@bli-cockpit/cli 0.2.250 → 0.2.252

package/dist/adapters/raw-evidence-prepare-memo.js

@@ -13,11 +13,14 @@
  *
  *   - the file's size, inode and mtime/ctime (nanoseconds) are exactly what they were when the
  *     entry was written, so the file has not been appended to, rewritten or replaced;
- *   - the recipe is the same: collector version (slim rules and secret rules ship with it), the
- *     machine's image-handling mode, and whether an image step ran at all;
- *   - the entry is not `volatile`: a transcript that still held an image with no stored
- *     description (or whose slim step failed) can change shape on the next tick without the file
- *     changing, so it is always rebuilt.
+ *   - the recipe is the same: a fingerprint of the rules that decide the bytes (the slim rules, the
+ *     secret rules, the image marker format; NOT the CLI version, BLI-22794), the machine's
+ *     image-handling mode, and whether an image step ran at all;
+ *   - the entry is not `volatile`, or it is volatile ONLY because of images with no stored
+ *     description (`pending_images`, BLI-22794). Such an entry is handed back to the caller, who
+ *     trusts it while none of those images has been described since; a described one rebuilds. A
+ *     volatile entry without `pending_images` (a failed slim step, or any older entry) is never
+ *     recalled: it can change shape on the next tick without the file changing.
  *
  * The caller then reuses the entry ONLY when the remembered content hash is already durable
  * remotely (`skipContentHashes`). Anything else, including every file that is not yet delivered,
@@ -25,11 +28,13 @@
  * known. A memo that is lost, corrupt, or wrong in the safe direction costs one slow tick, never
  * a missed upload. It holds ids, sizes, hashes and redaction metadata, never a byte of content.
  */
+import { createHash } from "node:crypto";
 import path from "node:path";
 import fs from "node:fs/promises";
 import { isMissingFileFailure } from "../health-detail.js";
 import { readJsonFile, writeJsonFileAtomic } from "../local-state-files.js";
-import { LOCAL_COLLECTOR_VERSION } from "../local-state-config.js";
+import { evidenceRulesFingerprint } from "@bli-cockpit/telemetry-core";
+import { SLIM_RULES_VERSION } from "../slim/slim-transcript.js";
 export const PREPARE_MEMO_FILENAME = "transcript-prepare-memo.json";
 const MEMO_FORMAT = "prepare-memo.v1";
 /** About 700 bytes an entry, so this is at most a few MB; the oldest leave first. */
@@ -54,14 +59,25 @@ export async function statSourceForMemo(filePath) {
         return null;
     }
 }
+/**
+ * What the prepared bytes depend on besides the file: the slim rules and the secret and image rules.
+ * Computed once per process. NOT the CLI version (BLI-22794): an upgrade that changes none of these
+ * keeps the memo warm. Changing a rule changes this, which rebuilds the memo once; that is intended
+ * (see `SLIM_RULES_VERSION` and telemetry-core's `evidenceRulesFingerprint`).
+ */
+let rulesFingerprint;
+export function rulesFingerprintOf(slimRulesVersion, evidenceRules) {
+    return createHash("sha256").update(`${slimRulesVersion}|${evidenceRules}`).digest("hex").slice(0, 32);
+}
 /** Everything besides the file that decides what its prepared bytes are. */
 export function prepareRecipe(options) {
-    return [MEMO_FORMAT, LOCAL_COLLECTOR_VERSION, options.kind, options.imageMode ?? "no_image_step"].join("|");
+    rulesFingerprint ??= rulesFingerprintOf(SLIM_RULES_VERSION, evidenceRulesFingerprint());
+    return [MEMO_FORMAT, rulesFingerprint, options.kind, options.imageMode ?? "no_image_step"].join("|");
 }
 export function prepareMemoPath(stateDir) {
     return path.join(stateDir, PREPARE_MEMO_FILENAME);
 }
-const EMPTY_COUNTS = { skipped_count: 0, skipped_bytes: 0, rebuilt_count: 0, rebuilt_bytes: 0, volatile_count: 0 };
+const EMPTY_COUNTS = { skipped_count: 0, skipped_bytes: 0, rebuilt_count: 0, rebuilt_bytes: 0, volatile_count: 0, volatile_reused_count: 0, volatile_reused_bytes: 0, rebuilt_image_now_described_count: 0 };
 export async function openPrepareMemo(stateDir) {
     const entries = await readMemoEntries(stateDir);
     let dirty = false;
@@ -73,7 +89,9 @@ export async function openPrepareMemo(stateDir) {
             const entry = entries.get(sourceKey);
             if (!entry)
                 return null;
-            if (entry.volatile || entry.recipe !== recipe)
+            if (entry.volatile && !entry.pending_images)
+                return null;
+            if (entry.recipe !== recipe)
                 return null;
             if (entry.size !== stamp.size || entry.identity !== stamp.identity)
                 return null;
@@ -85,6 +103,15 @@ export async function openPrepareMemo(stateDir) {
             dirty = true;
         },
         count(outcome, bytes) {
+            if (outcome === "rebuilt_image_now_described") {
+                counts.rebuilt_image_now_described_count += 1;
+                return;
+            }
+            if (outcome === "volatile_reused") {
+                counts.volatile_reused_count += 1;
+                counts.volatile_reused_bytes += bytes;
+                return;
+            }
             if (outcome === "skipped") {
                 counts.skipped_count += 1;
                 counts.skipped_bytes += bytes;
@@ -184,6 +211,9 @@ function parseEntry(value) {
         return null;
     if (typeof entry["volatile"] !== "boolean" || typeof entry["seen_ms"] !== "number")
         return null;
+    const pending = entry["pending_images"];
+    if (pending !== undefined && (!Array.isArray(pending) || !pending.every((item) => typeof item === "string")))
+        return null;
     const redacted = entry["redacted"];
     if (redacted !== null && (typeof redacted !== "object" || typeof redacted.completenessLabel !== "string"))
         return null;

