@bli-cockpit/cli 0.2.70 → 0.2.74
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/dist/commands/doctor-disk-words.js +59 -0
- package/dist/commands/doctor-mcp.js +98 -0
- package/dist/commands/doctor-pipeline.js +15 -21
- package/dist/commands/doctor.js +29 -8
- package/dist/commands/heartbeat.js +35 -0
- package/dist/commands/mcp-stdio-probe.js +267 -0
- package/dist/commands/public-root.js +1 -1
- package/dist/commands/status.js +8 -1
- package/dist/commands/sync-followups.js +64 -1
- package/dist/commands/sync.js +13 -2
- package/dist/disk-retention.js +5 -1
- package/dist/disk-usage.js +62 -9
- package/dist/evidence-redelivery-plan.js +277 -0
- package/dist/evidence-redelivery.js +280 -0
- package/dist/local-state-status.js +34 -0
- package/package.json +5 -5
|
@@ -2,7 +2,8 @@
|
|
|
2
2
|
* What the sync tick does AFTER collection's own outcome is decided and
|
|
3
3
|
* reported: keep this machine's CLI current on npm `latest` (BLI-2601), put a
|
|
4
4
|
* broken scheduler registration back (BLI-2721), keep BLI Memory registered
|
|
5
|
-
* with both agent hosts (BLI-3580),
|
|
5
|
+
* with both agent hosts (BLI-3580), re-offer staged evidence that never landed
|
|
6
|
+
* (BLI-3797), and stop the laptop filling up (BLI-3619).
|
|
6
7
|
*
|
|
7
8
|
* Split out of commands/sync.ts (BLI-3578), moved verbatim. They belong
|
|
8
9
|
* together because they share one rule, and it is the reason both are called
|
|
@@ -20,6 +21,7 @@ import { getCollectorRuntimePaths, LOCAL_COLLECTOR_VERSION, } from "../local-sta
|
|
|
20
21
|
import { runAutostartSelfHeal, } from "../autostart-self-heal.js";
|
|
21
22
|
import { runStagingPrune } from "../disk-prune.js";
|
|
22
23
|
import { runEvidenceReconcile, } from "../evidence-reconcile-client.js";
|
|
24
|
+
import { runEvidenceRedelivery, } from "../evidence-redelivery.js";
|
|
23
25
|
import { envWithNodeRuntimeOnPath, runScheduledSelfUpdate, } from "../scheduled-self-update.js";
|
|
24
26
|
/** How many reconcile batches (of up to 500 hashes each) one sync tick may spend. */
|
|
25
27
|
const RECONCILE_BATCHES_PER_TICK = 1;
|
|
@@ -215,12 +217,20 @@ function memoryInstallEvent(outcome) {
|
|
|
215
217
|
* costs one cheap local read and no network call at all, so it never competes
|
|
216
218
|
* with collection for the tick's time. A reconcile failure never blocks the
|
|
217
219
|
* prune that follows it.
|
|
220
|
+
*
|
|
221
|
+
* BLI-3797 sits between them: the reconcile has just established which staged
|
|
222
|
+
* objects the SERVER says it never received, so the redelivery drain re-offers
|
|
223
|
+
* exactly those, bounded, before the prune runs and can delete whatever landed.
|
|
224
|
+
* Neither of the two can block the prune, and none of the three can fail a tick.
|
|
218
225
|
*/
|
|
219
226
|
export async function runStagingPruneAfterSync(command, io, dashboardUrl, options = {}) {
|
|
220
227
|
const events = [];
|
|
221
228
|
const reconciled = await runReconcileFollowUp(command, io, dashboardUrl, options);
|
|
222
229
|
if (reconciled)
|
|
223
230
|
events.push(reconcileEvent(reconciled));
|
|
231
|
+
const redelivered = await runRedeliveryFollowUp(command, io, dashboardUrl, options);
|
|
232
|
+
if (redelivered)
|
|
233
|
+
events.push(redeliveryEvent(redelivered));
|
|
224
234
|
let result;
|
|
225
235
|
try {
|
|
226
236
|
result = await runStagingPrune(getCollectorRuntimePaths(command.homeDir), {
|
|
@@ -269,6 +279,59 @@ async function runReconcileFollowUp(command, io, dashboardUrl, options) {
|
|
|
269
279
|
});
|
|
270
280
|
return result.reason === "nothing_unknown" ? null : result;
|
|
271
281
|
}
|
|
282
|
+
/**
|
|
283
|
+
* BLI-3797, between the reconcile above and the prune below, and in that order
|
|
284
|
+
* for a reason: reconcile turns `unknown` into a server-backed answer, this
|
|
285
|
+
* re-offers what that answer says never landed, and the prune then deletes
|
|
286
|
+
* whatever this just made durable. Running the drain first would ask the server
|
|
287
|
+
* about hashes it is about to be told the truth about; running it after the
|
|
288
|
+
* prune would leave a tick's worth of freed cap unused.
