rungate 0.4.1 → 0.6.0
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/config/behavioral-sc-map.json +170 -0
- package/config/brief-policies.json +28 -0
- package/config/ceremony-profiles.json +207 -0
- package/config/check-registry.json +114 -0
- package/config/matcher-registry.json +142 -0
- package/config/prove-evidence.schema.json +65 -0
- package/config/spec-policies.json +36 -0
- package/config/universal-rules.yaml +15 -0
- package/config/workflow-schema.json +279 -0
- package/lib/agents-md-sections.ts +144 -0
- package/lib/branch-diff.ts +139 -0
- package/lib/bun-version.ts +268 -0
- package/lib/ci-checkout.ts +1 -1
- package/lib/cli-commands.ts +7 -1
- package/lib/compliance.ts +13 -11
- package/lib/config-loader.ts +12 -0
- package/lib/conformity.ts +1147 -116
- package/lib/context-docs.ts +143 -28
- package/lib/create-brief.ts +2 -1
- package/lib/create-sc.ts +90 -4
- package/lib/doc-scan-targets.ts +9 -0
- package/lib/evidence-command-transport.ts +160 -0
- package/lib/evidence-prevalidator.ts +22 -0
- package/lib/fallow-availability.ts +129 -0
- package/lib/gate-refusal-currency.ts +112 -0
- package/lib/gate-timeouts.ts +82 -0
- package/lib/generators/agent-briefs.ts +4 -1
- package/lib/generators/agents-md.ts +71 -23
- package/lib/generators/code-map.ts +2 -1
- package/lib/generators/types.ts +8 -0
- package/lib/git-dates.ts +55 -0
- package/lib/github.ts +195 -0
- package/lib/js-top-level-code.ts +27 -2
- package/lib/junit-checks.ts +484 -0
- package/lib/local-calendar-day.ts +78 -0
- package/lib/package-closure.ts +278 -2
- package/lib/project-state-shape.ts +59 -0
- package/lib/promote-outputs.ts +6 -2
- package/lib/release-preflight.ts +196 -0
- package/lib/report-provenance.ts +156 -0
- package/lib/sc-claim-forms.ts +75 -0
- package/lib/sc-id-cross-tree.ts +311 -0
- package/lib/scaffold/ai-review-step.ts +180 -0
- package/lib/scaffold/defaults.ts +40 -0
- package/lib/scaffold/managed-workflow.ts +125 -27
- package/lib/scaffold/owned-files.ts +48 -2
- package/lib/scaffold/profiles.ts +94 -0
- package/lib/scaffold/rehearsal.ts +232 -73
- package/lib/scaffold/steps.ts +777 -152
- package/lib/scaffold/write-steps.ts +56 -13
- package/lib/scanner.ts +38 -12
- package/lib/source-anchor.ts +144 -0
- package/lib/spec-listing.ts +148 -0
- package/lib/spec-requirements.ts +209 -0
- package/lib/success-doc.ts +779 -0
- package/lib/suite-run-reading.ts +364 -0
- package/lib/test-census.ts +199 -0
- package/lib/test-command.ts +133 -0
- package/lib/test-suite-lock.ts +166 -10
- package/lib/tool-probe.ts +120 -0
- package/lib/tracked-vs-ignored.ts +234 -0
- package/lib/trajectory.ts +377 -0
- package/lib/validators/spec-validators.ts +35 -6
- package/lib/workflow-security.ts +101 -0
- package/lib/workflow-test-harness.ts +349 -0
- package/package.json +6 -1
- package/scripts/audit-specs.ts +247 -12
- package/scripts/audit-transcript.ts +2 -1
- package/scripts/convert-spec.ts +2 -1
- package/scripts/create-sc.ts +17 -7
- package/scripts/create-spec.ts +2 -1
- package/scripts/create-success.ts +163 -0
- package/scripts/extract-constraints.ts +90 -19
- package/scripts/generate-code-map.ts +53 -12
- package/scripts/scaffold-project.ts +31 -4
- package/scripts/split-spec.ts +2 -1
- package/scripts/sync-sc-status.ts +57 -5
- package/scripts/sync-spec-tests.ts +2 -1
- package/specs/SPEC-TEMPLATE.md +200 -0
- package/templates/SUCCESS-TEMPLATE.md +72 -0
- package/templates/agent-briefs/_shared.md +9 -0
- package/templates/agent-briefs/aditi.md +46 -0
- package/templates/agent-briefs/discovery.md +37 -0
- package/templates/agent-briefs/marcus.md +39 -0
- package/templates/agent-briefs/quinn.md +67 -0
- package/templates/agent-briefs/rook.md +40 -0
- package/templates/agent-briefs/serena.md +39 -0
|
@@ -0,0 +1,156 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* The tree a report was measured against, stated by the reporter (#553).
|
|
3
|
+
*
|
|
4
|
+
* Two security reviews arrived from peer sessions on 2026-10-17. Both
|
|
5
|
+
* reported CONFIRMED findings, both were measured against a tree the branch
|
|
6
|
+
* had already moved past, and in both cases the report gave no way to tell.