package/dist/adapters/raw-evidence-transcript-file.js

@@ -26,6 +26,7 @@ import { mightHoldEmbeddedImages } from "@bli-cockpit/telemetry-core";
 import { markBudgetCapApplied, markCapApplied, } from "./raw-evidence-completeness.js";
 import { evidenceEntry } from "./raw-evidence-manifest.js";
 import { prepareRecipe, statSourceForMemo, } from "./raw-evidence-prepare-memo.js";
+import { logVolatileReuse, trustVolatileEntry } from "./raw-evidence-volatile-reuse.js";
 import { describeSnapshot, stubReplacementClaim, supersedeClaimFor, } from "./raw-evidence-snapshot-supersede.js";
 import { admitEvidence, stageEvidenceBytes, } from "./raw-evidence-collection-state.js";
 /**
@@ -63,9 +64,9 @@ export async function collectOneEvidenceFile(collection, options) {
         markCapApplied(collection, options.kind, "max_file_bytes");
         return skip("file_too_large");
     }
-    const { sanitized, contentHash, stubClaim, volatile, redactedEntry } = await prepareTranscriptForUpload(collection, options, names, read.bytes);
+    const { sanitized, contentHash, stubClaim, volatile, pendingImages, redactedEntry } = await prepareTranscriptForUpload(collection, options, names, read.bytes);
     await rememberPreparedTranscript(collection, options, names, {
-        stamp, recipe, readBytes: read.bytes.byteLength, contentHash, byteSize: sanitized.bytes.byteLength, volatile, redactedEntry,
+        stamp, recipe, readBytes: read.bytes.byteLength, contentHash, byteSize: sanitized.bytes.byteLength, volatile, pendingImages, redactedEntry,
     });
     if (collection.skipContentHashes.has(contentHash) && !stubClaim) {
         collection.reused.push({
@@ -121,7 +122,7 @@ async function prepareTranscriptForUpload(collection, options, names, bytes) {
         byteSize: sanitized.bytes.byteLength,
         alreadyDurable: collection.skipContentHashes.has(contentHash),
     });
-    return { sanitized, contentHash, stubClaim, volatile: images.volatile, redactedEntry };
+    return { sanitized, contentHash, stubClaim, volatile: images.volatile, pendingImages: images.pendingImages, redactedEntry };
 }
 /**
  * The transcript bytes after the image step. Unchanged (the very same buffer) unless this sync has
@@ -133,10 +134,10 @@ async function prepareEmbeddedImages(collection, file) {
     const session = collection.budget.images;
     const source = file.kind.startsWith("codex") ? "codex" : file.kind.startsWith("claude") ? "claude_code" : null;
     if (!session || !source)
-        return { bytes: file.bytes, volatile: false };
+        return { bytes: file.bytes, volatile: false, pendingImages: null };
     // Drop mode slims every transcript, image-bearing or not; keep mode only describes the ones that can hold an image.
     if (session.mode !== "describe_and_drop" && !mightHoldEmbeddedImages(file.bytes))
-        return { bytes: file.bytes, volatile: false };
+        return { bytes: file.bytes, volatile: false, pendingImages: null };
     try {
         const prepared = await session.prepareTranscript({
             text: file.bytes.toString("utf8"),
@@ -151,13 +152,16 @@ async function prepareEmbeddedImages(collection, file) {
             : prepared.imageCount > 0;
         const slimFailed = session.mode === "describe_and_drop" && prepared.slim === null;
         const volatile = imagesPending || slimFailed;
+        // BLI-22794: volatile ONLY because of images without a stored description (the slim step held
+        // up): the bytes change only when one of these gains one, so the memo may keep the entry.
+        const pendingImages = volatile && !slimFailed ? prepared.pendingImages : null;
         if (!prepared.changed)
-            return { bytes: file.bytes, volatile };
-        return { bytes: Buffer.from(prepared.text, "utf8"), volatile };
+            return { bytes: file.bytes, volatile, pendingImages };
+        return { bytes: Buffer.from(prepared.text, "utf8"), volatile, pendingImages };
     }
     catch (error) {
         console.error("[raw-evidence] image or slim step failed, uploading the transcript as it was", JSON.stringify({ reason: "image_step_failed", kind: file.kind, ...describeError(error) }));
-        return { bytes: file.bytes, volatile: true };
+        return { bytes: file.bytes, volatile: true, pendingImages: null };
     }
 }
 /**
@@ -178,8 +182,15 @@ async function reuseUnchangedTranscript(collection, options, names, stamp, recip
     if (options.maxFileBytes && stamp.size > options.maxFileBytes)
         return false;
     const known = memo.recall(names.sourceKey, stamp, recipe);
-    if (!known || !collection.skipContentHashes.has(known.content_hash))
+    if (!known)
         return false;
+    if (known.volatile && !trustVolatileEntry(collection, memo, known, options.kind, names.sourceKey, stamp.size))
+        return false;
+    if (!collection.skipContentHashes.has(known.content_hash)) {
+        if (known.volatile)
+            logVolatileReuse(options.kind, names.sourceKey, known.content_hash, "hash_not_durable_full_path");
+        return false;
+    }
     if (known.redacted) {
         collection.redacted.push({
             kind: options.kind,
@@ -194,6 +205,12 @@ async function reuseUnchangedTranscript(collection, options, names, stamp, recip
         content_hash_sha256: known.content_hash,
         codex_session_id: options.sessionId,
     });
+    if (known.volatile) {
+        // BLI-22794: only images without a description held it back, and none has one now.
+        memo.count("volatile_reused", stamp.size);
+        logVolatileReuse(options.kind, names.sourceKey, known.content_hash, "pending_images_still_undescribed");
+        return true;
+    }
     memo.count("skipped", stamp.size);
     return true;
 }
@@ -203,7 +220,7 @@ async function rememberPreparedTranscript(collection, options, names, built) {
     if (!memo)
         return;
     memo.count(built.volatile ? "volatile" : "rebuilt", built.readBytes);
-    if (built.volatile || !built.stamp)
+    if ((built.volatile && !built.pendingImages) || !built.stamp)
         return;
     const after = await statSourceForMemo(options.filePath);
     const stable = after !== null
@@ -212,10 +229,7 @@ async function rememberPreparedTranscript(collection, options, names, built) {
         && built.readBytes === built.stamp.size;
     if (!stable)
         return;
-    const known = memo.recall(names.sourceKey, built.stamp, built.recipe);
-    if (known?.content_hash === built.contentHash)
-        return;
-    memo.remember(names.sourceKey, {
+    const entry = {
         ...built.stamp,
         recipe: built.recipe,
         content_hash: built.contentHash,
@@ -223,8 +237,14 @@ async function rememberPreparedTranscript(collection, options, names, built) {
         redacted: built.redactedEntry
             ? { redaction: built.redactedEntry.redaction, completenessLabel: built.redactedEntry.completenessLabel }
             : null,
-        volatile: false,
-    });
+    };
+    const known = memo.recall(names.sourceKey, built.stamp, built.recipe);
+    const pending = built.volatile ? built.pendingImages : undefined;
+    if (known && known.volatile === built.volatile && known.content_hash === built.contentHash
+        && (known.pending_images ?? []).join() === (pending ?? []).join())
+        return;
+    // BLI-22794: a volatile entry is kept across ticks with the images it is waiting on.
+    memo.remember(names.sourceKey, { ...entry, volatile: built.volatile, ...(pending ? { pending_images: pending } : {}) });
 }
 function nameEvidenceFile(options) {
     const fileName = path.basename(options.filePath);