|
|
289
|
+
*
|
|
290
|
+
* Never throws (`runEvidenceRedelivery` already never does); returns `null` for
|
|
291
|
+
* the boring, common case — nothing on this disk is undelivered — so the steady
|
|
292
|
+
* state costs no receipt, the same rule the reconcile follow-up above and the
|
|
293
|
+
* prune's daily throttle below both follow.
|
|
294
|
+
*/
|
|
295
|
+
async function runRedeliveryFollowUp(command, io, dashboardUrl, options) {
|
|
296
|
+
const result = await runEvidenceRedelivery({
|
|
297
|
+
homeDir: command.homeDir,
|
|
298
|
+
dashboardUrl,
|
|
299
|
+
env: io.env,
|
|
300
|
+
fetch: io.fetch,
|
|
301
|
+
...(options.now ? { now: options.now } : {}),
|
|
302
|
+
});
|
|
303
|
+
return result.reason === "nothing_uncommitted" ? null : result;
|
|
304
|
+
}
|
|
305
|
+
/** Counts and byte totals only; no pack id and no hash travels in a receipt. */
|
|
306
|
+
function redeliveryEvent(result) {
|
|
307
|
+
const detail = [
|
|
308
|
+
`offered ${result.offered} object(s), ${result.offered_bytes}B, across ${result.packs} pack(s)`,
|
|
309
|
+
`uploaded ${result.uploaded} (${result.uploaded_bytes}B), reused ${result.reused}, failed ${result.failed}`,
|
|
310
|
+
`held ${result.held}, deferred ${result.deferred} (${result.deferred_bytes}B)`,
|
|
311
|
+
result.failure_reasons.length > 0
|
|
312
|
+
? `failure_reasons ${result.failure_reasons.join(",")}`
|
|
313
|
+
: null,
|
|
314
|
+
]
|
|
315
|
+
.filter((part) => Boolean(part))
|
|
316
|
+
.join("; ");
|
|
317
|
+
if (result.status === "fail") {
|
|
318
|
+
return {
|
|
319
|
+
step: "evidence_redelivery",
|
|
320
|
+
status: "fail",
|
|
321
|
+
error_code: result.reason,
|
|
322
|
+
error_detail: detail,
|
|
323
|
+
};
|
|
324
|
+
}
|
|
325
|
+
if (result.status === "skipped") {
|
|
326
|
+
return {
|
|
327
|
+
step: "evidence_redelivery",
|
|
328
|
+
status: "skipped",
|
|
329
|
+
error_code: result.reason,
|
|
330
|
+
error_detail: detail,
|
|
331
|
+
};
|
|
332
|
+
}
|
|
333
|
+
return { step: "evidence_redelivery", status: "ok", error_detail: detail };
|
|
334
|
+
}
|
|
272
335
|
function prunedNothing() {
|
|
273
336
|
return {
|
|
274
337
|
status: "skipped",
|
package/dist/commands/sync.js
CHANGED
|
@@ -2,7 +2,7 @@ import { writeLine } from "./cli-io.js";
|
|
|
2
2
|
import { attributedSyncRunStatus, cursorStatusLine, displayTicketId, rawEvidenceSyncLine, shortSha, worktreeSyncRow, writeAgentSessionSummary, } from "./collection-report.js";
|
|
3
3
|
import { collectionRootConsentAliases } from "./collection-roots.js";
|
|
4
4
|
import { discoverCommandWorktrees } from "./local-discovery.js";
|
|
5
|
-
import { sendCollectorHeartbeatBestEffort, } from "./heartbeat.js";
|
|
5
|
+
import { sendCollectorHeartbeatBestEffort, readHeartbeatStagingFacts, } from "./heartbeat.js";
|
|
6
6
|
import { classifySyncFailureRecords, classifySyncHealthError, redactedSyncErrorDetail, reportInstallEventsBestEffort, } from "./install-receipts.js";
|
|
7
7
|
import { runAttributedWorktreeSync, } from "./session-sync.js";
|
|
8
8
|
import { runAutostartSelfHealAfterSync, runMemoryInstallAfterSync, runScheduledSelfUpdateAfterSync, runStagingPruneAfterSync, } from "./sync-followups.js";
|
|
@@ -98,11 +98,22 @@ export async function runSync(command, io) {
|
|
|
98
98
|
*/
|
|
99
99
|
async function sendSyncHeartbeat(command, io, dashboardUrl, facts) {
|
|
100
100
|
const roots = await resolveSyncCollectionRoots(command).catch(() => []);
|
|
101
|
+
// BLI-3797: the backlog figure rides the same check-in as the root labels, so
|
|
102
|
+
// `cockpit ops` learns about undelivered evidence on the SAME tick that proves
|
|
103
|
+
// the machine is alive. Best-effort: nulls omit the fields rather than
|
|
104
|
+
// reporting a zero nobody measured.