|
|
7
|
+
* sec-492's three findings were measured at `7fdfa72a` (20:32); two commits
|
|
8
|
+
* landed about 75 minutes later and closed all three, and two of the
|
|
9
|
+
* "0 red" mutation rows are 3 and 1 on main.
|
|
10
|
+
*
|
|
11
|
+
* ## What is missing is the SENDING side
|
|
12
|
+
*
|
|
13
|
+
* `.claude/rules/parallel-sessions.md` already covers the receiving side —
|
|
14
|
+
* *"peer reports are evidence, not fact"* — and that rule worked both times:
|
|
15
|
+
* the reports were checked before anything was acted on, which is the only
|
|
16
|
+
* reason neither cost anything. But verifying every finding from scratch is
|
|
17
|
+
* the expensive path to a cheap fact that **the reviewer already holds**.
|
|
18
|
+
*
|
|
19
|
+
* The ship pipeline solves this inside itself: `reviewIsCurrent` pins a
|
|
20
|
+
* review to a commit and blocks when the branch has moved, which is #129,
|
|
21
|
+
* #169, #171 and #244 of accumulated work. A review conducted OUTSIDE
|
|
22
|
+
* `workflows/ship.js` inherits none of it.
|
|
23
|
+
*
|
|
24
|
+
* So this is that same binding, reachable from a terminal. It does not
|
|
25
|
+
* reimplement the comparison — `reviewIsCurrent` is imported, because a
|
|
26
|
+
* second decider answering the same question is behaviourally one decider
|
|
27
|
+
* and no test can tell them apart
|
|
28
|
+
* (`.claude/rules/checks-must-be-able-to-fail.md`).
|
|
29
|
+
*
|
|
30
|
+
* ## It fails closed, and an unanswerable question is not a pass
|
|
31
|
+
*
|
|
32
|
+
* A branch that cannot be resolved, a SHA that is not a SHA, a directory
|
|
33
|
+
* that is not a repository: all `current: false`, each with a reason naming
|
|
34
|
+
* what could not be read. A reviewer who cannot establish currency has
|
|
35
|
+
* learned something worth putting in the report, and a provenance block that
|
|
36
|
+
* went quiet on an error would be worse than none — it would be a claim.
|
|
37
|
+
*/
|
|
38
|
+
import { execFileSync } from "child_process";
|
|
39
|
+
|
|
40
|
+
import { reviewIsCurrent } from "./security-verdict";
|
|
41
|
+
|
|
42
|
+
/** The exit code of a stale or unmeasurable reading. One site, so it can be mutated. */
|
|
43
|
+
export const STALE_EXIT = 1;
|
|
44
|
+
|
|
45
|
+
export interface ProvenanceReading {
|
|
46
|
+
/** The commit the report describes. */
|
|
47
|
+
measured: string | null;
|
|
48
|
+
/** The commit the branch ends at now, or null when it could not be read. */
|
|
49
|
+
tip: string | null;
|
|
50
|
+
/** The ref the tip was read from, e.g. `origin/fix/553`. */
|
|
51
|
+
ref: string | null;
|
|
52
|
+
current: boolean;
|
|
53
|
+
/** Always present. Names both operands when there are two to name. */
|
|
54
|
+
reason: string;
|
|
55
|
+
}
|
|
56
|
+
|
|
57
|
+
function git(root: string, args: string[]): string | null {
|
|
58
|
+
try {
|
|
59
|
+
return execFileSync("git", args, {
|
|
60
|
+
cwd: root,
|
|
61
|
+
encoding: "utf-8",
|
|
62
|
+
timeout: 60_000,
|
|
63
|
+
stdio: ["ignore", "pipe", "pipe"],
|
|
64
|
+
}).trim();
|
|
65
|
+
} catch {
|
|
66
|
+
return null;
|
|
67
|
+
}
|
|
68
|
+
}
|
|
69
|
+
|
|
70
|
+
/**
|
|
71
|
+
* The ref whose tip a report should be aged against.
|
|
72
|
+
*
|
|
73
|
+
* `origin/<branch>` when it exists, because that is what another session can
|
|
74
|
+
* see; the local branch otherwise, which is the honest answer for work that
|
|
75
|
+
* has not been pushed. Detached HEAD has no branch and says so.
|
|
76
|
+
*/
|
|
77
|
+
export function defaultRef(root: string): string | null {
|
|
78
|
+
const branch = git(root, ["rev-parse", "--abbrev-ref", "HEAD"]);
|
|
79
|
+
if (branch === null || branch === "" || branch === "HEAD") return null;
|
|
80
|
+
if (git(root, ["rev-parse", "--verify", "--quiet", `refs/remotes/origin/${branch}`]) !== null) {
|
|
81
|
+
return `origin/${branch}`;
|
|
82
|
+
}
|
|
83
|
+
return branch;
|
|
84
|
+
}
|
|
85
|
+
|
|
86
|
+
/**
|
|
87
|
+
* Compare the tree a report describes against the tree the branch ends at.
