@dev-loops/core 1.0.0-rc.5 → 1.0.0-rc.7
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 +12 -1
- package/src/analysis/change-classifier.mjs +10 -0
- package/src/analysis/diff-analyzer.mjs +68 -1
- package/src/claude/hook-decisions.mjs +204 -5
- package/src/cli/primitives.mjs +51 -1
- package/src/config/config.mjs +307 -14
- package/src/config/extension-defaults.yaml +39 -1
- package/src/github/comment-id-guard.mjs +158 -0
- package/src/github/copilot-helpers.mjs +145 -5
- package/src/github/gh.mjs +94 -0
- package/src/github/issue-ops.mjs +13 -0
- package/src/loop/agent-stall.mjs +196 -0
- package/src/loop/bash-command-classify.mjs +277 -0
- package/src/loop/cache-telemetry-evidence.mjs +437 -0
- package/src/loop/copilot-loop-iterations.mjs +2 -1
- package/src/loop/default-branch-guard.mjs +35 -2
- package/src/loop/gate-carry-forward.mjs +19 -6
- package/src/loop/gate-fanin.mjs +190 -29
- package/src/loop/handoff-envelope.mjs +40 -20
- package/src/loop/issue-refinement-artifact.mjs +94 -0
- package/src/loop/lifecycle-state.mjs +21 -2
- package/src/loop/main-checkout-ff.mjs +73 -0
- package/src/loop/markdown-sections.mjs +40 -0
- package/src/loop/normalize.mjs +7 -0
- package/src/loop/plan-file-promote-contract.mjs +14 -1
- package/src/loop/plan-file-refine-contract.mjs +92 -8
- package/src/loop/policy-constants.mjs +9 -0
- package/src/loop/pr-gate-coordination.mjs +65 -12
- package/src/loop/primer-evidence.mjs +375 -0
- package/src/loop/public-dev-loop-routing.mjs +7 -15
- package/src/loop/queue-board-sync.mjs +1 -26
- package/src/loop/queue-driver.mjs +14 -1
- package/src/loop/refinement-grill-state.mjs +3 -5
- package/src/loop/review-dispatch-plan.mjs +1034 -0
- package/src/loop/review-lineage.mjs +588 -0
- package/src/loop/reviewer-loop-state.mjs +8 -13
- package/src/loop/run-post-merge-actions.mjs +148 -0
- package/src/loop/size-budget-merge-gate.mjs +121 -0
- package/src/loop/tracker-pr-state.mjs +5 -15
- package/src/loop/ui-designer-review-scoping.mjs +171 -0
- package/src/loop/ui-review-drive.mjs +3 -1
- package/src/loop/ui-review-report.mjs +2 -5
- package/src/loop/ui-review-teardown.mjs +3 -1
- package/src/loop/worktree-guard.mjs +80 -0
- package/src/projects/list-queue-items.mjs +1 -27
- package/src/projects/move-queue-item.mjs +38 -28
- package/src/security/secret-scan.mjs +330 -0
|
@@ -0,0 +1,437 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* cache-telemetry-evidence.mjs — cache telemetry adapter + before/after evidence
|
|
3
|
+
* artifact (issue #1468 slice 4).
|
|
4
|
+
*
|
|
5
|
+
* Slices 1-3 produced the deterministic request plan, request-prefix fingerprints,
|
|
6
|
+
* stable/volatile separation, per-model primer-group partitioning, and the
|
|
7
|
+
* primer-dispatch ordering evidence + fail-closed fan-in validation. That
|
|
8
|
+
* proves the ordering and request-fingerprint invariants a cache-aware dispatch
|
|
9
|
+
* relies on, but it does not by itself MEASURE the provider cache reuse: whether
|
|
10
|
+
* the N reviewers actually read the cache entry the primer wrote.
|
|
11
|
+
*
|
|
12
|
+
* This slice adds the harness-capability-aware telemetry surface (Section D of
|
|
13
|
+
* #1468). Where a harness exposes cache creation/read usage telemetry, we
|
|
14
|
+
* persist per-primer creation tokens and per-reviewer read tokens and emit an
|
|
15
|
+
* aggregate read:create report. Where the harness is opaque or telemetry is
|
|
16
|
+
* unavailable, we record `cacheReuseVerified: false` with the reason and NEVER
|
|
17
|
+
* describe the result as a verified `1 write + N reads` outcome — only the
|
|
18
|
+
* ordering + fingerprint invariants from the earlier slices may be claimed.
|
|
19
|
+
*
|
|
20
|
+
* This module is pure and offline (no GitHub, no harness, no clock). It owns:
|
|
21
|
+
*
|
|
22
|
+
* 1. before/after evidence artifact builder — a deterministic per-gate-run record
|
|
23
|
+
* ('<gate>-<headSha>.cache-telemetry.json') pairing the plan + capability
|
|
24
|
+
* record with the observed cache-creation (before) and cache-read (after)
|
|
25
|
+
* telemetry events, and deriving the aggregate read:create report.
|
|
26
|
+
* 2. fail-closed validator — checks, named individually, that verified reuse is
|
|
27
|
+
* never claimed without the capability + at least one creation and one read
|
|
28
|
+
* event that telemetry can actually observe; an opaque/unavailable harness
|
|
29
|
+
* must fail closed to `cacheReuseVerified: false` (Section D honesty gate).
|
|
30
|
+
* 3. deterministic path + writer — consumable by the gate ledger without
|
|
31
|
+
* re-derivation (GATE-EXEC-CACHE-TELEMETRY).