|
|
105
|
+
const staging = await readHeartbeatStagingFacts({
|
|
106
|
+
homeDir: command.homeDir,
|
|
107
|
+
}).catch(() => ({ bytes: null, reason: null }));
|
|
101
108
|
await sendCollectorHeartbeatBestEffort({
|
|
102
109
|
homeDir: command.homeDir,
|
|
103
110
|
dashboardUrl,
|
|
104
111
|
roots,
|
|
105
|
-
facts
|
|
112
|
+
facts: {
|
|
113
|
+
...facts,
|
|
114
|
+
stagingUncommittedBytes: staging.bytes,
|
|
115
|
+
stagingUncommittedReason: staging.reason,
|
|
116
|
+
},
|
|
106
117
|
io,
|
|
107
118
|
}).catch((error) => {
|
|
108
119
|
// The sender already swallows everything it knows about; this is the net
|
package/dist/disk-retention.js
CHANGED
|
@@ -131,7 +131,11 @@ function pushDeletion(plan, object, reason) {
|
|
|
131
131
|
function addToBucket(bucket, object) {
|
|
132
132
|
bucket.count += 1;
|
|
133
133
|
bucket.bytes += object.byte_size;
|
|
134
|
-
|
|
134
|
+
// Report the age of the BYTES, not of the last answer about them (BLI-3797).
|
|
135
|
+
// `age_ms` on a reconciled object is the age of the reconcile reply, which
|
|
136
|
+
// made `oldest_uncommitted_age_ms` read 0.94 days for a 45-day-old backlog.
|
|
137
|
+
// Deletion still keys off `age_ms` above — this bucket is report-only.
|
|
138
|
+
bucket.oldest_age_ms = Math.max(bucket.oldest_age_ms, object.disk_age_ms || object.age_ms);
|
|
135
139
|
}
|
|
136
140
|
function removeFromBucket(bucket, object) {
|
|
137
141
|
bucket.count -= 1;
|
package/dist/disk-usage.js
CHANGED
|
@@ -104,6 +104,7 @@ export async function readStagingInventory(paths, now = new Date()) {
|
|
|
104
104
|
inventory.packs.push(pack);
|
|
105
105
|
accumulatePack(inventory, pack);
|
|
106
106
|
}
|
|
107
|
+
rankUncommittedReasons(inventory);
|
|
107
108
|
return inventory;
|
|
108
109
|
}
|
|
109
110
|
/**
|
|
@@ -150,15 +151,23 @@ async function readPack(dir, packId, options) {
|
|
|
150
151
|
manifest_present: manifest !== null,
|
|
151
152
|
};
|
|
152
153
|
for (const relative of await listPackFiles(dir)) {
|
|
153
|
-
|
|
154
|
-
|
|
155
|
-
|
|
156
|
-
|
|
154
|
+
// One stat, two facts: how big the staged copy is and WHEN it was written.
|
|
155
|
+
// The mtime is the only unfalsifiable record of how long undelivered
|
|
156
|
+
// evidence has been sitting here (BLI-3797) — every other timestamp on the
|
|
157
|
+
// machine records when something last had an opinion about it.
