@bli-cockpit/cli 0.2.46 → 0.2.47

package/dist/commands/session-sync.js

@@ -143,17 +143,34 @@ function claudeAttributionReadFailureCount(scan) {
 }
 export function sourceScanRetryReason(source, scan) {
     const reasons = new Set();
-    const readFailureCount = source === "codex"
-        ? codexAttributionReadFailureCount(scan)
-        : claudeAttributionReadFailureCount(scan);
-    if (readFailureCount > 0) {
-        reasons.add(`${source}_session_store_read_failed`);
-    }
+    const failure = sourceScanFailureReason(source, scan);
+    if (failure)
+        reasons.add(failure);
+    // Kept HERE and nowhere else (BLI-3551): a repo that is not on disk is a
+    // reason to widen the next scan window, because the transcript fallback can
+    // still attribute it. It is not a reason to call this sync failed — see
+    // `sourceScanFailureReason`.
     if (scan.results.some((result) => result.reason === "repo_not_on_disk")) {
         reasons.add("repo_not_on_disk");
     }
     return reasons.size > 0 ? [...reasons].sort().join(",") : null;
 }
+/**
+ * The part of the scan outcome that is a genuine FAILURE: the session store
+ * itself could not be read, so sessions that exist were not seen.
+ *
+ * Split from {@link sourceScanRetryReason} in BLI-3551. The two used to be one
+ * function, so `repo_not_on_disk` — a label the attribution umbrella finding
+ * already established is not a defect (nothing was deleted; the transcript
+ * names a path git no longer tracks) — failed the sync on every tick for three
+ * operators. A retry hint and a failure are different claims.
+ */
+export function sourceScanFailureReason(source, scan) {
+    const readFailureCount = source === "codex"
+        ? codexAttributionReadFailureCount(scan)
+        : claudeAttributionReadFailureCount(scan);
+    return readFailureCount > 0 ? `${source}_session_store_read_failed` : null;
+}
 async function reconcileSourceScanRetry(options) {
     if (options.reason) {
         await recordSourceRetryFailure(options.paths, {
@@ -447,25 +464,47 @@ export async function runAttributedWorktreeSync(options) {
     });
     // Same conditions as before, one per line, each writing down its own reason.
     // The old version was a single boolean chain: correct, and completely mute.
-    const failureReasons = new Set();
-    const fail = (condition, reason) => {
+    //
+    // Since BLI-3551 each condition also writes down the LABEL it is classified
+    // by, beside the rendered string a person reads. The health receipt reads the
+    // label; nothing parses the sentence back apart.
+    const failureRecords = new Map();
+    const add = (record) => {
+        if (!failureRecords.has(record.rendered)) {
+            failureRecords.set(record.rendered, record);
+        }
+    };
+    const fail = (condition, label, rendered = label) => {
         if (condition)
-            failureReasons.add(reason);
+            add({ label, rendered });
     };
     for (const { worktree, sync } of outcomes) {
         if (sync.status !== "uploaded") {
             // The spooled reason is the most specific thing anyone has, so lead with
             // it and name the worktree it belongs to — a fleet failure is usually one
             // repo, and "which one" is the first question asked.
