 @@ learns the run produced none.`,
|
|
|
955
955
|
// dead-ending inside `scope_searched`.
|
|
956
956
|
const cwdAudit = [];
|
|
957
957
|
|
|
958
|
-
//
|
|
958
|
+
// Both prompts below carry the inherited-cwd rule, and it is stated in
|
|
959
959
|
// each rather than shared: review-eval.mjs's own header holds the measurement
|
|
960
960
|
// and the reason this is prompt prose at all (no dispatch primitive takes a
|
|
961
|
-
// per-call cwd), and
|
|
961
|
+
// per-call cwd), and the shared-source route is ruled out for
|
|
962
962
|
// exactly these blocks. Three parts, in this order, because the last two are
|
|
963
963
|
// inert without the first: the cwd the specialist starts in is NAMED as a
|
|
964
964
|
// tree it must not write to, `pwd` fixes which directory that is, and the
|
|
@@ -968,7 +968,7 @@ learns the run produced none.`,
|
|
|
968
968
|
phase("Review");
|
|
969
969
|
const reviewed = await pipeline(
|
|
970
970
|
dimensions,
|
|
971
|
-
//
|
|
971
|
+
// A crashed specialist (null or thrown) is re-dispatched once
|
|
972
972
|
// before it counts as unrun. A specialist that RETURNED is never re-run
|
|
973
973
|
// here — the shared run, not the specialist, is what `unrunEntries` judges.
|
|
974
974
|
(d) =>
|
|
@@ -1022,7 +1022,7 @@ beginning \`CWD-AUDIT:\` — \`CWD-AUDIT: clean <path>\` when it printed nothing
|
|
|
1022
1022
|
every run, clean or not: a clean tree is the result this check exists to
|
|
1023
1023
|
produce, and an omitted line reads exactly like a check never run. Three PRs
|
|
1024
1024
|
reviewed from one cell left four files modified in that checkout with nothing in
|
|
1025
|
-
any payload saying so
|
|
1025
|
+
any payload saying so, so a path you cannot account for is still yours
|
|
1026
1026
|
to name.`,
|
|
1027
1027
|
{ label: `review:${d.key}${forPr}`, phase: "Review", agentType: d.agentType, schema: FINDINGS_SCHEMA },
|
|
1028
1028
|
),
|
|
@@ -1036,7 +1036,7 @@ to name.`,
|
|
|
1036
1036
|
(review && review.findings ? review.findings : []).map((f, fi) => () => {
|
|
1037
1037
|
const n = verifiersForRun(f.severity);
|
|
1038
1038
|
if (n === 0) return Promise.resolve({ ...f, dimension: d.key, ...verdictFor(0, []) });
|
|
1039
|
-
//
|
|
1039
|
+
// A pair whose EVERY vote died (or whose dispatch threw) is
|
|
1040
1040
|
// re-dispatched once, as a pair, before `verdictFor` reads it — one
|
|
1041
1041
|
// live vote means the pair did not crash and is ruled on that vote.