|
|
158
|
+
const info = await fs.stat(path.join(dir, relative)).catch(() => null);
|
|
159
|
+
const size = info?.isFile() ? info.size : 0;
|
|
157
160
|
if (relative === "manifest.json") {
|
|
158
161
|
pack.manifest_bytes += size;
|
|
159
162
|
continue;
|
|
160
163
|
}
|
|
161
|
-
|
|
164
|
+
const stagedOnDiskAt = info && Number.isFinite(info.mtimeMs)
|
|
165
|
+
? new Date(info.mtimeMs).toISOString()
|
|
166
|
+
: null;
|
|
167
|
+
pack.objects.push(await classifyObject(dir, packId, relative, size, manifest, {
|
|
168
|
+
...options,
|
|
169
|
+
stagedOnDiskAt,
|
|
170
|
+
}));
|
|
162
171
|
}
|
|
163
172
|
return pack;
|
|
164
173
|
}
|
|
@@ -167,7 +176,13 @@ async function classifyObject(dir, packId, relative, size, manifest, options) {
|
|
|
167
176
|
(await hashUnnamedFile(path.join(dir, relative), size, options.budget));
|
|
168
177
|
const stagedAt = options.stagedAt.get(`${packId}/${relative}`) ?? null;
|
|
169
178
|
const committedAt = hash ? options.ledger.committed.get(hash) : undefined;
|
|
170
|
-
const base = {
|
|
179
|
+
const base = {
|
|
180
|
+
pack_id: packId,
|
|
181
|
+
relative_path: relative,
|
|
182
|
+
byte_size: size,
|
|
183
|
+
staged_on_disk_at: options.stagedOnDiskAt,
|
|
184
|
+
disk_age_ms: ageMs(options.stagedOnDiskAt, options.now),
|
|
185
|
+
};
|
|
171
186
|
if (committedAt) {
|
|
172
187
|
return {
|
|
173
188
|
...base,
|
|
@@ -359,13 +374,49 @@ function accumulatePack(inventory, pack) {
|
|
|
359
374
|
}
|
|
360
375
|
inventory.uncommitted_count += 1;
|
|
361
376
|
inventory.uncommitted_bytes += object.byte_size;
|
|
362
|
-
|
|
377
|
+
tallyUncommittedReason(inventory, object);
|
|
378
|
+
// The age of the BYTES, not of the answer about them (BLI-3797). Falls back
|
|
379
|
+
// to `decided_at` only when the file has no readable mtime, which is the
|
|
380
|
+
// one case where the disk itself cannot say.
|
|
381
|
+
const stagedAt = object.staged_on_disk_at ?? object.decided_at;
|
|
382
|
+
if (stagedAt &&
|
|
363
383
|
(!inventory.oldest_uncommitted_at ||
|
|
364
|
-
|
|
365
|
-
inventory.oldest_uncommitted_at =
|
|
384
|
+
stagedAt.localeCompare(inventory.oldest_uncommitted_at) < 0)) {
|
|
385
|
+
inventory.oldest_uncommitted_at = stagedAt;
|
|
366
386
|
}
|
|
367
387
|
}
|
|
368
388
|
}
|
|
389
|
+
/**
|
|
390
|
+
* One more object against its reason label. Kept as a mutable array rather than
|
|
391
|
+
* a Map so the inventory stays a plain serializable value — it rides `--json`
|
|
392
|
+
* output and a receipt, and a Map would silently become `{}` in both.
|
|
393
|
+
*/
|
|
394
|
+
function tallyUncommittedReason(inventory, object) {
|
|
395
|
+
const existing = inventory.uncommitted_reasons.find((tally) => tally.reason === object.reason);
|
|
396
|
+
const tally = existing ?? {
|
|
397
|
+
reason: object.reason,
|
|
398
|
+
count: 0,
|
|
399
|
+
bytes: 0,
|
|
400
|
+
oldest_disk_age_ms: 0,
|
|
401
|
+
};
|
|
402
|
+
tally.count += 1;
|
|
403
|
+
tally.bytes += object.byte_size;
|
|
404
|
+
tally.oldest_disk_age_ms = Math.max(tally.oldest_disk_age_ms, object.disk_age_ms || object.age_ms);
|
|
405
|
+
if (!existing)
|
|
406
|
+
inventory.uncommitted_reasons.push(tally);
|
|
407
|
+
}
|
|
408
|
+
/**
|
|
409
|
+
* Rank the reasons by BYTES, then name the biggest one.
|
|
410
|
+
*
|
|
411
|
+
* Bytes, not object count, because the question the ranking answers is "what is
|
|
412
|
+
* holding this laptop over its cap" — 234 objects at 1,836 MB and 17 at 81 MB
|
|
413
|
+
* are not the same finding, and counting rows would have called them close.
|
|
414
|
+
*/
|
|
415
|
+
function rankUncommittedReasons(inventory) {
|
|
416
|
+
inventory.uncommitted_reasons.sort((left, right) => right.bytes - left.bytes || right.count - left.count);
|
|
417
|
+
inventory.dominant_uncommitted_reason =
|
|
418
|
+
inventory.uncommitted_reasons[0] ?? null;
|
|
419
|
+
}
|
|
369
420
|
function emptyStagingInventory() {
|
|
370
421
|
return {
|
|
371
422
|
pack_count: 0,
|
|
@@ -376,6 +427,8 @@ function emptyStagingInventory() {
|
|
|
376
427
|
uncommitted_count: 0,
|
|
377
428
|
uncommitted_bytes: 0,
|
|
378
429
|
oldest_uncommitted_at: null,
|
|
430
|
+
uncommitted_reasons: [],
|
|
431
|
+
dominant_uncommitted_reason: null,
|
|
379
432
|
unknown_count: 0,
|
|
380
433
|
unknown_bytes: 0,
|
|
381
434
|
manifest_bytes: 0,
|
|
@@ -0,0 +1,277 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Rebuilding an upload offer for evidence that is already on this disk.