-            failureReasons.add(sync.status === "spooled" && sync.failure_reason
-                ? `${worktree.worktree_label}:${sync.failure_reason}`
-                : `${worktree.worktree_label}:upload_${sync.status}`);
+            add(sync.status === "spooled" && sync.failure_reason
+                ? {
+                    label: sync.failure_class,
+                    rendered: `${worktree.worktree_label}:${sync.failure_reason}`,
+                    http_status: sync.failure_http_status,
+                }
+                : {
+                    label: "upload_not_completed",
+                    rendered: `${worktree.worktree_label}:upload_${sync.status}`,
+                });
         }
         for (const reason of sync.raw_evidence_failure_reasons ?? []) {
-            failureReasons.add(`raw_evidence:${reason}`);
+            add({
+                label: "raw_evidence_upload_failed",
+                rendered: `raw_evidence:${reason}`,
+            });
         }
         for (const reason of sync.raw_evidence_retry_reasons ?? []) {
-            failureReasons.add(`raw_evidence_retry:${reason}`);
+            add({
+                label: "raw_evidence_retry_required",
+                rendered: `raw_evidence_retry:${reason}`,
+            });
         }
         fail(sync.raw_evidence_deferred_byte_budget > 0, "deferred_byte_budget");
         fail(sync.raw_evidence_deferred_object_budget > 0, "deferred_object_budget");
@@ -474,28 +513,110 @@ export async function runAttributedWorktreeSync(options) {
     fail(claudeAttribution.session_limit_applied, "claude_session_limit_applied");
     fail(codexAttributionReadFailureCount(codexAttribution) > 0, "codex_session_read_failed");
     fail(claudeAttributionReadFailureCount(claudeAttribution) > 0, "claude_session_read_failed");
-    const codexScanRetry = sourceScanRetryReason("codex", codexAttribution);
-    if (codexScanRetry)
-        failureReasons.add(`codex_scan:${codexScanRetry}`);
-    const claudeScanRetry = sourceScanRetryReason("claude_code", claudeAttribution);
-    if (claudeScanRetry)
-        failureReasons.add(`claude_scan:${claudeScanRetry}`);
-    fail(reportRequired && !report.posted, `session_report_unposted:${report.reason ?? "unknown"}`);
+    // BLI-3551: the scan's RETRY reason and the scan's FAILURE reason are two
+    // different questions, and answering both with one function is what put
+    // `claude_scan:repo_not_on_disk` on every tick of three machines. A repo that
+    // is not on disk is a label on the session (the attribution umbrella finding:
+    // nothing was deleted, the transcript simply names a path git no longer
+    // knows). It still widens the next scan window; it is not a failed sync.
+    const codexScanFailure = sourceScanFailureReason("codex", codexAttribution);
+    if (codexScanFailure) {
+        add({ label: "codex_scan_read_failed", rendered: `codex_scan:${codexScanFailure}` });
+    }
+    const claudeScanFailure = sourceScanFailureReason("claude_code", claudeAttribution);
+    if (claudeScanFailure) {
+        add({
+            label: "claude_scan_read_failed",
+            rendered: `claude_scan:${claudeScanFailure}`,
+        });
+    }
+    // BLI-3551: a tick that observed only sessions from outside the operator's
+    // approved roots has nothing to post, and that is the consent boundary
+    // working — not a failure. It used to fail as
+    // `session_report_unposted:no_successful_sync`, whose word "session" then
+    // classified as `auth_failed`; one machine reported a broken credential 377
+    // times in 38 hours while its token had eleven weeks left. The withhold
+    // decision itself is untouched (adapters/attribution-core.ts) — only what it
+    // is CALLED.
+    const sessionsOutsideRoot = sessions.filter((session) => OUTSIDE_APPROVED_ROOT_REASONS.has(session.attribution_reason)).length;
+    const nothingInRoot = nothingInRootCount({
+        sessionCount: sessions.length,
+        outsideRootCount: sessionsOutsideRoot,
+        outcomes,
+        reportPosted: report.posted,
+        reportReason: report.reason,
+    });
+    fail(reportRequired && !report.posted && nothingInRoot === null, "session_report_unposted", `session_report_unposted:${report.reason ?? "unknown"}`);
     // `ok` may already be false from the per-worktree loop above; the outcome
     // scan re-derives that, so the two agree by construction.