package/dist/adapters/raw-evidence-volatile-reuse.js

@@ -0,0 +1,27 @@
+/**
+ * BLI-22794: may a remembered image-pending entry stand in for a rebuild? Not when one of its
+ * images has a stored description now (the bytes would differ, counted as
+ * `rebuilt_image_now_described_count`), and not while the describer could still take one of them
+ * this tick: a skipped transcript never reaches the describer, so reusing it would starve the
+ * images. Once the tick's budget is spent, the rest are reused until the next tick.
+ */
+export function trustVolatileEntry(collection, memo, known, kind, sourceKey, bytes) {
+    const session = collection.budget.images;
+    const pending = known.pending_images ?? [];
+    if (!session)
+        return false;
+    if (pending.some((sha256) => session.isDescribed(sha256))) {
+        memo.count("rebuilt_image_now_described", bytes);
+        logVolatileReuse(kind, sourceKey, known.content_hash, "rebuilt_image_now_described");
+        return false;
+    }
+    if (session.hasDescribeRoom(pending)) {
+        logVolatileReuse(kind, sourceKey, known.content_hash, "rebuilt_describer_has_room");
+        return false;
+    }
+    return true;
+}
+export function logVolatileReuse(kind, sourceKey, contentHash, reason) {
+    const reused = reason === "pending_images_still_undescribed";
+    console.error("[raw-evidence] volatile transcript " + (reused ? "reused across ticks" : "rebuilt"), JSON.stringify({ reason, kind, source_key: sourceKey, hash_prefix: contentHash.slice(0, 12) }));
+}

package/dist/adapters/raw-evidence.js

@@ -107,15 +107,18 @@ export async function collectRawEvidencePack(context, options) {
  */
 async function reportPrepareMemo(collection) {
     const counts = await collection.budget.prepareMemo?.settle();
-    if (!counts || counts.skipped_count + counts.rebuilt_count === 0)
+    if (!counts || counts.skipped_count + counts.rebuilt_count + counts.volatile_reused_count === 0)
         return;
-    console.error(counts.skipped_count > 0 ? "[raw-evidence] slim skipped unchanged" : "[raw-evidence] no unchanged transcripts to skip", JSON.stringify({
+    console.error(counts.skipped_count + counts.volatile_reused_count > 0 ? "[raw-evidence] slim skipped unchanged" : "[raw-evidence] no unchanged transcripts to skip", JSON.stringify({
         reason: "source_unchanged_since_last_prepare",
         count: counts.skipped_count,
         bytes: counts.skipped_bytes,
         rebuilt_count: counts.rebuilt_count,
         rebuilt_bytes: counts.rebuilt_bytes,
         rebuilt_volatile_count: counts.volatile_count,
+        rebuilt_volatile_reused_count: counts.volatile_reused_count,
+        rebuilt_volatile_reused_bytes: counts.volatile_reused_bytes,
+        rebuilt_image_now_described_count: counts.rebuilt_image_now_described_count,
     }));
 }
 /**