|
|
32
|
+
*/
|
|
33
|
+
import { mkdir, writeFile } from "node:fs/promises";
|
|
34
|
+
import path from "node:path";
|
|
35
|
+
|
|
36
|
+
import {
|
|
37
|
+
CACHE_BOUNDARY_AFTER_SHARED_PREFIX,
|
|
38
|
+
cacheReuseVeracity,
|
|
39
|
+
normalizeHarnessCapabilities,
|
|
40
|
+
} from "./review-dispatch-plan.mjs";
|
|
41
|
+
|
|
42
|
+
export const CACHE_TELEMETRY_SCHEMA_VERSION = 1;
|
|
43
|
+
|
|
44
|
+
/**
|
|
45
|
+
* Deterministic artifact path for a gate run's cache-telemetry evidence.
|
|
46
|
+
*
|
|
47
|
+
* @param {object} input
|
|
48
|
+
* @param {string} input.dir - directory to write under (e.g. the gate-context dir).
|
|
49
|
+
* @param {string} input.gate - gate name (pre_approval_gate, draft_gate, ...).
|
|
50
|
+
* @param {string} input.headSha - reviewed head SHA (hex, 7-64).
|
|
51
|
+
* @returns {string} absolute-style path joined under `dir`.
|
|
52
|
+
*/
|
|
53
|
+
export function cacheTelemetryPath({ dir, gate, headSha } = {}) {
|
|
54
|
+
if (typeof dir !== "string" || dir.length === 0) throw new Error("cacheTelemetryPath requires a dir");
|
|
55
|
+
if (typeof gate !== "string" || gate.length === 0) throw new Error("cacheTelemetryPath requires a gate");
|
|
56
|
+
if (typeof headSha !== "string" || !/^[0-9a-f]{7,64}$/i.test(headSha.trim())) {
|
|
57
|
+
throw new Error("cacheTelemetryPath requires a hex headSha");
|
|
58
|
+
}
|
|
59
|
+
return path.join(dir, `${gate}-${headSha.trim().toLowerCase()}.cache-telemetry.json`);
|
|
60
|
+
}
|
|
61
|
+
|
|
62
|
+
/** Count of entries whose `tokens` is a finite non-negative number. Non-object/null entries are skipped so a malformed artifact element can never throw. */
|
|
63
|
+
function countTokens(entries) {
|
|
64
|
+
return entries.filter(
|
|
65
|
+
(e) => e != null && typeof e === "object" && Number.isFinite(e.tokens) && e.tokens >= 0,
|
|
66
|
+
).length;
|
|
67
|
+
}
|
|
68
|
+
|
|
69
|
+
/** Sum of finite non-negative `tokens` across entries. Non-object/null entries are skipped. */
|
|
70
|
+
function sumTokens(entries) {
|
|
71
|
+
return entries.reduce((acc, e) => {
|
|
72
|
+
const t =
|
|
73
|
+
e != null && typeof e === "object" && Number.isFinite(e.tokens) && e.tokens >= 0 ? e.tokens : 0;
|
|
74
|
+
return acc + t;
|
|
75
|
+
}, 0);
|
|
76
|
+
}
|
|
77
|
+
|
|
78
|
+
/**
|
|
79
|
+
* Build the before/after cache-telemetry evidence artifact for a gate run.
|
|
80
|
+
*
|
|
81
|
+
* "Before" = the cache-creation events (the primer write side); "after" = the
|
|
82
|
+
* cache-read events (the reviewer read side). Where a harness does not expose
|
|
83
|
+
* per-event token counts, callers may record the event with `tokens: null`; the
|
|
84
|
+
* event still counts toward the create/read tally but not the token aggregates.
|
|
85
|
+
*
|
|
86
|
+
* @param {object} input
|
|
87
|
+
* @param {object} input.plan - the dispatch plan from buildReviewDispatchPlan().
|
|
88
|
+
* @param {object} [input.capabilities] - normalized harness capabilities (used
|
|
89
|
+
* to drive the veracity gate; defaults to the plan's capabilities when present).
|
|
90
|
+
* @param {Array<object>} input.primerCacheCreations - [{ model, primerForm, tokens|null }]
|
|
91
|
+
* observed cache creations (the "before" write events).
|
|
92
|
+
* @param {Array<object>} input.reviewerCacheReads - [{ model, angle, tokens|null }]
|
|
93
|
+
* observed cache reads (the "after" read events).
|
|
94
|
+
* @returns {object} canonical evidence artifact.
|
|
95
|
+
*/
|
|
96
|
+
export function buildCacheTelemetryEvidence({
|
|
97
|
+
plan,
|
|
98
|
+
capabilities,
|
|
99
|
+
primerCacheCreations = [],
|
|
100
|
+
reviewerCacheReads = [],
|
|
101
|
+
} = {}) {
|
|
102
|
+
if (!plan || typeof plan !== "object" || !Array.isArray(plan.requestGroups)) {
|
|
103
|
+
throw new Error("buildCacheTelemetryEvidence requires a plan with requestGroups");
|
|
104
|
+
}
|
|
105
|
+
if (typeof plan.gate !== "string" || plan.gate.length === 0) {
|
|
106
|
+
throw new Error("buildCacheTelemetryEvidence requires a plan with a non-empty gate");
|
|
107
|
+
}
|
|
108
|
+
if (typeof plan.headSha !== "string" || !/^[0-9a-f]{7,64}$/i.test(plan.headSha.trim())) {
|
|
109
|
+
throw new Error("buildCacheTelemetryEvidence requires a plan with a hex headSha");
|
|
110
|
+
}
|
|
111
|
+
if (typeof plan.planHash !== "string" || plan.planHash.length === 0) {
|
|
112
|
+
throw new Error("buildCacheTelemetryEvidence requires a plan with a non-empty planHash");
|
|
113
|
+
}
|
|
114
|
+
if (!Array.isArray(primerCacheCreations)) throw new Error("primerCacheCreations must be an array");
|
|
115
|
+
if (!Array.isArray(reviewerCacheReads)) throw new Error("reviewerCacheReads must be an array");
|
|
116
|
+
|
|
117
|
+
// Resolve the capability record: an explicitly passed one wins, else fall back
|
|
118
|
+
// to the plan's `capabilities` field. A missing capability record is itself a
|
|
119
|
+
// fail-closed truth (cannot claim verified reuse without a capability record).