|
|
88
|
+
*
|
|
89
|
+
* `measured` defaults to this working tree's HEAD, which is what a reviewer
|
|
90
|
+
* who just finished reading has in front of them.
|
|
91
|
+
*/
|
|
92
|
+
export function provenanceReading(
|
|
93
|
+
root: string,
|
|
94
|
+
measured?: string,
|
|
95
|
+
ref?: string,
|
|
96
|
+
): ProvenanceReading {
|
|
97
|
+
const measuredSha = measured ?? git(root, ["rev-parse", "HEAD"]);
|
|
98
|
+
const resolvedRef = ref ?? defaultRef(root);
|
|
99
|
+
|
|
100
|
+
if (resolvedRef === null) {
|
|
101
|
+
return {
|
|
102
|
+
measured: measuredSha,
|
|
103
|
+
tip: null,
|
|
104
|
+
ref: null,
|
|
105
|
+
current: false,
|
|
106
|
+
reason:
|
|
107
|
+
`UNMEASURED: no branch to age this report against — ${root} is not a repository, ` +
|
|
108
|
+
`or HEAD is detached. State the commit you measured and say that the branch could ` +
|
|
109
|
+
`not be resolved.`,
|
|
110
|
+
};
|
|
111
|
+
}
|
|
112
|
+
|
|
113
|
+
// Fetch nothing. A reviewer who has not fetched is comparing against a
|
|
114
|
+
// stale remote-tracking ref, which is a different failure from a moved
|
|
115
|
+
// branch and must not be hidden by a network call this function made on
|
|
116
|
+
// its own.
|
|
117
|
+
const tip = git(root, ["rev-parse", "--verify", `${resolvedRef}^{commit}`]);
|
|
118
|
+
if (tip === null) {
|
|
119
|
+
return {
|
|
120
|
+
measured: measuredSha,
|
|
121
|
+
tip: null,
|
|
122
|
+
ref: resolvedRef,
|
|
123
|
+
current: false,
|
|
124
|
+
reason:
|
|
125
|
+
`UNMEASURED: ${resolvedRef} could not be resolved, so there is nothing to compare ` +
|
|
126
|
+
`${measuredSha ?? "an unreadable HEAD"} against.`,
|
|
127
|
+
};
|
|
128
|
+
}
|
|
129
|
+
|
|
130
|
+
// The one binding. `workflows/ship.js` inlines a copy of this function for
|
|
131
|
+
// the sandbox (#69) and `test/security-verdict-blocks.test.ts` holds the
|
|
132
|
+
// two to parity; this is the library side, imported rather than repeated.
|
|
133
|
+
const currency = reviewIsCurrent(measuredSha, tip);
|
|
134
|
+
return {
|
|
135
|
+
measured: measuredSha,
|
|
136
|
+
tip,
|
|
137
|
+
ref: resolvedRef,
|
|
138
|
+
current: currency.current,
|
|
139
|
+
reason: currency.current
|
|
140
|
+
? `the report describes ${measuredSha}, which is the tip of ${resolvedRef}`
|
|
141
|
+
: (currency.reason ??
|
|
142
|
+
`the report describes ${measuredSha}, which is not the tip of ${resolvedRef} (${tip})`),
|
|
143
|
+
};
|
|
144
|
+
}
|
|
145
|
+
|
|
146
|
+
/** The block to paste at the top of a report. Always states both commits. */
|
|
147
|
+
export function provenanceBlock(reading: ProvenanceReading): string {
|
|
148
|
+
const lines = [
|
|
149
|
+
`Measured against: ${reading.measured ?? "UNKNOWN"}`,
|
|
150
|
+
`Branch tip (${reading.ref ?? "unresolved"}): ${reading.tip ?? "UNKNOWN"}`,
|
|
151
|
+
reading.current
|
|
152
|
+
? `Currency: CURRENT — ${reading.reason}`
|
|
153
|
+
: `Currency: NOT CURRENT — ${reading.reason}`,
|
|
154
|
+
];
|
|
155
|
+
return lines.join("\n");
|
|
156
|
+
}
|
package/lib/sc-claim-forms.ts
CHANGED
|
@@ -169,11 +169,86 @@ const IDS_ONLY =
|
|
|
169
169
|
*/
|
|
170
170
|
const PAREN_LABEL = /^\s*(?:SC|AC)-[A-Z]?\d+\s*:/;
|
|
171
171
|
|
|
172
|
+
/**
|
|
173
|
+
* A RANGE inside a parenthetical — `(SC-1930..SC-1948)`.
|
|
174
|
+
*
|
|
175
|
+
* An attribution names THE criterion a statement is about. A range names a
|
|
176
|
+
* group, which is how prose refers to a set it is discussing rather than
|
|
177
|
+
* claiming, and the scanner can only see the two ids that are literally
|
|
178
|
+
* written: a range of nineteen credits its two endpoints and silently not
|
|
179
|
+
* the seventeen between them. A form that credits 2 of the 19 ids it names
|
|
180
|
+
* is not measuring anything.