package/dist/commands/backfill-live-window.js

@@ -0,0 +1,41 @@
+/**
+ * BLI-22790: backfill is history. A session file modified inside the live window belongs to the
+ * normal sync tick, which collects the same `DEFAULT_SINCE_MINUTES` (24 h) every ~15 minutes and
+ * already uploads these sessions (`riding_already_durable`). Re-preparing a live, still-growing
+ * session in backfill held the collection lock for 232 s per tick and kept the run "partial"
+ * forever, so doctor re-ran the backfill on every tick.
+ */
+import { DEFAULT_SINCE_MINUTES } from "../adapters/raw-evidence-collection-state.js";
+import { stubSessionKey } from "../evidence-stubs-client.js";
+export const BACKFILL_LIVE_WINDOW_MS = DEFAULT_SINCE_MINUTES * 60 * 1000;
+/** The window this run uses: none for a stub pass (it names its sessions), else the test seam or the default. */
+export function liveWindowMsFor(command) {
+    return command.stubPass ? 0 : (command.liveWindowMs ?? BACKFILL_LIVE_WINDOW_MS);
+}
+/** `onlySessions` (`codex:<id>` / `claude_code:<id>`) when set, then the live window. */
+export function scopeBackfillCandidates(candidates, options) {
+    const key = (c) => stubSessionKey(c.source, c.session_id);
+    const scoped = options.onlySessions ? candidates.filter((c) => options.onlySessions.has(key(c))) : candidates;
+    return leaveLiveSessionsToTheTick(scoped, options.live);
+}
+/**
+ * Drop the candidates the tick owns: modified within `liveWindowMs` (0 keeps all), unless the
+ * server names the session a redaction stub (BLI-11102), which is re-offered even while live.
+ * Logs the count either way.
+ */
+export function leaveLiveSessionsToTheTick(candidates, options) {
+    let liveExcluded = 0;
+    const kept = options.liveWindowMs <= 0 ? candidates : candidates.filter((candidate) => {
+        const live = options.now.getTime() - candidate.session_file_mtime_ms < options.liveWindowMs
+            && !options.stubbedSessions.has(stubSessionKey(candidate.source, candidate.session_id));
+        if (live)
+            liveExcluded += 1;
+        return !live;
+    });
+    console.error("[backfill] live sessions left to the tick", JSON.stringify({
+        reason: liveExcluded > 0 ? "live_owned_by_tick" : "no_live_sessions",
+        live_excluded: liveExcluded,
+        live_window_minutes: Math.round(options.liveWindowMs / 60_000),
+    }));
+    return { kept, liveExcluded };
+}

package/dist/commands/backfill-report.js

@@ -228,7 +228,17 @@ function countRemainingHistory(options) {
 async function writeAllHistoryCompletionMarker(ctx) {
     const { paths, cursor, sources, scopedCursor, scan, oversizedCandidateKeys, now } = ctx;
     const cappedSidecars = cappedSidecarCount(scan.candidates);
-    recordBackfillScanCoverage(cursor, sources, now, now);
+    // BLI-22790: sessions left to the tick were NOT covered here. Mark coverage only up to the
+    // start of the live window, so once they age out a later backfill still sees them.
+    const coveredThrough = scan.live_excluded_count > 0
+        ? new Date(now.getTime() - scan.live_window_ms)
+        : now;
+    console.error("[backfill] history coverage recorded", JSON.stringify({
+        reason: scan.live_excluded_count > 0 ? "covered_through_live_window_start" : "covered_through_now",
+        live_excluded: scan.live_excluded_count,
+        live_window_minutes: Math.round(scan.live_window_ms / 60_000),
+    }));
+    recordBackfillScanCoverage(cursor, sources, coveredThrough, now);
     await writeBackfillCursor(paths, cursor);
     const oversizedCandidates = scan.candidates.filter((candidate) => oversizedCandidateKeys.has(candidateCursorKey(candidate)));
     await writeBackfillCompletionMarker(paths, {

package/dist/commands/backfill-scan.js

@@ -14,6 +14,7 @@ import { scanAndAttributeClaudeSessions, } from "../adapters/claude-attribution.
 import { defaultCodexSessionDirs, scanAndAttributeCodexSessions, } from "../adapters/codex-attribution.js";
 import { emptyBackfillCursorState, prepareBackfillCursorForScope, readBackfillCursor, } from "../cursors/backfill-cursor.js";
 import { CLAUDE_CURSOR_FILENAME, readRawEvidenceCursor, } from "../cursors/raw-evidence-cursor.js";
+import { liveWindowMsFor, scopeBackfillCandidates } from "./backfill-live-window.js";
 import { getCollectorRuntimePaths } from "../local-state.js";
 import { candidateCursorKey, compareBackfillCandidates, isAfterCursor, } from "./backfill-candidates.js";
 import { addReadOnlyGuardIssues, addRepoDiscoveryIssues, backfillScanIssues, countReadOnlyGuards, deterministicSkipCandidateKeys, oversizedBackfillCandidateKeys, retryableCandidateKeys, sortScanIssuesByPriority, } from "./backfill-issues.js";
@@ -23,7 +24,6 @@ import { blockedBackfillResult } from "./backfill-result.js";
 import { emptyStubReofferPlan, loadStubReofferPlan, reopenStubbedSessions, } from "./backfill-stub-reoffer.js";
 import { loadStubCursorHashes, locateStubbedSessions } from "./backfill-stub-locate.js";
 import { resolveBackfillScope } from "./backfill-scope.js";
-import { stubSessionKey } from "../evidence-stubs-client.js";
 /**
  * SCAN: is this machine paired, what roots/sources/window apply, and what did
  * the archived Codex + Claude session stores actually contain. Returns either
@@ -149,11 +149,13 @@ async function scanBackfillSessions(options) {
         cursorHashes: options.stubCursorHashes,
         sinceMinutes: options.window.since_minutes,
     });
+    const liveWindowMs = liveWindowMsFor(options.command);
     const selection = selectBackfillCandidates({
         scanned,
         cursor: options.cursor,
         pointerlessTerminalSessions: options.pointerlessTerminalSessions,
         maxFiles: options.command.maxFiles,
+        live: { liveWindowMs, now: options.now, stubbedSessions: options.stubReoffer.sessions },
         onlySessions: options.command.stubPass
             ? new Set(options.command.stubPass.plan.sessions.keys())
             : undefined,
@@ -177,6 +179,8 @@ async function scanBackfillSessions(options) {
         issues,
         retryable_candidate_keys: retryableCandidateKeys(candidates),
         omitted_candidate_count: selection.omittedCandidateCount,
+        live_excluded_count: selection.liveExcluded,
+        live_window_ms: liveWindowMs,
     };
 }
 /**
@@ -220,9 +224,8 @@ async function scanClaudeHistory(request, projectsDir, scanLimit) {
  * so a `--max-files` cap spends its budget on them before newer history.
  */
 function selectBackfillCandidates(options) {
-    const allCandidates = options.scanned
-        .filter((candidate) => !options.onlySessions ||
-        options.onlySessions.has(stubSessionKey(candidate.source, candidate.session_id)))
+    const { kept, liveExcluded } = scopeBackfillCandidates(options.scanned, options);
+    const allCandidates = kept
         .filter((candidate) => isAfterCursor(candidate, options.cursor) ||
         isPointerlessTerminalRetry(candidate, options.pointerlessTerminalSessions))
         .sort((a, b) => compareBackfillCandidatesForRetry(a, b, options.pointerlessTerminalSessions));
@@ -230,6 +233,7 @@ function selectBackfillCandidates(options) {
     return {
         candidates,
         omittedCandidateCount: allCandidates.length - candidates.length,
+        liveExcluded,
     };
 }
 async function loadPointerlessTerminalSessions(paths, sources) {