|
|
1042
1042
|
return retryCrashed(
|
|
@@ -1113,8 +1113,7 @@ Report it in \`reason\` as one line beginning \`CWD-AUDIT:\` —
|
|
|
1113
1113
|
<what it printed>\` when it printed anything, \`CWD-AUDIT: unrepo <path>\` when
|
|
1114
1114
|
git answered \`fatal: not a git repository\` — every run, clean or not: an
|
|
1115
1115
|
omitted line reads exactly like a check never run, and applying a mutation is
|
|
1116
|
-
how three reviews from one cell left four files modified in that checkout
|
|
1117
|
-
(#1433).`,
|
|
1116
|
+
how three reviews from one cell left four files modified in that checkout.`,
|
|
1118
1117
|
{ label: `verify:${d.key}${forPr}`, phase: "Verify", agentType: VERIFIER_AGENT_TYPE, schema: VERDICT_SCHEMA },
|
|
1119
1118
|
),
|
|
1120
1119
|
),
|
|
@@ -1122,7 +1121,7 @@ how three reviews from one cell left four files modified in that checkout
|
|
|
1122
1121
|
(votes) => !(votes && votes.some(Boolean)),
|
|
1123
1122
|
).then(
|
|
1124
1123
|
(votes) => ({ ...f, dimension: d.key, ...verdictFor(n, votes) }),
|
|
1125
|
-
//
|
|
1124
|
+
// A rejection here means retryCrashed's OWN final attempt
|
|
1126
1125
|
// rejected — both dispatches of this finding's refuter pair
|
|
1127
1126
|
// crashed, not just returned no votes. Left unhandled, that
|
|
1128
1127
|
// rejection propagates into the shared `parallel()` above (a bare
|
|
@@ -1147,7 +1146,7 @@ how three reviews from one cell left four files modified in that checkout
|
|
|
1147
1146
|
const refuted = all.filter((f) => f.verdict === "refuted");
|
|
1148
1147
|
const unverified = all.filter((f) => f.verdict === "unverified");
|
|
1149
1148
|
|
|
1150
|
-
//
|
|
1149
|
+
// One verdict on the shared run, for every dimension whose chain did
|
|
1151
1150
|
// not die — a crashed one is named below with its own reason instead, so no
|
|
1152
1151
|
// key is listed twice. Read only once every dimension has been verified: the
|
|
1153
1152
|
// failing-tests check reads the findings that reached the payload, across
|
|
@@ -1174,10 +1173,10 @@ how three reviews from one cell left four files modified in that checkout
|
|
|
1174
1173
|
// `dimensionsRun` names what was DISPATCHED after the size trim: a trimmed
|
|
1175
1174
|
// fan-out must say so, never read as full coverage. It is not a coverage claim
|
|
1176
1175
|
// on its own and never was — a specialist can be dispatched and die, or the
|
|
1177
|
-
// review's one shared test run can execute nothing
|
|
1176
|
+
// review's one shared test run can execute nothing — so
|
|
1178
1177
|
// `dimensionsUnrun` names which of those keys did not cover their ground,
|
|
1179
1178
|
// and why. A key in the first and NOT in the second ran a suite — not that
|
|
1180
|
-
// it is covered
|
|
1179
|
+
// it is covered. `unrunReason` reads the shared run's counts and
|
|
1181
1180
|
// quotes its command into its message; it never checks that the test-run
|
|
1182
1181
|
// agent ran the command it was handed rather than a narrower one, so a
|
|
1183
1182
|
// substituted runner is not classified unrun.
|
|
@@ -1187,7 +1186,7 @@ how three reviews from one cell left four files modified in that checkout
|
|
|
1187
1186
|
// make a crashed dimension indistinguishable from one the size tier never
|
|
1188
1187
|
// dispatched — this ticket set's own defect, moved one field over.
|
|
1189
1188
|
//
|
|
1190
|
-
//
|
|
1189
|
+
// Digest first, in DIGEST_KEYS order, bulk last — see DIGEST_KEYS
|
|
1191
1190
|
// above for why the order is the contract.
|
|
1192
1191
|
return {
|
|
1193
1192
|
pr,
|
|
@@ -1195,13 +1194,13 @@ how three reviews from one cell left four files modified in that checkout
|
|
|
1195
1194
|
// The recovery, next to the fields a controller reads first; null unless
|
|
1196
1195
|
// a refuter pair crashed again after the in-run retry.
|
|
1197
1196
|
resume,
|
|
1198
|
-
//
|
|
1197
|
+
// Always present, in both regimes: a reader of this payload can
|
|
1199
1198
|
// never be left unable to tell an environment artifact from a regression,
|
|
1200
1199
|
// and that costs nothing when there is nothing wrong to report.