|
|
181
|
+
*
|
|
182
|
+
* Found in #574, on this sentence:
|
|
183
|
+
*
|
|
184
|
+
* * The spec's 19 behavioural criteria (SC-1930..SC-1948) are NOT
|
|
185
|
+
* * covered here.
|
|
186
|
+
*
|
|
187
|
+
* — which credited SC-1930 and SC-1948 and ticked both boxes. The
|
|
188
|
+
* self-reference this area keeps producing, now in its sharpest form: a
|
|
189
|
+
* comment saying the criteria are not covered was the evidence that they
|
|
190
|
+
* were.
|
|
191
|
+
*
|
|
192
|
+
* Measured before narrowing, because narrowing a detector is the change
|
|
193
|
+
* most likely to buy green with blindness: of 32 attribution credits in the
|
|
194
|
+
* tree, 30 are plain and 2 are ranges — and the 2 are these. A range at the
|
|
195
|
+
* START of a line is still a claim, by the `line` form, which is where
|
|
196
|
+
* `* SC-561..SC-564 (SPEC.md)` is credited and is unaffected.
|
|
197
|
+
*/
|
|
198
|
+
const PAREN_RANGE = /\.\.+/;
|
|
199
|
+
|
|
172
200
|
function inIdOnlyParenthetical(line: string, at: number): boolean {
|
|
173
201
|
const open = line.lastIndexOf("(", at);
|
|
174
202
|
if (open < 0) return false;
|
|
175
203
|
const close = line.indexOf(")", open);
|
|
176
204
|
if (close < 0 || close < at) return false;
|
|
177
205
|
const inner = line.slice(open + 1, close);
|
|
206
|
+
if (PAREN_RANGE.test(inner)) return false;
|
|
178
207
|
return IDS_ONLY.test(inner) || PAREN_LABEL.test(inner);
|
|
179
208
|
}
|
|
209
|
+
|
|
210
|
+
/**
|
|
211
|
+
* A `describe`/`test`/`it` call that opens a title literal, and whether it
|
|
212
|
+
* will ever run (#574).
|
|
213
|
+
*
|
|
214
|
+
* A test title is the project's oldest coverage convention and the one
|
|
215
|
+
* position that needed no rule — until `test.todo("SC-1930: …")` turned out
|
|
216
|
+
* to be a title too. A todo is the one test guaranteed never to fail, so
|
|
217
|
+
* crediting it produces a tick indistinguishable from a tick earned by a
|
|
218
|
+
* test that can go red. Measured when this was found: 64 ids named only
|
|
219
|
+
* inside a `.todo`, 37 of them ticked `[x]` in a spec, and 19 more about to
|
|
220
|
+
* flip on the next run of the sync — one of them from a file whose own
|
|
221
|
+
* comment says, four lines from the top, that none of its criteria are
|
|
222
|
+
* covered here.
|
|
223
|
+
*
|
|
224
|
+
* `.skip` is the same fact with a different word, and there are five of
|
|
225
|
+
* those. `.skipIf` and `.todoIf` take a condition, so whether they ran is
|
|
226
|
+
* not answerable from the source at all; there are none today and they are
|
|
227
|
+
* named anyway, because a predicate narrower than what it detects is this
|
|
228
|
+
* repo's third defect shape and the cost of naming them is one word each.
|
|
229
|
+
*
|
|
230
|
+
* `.only`, `.each`, `.failing` and `.concurrent` all RUN, so they still
|
|
231
|
+
* credit — `.failing` included, because a test expected to fail is still a
|
|
232
|
+
* test that goes red when the expectation stops holding.
|
|
233
|
+
*
|
|
234
|
+
* Returns `null` when the text does not open a title call at all. The three
|
|
235
|
+
* answers are distinct on purpose: a non-running title must be reported as
|
|
236
|
+
* UNMEASURED rather than dropped, and dropping it would make "nobody has
|
|
237
|
+
* written this test" look exactly like "nobody has mentioned this id".
|
|
238
|
+
*/
|
|
239
|
+
export type TitleCall = "runs" | "never-runs";
|
|
240
|
+
|
|
241
|
+
const NON_RUNNING = /\.(?:todo|skip|skipIf|todoIf)\b/;
|
|
242
|
+
|
|
243
|
+
/** Opens a `describe`/`test`/`it` title, allowing `.each(...)`, `.only`, `.skip`, `.todo`. */
|
|
244
|
+
const TITLE_CALL = /\b(?:describe|test|it)\s*((?:\.\w+\s*(?:\([^)]*\))?\s*)*)\(\s*$/;
|
|
245
|
+
|
|
246
|
+
/**
|
|
247
|
+
* THE single exported predicate for whether a title call declares coverage,
|
|
248
|
+
* for the same reason `commentClaimForm` is: one binding to mutate.