package/dist/commands/public-root.js

@@ -21,7 +21,7 @@ export async function runCockpitCli(argv, io) {
   }
 
   if (command === "--version" || command === "-V" || command === "version") {
-    writeLine(io?.stdout ?? process.stdout, "0.2.250");
+    writeLine(io?.stdout ?? process.stdout, "0.2.252");
     return 0;
   }
 

package/dist/crash-guard.js

@@ -1,3 +1,4 @@
+import { releaseHeldSyncLocksOnExit } from "./sync-lock.js";
 /**
  * A run that dies must not be able to say it worked (BLI-4110).
  *
@@ -125,6 +126,14 @@ export function installCrashGuard(options = {}) {
         }));
     });
     target.on("exit", () => {
+        // A lock still held at exit is an orphan, whatever the exit code: release
+        // it synchronously so the next tick is not blocked (BLI-22795).
+        try {
+            releaseHeldSyncLocksOnExit();
+        }
+        catch (error) {
+            console.error("[sync-lock] exit release threw", JSON.stringify({ reason: "exit_release_threw", ...describeCrash(error) }));
+        }
         if (commandReturned)
             return;
         // `EXIT_DIED_MID_RUN` has to be in this set, and leaving it out made this

package/dist/evidence-upload-deadline.js

@@ -0,0 +1,36 @@
+/**
+ * A deadline that keeps the event loop alive (BLI-22795).
+ *
+ * `AbortSignal.timeout()` is documented to use an unref'd timer. A fetch whose
+ * socket has gone (network dropped mid-upload) then has nothing ref'd behind
+ * it: the loop drains, Node exits with the await still pending, and the
+ * crash guard logs `died_mid_run` with the collection lock still held. Here
+ * the timer is ref'd, and the race rejects even when an injected or broken
+ * fetch ignores the abort, so the caller always gets a named failure.
+ */
+export const UPLOAD_DEADLINE_ERROR_NAME = "EvidenceUploadDeadlineError";
+export function isUploadDeadlineError(error) {
+    return error instanceof Error && error.name === UPLOAD_DEADLINE_ERROR_NAME;
+}
+export async function fetchWithDeadline(fetchImpl, url, init, deadlineMs) {
+    const controller = new AbortController();
+    let timer;
+    const deadline = new Promise((_, reject) => {
+        timer = setTimeout(() => {
+            const error = new Error("upload request exceeded its deadline");
+            error.name = UPLOAD_DEADLINE_ERROR_NAME;
+            controller.abort();
+            reject(error);
+        }, deadlineMs);
+    });
+    try {
+        const pending = fetchImpl(url, { ...init, signal: controller.signal });
+        // If the deadline wins, the abandoned fetch may reject later; that must not
+        // become an unhandled rejection.
+        pending.catch(() => undefined);
+        return await Promise.race([pending, deadline]);
+    }
+    finally {
+        clearTimeout(timer);
+    }
+}

package/dist/evidence-upload-direct.js

@@ -1,4 +1,5 @@
 import { encodeEvidenceBytes } from "@bli-cockpit/telemetry-core/evidence-storage-encoding";
+import { fetchWithDeadline, isUploadDeadlineError } from "./evidence-upload-deadline.js";
 import { noteSyncProgress } from "./sync-lock.js";
 /** The signed capability is the only credential sent to Storage. Never log it,
  * its query string, provider bodies, or thrown errors that might contain it. */
@@ -13,10 +14,10 @@ export async function putSignedChunk(options, signedUrl, bytes, uploadId, index)
     for (let attempt = 1; attempt <= attempts; attempt++) {
         let status = 0;
         try {
-            const response = await options.fetchImpl(signedUrl, {
+            const response = await fetchWithDeadline(options.fetchImpl, signedUrl, {
                 method: "PUT", headers: { "Content-Type": "application/octet-stream", "x-upsert": "false" },
-                body: new Uint8Array(encoded), signal: AbortSignal.timeout(60_000), redirect: "error",
-            });
+                body: new Uint8Array(encoded), redirect: "error",
+            }, 60_000);
             status = response.status;
             await response.body?.cancel();
             if (response.ok) {
@@ -28,8 +29,8 @@ export async function putSignedChunk(options, signedUrl, bytes, uploadId, index)
             if (status < 500 && status !== 429)
                 return false;
         }
-        catch {
-            log("signed_upload_network_failure", uploadId, index, attempt, status);
+        catch (error) {
+            log(isUploadDeadlineError(error) ? "signed_upload_deadline_exceeded" : "signed_upload_network_failure", uploadId, index, attempt, status);
         }
         if (attempt < attempts)
             await (options.sleep ?? sleep)(250 * attempt);