|
|
120
|
+
let caps = capabilities ?? plan.capabilities ?? null;
|
|
121
|
+
if (caps != null) {
|
|
122
|
+
// A capability spec may carry a `harness` key plus dimension overrides
|
|
123
|
+
// (parity with buildReviewDispatchPlan's handling).
|
|
124
|
+
const hasHarness = typeof caps === "object" && !Array.isArray(caps) && typeof caps.harness === "string";
|
|
125
|
+
if (hasHarness) {
|
|
126
|
+
const { harness: harnessName, ...dims } = caps;
|
|
127
|
+
caps = normalizeHarnessCapabilities({ harness: harnessName, capabilities: dims });
|
|
128
|
+
} else {
|
|
129
|
+
caps = normalizeHarnessCapabilities({ capabilities: caps });
|
|
130
|
+
}
|
|
131
|
+
}
|
|
132
|
+
|
|
133
|
+
const normalizeTokens = (tokens, label) => {
|
|
134
|
+
// Fail CLOSED on a non-null token that is not a finite non-negative
|
|
135
|
+
// number (numeric strings, NaN, Infinity, negatives, other types): the
|
|
136
|
+
// reported value is telemetry evidence and a silently-coerced token would
|
|
137
|
+
// under-report the create/read contribution with no signal. `null` is the
|
|
138
|
+
// honest "not observable" marker.
|
|
139
|
+
if (tokens == null) return null;
|
|
140
|
+
if (typeof tokens !== "number" || !Number.isFinite(tokens) || tokens < 0) {
|
|
141
|
+
throw new Error(`${label} tokens must be a finite non-negative number or null, got ${JSON.stringify(tokens)}`);
|
|
142
|
+
}
|
|
143
|
+
return tokens;
|
|
144
|
+
};
|
|
145
|
+
|
|
146
|
+
const normCreations = Object.freeze(
|
|
147
|
+
primerCacheCreations.map((c, i) => {
|
|
148
|
+
if (c == null || typeof c !== "object" || typeof c.model !== "string" || c.model.length === 0) {
|
|
149
|
+
throw new Error(`primerCacheCreations[${i}] must be an object with a non-empty concrete model`);
|
|
150
|
+
}
|
|
151
|
+
return Object.freeze({
|
|
152
|
+
model: c.model,
|
|
153
|
+
primerForm: c.primerForm ?? null,
|
|
154
|
+
tokens: normalizeTokens(c.tokens, `primerCacheCreations[${i}]`),
|
|
155
|
+
});
|
|
156
|
+
}),
|
|
157
|
+
);
|
|
158
|
+
const normReads = Object.freeze(
|
|
159
|
+
reviewerCacheReads.map((r, i) => {
|
|
160
|
+
if (r == null || typeof r !== "object" || typeof r.model !== "string" || r.model.length === 0) {
|
|
161
|
+
throw new Error(`reviewerCacheReads[${i}] must be an object with a non-empty concrete model`);
|
|
162
|
+
}
|
|
163
|
+
return Object.freeze({
|
|
164
|
+
model: r.model,
|
|
165
|
+
angle: r.angle ?? null,
|
|
166
|
+
tokens: normalizeTokens(r.tokens, `reviewerCacheReads[${i}]`),
|
|
167
|
+
});
|
|
168
|
+
}),
|
|
169
|
+
);
|
|
170
|
+
|
|
171
|
+
// Honesty gate (Section D): an opaque / unavailable telemetry harness can never
|
|
172
|
+
// be described as verified 1 write + N reads. cacheReuseVeracity() refuses to
|
|
173
|
+
// claim verified reuse unless usageTelemetry === "available".
|
|
174
|
+
const veracity = cacheReuseVeracity(caps);
|
|
175
|
+
|
|
176
|
+
const telemetryAvailable = caps != null && caps.usageTelemetry === "available";
|
|
177
|
+
// Verified reuse requires BOTH the capability truth AND at least one observed
|
|
178
|
+
// creation and one observed read. Without a create then a read for a
|
|
179
|
+
// multi-reviewer group, even a telemetry-capable harness has no measured
|
|
180
|
+
// reuse to report.