|
|
3
|
+
*
|
|
4
|
+
* BLI-3797. The live sync only ever offers what THIS tick collected: it reads
|
|
5
|
+
* a transcript, hashes it, stages a copy, uploads that copy, and moves on. If
|
|
6
|
+
* the upload never happens — the process is killed between staging and
|
|
7
|
+
* delivery, the envelope is refused, the source file grows so next tick mints a
|
|
8
|
+
* different hash into a different pack — the staged bytes are orphaned. Nothing
|
|
9
|
+
* offers them again, ever. `disk-retention.ts` will not delete them (correctly:
|
|
10
|
+
* the ledger cannot vouch for them), so they sit until the disk fills.
|
|
11
|
+
*
|
|
12
|
+
* Measured on the reference Mac 2026-09-06, not reasoned about: 251 objects and
|
|
13
|
+
* 1,916.9 MB uncommitted against a 2,048 MB cap, 234 of them (1,835.8 MB)
|
|
14
|
+
* carrying `reconciled_unknown_to_server` — the SERVER was asked and answered
|
|
15
|
+
* that it has never held those hashes. `raw-evidence-staging.json` held six
|
|
16
|
+
* delivery-attempt rows in total, so those objects had not failed delivery;
|
|
17
|
+
* they had never been offered. The oldest pack was written 2026-07-23, 45 days
|
|
18
|
+
* earlier, and `cockpit doctor`'s fix sentence said "run `cockpit sync` to
|
|
19
|
+
* deliver them" — a sentence that could not come true.
|
|
20
|
+
*
|
|
21
|
+
* The repair is possible because each pack's own `manifest.json` already
|
|
22
|
+
* carries everything an offer needs: `object_key`, `content_hash_sha256`,
|
|
23
|
+
* `byte_size`, `media_type` and `redacted_summary` per file, and the pack's
|
|
24
|
+
* `work_context_id` / `session_id` / `operator_id` / `repo_basename` / `branch`
|
|
25
|
+
* beside them. This module rebuilds a `RawEvidenceUploadFile` from those
|
|
26
|
+
* recorded facts and NEVER re-derives them: the bytes on disk were sanitized
|
|
27
|
+
* when they were staged, and re-reading the original source would both cost the
|
|
28
|
+
* IO and risk offering content the redactor of the day would have masked
|
|
29
|
+
* differently.
|
|
30
|
+
*
|
|
31
|
+
* Metadata only, as everywhere in this family: pack ids, pack-relative paths,
|
|
32
|
+
* hashes, byte counts, reason labels. Never content, never an operator path.
|
|
33
|
+
*/
|
|
34
|
+
import fs from "node:fs/promises";
|
|
35
|
+
import path from "node:path";
|
|
36
|
+
import { RAW_EVIDENCE_BUCKET, RAW_EVIDENCE_RETENTION_MODE, } from "./adapters/raw-evidence-manifest.js";
|
|
37
|
+
import { RAW_EVIDENCE_DIR } from "./disk-usage.js";
|
|
38
|
+
import { describeError } from "./health-detail.js";
|
|
39
|
+
import { LOCAL_COLLECTOR_VERSION } from "./local-state.js";
|
|
40
|
+
import { deliveryHold, } from "./raw-evidence-staging.js";
|
|
41
|
+
/**
|
|
42
|
+
* How much undelivered backlog one tick may re-offer.
|
|
43
|
+
*
|
|
44
|
+
* A tick is fifteen minutes and the backlog is measured in gigabytes, so this
|
|
45
|
+
* is a drain rate, not a ceiling on what will eventually be sent: 512 MB a tick
|
|
46
|
+
* clears the reference Mac's 1.9 GB in four ticks (one hour) without competing
|
|
47
|
+
* with the tick's own collection, which owns a separate 2 GB budget. The object
|
|
48
|
+
* count bounds the number of `begin` round-trips for the opposite case — a
|
|
49
|
+
* backlog of thousands of small objects.