package/dist/evidence-upload-transport.js

@@ -1,5 +1,6 @@
 import crypto from "node:crypto";
 import { describeError } from "./health-detail.js";
+import { fetchWithDeadline, isUploadDeadlineError } from "./evidence-upload-deadline.js";
 import { noteSyncProgress } from "./sync-lock.js";
 /**
  * The `send` stage: one retrying POST to an evidence-upload route, plus the
@@ -10,6 +11,7 @@ import { noteSyncProgress } from "./sync-lock.js";
  */
 const DEFAULT_MAX_ATTEMPTS = 3;
 const RETRY_DELAY_MS = 250;
+const REQUEST_DEADLINE_MS = 120_000;
 export async function requestJson(options, routePath, body) {
     const maxAttempts = options.maxAttemptsPerRequest ?? DEFAULT_MAX_ATTEMPTS;
     const sleep = options.sleep ?? defaultSleep;
@@ -17,7 +19,7 @@ export async function requestJson(options, routePath, body) {
     let lastBody = null;
     for (let attempt = 1; attempt <= maxAttempts; attempt += 1) {
         try {
-            const response = await options.fetchImpl(`${options.dashboardUrl}${routePath}`, {
+            const response = await fetchWithDeadline(options.fetchImpl, `${options.dashboardUrl}${routePath}`, {
                 method: "POST",
                 headers: {
                     "Authorization": `Bearer ${options.deviceToken}`,
@@ -26,7 +28,7 @@ export async function requestJson(options, routePath, body) {
                         ? { "x-tower-evidence-direct": "staged-gzip-v1" } : {}),
                 },
                 body: JSON.stringify(body),
-            });
+            }, REQUEST_DEADLINE_MS);
             lastStatus = response.status;
             lastBody = await readResponseJson(response);
             if (response.ok || (response.status < 500 && response.status !== 429)) {
@@ -42,7 +44,9 @@ export async function requestJson(options, routePath, body) {
             // the request left the machine. This is the exact silence BLI-2528 sat
             // behind for 57 days; the route is named, the body never is.
             console.error("[evidence-upload] request failed before a status came back", JSON.stringify({
-                reason: "upload_request_transport_error",
+                reason: isUploadDeadlineError(error)
+                    ? "upload_request_deadline_exceeded"
+                    : "upload_request_transport_error",
                 route: routePath,
                 attempt,
                 max_attempts: maxAttempts,

package/dist/image-description-session.js

@@ -10,9 +10,14 @@ import { postImageToDescribeDoor } from "./image-describe-client.js";
 import { resolveImageHandlingMode } from "./image-mode-state.js";
 import { slimOrKeep, slimTranscript } from "./slim/slim-transcript.js";
 import { imageStateKey, imageRetryDelayMs, readDescribedImageKeys, readFailedImages, writeImageDescriptionState, } from "./image-description-state.js";
-/** New images sent to the door in one sync (all worktrees together). */
+/** New images sent to the door in one sync (all worktrees together); the count never refills within a tick. */
 export const IMAGE_DESCRIBE_MAX_PER_RUN = 100;
-/** After this long in one sync, remaining images wait for the next tick. */
+/**
+ * After this long spent DESCRIBING, remaining images wait for the next tick. Only time inside the
+ * sends counts: collection reaches image-bearing transcripts minutes into a tick and keeps slimming
+ * other transcripts between sends, and a wall clock from the session open (or from the first send)
+ * spent the budget on that work, so almost nothing was described (BLI-22789: 1 of 1,874, then 10).
+ */
 export const IMAGE_DESCRIBE_MAX_RUN_MS = 120_000;
 const CONCURRENCY = 3;
 const TAG = "[image-descriptions]";
@@ -25,13 +30,14 @@ export async function openImageDescriptions(options) {
     const stateDirty = false;
     const settledThisRun = new Map();
     const now = options.now ?? Date.now;
-    const startedAt = now();
     const maxPerRun = options.maxPerRun ?? IMAGE_DESCRIBE_MAX_PER_RUN;
     const maxRunMs = options.maxRunMs ?? IMAGE_DESCRIBE_MAX_RUN_MS;
     const sentThisRun = 0;
-    const run = { options, mode, fetchImpl, described, orderedKeys, failedImages, stateDirty, settledThisRun, now, startedAt, maxPerRun, maxRunMs, sentThisRun };
+    const run = { options, mode, fetchImpl, described, orderedKeys, failedImages, stateDirty, settledThisRun, now, describingMs: 0, maxPerRun, maxRunMs, sentThisRun };
     return {
         mode,
+        isDescribed: (sha256) => described.has(imageStateKey(sha256)),
+        hasDescribeRoom: (sha256s) => describeRoomLeft(run, sha256s),
         async prepareTranscript(input) {
             const afterImages = await describeAndReplace(run, input);
             if (mode !== "describe_and_drop")
@@ -41,6 +47,19 @@ export async function openImageDescriptions(options) {
         },
     };
 }
+function describeRoomLeft(run, sha256s) {
+    if (run.sentThisRun >= run.maxPerRun)
+        return false;
+    if (run.describingMs > run.maxRunMs)
+        return false;
+    return sha256s.some((sha256) => {
+        // Already sent (or found too large) this run: it will not be sent again before the next tick.
+        if (run.settledThisRun.has(sha256))
+            return false;
+        const failure = run.failedImages[imageStateKey(sha256)];
+        return !(failure && failure.retry_after > run.now());
+    });
+}
 function markDescribed(run, sha256) {
     const { described, orderedKeys } = run;
     const key = imageStateKey(sha256);
@@ -92,12 +111,12 @@ async function describeAndReplace(run, input) {
     const { mode, described, options, orderedKeys, failedImages } = run;
     const images = findEmbeddedImages(input.text);
     if (images.length === 0)
-        return { text: input.text, imageCount: 0, uniqueCount: 0, replacedCount: 0, charsSaved: 0 };
+        return { text: input.text, pendingImages: [], imageCount: 0, uniqueCount: 0, replacedCount: 0, charsSaved: 0, deferredOnly: false };
     const unique = new Map();
     for (const image of images)
         if (!unique.has(image.sha256))
             unique.set(image.sha256, image);
-    const counts = { already: 0, described_now: 0, known: 0, too_large: 0, over_budget: 0, failed: 0, backing_off: 0 };
+    const counts = { settled_earlier: 0, already: 0, described_now: 0, known: 0, too_large: 0, over_budget: 0, failed: 0, backing_off: 0 };
     const toSend = selectImagesToDescribe(run, unique, counts);
     await sendPendingImages(run, toSend, counts, input);
     const failedReasons = countImageOutcomes(run, toSend, counts);
@@ -105,15 +124,19 @@ async function describeAndReplace(run, input) {
         await writeImageDescriptionState(options.stateDir, { described: orderedKeys, failed: failedImages });
         run.stateDirty = false;
     }
+    const deferredOnly = counts.over_budget + counts.backing_off > 0
+        && counts.described_now + counts.known + counts.failed + counts.too_large + counts.settled_earlier === 0;
     const replaced = mode === "describe_and_drop"
         ? replaceImagesWithMarkers(input.text, images, (image) => described.has(imageStateKey(image.sha256)))
         : { text: input.text, replacedCount: 0, charsSaved: 0 };
+    const pendingImages = [...unique.keys()].filter((sha256) => !described.has(imageStateKey(sha256)));
     console.error(TAG + " transcript prepared", JSON.stringify({
         mode,
         source: input.source,
         images: images.length,
         unique: unique.size,
         ...counts,
+        deferred_only: deferredOnly,
         failed_reasons: failedReasons,
         replaced: replaced.replacedCount,
         chars_saved: replaced.charsSaved,
@@ -121,10 +144,12 @@ async function describeAndReplace(run, input) {
     }));
     return {
         text: replaced.text,
+        pendingImages,
         imageCount: images.length,
         uniqueCount: unique.size,
         replacedCount: replaced.replacedCount,
         charsSaved: replaced.charsSaved,
+        deferredOnly,
     };
 }
 function selectImagesToDescribe(run, unique, counts) {
@@ -136,8 +161,10 @@ function selectImagesToDescribe(run, unique, counts) {
             continue;
         }
         const earlier = settledThisRun.get(image.sha256);
-        if (earlier)
+        if (earlier) {
+            counts.settled_earlier += 1;
             continue;
+        }
         if (image.byteSize > IMAGE_DESCRIBE_MAX_BYTES) {
             settledThisRun.set(image.sha256, "too_large");
             counts.too_large += 1;
@@ -153,12 +180,14 @@ function selectImagesToDescribe(run, unique, counts) {
     return toSend;
 }
 async function sendPendingImages(run, toSend, counts, input) {
-    const { now, maxPerRun, startedAt, maxRunMs } = run;
+    const { now, maxPerRun, maxRunMs } = run;
+    const callStartedAt = now();
     let next = 0;
     const worker = async () => {
         while (next < toSend.length) {
             const image = toSend[next++];
-            if (run.sentThisRun >= maxPerRun || now() - startedAt > maxRunMs) {
+            const timeUp = run.describingMs + (now() - callStartedAt) > maxRunMs;
+            if (run.sentThisRun >= maxPerRun || timeUp) {
                 counts.over_budget += 1;
                 continue;
             }
@@ -166,6 +195,7 @@ async function sendPendingImages(run, toSend, counts, input) {
         }
     };
     await Promise.all(Array.from({ length: Math.min(CONCURRENCY, toSend.length) }, worker));
+    run.describingMs += now() - callStartedAt;
 }
 function countImageOutcomes(run, toSend, counts) {
     const { settledThisRun } = run;