|
|
181
|
+
const hasMeasuredSequence =
|
|
182
|
+
normCreations.length > 0 && normReads.length > 0 && telemetryAvailable;
|
|
183
|
+
const cacheReuseVerified = hasMeasuredSequence && veracity.verified;
|
|
184
|
+
|
|
185
|
+
const creationsWithTokens = countTokens(normCreations);
|
|
186
|
+
const readsWithTokens = countTokens(normReads);
|
|
187
|
+
|
|
188
|
+
const baseReason = cacheReuseVeracity(caps).reason;
|
|
189
|
+
const veracityReason = hasMeasuredSequence
|
|
190
|
+
? baseReason ?? null
|
|
191
|
+
: (baseReason ?? "no measured create-then-read sequence observed");
|
|
192
|
+
|
|
193
|
+
return Object.freeze({
|
|
194
|
+
schemaVersion: CACHE_TELEMETRY_SCHEMA_VERSION,
|
|
195
|
+
gate: plan.gate,
|
|
196
|
+
headSha: String(plan.headSha).trim().toLowerCase(),
|
|
197
|
+
planHash: plan.planHash,
|
|
198
|
+
sharedPrefixHash: plan.sharedPrefixHash ?? null,
|
|
199
|
+
cacheBoundary: plan.requestGroups[0]?.cacheBoundary ?? CACHE_BOUNDARY_AFTER_SHARED_PREFIX,
|
|
200
|
+
capabilities: caps ? Object.freeze({ ...caps }) : null,
|
|
201
|
+
telemetryAvailable,
|
|
202
|
+
cacheReuseVerified,
|
|
203
|
+
veracityReason,
|
|
204
|
+
// "before" side
|
|
205
|
+
primerCacheCreations: normCreations,
|
|
206
|
+
creationCount: normCreations.length,
|
|
207
|
+
creationTokens: sumTokens(normCreations),
|
|
208
|
+
// "after" side
|
|
209
|
+
reviewerCacheReads: normReads,
|
|
210
|
+
readCount: normReads.length,
|
|
211
|
+
readTokens: sumTokens(normReads),
|
|
212
|
+
// aggregate read:create report
|
|
213
|
+
aggregate: Object.freeze({
|
|
214
|
+
creates: normCreations.length,
|
|
215
|
+
reads: normReads.length,
|
|
216
|
+
readToCreateRatio: normCreations.length > 0 ? normReads.length / normCreations.length : 0,
|
|
217
|
+
creationsWithTokens,
|
|
218
|
+
readsWithTokens,
|
|
219
|
+
measured: cacheReuseVerified,
|
|
220
|
+
report:
|
|
221
|
+
cacheReuseVerified
|
|
222
|
+
? `verified ${normReads.length} cache read${normReads.length === 1 ? "" : "s"} after ${normCreations.length} cache creation${normCreations.length === 1 ? "" : "s"} (measurable read:create = ${normReads.length}:${normCreations.length})`
|
|
223
|
+
: `provider reuse could not be verified (usageTelemetry=${caps?.usageTelemetry ?? "missing"}) — only ordering + request-fingerprint invariants may be claimed`,
|
|
224
|
+
}),
|
|
225
|
+
});
|
|
226
|
+
}
|
|
227
|
+
|
|
228
|
+
/**
|
|
229
|
+
* Fail-closed validation of cache-telemetry evidence (Section D / GATE-EXEC-
|
|
230
|
+
* CACHE-TELEMETRY). The honesty invariant: no code path may describe an opaque
|
|
231
|
+
* harness's behaviour as a verified `1 write + N reads` result.
|
|
232
|
+
*
|
|
233
|
+
* @param {object} input
|
|
234
|
+
* @param {object} input.evidence - artifact from buildCacheTelemetryEvidence().
|
|
235
|
+
* @returns {{ ok: boolean, failures: Array<{check: string, reason: string}> }}
|
|
236
|
+
*/
|
|
237
|
+
export function validateCacheTelemetryEvidence({ evidence } = {}) {
|
|
238
|
+
const failures = [];
|
|
239
|
+
if (!evidence || typeof evidence !== "object") {
|
|
240
|
+
return { ok: false, failures: [{ check: "artifact", reason: "missing cache-telemetry evidence artifact" }] };
|
|
241
|
+
}
|
|
242
|
+
|
|
243
|
+
// Identity fields (gate/headSha/planHash/schemaVersion) must be present and
|
|
244
|
+
// well-formed for the artifact to be accepted — fail closed on a structurally
|
|
245
|
+
// incomplete artifact (missing/dropped identity fields) even when the event
|
|
246
|
+
// arrays and aggregates are internally self-consistent. A hand-edited or
|
|
247
|
+
// truncated artifact carrying none of its round identity is not
|
|
248
|
+
// cache-telemetry evidence for any gate/head and must never be accepted by
|
|
249
|
+
// fan-in as a valid artifact.
|
|
250
|
+
const identityChecks = [
|
|
251
|
+
{ field: "gate", valid: (v) => typeof v === "string" && v.trim().length > 0 },
|
|
252
|
+
{ field: "headSha", valid: (v) => typeof v === "string" && /^[0-9a-f]{7,64}$/i.test(v.trim()) },
|
|
253
|
+
{ field: "planHash", valid: (v) => typeof v === "string" && v.trim().length > 0 },
|
|
254
|
+
{ field: "schemaVersion", valid: (v) => typeof v === "number" && Number.isInteger(v) },
|
|
255
|
+
];
|
|
256
|
+
for (const { field, valid } of identityChecks) {
|
|
257
|
+
if (!valid(evidence[field])) {
|
|
258
|
+
failures.push({
|
|
259
|
+
check: "identity_field",
|
|
260
|
+
reason: `cache-telemetry evidence missing or malformed identity field "${field}" (${JSON.stringify(
|
|
261
|
+
evidence[field],
|
|
262
|
+
)}) — a structurally incomplete artifact must fail closed`,
|
|
263
|
+
});
|
|
264
|
+
}
|
|
265
|
+
}
|
|
266
|
+
|
|
267
|
+
// Capability record must be present to reason about veracity.