|
|
249
|
+
*/
|
|
250
|
+
export function titleCall(before: string): TitleCall | null {
|
|
251
|
+
const m = TITLE_CALL.exec(before);
|
|
252
|
+
if (m === null) return null;
|
|
253
|
+
return NON_RUNNING.test(m[1] ?? "") ? "never-runs" : "runs";
|
|
254
|
+
}
|
|
@@ -0,0 +1,311 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* The SC id namespace as it exists on this machine, not in this checkout (#503).
|
|
3
|
+
*
|
|
4
|
+
* `nextSCNumber` allocates "highest + 1", and #392 taught it to read the whole
|
|
5
|
+
* registry rather than one spec's content. Both fixes are about the same
|
|
6
|
+
* mistake at different widths, and the width was still wrong: the registry is
|
|
7
|
+
* `specs/` **in the working tree the command happens to run in**.
|
|
8
|
+
*
|
|
9
|
+
* On 2026-10-16 two sessions allocated `SC-1706..1710` within minutes of each
|
|
10
|
+
* other — one in `specs/HARNESS-STANDARD.md`, one in
|
|
11
|
+
* `specs/CONFIG-DIRECTORY-STRUCTURE-SPEC.md`, in two worktrees of the same
|
|
12
|
+
* repository. Both read the highest id they could see. Both were right about
|
|
13
|
+
* their own tree. `test/spec-sc-ids-unique.test.ts` was green in each,
|
|
14
|
+
* independently, until a rebase put them in one place.
|
|
15
|
+
*
|
|
16
|
+
* That is the second shape in `.claude/rules/checks-must-be-able-to-fail.md`
|
|
17
|
+
* — a detector narrower than what it detects. The question is "is this number
|
|
18
|
+
* free in the project"; the detector answered "is this number free in this
|
|
19
|
+
* checkout". The two differ exactly when more than one session is working,
|
|
20
|
+
* which here is the normal case: this repository has 110 worktrees on disk.
|
|
21
|
+
*
|
|
22
|
+
* ## Two sources, because the collision lives in the gap between them
|
|
23
|
+
*
|
|
24
|
+
* | Source | Catches | Misses |
|
|
25
|
+
* |---|---|---|
|
|
26
|
+
* | sibling worktrees' `specs/` | **uncommitted** ids in another session's tree | a branch whose worktree was removed |
|
|
27
|
+
* | every local and remote-tracking ref | committed ids on any branch, pushed or not | work that exists only on another machine |
|
|
28
|
+
*
|
|
29
|
+
* Neither alone is enough. The #503 collision was an unpushed branch in
|
|
30
|
+
* another worktree, which the ref scan sees only once it is committed —
|
|
31
|
+
* and `.claude/rules/parallel-sessions.md` exists because that window is
|
|
32
|
+
* real and people sit in it for an hour.
|
|
33
|
+
*
|
|
34
|
+
* ## This errs toward "taken", deliberately
|
|
35
|
+
*
|
|
36
|
+
* A stale worktree holding an abandoned branch contributes ids that nothing
|
|
37
|
+
* will ever claim, and the allocator will skip them. That costs a gap in the
|
|
38
|
+
* numbering, which costs nothing: SC ids are opaque join keys, not an
|
|
39
|
+
* ordering anyone reads. The other direction costs a collision, which is two
|
|
40
|
+
* spec lines a test can bind to and no way to see which one it bound to.
|
|
41
|
+
*
|
|
42
|
+
* ## What this is NOT
|
|
43
|
+
*
|
|
44
|
+
* It is not a guard that refuses on a cross-tree duplicate, and that is a
|
|
45
|
+
* measurement rather than a shrug. Across the 110 worktrees on this machine,
|
|
46
|
+
* 30 ids are declared in more than one spec FILE: `SC-A1`/`SC-A2`, which are
|
|
47
|
+
* anti-criteria that several specs legitimately carry, and `SC-694..703`,
|
|
48
|
+
* which is a REAL collision from months ago, resolved on main by renumbering
|
|
49
|
+
* and preserved forever in the stale trees that still hold the losing side.
|
|
50
|
+
* A guard over that population is permanently red for reasons nobody can act
|
|
51
|
+
* on, which is the fastest way to build a refusal everyone learns to click
|
|
52
|
+
* past. Fix the allocator that mints collisions; do not alarm on the fossils.
|
|
53
|
+
*/
|
|
54
|
+
import { execFileSync } from "child_process";
|
|
55
|
+
import { existsSync, readdirSync, readFileSync, statSync } from "fs";
|
|
56
|
+
import { join } from "path";
|
|
57
|
+
|
|
58
|
+
import { isFixtureScId, scDeclarationPattern, scHeadingDeclarationPattern } from "./sc-ids";
|
|
59
|
+
|
|
60
|
+
/**
|
|
61
|
+
* One place ids were looked for, and whether the look succeeded.
|
|
62
|
+
*
|
|
63
|
+
* `worktree` and `refs` are this module's own two sources. The other three
|
|
64
|
+
* are the legs `allocateScNumber` composes on top of them (#593) — the
|
|
65
|
+
* spec's own content, the spec registry, and the ids spelled in source. They
|
|
66
|
+
* live in the same shape because the question a reader asks is one question:
|
|
67
|
+
* which widths produced this number, and did any of them fail to run.