package/dist/slim/slim-transcript.js

@@ -25,6 +25,17 @@
 import { newClaudeSlimState, slimClaudeRecord } from "./claude-rules.js";
 import { newCodexSlimState, slimCodexRecord } from "./codex-rules.js";
 import { LineEdits, SeenBlocks, isJsonObject } from "./slim-support.js";
+/**
+ * The version of the slim rule set (claude-rules.ts, codex-rules.ts, claude-tool-results.ts,
+ * claude-attachments.ts, slim-support.ts and this file), BLI-22794. The local prepare memo keys what a
+ * transcript became on this value (plus the secret and image rules, see telemetry-core's
+ * `evidenceRulesFingerprint`) instead of on the CLI version.
+ *
+ * CHANGING ANY SLIM RULE MEANS BUMPING THIS CONSTANT. That rebuilds every collector's prepare memo
+ * once, which is intended: the stored bytes are no longer what the new rules would produce.
+ * `slim-rules-version.test.ts` fails when a rule changes without this moving.
+ */
+export const SLIM_RULES_VERSION = "slim-rules.v1";
 /** The text an image marker starts with; a line carrying one is a modified line (an image became a marker). */
 const IMAGE_MARKER_START = "[image sha256:";
 export function createSlimmer(kind, rulesForTests) {