|
|
268
|
+
if (!evidence.capabilities) {
|
|
269
|
+
failures.push({
|
|
270
|
+
check: "capability_record",
|
|
271
|
+
reason: "cache-telemetry evidence has no capability record — provider cache reuse cannot be classified",
|
|
272
|
+
});
|
|
273
|
+
}
|
|
274
|
+
|
|
275
|
+
// Events fields must be arrays (a malformed JSON artifact with a non-array
|
|
276
|
+
// events field would otherwise make sumTokens/countTokens below throw a raw
|
|
277
|
+
// TypeError instead of a structured fail-closed failure). Downstream the
|
|
278
|
+
// enforce call re-throws as GATE-EXEC-CACHE-TELEMETRY, but the validator's
|
|
279
|
+
// documented contract is "return failures, never throw".
|
|
280
|
+
if (!Array.isArray(evidence.primerCacheCreations)) {
|
|
281
|
+
failures.push({
|
|
282
|
+
check: "aggregate_consistency",
|
|
283
|
+
reason: `primerCacheCreations must be an array, got ${JSON.stringify(evidence.primerCacheCreations)}`,
|
|
284
|
+
});
|
|
285
|
+
}
|
|
286
|
+
if (!Array.isArray(evidence.reviewerCacheReads)) {
|
|
287
|
+
failures.push({
|
|
288
|
+
check: "aggregate_consistency",
|
|
289
|
+
reason: `reviewerCacheReads must be an array, got ${JSON.stringify(evidence.reviewerCacheReads)}`,
|
|
290
|
+
});
|
|
291
|
+
}
|
|
292
|
+
// Safe, array-only view of the events fields for every downstream consumer
|
|
293
|
+
// (length/sum/count checks). A truthy non-array field never reaches
|
|
294
|
+
// sumTokens/countTokens (`.reduce`/`.filter` would throw on it) and its
|
|
295
|
+
// `.length` is never read (a string would report its char count): the
|
|
296
|
+
// validator keeps its documented "return failures, never throw" contract
|
|
297
|
+
// even for an malformed artifact whose events field is a string/object/number.
|
|
298
|
+
const creations = Array.isArray(evidence.primerCacheCreations)
|
|
299
|
+
? evidence.primerCacheCreations
|
|
300
|
+
: [];
|
|
301
|
+
const reads = Array.isArray(evidence.reviewerCacheReads) ? evidence.reviewerCacheReads : [];
|
|
302
|
+
|
|
303
|
+
// Honesty gate: verified reuse requires the capability record's usage
|
|
304
|
+
// telemetry to be available. The verdict is re-derived from
|
|
305
|
+
// evidence.capabilities.usageTelemetry — NOT from the stored
|
|
306
|
+
// evidence.telemetryAvailable boolean — so a hand-edited / forged artifact
|
|
307
|
+
// that flips BOTH cacheReuseVerified AND telemetryAvailable to true still
|
|
308
|
+
// fails closed (the capability record is the source of truth). The builder
|
|
309
|
+
// always normalizes capabilities (usageTelemetry included), so this is
|
|
310
|
+
// re-derivable here.
|
|
311
|
+
const derivedTelemetryAvailable =
|
|
312
|
+
evidence.capabilities != null && evidence.capabilities.usageTelemetry === "available";
|
|
313
|
+
if (evidence.cacheReuseVerified && !derivedTelemetryAvailable) {
|
|
314
|
+
failures.push({
|
|
315
|
+
check: "opaque_veracity",
|
|
316
|
+
reason: `cacheReuseVerified=true but usageTelemetry=${evidence.capabilities?.usageTelemetry ?? "missing"} — an opaque/unavailable harness must never claim verified provider reuse`,
|
|
317
|
+
});
|
|
318
|
+
}
|
|
319
|
+
|
|
320
|
+
// Verified reuse requires a measured create-then-read sequence (at least one
|
|
321
|
+
// creation and at least one read for a group). A claim of verified reuse with
|
|
322
|
+
// no measured sequence is not derived from evidence.
|
|
323
|
+
if (evidence.cacheReuseVerified) {
|
|
324
|
+
if (!(evidence.creationCount > 0 && evidence.readCount > 0)) {
|
|
325
|
+
failures.push({
|
|
326
|
+
check: "measured_sequence",
|
|
327
|
+
reason: `cacheReuseVerified=true but no measured create-then-read sequence (creations=${evidence.creationCount}, reads=${evidence.readCount})`,
|
|
328
|
+
});
|
|
329
|
+
}
|
|
330
|
+
}
|
|
331
|
+
|
|
332
|
+
// Self-consistency of the aggregate report against the recorded events.
|
|
333
|
+
if (evidence.creationCount !== creations.length) {
|
|
334
|
+
failures.push({
|
|
335
|
+
check: "aggregate_consistency",
|
|
336
|
+
reason: `creationCount ${evidence.creationCount} != recorded primerCacheCreations.length ${creations.length}`,
|
|
337
|
+
});
|
|
338
|
+
}
|
|
339
|
+
if (evidence.readCount !== reads.length) {
|
|
340
|
+
failures.push({
|
|
341
|
+
check: "aggregate_consistency",
|
|
342
|
+
reason: `readCount ${evidence.readCount} != recorded reviewerCacheReads.length ${reads.length}`,
|
|
343
|
+
});
|
|
344
|
+
}
|
|
345
|
+
|
|
346
|
+
// Token aggregates must equal the sum over finite event tokens.