|
|
1201
1200
|
testEnvironment: environmentNote(snap),
|
|
1202
1201
|
dimensionsRun: dimensions.map((d) => d.key),
|
|
1203
1202
|
dimensionsUnrun,
|
|
1204
|
-
//
|
|
1203
|
+
// `cwdAuditFrom`'s per-dimension read of the specialist's own
|
|
1205
1204
|
// CWD-AUDIT line — the fact a dirty or unrepo'd inherited checkout is
|
|
1206
1205
|
// otherwise reported into `scope_searched` and read by nothing.
|
|
1207
1206
|
cwdAudit,
|
package/scripts/review-eval.mjs
CHANGED
|
@@ -1,5 +1,5 @@
|
|
|
1
|
-
// review-eval.mjs — the shim for the PR review port
|
|
2
|
-
//
|
|
1
|
+
// review-eval.mjs — the shim for the PR review port. It is loaded (never
|
|
2
|
+
// review-core.mjs directly) through
|
|
3
3
|
// the Resolver, from an `eval` cell:
|
|
4
4
|
//
|
|
5
5
|
// const path = (await Bun.$`~/.fleet/bin/fleet-run --path review-eval.mjs`.text()).trim();
|
|
@@ -8,9 +8,9 @@
|
|
|
8
8
|
//
|
|
9
9
|
// `scratch` is the scratch ROOT, never the review side's `<root>/pr<N>`:
|
|
10
10
|
// runReview creates `pr<N>/` under it itself, and refuses a scratch that
|
|
11
|
-
// already ends in one
|
|
11
|
+
// already ends in one.
|
|
12
12
|
//
|
|
13
|
-
//
|
|
13
|
+
// The cell that does this is the `review-pr-<pr#>` member's own —
|
|
14
14
|
// agents/fleet-review-runner.agent.md, off the controller's turn — and it calls
|
|
15
15
|
// `runReviewToFile` (bottom of this file), which wraps `runReviewOnOmp` with the
|
|
16
16
|
// one retry and writes the result file. A controller holding its own turn can
|
|
@@ -20,12 +20,12 @@
|
|
|
20
20
|
// together under the same Install root) with a RELATIVE specifier, so the
|
|
21
21
|
// Resolver is only needed ONCE, to find this file itself — the relationship
|
|
22
22
|
// between the two is an ordinary same-install sibling import, which eval's
|
|
23
|
-
// Bun VM permits without restriction
|
|
23
|
+
// Bun VM permits without restriction. Never resolve
|
|
24
24
|
// review-core.mjs's own path through the Resolver a second time — that would
|
|
25
25
|
// be two doors where CONTEXT.md's Resolver entry says there is exactly one.
|
|
26
26
|
//
|
|
27
|
-
// No `model`/`effort` appears anywhere below
|
|
28
|
-
//
|
|
27
|
+
// No `model`/`effort` appears anywhere below, and a suite audit holds it
|
|
28
|
+
// there. Every dispatch names a bare fleet-owned
|
|
29
29
|
// `agentType` from review-core.mjs's DEFAULT_DIMENSIONS or from
|
|
30
30
|
// runReview's own snapshot/verifier dispatch, unmodified — omp's `agent()`
|
|
31
31
|
// resolves a bare frontmatter `name:` exactly, which is already what those
|
|
@@ -34,7 +34,7 @@ import { mkdir, writeFile } from "node:fs/promises";
|
|
|
34
34
|
import { isAbsolute, join } from "node:path";
|
|
35
35
|
import { digestOf, runReview, runnerPrRefusal, runnerScratchRefusal } from "./review-core.mjs";
|
|
36
36
|
|
|
37
|
-
// eval's `agent()` returns a HANDLE, not data
|
|
37
|
+
// eval's `agent()` returns a HANDLE, not data: `agent(prompt,
|
|
38
38
|
// opts)` resolves near-instantly to an `AgentHandle` with `.wait()`, and only
|
|
39
39
|
// `.wait()` unwraps the schema-validated `structuredOutput.data`. This
|
|
40
40
|
// wrapper is what makes review-core.mjs's `await host.agent(...)` read like
|
|
@@ -45,7 +45,7 @@ import { digestOf, runReview, runnerPrRefusal, runnerScratchRefusal } from "./re