|
|
68
|
+
*/
|
|
69
|
+
export interface SourceReading {
|
|
70
|
+
label: string;
|
|
71
|
+
kind: "worktree" | "refs" | "spec-content" | "registry" | "source-mentions";
|
|
72
|
+
/** `read` means the ids below came from here; `skipped` means they did not. */
|
|
73
|
+
status: "read" | "skipped";
|
|
74
|
+
/** Why it was skipped. Absent on a successful read. */
|
|
75
|
+
reason?: string;
|
|
76
|
+
/** How many distinct ids this source contributed. */
|
|
77
|
+
found: number;
|
|
78
|
+
}
|
|
79
|
+
|
|
80
|
+
export interface ScNamespaceReading {
|
|
81
|
+
/** id → the labels that declared it. */
|
|
82
|
+
ids: Map<string, string[]>;
|
|
83
|
+
sources: SourceReading[];
|
|
84
|
+
/**
|
|
85
|
+
* Why this reading is a floor rather than a fact. Never empty: a clone on
|
|
86
|
+
* another machine is outside anything this process can see, and saying so
|
|
87
|
+
* is the difference between an unmeasured half and a silent pass
|
|
88
|
+
* (`.claude/rules/config-access.md`).
|
|
89
|
+
*/
|
|
90
|
+
unmeasured: string[];
|
|
91
|
+
}
|
|
92
|
+
|
|
93
|
+
function git(root: string, args: string[]): string {
|
|
94
|
+
// 64 MB, because the default is 1 MB and a grep over a thousand refs blows
|
|
95
|
+
// through it. Measured: the first run of this module reported
|
|
96
|
+
// `ENOBUFS (stdout or stderr buffer reached maxBuffer size limit)` as a
|
|
97
|
+
// skipped source, which is the reading behaving correctly over a bug.
|
|
98
|
+
return execFileSync("git", args, {
|
|
99
|
+
cwd: root,
|
|
100
|
+
encoding: "utf-8",
|
|
101
|
+
timeout: 120_000,
|
|
102
|
+
maxBuffer: 64 * 1024 * 1024,
|
|
103
|
+
// Captured, not inherited. A failed `git` here is a source this reading
|
|
104
|
+
// records as skipped, and letting its stderr through would put
|
|
105
|
+
// "fatal: not a git repository" in the middle of a suite's output for a
|
|
106
|
+
// case that is handling exactly that (#60: a test must not corrupt the
|
|
107
|
+
// output stream of the suite containing it).
|
|
108
|
+
stdio: ["ignore", "pipe", "pipe"],
|
|
109
|
+
});
|
|
110
|
+
}
|
|
111
|
+
|
|
112
|
+
/**
|
|
113
|
+
* Every worktree of this repository except `root` itself.
|
|
114
|
+
*
|
|
115
|
+
* `failure` is non-null when git could not be asked at all — which is a
|
|
116
|
+
* different fact from "there are no siblings", and conflating the two is how
|
|
117
|
+
* a reading of nothing reads as a reading of everything.
|
|
118
|
+
*/
|
|
119
|
+
export function siblingWorktrees(root: string): { paths: string[]; failure: string | null } {
|
|
120
|
+
let out: string;
|
|
121
|
+
try {
|
|
122
|
+
out = git(root, ["worktree", "list", "--porcelain"]);
|
|
123
|
+
} catch (err) {
|
|
124
|
+
return { paths: [], failure: (err as Error).message.split("\n")[0] };
|
|
125
|
+
}
|
|
126
|
+
const paths: string[] = [];
|
|
127
|
+
let current: string | null = null;
|
|
128
|
+
let prunable = false;
|
|
129
|
+
const flush = (): void => {
|
|
130
|
+
if (current && !prunable && current !== root) paths.push(current);
|
|
131
|
+
current = null;
|
|
132
|
+
prunable = false;
|
|
133
|
+
};
|
|
134
|
+
for (const line of out.split("\n")) {
|
|
135
|
+
if (line.startsWith("worktree ")) {
|
|
136
|
+
flush();
|
|
137
|
+
current = line.slice("worktree ".length);
|
|
138
|
+
} else if (line.startsWith("prunable")) {
|
|
139
|
+
prunable = true;
|
|
140
|
+
}
|
|
141
|
+
}
|
|
142
|
+
flush();
|
|
143
|
+
return { paths, failure: null };
|
|
144
|
+
}
|
|
145
|
+
|
|
146
|
+
/** The ids a single `specs/` directory declares, by checkbox or by heading. */
|
|
147
|
+
function idsInSpecsDir(specsDir: string): string[] {
|
|
148
|
+
const ids: string[] = [];
|
|
149
|
+
for (const name of readdirSync(specsDir)) {
|
|
150
|
+
if (!name.endsWith(".md")) continue;
|
|
151
|
+
let content: string;
|
|
152
|
+
try {
|
|
153
|
+
content = readFileSync(join(specsDir, name), "utf-8");
|
|
154
|
+
} catch {
|
|
155
|
+
// A spec we cannot read contributes nothing, and the caller records the
|
|
156
|
+
// tree as read regardless: one unreadable file is not an unmeasured
|
|
157
|
+
// tree, and pretending otherwise would make every reading unmeasured.