|
|
347
|
+
const expectedCreationTokens = sumTokens(creations);
|
|
348
|
+
if (evidence.creationTokens !== expectedCreationTokens) {
|
|
349
|
+
failures.push({
|
|
350
|
+
check: "token_aggregate",
|
|
351
|
+
reason: `creationTokens ${evidence.creationTokens} != sum of primer creations ${expectedCreationTokens}`,
|
|
352
|
+
});
|
|
353
|
+
}
|
|
354
|
+
const expectedReadTokens = sumTokens(reads);
|
|
355
|
+
if (evidence.readTokens !== expectedReadTokens) {
|
|
356
|
+
failures.push({
|
|
357
|
+
check: "token_aggregate",
|
|
358
|
+
reason: `readTokens ${evidence.readTokens} != sum of reviewer reads ${expectedReadTokens}`,
|
|
359
|
+
});
|
|
360
|
+
}
|
|
361
|
+
|
|
362
|
+
// Aggregate reads/creates must mirror the counts, and every derived aggregate
|
|
363
|
+
// value (readToCreateRatio, creationsWithTokens, readsWithTokens, measured)
|
|
364
|
+
// must be re-derivable from the recorded events — a forged / hand-edited
|
|
365
|
+
// aggregate that contradicts the evidence fails closed.
|
|
366
|
+
const record = { creates: evidence.creationCount, reads: evidence.readCount };
|
|
367
|
+
const expectedRatio = record.creates > 0 ? record.reads / record.creates : 0;
|
|
368
|
+
// The human-readable report must agree with the machine verdict: a verified
|
|
369
|
+
// report only when cacheReuseVerified is true, a could-not-verify report
|
|
370
|
+
// otherwise. This closes the over-claim surface where a hand-edited artifact
|
|
371
|
+
// keeps the numeric aggregates consistent (measured:false) while the report
|
|
372
|
+
// prose claims "verified N reads after M creations" (informational, not a
|
|
373
|
+
// gate-number bypass, but the human-facing surface must not contradict it).
|
|
374
|
+
const expectedReportVerified = /^verified .+ cache read/.test(evidence.aggregate?.report ?? "");
|
|
375
|
+
if (expectedReportVerified !== evidence.cacheReuseVerified) {
|
|
376
|
+
failures.push({
|
|
377
|
+
check: "aggregate_consistency",
|
|
378
|
+
reason: `aggregate.report prose (${JSON.stringify(evidence.aggregate?.report)}) contradicts cacheReuseVerified=${evidence.cacheReuseVerified}`,
|
|
379
|
+
});
|
|
380
|
+
}
|
|
381
|
+
if (
|
|
382
|
+
evidence.aggregate?.creates !== evidence.creationCount ||
|
|
383
|
+
evidence.aggregate?.reads !== evidence.readCount ||
|
|
384
|
+
evidence.aggregate?.readToCreateRatio !== expectedRatio ||
|
|
385
|
+
evidence.aggregate?.creationsWithTokens !== countTokens(creations) ||
|
|
386
|
+
evidence.aggregate?.readsWithTokens !== countTokens(reads) ||
|
|
387
|
+
evidence.aggregate?.measured !== evidence.cacheReuseVerified
|
|
388
|
+
) {
|
|
389
|
+
failures.push({
|
|
390
|
+
check: "aggregate_consistency",
|
|
391
|
+
reason: `aggregate report does not mirror the recorded events (creates=${evidence.aggregate?.creates}, reads=${evidence.aggregate?.reads}, readToCreateRatio=${evidence.aggregate?.readToCreateRatio}, creationsWithTokens=${evidence.aggregate?.creationsWithTokens}, readsWithTokens=${evidence.aggregate?.readsWithTokens}, measured=${evidence.aggregate?.measured}); expected creates=${record.creates}, reads=${record.reads}, ratio=${expectedRatio}, creationsWithTokens=${countTokens(creations)}, readsWithTokens=${countTokens(reads)}, measured=${evidence.cacheReuseVerified}`,
|
|
392
|
+
});
|
|
393
|
+
}
|
|
394
|
+
|
|
395
|
+
return { ok: failures.length === 0, failures };
|
|
396
|
+
}
|
|
397
|
+
|
|
398
|
+
/**
|
|
399
|
+
* Strict fail-closed enforcement surface (GATE-EXEC-CACHE-TELEMETRY): throws
|
|
400
|
+
* when cache-telemetry evidence is missing or invalid, naming the failing check.
|
|
401
|
+
* This is the refusal path a gate conductor calls after
|
|
402
|
+
* validateCacheTelemetryEvidence returns ok:false — it turns a reported failure
|
|
403
|
+
* into a hard stop.
|
|
404
|
+
*
|
|
405
|
+
* @param {object} input
|
|
406
|
+
* @param {object} input.evidence - artifact from buildCacheTelemetryEvidence().
|
|
407
|
+
* @returns {true}
|
|
408
|
+
* @throws {Error} when any cache-telemetry check fails.
|
|
409
|
+
*/
|
|
410
|
+
export function enforceCacheTelemetryEvidence({ evidence } = {}) {
|
|
411
|
+
const r = validateCacheTelemetryEvidence({ evidence });
|
|
412
|
+
if (!r.ok) {
|
|
413
|
+
throw new Error(
|
|
414
|
+
`GATE-EXEC-CACHE-TELEMETRY: cache-telemetry evidence failed validation; refusing to proceed (${r.failures.map((f) => `${f.check}: ${f.reason}`).join("; ")})`,
|
|
415
|
+
);
|
|
416
|
+
}
|
|
417
|
+
return true;
|
|
418
|
+
}
|
|
419
|
+
|
|
420
|
+
/**
|
|
421
|
+
* Persist the evidence artifact to its deterministic path.