|
|
50
|
+
*/
|
|
51
|
+
export const REDELIVERY_MAX_BYTES_PER_TICK = 512 * 1024 * 1024;
|
|
52
|
+
export const REDELIVERY_MAX_OBJECTS_PER_TICK = 64;
|
|
53
|
+
/**
|
|
54
|
+
* Which undelivered objects this tick will re-offer, oldest bytes first.
|
|
55
|
+
*
|
|
56
|
+
* Oldest first is deliberate: the objects that have been undelivered longest
|
|
57
|
+
* are the ones most likely to be the only remaining copy of a session whose
|
|
58
|
+
* source transcript has since been rotated away by its agent host, and they are
|
|
59
|
+
* also the ones the disk cap will start refusing to keep. A newest-first drain
|
|
60
|
+
* would leave a 45-day-old orphan permanently at the back of the queue.
|
|
61
|
+
*
|
|
62
|
+
* `unknown` objects are deliberately NOT planned. "The ledger cannot say" is
|
|
63
|
+
* `cockpit clean --reconcile`'s question, it already runs one bounded batch per
|
|
64
|
+
* tick ahead of this, and it turns each answer into `committed` or
|
|
65
|
+
* `uncommitted` — so an unknown object reaches this planner on a later tick,
|
|
66
|
+
* with the server's own answer behind it, rather than being uploaded again on a
|
|
67
|
+
* guess.
|
|
68
|
+
*/
|
|
69
|
+
export async function planEvidenceRedelivery(options) {
|
|
70
|
+
const plan = {
|
|
71
|
+
packs: [],
|
|
72
|
+
object_count: 0,
|
|
73
|
+
bytes: 0,
|
|
74
|
+
held_count: 0,
|
|
75
|
+
held_bytes: 0,
|
|
76
|
+
unplannable: {},
|
|
77
|
+
deferred_count: 0,
|
|
78
|
+
deferred_bytes: 0,
|
|
79
|
+
};
|
|
80
|
+
const maxBytes = options.maxBytes ?? REDELIVERY_MAX_BYTES_PER_TICK;
|
|
81
|
+
const maxObjects = options.maxObjects ?? REDELIVERY_MAX_OBJECTS_PER_TICK;
|
|
82
|
+
const root = path.join(options.stateDir, RAW_EVIDENCE_DIR);
|
|
83
|
+
const candidates = uncommittedOldestFirst(options.inventory);
|
|
84
|
+
const manifests = new Map();
|
|
85
|
+
const packPlans = new Map();
|
|
86
|
+
for (const object of candidates) {
|
|
87
|
+
if (!object.content_hash) {
|
|
88
|
+
countUnplannable(plan, "no_content_hash");
|
|
89
|
+
continue;
|
|
90
|
+
}
|
|
91
|
+
if (deliveryHold(options.staging, object.content_hash, options.now)) {
|
|
92
|
+
plan.held_count += 1;
|
|
93
|
+
plan.held_bytes += object.byte_size;
|
|
94
|
+
continue;
|
|
95
|
+
}
|
|
96
|
+
if (plan.object_count >= maxObjects ||
|
|
97
|
+
plan.bytes + object.byte_size > maxBytes) {
|
|
98
|
+
plan.deferred_count += 1;
|
|
99
|
+
plan.deferred_bytes += object.byte_size;
|
|
100
|
+
continue;
|
|
101
|
+
}
|
|
102
|
+
if (!manifests.has(object.pack_id)) {
|
|
103
|
+
manifests.set(object.pack_id, await readPackManifest(path.join(root, object.pack_id), object.pack_id));
|
|
104
|
+
}
|
|
105
|
+
const manifest = manifests.get(object.pack_id) ?? null;
|
|
106
|
+
if (!manifest) {
|
|
107
|
+
countUnplannable(plan, "manifest_unreadable");
|
|
108
|
+
continue;
|
|
109
|
+
}
|
|
110
|
+