|
|
158
|
+
continue;
|
|
159
|
+
}
|
|
160
|
+
for (const m of content.matchAll(scDeclarationPattern())) ids.push(m[2]);
|
|
161
|
+
for (const m of content.matchAll(scHeadingDeclarationPattern())) ids.push(m[1]);
|
|
162
|
+
}
|
|
163
|
+
return ids;
|
|
164
|
+
}
|
|
165
|
+
|
|
166
|
+
/**
|
|
167
|
+
* Ids declared in `specs/` on any ref this repository holds — local branches
|
|
168
|
+
* and remote-tracking branches both.
|
|
169
|
+
*
|
|
170
|
+
* One `git grep` over every ref at once rather than a checkout each: measured
|
|
171
|
+
* at 0.17s over 1087 refs, against minutes for the obvious implementation.
|
|
172
|
+
*/
|
|
173
|
+
function idsInRefs(root: string): { ids: string[]; refs: number; failure: string | null } {
|
|
174
|
+
let refList: string;
|
|
175
|
+
try {
|
|
176
|
+
refList = git(root, ["for-each-ref", "--format=%(refname)", "refs/heads", "refs/remotes"]);
|
|
177
|
+
} catch (err) {
|
|
178
|
+
return { ids: [], refs: 0, failure: (err as Error).message.split("\n")[0] };
|
|
179
|
+
}
|
|
180
|
+
const refs = refList.split("\n").filter(r => r.trim() !== "");
|
|
181
|
+
if (refs.length === 0) return { ids: [], refs: 0, failure: null };
|
|
182
|
+
|
|
183
|
+
const ids: string[] = [];
|
|
184
|
+
// Batched: a ref list this long overruns the argument limit in one call.
|
|
185
|
+
for (let i = 0; i < refs.length; i += 100) {
|
|
186
|
+
let out: string;
|
|
187
|
+
try {
|
|
188
|
+
out = git(root, [
|
|
189
|
+
"grep",
|
|
190
|
+
"-h",
|
|
191
|
+
"-E",
|
|
192
|
+
// Declaration-shaped lines only — a checkbox or a heading. Every
|
|
193
|
+
// mention would be correct and is ~30x the output, which is how the
|
|
194
|
+
// first version hit the buffer limit.
|
|
195
|
+
"^(- \\[.\\] |#+ )SC-[A-Z]?[0-9]+",
|
|
196
|
+
...refs.slice(i, i + 100),
|
|
197
|
+
"--",
|
|
198
|
+
"specs/*.md",
|
|
199
|
+
]);
|
|
200
|
+
} catch (err) {
|
|
201
|
+
// git grep exits 1 when nothing matched, which is a real answer and not
|
|
202
|
+
// a failure. Anything else is a failure, and the caller must hear it.
|
|
203
|
+
const status = (err as { status?: number }).status;
|
|
204
|
+
if (status === 1) continue;
|
|
205
|
+
return { ids, refs: refs.length, failure: (err as Error).message.split("\n")[0] };
|
|
206
|
+
}
|
|
207
|
+
for (const m of out.matchAll(scDeclarationPattern())) ids.push(m[2]);
|
|
208
|
+
for (const m of out.matchAll(scHeadingDeclarationPattern())) ids.push(m[1]);
|
|
209
|
+
}
|
|
210
|
+
return { ids, refs: refs.length, failure: null };
|
|
211
|
+
}
|
|
212
|
+
|
|
213
|
+
/**
|
|
214
|
+
* Every SC id this machine can see, with an account of where it looked.
|
|
215
|
+
*
|
|
216
|
+
* Fixture ids — the reserved 9000s — are dropped. The reservation exists so a
|
|
217
|
+
* fixture can spell an id without claiming it, and counting one as taken
|
|
218
|
+
* would push every future criterion past 9000 and destroy the reservation it
|
|
219
|
+
* was built on. That is `nextSCNumber`'s rule for `spelledInSource`, applied
|
|
220
|
+
* at the same width.