|
|
422
|
+
*
|
|
423
|
+
* @param {object} input
|
|
424
|
+
* @param {string} input.dir
|
|
425
|
+
* @param {object} input.evidence
|
|
426
|
+
* @returns {Promise<{ path: string }>}
|
|
427
|
+
*/
|
|
428
|
+
export async function writeCacheTelemetryEvidence({ dir, evidence } = {}) {
|
|
429
|
+
const target = cacheTelemetryPath({
|
|
430
|
+
dir,
|
|
431
|
+
gate: evidence.gate,
|
|
432
|
+
headSha: evidence.headSha,
|
|
433
|
+
});
|
|
434
|
+
await mkdir(path.dirname(target), { recursive: true });
|
|
435
|
+
await writeFile(target, `${JSON.stringify(evidence, null, 2)}\n`, "utf8");
|
|
436
|
+
return { path: target };
|
|
437
|
+
}
|
|
@@ -1,4 +1,5 @@
|
|
|
1
1
|
import { SUBMITTED_REVIEW_STATES, isCopilotLogin, normalizeTimestamp } from "../github/copilot-helpers.mjs";
|
|
2
|
+
import { trimmedOrNull } from "./normalize.mjs";
|
|
2
3
|
|
|
3
4
|
const ACTIVE_COPILOT_REVIEW_REQUEST_STATUSES = new Set(["requested", "already-requested"]);
|
|
4
5
|
|
|
@@ -70,7 +71,7 @@ function normalizeCommits(commits) {
|
|
|
70
71
|
sortKey: index,
|
|
71
72
|
committedAtMs: normalizeTimestamp(commit?.committedAt),
|
|
72
73
|
authorLogin: typeof commit?.authorLogin === "string" ? commit.authorLogin.trim() : "",
|
|
73
|
-
sha:
|
|
74
|
+
sha: trimmedOrNull(commit?.sha),
|
|
74
75
|
}))
|
|
75
76
|
.filter((commit) => commit.committedAtMs !== null)
|
|
76
77
|
.sort((left, right) => left.committedAtMs - right.committedAtMs || left.sortKey - right.sortKey);
|
|
@@ -22,7 +22,7 @@ export const GUARD_OVERRIDE_ENV = "DEVLOOPS_ALLOW_MAIN";
|
|
|
22
22
|
// guarded branch land while a plain commit on the same branch was refused.
|
|
23
23
|
export const GUARDED_HOOKS = Object.freeze(["pre-commit", "pre-merge-commit", "pre-push"]);
|
|
24
24
|
|
|
25
|
-
const REFUSAL_BODY = (what, branchExpr) => ` echo "dev-loops: refusing to ${what} ($${branchExpr}) from this checkout." >&2
|
|
25
|
+
const REFUSAL_BODY = (what, branchExpr) => ` echo "dev-loops: WORKTREE-DEFAULT-BRANCH-GUARD refusing to ${what} ($${branchExpr}) from this checkout." >&2
|
|
26
26
|
echo " The dev-loop works in a linked worktree; a cwd that silently reset to the" >&2
|
|
27
27
|
echo " primary checkout is the usual cause. Re-run from the worktree, addressing it" >&2
|
|
28
28
|
echo " explicitly (git -C <absolute-worktree-path> ...)." >&2
|
|
@@ -91,12 +91,45 @@ export function renderGuardHook(hookName, defaultBranches = null, explicitBranch
|
|
|
91
91
|
}
|
|
92
92
|
const defaults = branches.join(" ");
|
|
93
93
|
const explicitDefaults = explicits.join(" ");
|
|
94
|
+
// pre-commit ONLY: runs the fail-closed secret scan BEFORE anything else in
|
|
95
|
+
// this hook — before the override-env check below (that override is a
|
|
96
|
+
// default-BRANCH escape hatch for a sanctioned release; it must never
|
|
97
|
+
// double as a secret-scan bypass, there is none) and before the
|
|
98
|
+
// unresolved-defaults early exit (a secret-scan must never go quiet just
|
|
99
|
+
// because branch resolution failed at install time). Git always invokes a
|
|
100
|
+
// hook with the working tree ROOT as its cwd (githooks(5)), so a plain
|
|
101
|
+
// cwd-relative path here already resolves against each worktree's OWN
|
|
102
|
+
// checked-out copy of the scanner — no --show-toplevel round trip needed.
|
|
103
|
+
// A missing scanner file is a checkout that predates this feature (nothing
|
|
104
|
+
// to run), not a bypass attempt, so it is skipped rather than blocked; a
|
|
105
|
+
// scanner that IS present and errors, or finds a hit, always blocks — see
|
|
106
|
+
// scripts/security/scan-staged-diff.mjs. Its stdout carries only the clean
|
|
107
|
+
// "{ok:true,hits:[]}" success payload (a hit's diagnostics go to STDERR,
|
|
108
|
+
// per the CLI's own `payload.ok ? stdout : stderr` split above) — silenced
|
|
109
|
+
// here so a normal commit prints no JSON noise, while a blocked commit's
|
|
110
|
+
// stderr (file/line/detector-class) still reaches the developer untouched.
|
|
111
|
+
// Shell var deliberately named "scan_cli" rather than e.g. "secret_scanner":
|
|
112
|
+
// this repo's own pre-commit hook (below) runs a heuristic sink-pattern
|
|
113
|
+
// detector that fires on a credential-shaped variable name feeding any
|
|
114
|
+
// stdout redirect, even one that discards output; naming it around a
|
|
115
|
+
// scanner CLI path rather than a credential keeps this file's own diff
|
|
116
|
+
// clean through its own hook.