const row = manifest.files.get(object.relative_path);
|
|
111
|
+
if (!row) {
|
|
112
|
+
countUnplannable(plan, "no_manifest_row");
|
|
113
|
+
continue;
|
|
114
|
+
}
|
|
115
|
+
let packPlan = packPlans.get(object.pack_id);
|
|
116
|
+
if (!packPlan) {
|
|
117
|
+
packPlan = {
|
|
118
|
+
pack_id: object.pack_id,
|
|
119
|
+
provenance: manifest.provenance,
|
|
120
|
+
files: [],
|
|
121
|
+
bytes: 0,
|
|
122
|
+
};
|
|
123
|
+
packPlans.set(object.pack_id, packPlan);
|
|
124
|
+
plan.packs.push(packPlan);
|
|
125
|
+
}
|
|
126
|
+
packPlan.files.push({
|
|
127
|
+
pointer: {
|
|
128
|
+
raw_evidence_pointer_id: row.object_key,
|
|
129
|
+
privacy_classification: "remote_durable_raw_evidence",
|
|
130
|
+
retention_policy: {
|
|
131
|
+
mode: RAW_EVIDENCE_RETENTION_MODE,
|
|
132
|
+
privacy_classification: "remote_durable_raw_evidence",
|
|
133
|
+
},
|
|
134
|
+
storage_scope: "remote_object",
|
|
135
|
+
storage_bucket: RAW_EVIDENCE_BUCKET,
|
|
136
|
+
object_key: row.object_key,
|
|
137
|
+
content_hash_sha256: row.content_hash_sha256,
|
|
138
|
+
byte_size: row.byte_size,
|
|
139
|
+
media_type: row.media_type,
|
|
140
|
+
redacted_summary: row.redacted_summary,
|
|
141
|
+
},
|
|
142
|
+
local_path: path.join(root, object.pack_id, ...object.relative_path.split("/").filter(Boolean)),
|
|
143
|
+
kind: row.kind,
|
|
144
|
+
codex_session_id: null,
|
|
145
|
+
});
|
|
146
|
+
packPlan.bytes += object.byte_size;
|
|
147
|
+
plan.object_count += 1;
|
|
148
|
+
plan.bytes += object.byte_size;
|
|
149
|
+
}
|
|
150
|
+
return plan;
|
|
151
|
+
}
|
|
152
|
+
/**
|
|
153
|
+
* Every uncommitted object, oldest staged bytes first.
|
|
154
|
+
*
|
|
155
|
+
* Sorted on `disk_age_ms` (the file's own mtime) rather than `age_ms`, because
|
|
156
|
+
* `age_ms` for a reconciled object is the age of the reconcile ANSWER — the
|
|
157
|
+
* same confusion that had 45-day-old orphans reading as one day old.
|
|
158
|
+
*/
|
|
159
|
+
function uncommittedOldestFirst(inventory) {
|
|
160
|
+
return inventory.packs
|
|
161
|
+
.flatMap((pack) => pack.objects)
|
|
162
|
+
.filter((object) => object.state === "uncommitted")
|
|
163
|
+
.sort((left, right) => (right.disk_age_ms || right.age_ms) - (left.disk_age_ms || left.age_ms));
|
|
164
|
+
}
|
|
165
|
+
function countUnplannable(plan, reason) {
|
|
166
|
+
plan.unplannable[reason] = (plan.unplannable[reason] ?? 0) + 1;
|
|
167
|
+
}
|
|
168
|
+
/**
|
|
169
|
+
* A pack's recorded provenance and file rows.
|
|
170
|
+
*
|
|
171
|
+
* Every field is read back from what the collector WROTE when it staged the
|
|
172
|
+
* pack; nothing is re-derived from the machine's current state. A pack staged
|
|
173
|
+
* six weeks ago under a work context that no longer exists still knows which
|
|
174
|
+
* repo, branch, operator and work context it belonged to, and that is the only
|
|
175
|
+
* honest label for those bytes.