|
|
221
|
+
*/
|
|
222
|
+
export function readScNamespace(root: string): ScNamespaceReading {
|
|
223
|
+
const ids = new Map<string, string[]>();
|
|
224
|
+
const sources: SourceReading[] = [];
|
|
225
|
+
const unmeasured: string[] = [];
|
|
226
|
+
|
|
227
|
+
const record = (id: string, label: string): void => {
|
|
228
|
+
if (isFixtureScId(id)) return;
|
|
229
|
+
const seen = ids.get(id) ?? [];
|
|
230
|
+
if (!seen.includes(label)) seen.push(label);
|
|
231
|
+
ids.set(id, seen);
|
|
232
|
+
};
|
|
233
|
+
|
|
234
|
+
const { paths, failure } = siblingWorktrees(root);
|
|
235
|
+
if (failure !== null) {
|
|
236
|
+
sources.push({ label: "git worktree list", kind: "worktree", status: "skipped", reason: failure, found: 0 });
|
|
237
|
+
unmeasured.push(`sibling worktrees could not be enumerated: ${failure}`);
|
|
238
|
+
}
|
|
239
|
+
for (const path of paths) {
|
|
240
|
+
const specsDir = join(path, "specs");
|
|
241
|
+
let usable = false;
|
|
242
|
+
try {
|
|
243
|
+
usable = existsSync(specsDir) && statSync(specsDir).isDirectory();
|
|
244
|
+
} catch {
|
|
245
|
+
usable = false;
|
|
246
|
+
}
|
|
247
|
+
if (!usable) {
|
|
248
|
+
// A worktree git lists but whose tree is gone, or which predates
|
|
249
|
+
// specs/. Recorded rather than dropped: a skipped source is the thing
|
|
250
|
+
// that makes the count below a floor.
|
|
251
|
+
sources.push({ label: path, kind: "worktree", status: "skipped", reason: "no readable specs/ directory", found: 0 });
|
|
252
|
+
unmeasured.push(`${path}: no readable specs/ directory`);
|
|
253
|
+
continue;
|
|
254
|
+
}
|
|
255
|
+
let found: string[];
|
|
256
|
+
try {
|
|
257
|
+
found = idsInSpecsDir(specsDir);
|
|
258
|
+
} catch (err) {
|
|
259
|
+
sources.push({ label: path, kind: "worktree", status: "skipped", reason: (err as Error).message, found: 0 });
|
|
260
|
+
unmeasured.push(`${path}: ${(err as Error).message}`);
|
|
261
|
+
continue;
|
|
262
|
+
}
|
|
263
|
+
for (const id of found) record(id, path);
|
|
264
|
+
sources.push({ label: path, kind: "worktree", status: "read", found: new Set(found).size });
|
|
265
|
+
}
|
|
266
|
+
|
|
267
|
+
const refs = idsInRefs(root);
|
|
268
|
+
if (refs.failure !== null) {
|
|
269
|
+
sources.push({ label: "refs", kind: "refs", status: "skipped", reason: refs.failure, found: 0 });
|
|
270
|
+
unmeasured.push(`branch refs could not be read: ${refs.failure}`);
|
|
271
|
+
} else {
|
|
272
|
+
for (const id of refs.ids) record(id, "refs");
|
|
273
|
+
sources.push({
|
|
274
|
+
label: `refs (${refs.refs} local and remote-tracking)`,
|
|
275
|
+
kind: "refs",
|
|
276
|
+
status: "read",
|
|
277
|
+
found: new Set(refs.ids).size,
|
|
278
|
+
});
|
|
279
|
+
}
|
|
280
|
+
|
|
281
|
+
// Always present, and not boilerplate: it is the one half of the namespace
|
|
282
|
+
// no process on this machine can reach, and an allocator that did not say
|
|
283
|
+
// so would be claiming a guarantee it cannot give.
|
|
284
|
+
unmeasured.push(
|
|
285
|
+
"ids allocated in a clone on another machine, and not yet pushed or fetched here, are not visible to any of the above",
|
|
286
|
+
);
|
|
287
|
+
|
|
288
|
+
return { ids, sources, unmeasured };
|
|
289
|
+
}
|
|
290
|
+
|
|
291
|
+
/**
|
|
292
|
+
* A one-line account of what a reading covered, for a CLI to print.
|
|
293
|
+
*
|
|
294
|
+
* It NAMES the legs rather than counting them (#593). A count says how many
|
|
295
|
+
* sources answered; it cannot say which width went unasked, and a number
|
|
296
|
+
* allocated without consulting the spec registry printed the same
|
|
297
|
+
* "2 source(s)" as one that had. This is the single binding every caller
|
|
298
|
+
* prints through, so there is no second renderer to drift from it.
|
|
299
|
+
*/
|
|
300
|
+
export function describeReading(reading: ScNamespaceReading): string {
|
|
301
|
+
const read = reading.sources.filter(s => s.status === "read");
|
|
302
|
+
const skipped = reading.sources.filter(s => s.status === "skipped");
|
|
303
|
+
const legs = [...new Set(reading.sources.map(s => s.kind))];
|
|
304
|
+
const parts = [
|
|
305
|
+
`${reading.ids.size} ids across ${read.length} source(s)`,
|
|
306
|
+
`legs: ${legs.join(", ")}`,
|
|
307
|
+
skipped.length > 0 ? `${skipped.length} skipped` : null,
|
|
308
|
+
`${reading.unmeasured.length} unmeasured`,
|
|
309
|
+
].filter(Boolean);
|
|
310
|
+
return parts.join(", ");
|
|
311
|
+
}
|