package/dist/sync-lock.js

@@ -1,4 +1,5 @@
 import crypto from "node:crypto";
+import fsSync from "node:fs";
 import fs from "node:fs/promises";
 import path from "node:path";
 import { describeError } from "./health-detail.js";
@@ -44,23 +45,25 @@ export async function acquireSyncLock(paths, now = new Date(), options = {}) {
         marks: 0,
         lastMarkMs: now.getTime(),
         longestGapMs: 0,
+        lockPath,
+        token,
     };
     const existing = await readLock(lockPath);
-    if (existing && now.getTime() - existing.heartbeat_ms <= STALE_TAKEOVER_MS) {
+    if (existing && isFreshAndAlive(existing, now)) {
         logHeldByLiveOwner(existing, now);
         return { acquired: false, held_since: existing.heartbeat_at };
     }
     // Free or stale: try an exclusive create first; if another process created it
     // in the gap, re-check staleness before taking over.
-    let takeover = existing ? "stale" : null;
+    let takeover = existing ? staleReason(existing) : null;
     let owned = await tryExclusiveCreate(lockPath, token, now, held);
     if (!owned) {
         const current = await readLock(lockPath);
-        if (current && now.getTime() - current.heartbeat_ms <= STALE_TAKEOVER_MS) {
+        if (current && isFreshAndAlive(current, now)) {
             logHeldByLiveOwner(current, now);
             return { acquired: false, held_since: current.heartbeat_at };
         }
-        takeover = current ? "stale" : "unreadable";
+        takeover = current ? staleReason(current) : "unreadable";
         await writeLock(lockPath, token, now, held); // take over the stale lock
         owned = true;
     }
@@ -68,6 +71,63 @@ export async function acquireSyncLock(paths, now = new Date(), options = {}) {
     heldStates.add(held);
     return maintainOwnedLock(lockPath, token, held);
 }
+/**
+ * Does a process with this pid exist? `kill(pid, 0)` sends nothing: ESRCH means
+ * no such process, EPERM means it exists but is not ours (alive). Same on
+ * Windows. An unknown pid (-1 from an older record) is reported alive, so only
+ * the heartbeat decides for it.
+ */
+export function isProcessAlive(pid) {
+    if (!Number.isInteger(pid) || pid <= 0)
+        return true;
+    try {
+        process.kill(pid, 0);
+        return true;
+    }
+    catch (error) {
+        return error.code !== "ESRCH";
+    }
+}
+/** Blocks only when the heartbeat is fresh AND the owner process still exists. */
+function isFreshAndAlive(record, now) {
+    return now.getTime() - record.heartbeat_ms <= STALE_TAKEOVER_MS && isProcessAlive(record.pid);
+}
+/** A fresh heartbeat reaching here means the owner pid is gone; otherwise it aged out. */
+function staleReason(record) {
+    return isProcessAlive(record.pid) ? "stale" : "dead";
+}
+/**
+ * Synchronous release of every lock this process still holds, for the `exit`
+ * handler (BLI-22795): a process that ends without `release()` must not keep
+ * the next tick out until the heartbeat ages. Ownership-safe by token.
+ */
+export function releaseHeldSyncLocksOnExit() {
+    let released = 0;
+    for (const held of [...heldStates]) {
+        heldStates.delete(held);
+        let outcome = "not_ours_or_gone";
+        try {
+            if (held.lockPath && held.token) {
+                const raw = JSON.parse(fsSync.readFileSync(held.lockPath, "utf8"));
+                if (raw.token === held.token) {
+                    fsSync.rmSync(held.lockPath, { force: true });
+                    outcome = "removed";
+                    released += 1;
+                }
+            }
+        }
+        catch (error) {
+            outcome = `release_failed_${error.code ?? "unknown"}`;
+        }
+        console.error("[sync-lock] lock released on exit", JSON.stringify({
+            reason: "holder_exited_holding_lock",
+            outcome,
+            role: held.role,
+            held_ms: Math.max(0, Date.now() - held.startedMs),
+        }));
+    }
+    return released;
+}
 function maintainOwnedLock(lockPath, token, held) {
     const timer = setInterval(() => {
         // Refresh only while the lock is still ours; if we were taken over, stop
@@ -132,7 +192,13 @@ function logAcquired(held, takeover, previous, now) {
         ? "[sync-lock] lock acquired"
         : "[sync-lock] took over a lock whose owner stopped answering", JSON.stringify({
         role: held.role,
-        reason: takeover === "stale" ? "heartbeat_stale" : takeover === "unreadable" ? "lock_file_unreadable" : "lock_free",
+        reason: takeover === "stale"
+            ? "heartbeat_stale"
+            : takeover === "dead"
+                ? "previous_owner_dead"
+                : takeover === "unreadable"
+                    ? "lock_file_unreadable"
+                    : "lock_free",
         previous_pid: previous && previous.pid > 0 ? previous.pid : null,
         previous_heartbeat_age_ms: previous ? Math.max(0, now.getTime() - previous.heartbeat_ms) : null,
     }));
@@ -222,7 +288,10 @@ export async function inspectSyncLock(paths, now = Date.now()) {
     const record = await readLock(path.join(paths.state_dir, LOCK_FILENAME));
     if (!record)
         return null;
-    return { ...record, held: now - record.heartbeat_ms <= STALE_TAKEOVER_MS };
+    return {
+        ...record,
+        held: now - record.heartbeat_ms <= STALE_TAKEOVER_MS && isProcessAlive(record.pid),
+    };
 }
 /**
  * Remove the lock file only if it still carries `token`: the holder this caller

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bli-cockpit/cli",
-  "version": "0.2.250",
+  "version": "0.2.252",
   "private": false,
   "type": "module",
   "bin": {
@@ -27,9 +27,9 @@
     "test": "node dist/cli.js --help && node ../../scripts/assert-public-cli-routing.mjs && node ../../scripts/assert-public-cli-verb-help.mjs && node ../../scripts/assert-public-cli-surface-doors.mjs && node ../../scripts/assert-public-cli-runtime-files.mjs && node ../../scripts/assert-public-cli-exit-contract.mjs && node ../../scripts/assert-public-cli-no-fleet-posts.mjs && node ../../scripts/assert-public-cli-doctor-json.mjs && node ../../scripts/assert-public-package-pack.mjs --workspace=@bli-cockpit/cli"
   },
   "dependencies": {
-    "@bli-cockpit/memory-mcp": "0.1.73",
-    "@bli-cockpit/mcp": "0.1.154",
-    "@bli-cockpit/telemetry-core": "0.1.80"
+    "@bli-cockpit/memory-mcp": "0.1.74",
+    "@bli-cockpit/mcp": "0.1.155",
+    "@bli-cockpit/telemetry-core": "0.1.81"
   },
   "repository": {
     "type": "git",