|
|
|
45
45
|
// in review-core.mjs (`if (snap) {...}`, `unrunCrashed`, `verdictFor`) is
|
|
46
46
|
// written against. eval's DEFAULT `schemaMode` is "permissive": an exhausted
|
|
47
47
|
// structured-output retry is ACCEPTED anyway, carrying `schemaOverridden:
|
|
48
|
-
// true` on invalid data
|
|
48
|
+
// true` on invalid data rather than nulling out. `schemaMode:
|
|
49
49
|
// "strict"` turns that same exhaustion into a thrown/rejected result
|
|
50
50
|
// instead — still not a null — so this wrapper treats a REJECTED `.wait()`
|
|
51
51
|
// as the crash signal (`agent()`'s only way to fail loudly under omp) and
|
|
@@ -56,7 +56,7 @@ import { digestOf, runReview, runnerPrRefusal, runnerScratchRefusal } from "./re
|
|
|
56
56
|
// gave up correcting is not this shim's contract to relitigate.
|
|
57
57
|
//
|
|
58
58
|
// NO PER-CALL WORKING DIRECTORY HERE, AND WHY THE ISOLATION RULE LIVES IN THE
|
|
59
|
-
// PROMPTS INSTEAD
|
|
59
|
+
// PROMPTS INSTEAD. Three PRs reviewed back-to-back from one eval cell
|
|
60
60
|
// left FOUR files modified in the checkout that cell was standing in — a
|
|
61
61
|
// specialist's mutation-test experiment reached through a RELATIVE path, in the
|
|
62
62
|
// tree the controller reads instruments.sh, ci-state.mjs and every gate
|
|
@@ -101,7 +101,7 @@ async function ompAgent(prompt, opts) {
|
|
|
101
101
|
}
|
|
102
102
|
}
|
|
103
103
|
|
|
104
|
-
// `pipeline()`/`parallel()` have no omp counterpart
|
|
104
|
+
// `pipeline()`/`parallel()` have no omp counterpart — this is the
|
|
105
105
|
// same hand-rolled pair review-core.mjs exports as its own default (see
|
|
106
106
|
// `defaultPipeline`/`defaultParallel` there), passed through explicitly here
|
|
107
107
|
// rather than left to the default so a reader of THIS file, the one the
|
|
@@ -137,9 +137,8 @@ export async function runReviewOnOmp(args) {
|
|
|
137
137
|
return runReview({ agent: ompAgent, phase, log, pipeline, parallel }, args);
|
|
138
138
|
}
|
|
139
139
|
|
|
140
|
-
// #1802 (spec 2026-09-24-slot-based-fleet-loop-design.md § 3 §1, §2, §5, §7).
|
|
141
140
|
// The whole of the omp review runner's job, so the agent's own cell is three
|
|
142
|
-
// lines and this contract can be run
|
|
141
|
+
// lines and this contract can be run by a test: the full result
|
|
143
142
|
// object goes to `<scratch>/review-<pr>.json` — the one artefact handed to
|
|
144
143
|
// the fix-applier — and only the digest comes back, because the digest is
|
|
145
144
|
// all the controller reads and a 26–61 KB result is what the file exists to
|
|
@@ -150,7 +149,7 @@ export async function runReviewOnOmp(args) {
|
|
|
150
149
|
// reviewer's name (`review-pr-<pr>-b`), and writes no file, so nothing reads a
|
|
151
150
|
// half-review as a review. A dispatch mistake — a pr that is not a PR number, a
|
|
152
151
|
// scratch that is not absolute (eval's cwd is the MAIN CHECKOUT, so a relative
|
|
153
|
-
// one would put the file there), a scratch that already ends in `pr<N>`
|
|
152
|
+
// one would put the file there), a scratch that already ends in `pr<N>`
|
|
154
153
|
// — throws before any run: it is not a review failure, and retrying or falling
|
|
155
154
|
// back would only repeat it. `run` is the seam the test injects; nothing else
|
|
156
155
|
// passes it.