|
|
117
|
+
const secretScanBlock = hookName === "pre-commit"
|
|
118
|
+
? `scan_cli="scripts/security/scan-staged-diff.mjs"
|
|
119
|
+
if [ -f "$scan_cli" ]; then
|
|
120
|
+
node "$scan_cli" >/dev/null
|
|
121
|
+
if [ "$?" != "0" ]; then
|
|
122
|
+
exit 1
|
|
123
|
+
fi
|
|
124
|
+
fi
|
|
125
|
+
`
|
|
126
|
+
: "";
|
|
94
127
|
const header = `#!/bin/sh
|
|
95
128
|
# ${GUARD_MARKER}
|
|
96
129
|
# Refuses a ${hookName} that would land on a guarded default branch. Installed
|
|
97
130
|
# in the common hook directory, so linked worktrees run it too — their branch
|
|
98
131
|
# is not one of the guarded ones, which is what lets their work through.
|
|
99
|
-
if [ "\${${GUARD_OVERRIDE_ENV}}" = "1" ]; then
|
|
132
|
+
${secretScanBlock}if [ "\${${GUARD_OVERRIDE_ENV}}" = "1" ]; then
|
|
100
133
|
exit 0
|
|
101
134
|
fi
|
|
102
135
|
defaults="${defaults}"
|
|
@@ -57,7 +57,10 @@ import { ALWAYS_INCLUDE, CATEGORY_ANGLE_MAP } from "../analysis/change-classifie
|
|
|
57
57
|
* @type {Record<string, string[]>}
|
|
58
58
|
*/
|
|
59
59
|
const KIND_TO_CATEGORIES = {
|
|
60
|
-
|
|
60
|
+
// #1442: a docs file is PROSE_PRESENT when it lands on the prose surface, so
|
|
61
|
+
// deslop's carry-forward surface is the `docs` kind (a non-docs delta never
|
|
62
|
+
// re-runs a clean deslop verdict).
|
|
63
|
+
docs: ["DOCS_ONLY", "PROSE_PRESENT"],
|
|
61
64
|
config: ["CONFIG_ONLY"],
|
|
62
65
|
test: ["TEST_ONLY"],
|
|
63
66
|
ci: ["CI_ONLY"],
|
|
@@ -108,6 +111,14 @@ const ANGLE_SURFACE_KINDS = (() => {
|
|
|
108
111
|
|
|
109
112
|
/**
|
|
110
113
|
* Resolve an angle's declared review surface (the pure angle -> surface mapping).
|
|
114
|
+
* `angle` is matched trim+lowercase against ALL THREE lookups below (the
|
|
115
|
+
* hardcoded ALWAYS_INCLUDE set, the caller-supplied `alwaysRerun` set, and the
|
|
116
|
+
* `kinds` map) — deliberately: this makes a case/whitespace-drifted name (e.g.
|
|
117
|
+
* "Correctness", "PR-Description") resolve the SAME surface as its canonical
|
|
118
|
+
* form, rather than falling through to `{ kind: "unknown" }`. That is a
|
|
119
|
+
* decision-affecting default: a case-drifted MAPPED angle now becomes
|
|
120
|
+
* carry-forward-eligible (via `resolveAngleCarryForward`) where an unnormalized
|
|
121
|
+
* lookup would have fail-closed it as unknown.
|
|
111
122
|
*
|
|
112
123
|
* - ALWAYS_INCLUDE angles (gate-evidence, renderer-security, pr-description) plus
|
|
113
124
|
* any explicit alwaysRerun angle -> `{ kind: "always" }`. These review a surface
|
|
@@ -122,15 +133,17 @@ const ANGLE_SURFACE_KINDS = (() => {
|
|
|
122
133
|
* @returns {AngleReviewSurface}
|
|
123
134
|
*/
|
|
124
135
|
export function angleReviewSurface(angle, { alwaysRerun } = {}) {
|
|
125
|
-
|
|
136
|
+
// Normalized ONCE here (trim+lowercase) so every caller — the hardcoded
|
|
137
|
+
// ALWAYS_INCLUDE check below, the configured alwaysRerun match, and the
|
|
138
|
+
// kinds lookup — agrees on one predicate regardless of whether the caller
|
|
139
|
+
// pre-lowercases (consolidate-fanin.mjs does; write-gate-context.mjs's
|
|
140
|
+
// mandatory-angle refusal does not, and must not have to).
|
|
141
|
+
const name = typeof angle === "string" ? angle.trim().toLowerCase() : "";
|
|
126
142
|
if (name.length === 0) return { kind: "unknown" };
|
|
127
143
|
if (ALWAYS_INCLUDE.has(name)) return { kind: "always" };
|
|
128
|
-
// Case-insensitive: callers key angles as base+lowercase while configs may
|
|
129
|
-
// carry case drift ("Correctness"); normalizing HERE keeps producer and
|
|
130
|
-
// consumer on one predicate instead of each caller pre-lowercasing.
|
|
131
144
|
if (alwaysRerun) {
|
|
132
145
|
const normalized = new Set([...alwaysRerun].map((entry) => String(entry).trim().toLowerCase()));
|
|
133
|
-
if (normalized.has(name
|
|
146
|
+
if (normalized.has(name)) return { kind: "always" };
|
|
134
147
|
}
|
|
135
148
|
const kinds = ANGLE_SURFACE_KINDS.get(name);
|
|
136
149
|
if (!kinds || kinds.size === 0) return { kind: "unknown" };
|