|
|
176
|
+
*/
|
|
177
|
+
async function readPackManifest(dir, packId) {
|
|
178
|
+
const raw = await fs
|
|
179
|
+
.readFile(path.join(dir, "manifest.json"), "utf8")
|
|
180
|
+
.catch(() => null);
|
|
181
|
+
if (!raw)
|
|
182
|
+
return null;
|
|
183
|
+
try {
|
|
184
|
+
const parsed = JSON.parse(raw);
|
|
185
|
+
const provenance = provenanceFromManifest(parsed, packId);
|
|
186
|
+
if (!provenance)
|
|
187
|
+
return null;
|
|
188
|
+
const files = new Map();
|
|
189
|
+
for (const entry of asArray(parsed["files"])) {
|
|
190
|
+
const row = manifestFileRow(entry);
|
|
191
|
+
if (row)
|
|
192
|
+
files.set(row.relative_path, row.value);
|
|
193
|
+
}
|
|
194
|
+
return { provenance, files };
|
|
195
|
+
}
|
|
196
|
+
catch (error) {
|
|
197
|
+
console.error("[evidence-redelivery] pack manifest unreadable; its objects cannot be re-offered", JSON.stringify({
|
|
198
|
+
reason: "manifest_unreadable",
|
|
199
|
+
pack_id: packId,
|
|
200
|
+
...describeError(error),
|
|
201
|
+
}));
|
|
202
|
+
return null;
|
|
203
|
+
}
|
|
204
|
+
}
|
|
205
|
+
function provenanceFromManifest(manifest, packId) {
|
|
206
|
+
const workContextId = asText(manifest["work_context_id"]);
|
|
207
|
+
const sessionId = asText(manifest["session_id"]);
|
|
208
|
+
const operatorId = asText(manifest["operator_id"]);
|
|
209
|
+
const repo = asText(manifest["repo_basename"]) ?? asText(manifest["repo_label"]);
|
|
210
|
+
if (!workContextId || !sessionId || !operatorId || !repo) {
|
|
211
|
+
console.error("[evidence-redelivery] pack manifest names no work context; its objects cannot be re-offered", JSON.stringify({
|
|
212
|
+
reason: "manifest_provenance_incomplete",
|
|
213
|
+
pack_id: packId,
|
|
214
|
+
has_work_context: Boolean(workContextId),
|
|
215
|
+
has_session: Boolean(sessionId),
|
|
216
|
+
has_operator: Boolean(operatorId),
|
|
217
|
+
has_repo: Boolean(repo),
|
|
218
|
+
}));
|
|
219
|
+
return null;
|
|
220
|
+
}
|
|
221
|
+
const repoLabel = asText(manifest["repo_label"]);
|
|
222
|
+
const worktreeLabel = asText(manifest["worktree_label"]);
|
|
223
|
+
return {
|
|
224
|
+
capture_source: "collector_runtime",
|
|
225
|
+
// The version that is re-offering, not the one that staged: this IS a
|
|
226
|
+
// delivery by today's collector, and pretending otherwise would misreport
|
|
227
|
+
// which build put the object in the bucket.
|
|
228
|
+
capture_adapter_version: LOCAL_COLLECTOR_VERSION,
|
|
229
|
+
collector_version: LOCAL_COLLECTOR_VERSION,
|
|
230
|
+
repo,
|
|
231
|
+
// A pack staged on a detached HEAD or a deleted worktree can have no branch
|
|
232
|
+
// recorded. `unknown` is a label, not a guess at which branch it was.
|
|
233
|
+
branch: asText(manifest["branch"]) ?? "unknown",
|
|
234
|
+
...(repoLabel ? { repo_label: repoLabel } : {}),
|
|
235
|
+
...(worktreeLabel ? { worktree_label: worktreeLabel } : {}),
|
|
236
|
+
operator_id: operatorId,
|
|
237
|
+
session_id: sessionId,
|
|
238
|
+
work_context_id: workContextId,
|
|
239
|
+
};
|
|
240
|
+
}
|
|
241
|
+
function manifestFileRow(entry) {
|
|
242
|
+
if (!entry || typeof entry !== "object")
|
|
243
|
+
return null;
|
|
244
|
+
const row = entry;
|
|
245
|
+
const relativePath = asText(row["relative_path"]);
|
|
246
|
+
const objectKey = asText(row["object_key"]);
|
|
247
|
+
const hash = asText(row["content_hash_sha256"]);
|
|
248
|
+
const mediaType = asText(row["media_type"]);
|
|
249
|
+
const summary = asText(row["redacted_summary"]);
|
|
250
|
+
const byteSize = row["byte_size"];
|
|
251
|
+
if (!relativePath ||
|
|
252
|
+
!objectKey ||
|
|
253
|
+
!hash ||
|
|
254
|
+
hash.length !== 64 ||
|
|
255
|
+
!mediaType ||
|
|
256
|
+
!summary ||
|
|
257
|
+
typeof byteSize !== "number") {
|
|
258
|
+
return null;
|
|
259
|
+
}
|
|
260
|
+
return {
|
|
261
|
+
relative_path: relativePath,
|
|
262
|
+
value: {
|
|
263
|
+
object_key: objectKey,
|
|
264
|
+
content_hash_sha256: hash,
|
|
265
|
+
byte_size: byteSize,
|
|
266
|
+
media_type: mediaType,
|
|
267
|
+
redacted_summary: summary,
|
|
268
|
+
kind: asText(row["kind"]) ?? "unknown",
|
|
269
|
+
},
|
|
270
|
+
};
|
|
271
|
+
}
|
|
272
|
+
function asArray(value) {
|
|
273
|
+
return Array.isArray(value) ? value : [];
|
|
274
|
+
}
|
|
275
|
+
function asText(value) {
|
|
276
|
+
return typeof value === "string" && value.trim() ? value : null;
|
|
277
|
+
}
|