|
|
@@ -173,7 +172,7 @@ export async function runReviewToFile(args, run = runReviewOnOmp) {
|
|
|
173
172
|
errors.push(`attempt ${attempt}: ${e?.message ?? String(e)}`);
|
|
174
173
|
continue;
|
|
175
174
|
}
|
|
176
|
-
//
|
|
175
|
+
// `result.counts` is read below to build the digest and the ledger
|
|
177
176
|
// token, but that read happens AFTER `mkdir`/`writeFile` have already put
|
|
178
177
|
// the file on disk — so a `run` that resolves to a truthy, object-shaped
|
|
179
178
|
// result missing `.counts` would otherwise throw on the destructure below
|
package/scripts/verify-sha.sh
CHANGED
|
@@ -11,7 +11,7 @@ set -eu
|
|
|
11
11
|
NAME=verify-sha
|
|
12
12
|
die() { printf '%s: %s\n' "$NAME" "$1" >&2; exit 2; }
|
|
13
13
|
|
|
14
|
-
# The escaping helpers
|
|
14
|
+
# The escaping helpers. json.sh's header holds the sourcing contract and
|
|
15
15
|
# the measurements behind it. Exit 1 out of THIS script means "the sha is not
|
|
16
16
|
# reachable", the one distinction it exists to make, so a missing library would
|
|
17
17
|
# report a member's PR as sitting somewhere it does not — which is why `[ -r ]`
|
|
@@ -22,7 +22,7 @@ json_lib="$(dirname "$0")/json.sh"
|
|
|
22
22
|
# shellcheck source=json.sh
|
|
23
23
|
. "$json_lib" || die "$json_lib failed to load"
|
|
24
24
|
|
|
25
|
-
# The bounded, prompt-suppressed git transport
|
|
25
|
+
# The bounded, prompt-suppressed git transport. The fetch
|
|
26
26
|
# below is unattended: with no bound it can prompt for a credential or a host
|
|
27
27
|
# key, or stall on a transport that connects and then goes quiet, and either
|
|
28
28
|
# holds a fleet slot until something outside kills it. net.sh's header holds the
|
|
@@ -43,7 +43,7 @@ sha=$2
|
|
|
43
43
|
# leaves the die message asserting a cause the script cannot know. A bad remote
|
|
44
44
|
# URL and a removed remote are both "cannot fetch"; only git can say which.
|
|
45
45
|
#
|
|
46
|
-
# `printf`, not `echo`, on THIS line alone
|
|
46
|
+
# `printf`, not `echo`, on THIS line alone: `$branch` is still raw argv
|
|
47
47
|
# here — `[ $# -eq 2 ]` is the only guard it has passed — so a `\c` in it
|
|
48
48
|
# truncated the trace AND swallowed its newline, welding git's own `fatal:`
|
|
49
49
|
# onto the tail of a line that had already lied about the command being run.
|
|
@@ -65,7 +65,7 @@ if [ "$fetch_rc" -ne 0 ]; then
|
|
|
65
65
|
die "cannot fetch origin/$branch"
|
|
66
66
|
fi
|
|
67
67
|
|
|
68
|
-
# --verify
|
|
68
|
+
# --verify: without it, an unresolvable "origin/$branch" falls back to
|
|
69
69
|
# treating the argument as a PATH — if a file or dir of that name sits in the
|
|
70
70
|
# cwd, rev-parse prints it and exits 0, and this guard's `|| die` never fires.
|
|
71
71
|
# Measured: a real ref still wins over a same-named path either way, so
|