-    ok = ok && failureReasons.size === 0;
-    if (!ok && failureReasons.size === 0) {
-        failureReasons.add(SYNC_FAILED_WITHOUT_REASON);
+    ok = ok && failureRecords.size === 0;
+    if (!ok && failureRecords.size === 0) {
+        add({
+            label: SYNC_FAILED_WITHOUT_REASON,
+            rendered: SYNC_FAILED_WITHOUT_REASON,
+        });
+    }
+    const notice = ok && nothingInRoot !== null ? `nothing_in_root:${nothingInRoot}` : null;
+    if (notice) {
+        // The success branch says something too: this is the receipt that proves a
+        // quiet machine is a working machine, and the count is what tells a coach
+        // that someone is working entirely outside the approved boundary.
+        console.error("[session-sync] nothing to collect inside the approved roots", JSON.stringify({
+            reason: "nothing_in_root",
+            sessions_outside_root: nothingInRoot,
+            collection_root_count: collectionRoots.length,
+            next_action: "widen the approved roots (an operator decision) if this machine should be collecting here",
+        }));
     }
+    const records = [...failureRecords.values()].sort((a, b) => a.rendered.localeCompare(b.rendered));
     return {
         ok,
-        failure_reasons: [...failureReasons].sort(),
+        failure_reasons: records.map((record) => record.rendered),
+        failure_records: records,
+        notice,
+        sessions_observed: sessions.length,
+        sessions_outside_root: sessionsOutsideRoot,
         outcomes,
         codexAttribution,
         claudeAttribution,
         summary,
     };
 }
+/**
+ * Reasons attribution gives when a session's working directory is not inside
+ * any approved collection root.
+ *
+ * Exact labels, not a pattern — the same discipline the classifier now follows.
+ * `attribution-core.ts` writes both of these and nothing else means
+ * "outside the boundary".
+ */
+const OUTSIDE_APPROVED_ROOT_REASONS = new Set([
+    "cwd_outside_scanned_worktrees",
+    "no_matching_worktree_signals",
+]);
+/**
+ * How many observed sessions were outside the approved roots, when that
+ * accounts for ALL of them and nothing else went wrong — otherwise `null`.
+ *
+ * Deliberately narrow. It requires that no worktree was synced at all (so no
+ * upload could have succeeded or failed), that every session observed this tick
+ * names an outside-the-root reason, and that the unposted report is the
+ * `no_successful_sync` shape rather than a spooled report that failed to flush.
+ * Anything else keeps its failure.
+ */
+export function nothingInRootCount(options) {
+    if (options.reportPosted)
+        return null;
+    if (options.reportReason !== "no_successful_sync")
+        return null;
+    if (options.outcomes.length > 0)
+        return null;
+    if (options.sessionCount === 0)
+        return null;
+    return options.outsideRootCount === options.sessionCount
+        ? options.outsideRootCount
+        : null;
+}
 export const ATTRIBUTION_STATE_RANK = CODEX_SESSION_ATTRIBUTION_STATE_RANK;
 function normalizeCodexResult(result) {
     return {

package/dist/evidence-upload-client.js

@@ -1,6 +1,7 @@
-import { COMMIT_CRASHED_PLATFORM, RAW_EVIDENCE_UPLOAD_DEFAULT_CHUNK_BYTES, RAW_EVIDENCE_UPLOAD_MAX_FILE_BYTES, RAW_EVIDENCE_UPLOAD_MAX_OBJECTS_PER_BEGIN, RawEvidenceLegacyUploadResponseSchema, RawEvidenceRedactionMetadataSchema, RawEvidenceUploadBeginResponseSchema, RawEvidenceUploadCommitResponseSchema, isPermanentUploadFailure, } from "@bli-cockpit/telemetry-core";
+import { COMMIT_CRASHED_PLATFORM, RAW_EVIDENCE_UPLOAD_DEFAULT_CHUNK_BYTES, RAW_EVIDENCE_UPLOAD_MAX_FILE_BYTES, RAW_EVIDENCE_UPLOAD_MAX_OBJECTS_PER_BEGIN, RawEvidenceLegacyUploadResponseSchema, RawEvidenceRedactionMetadataSchema, RawEvidenceUploadBeginResponseSchema, RawEvidenceUploadCommitResponseSchema, isPermanentUploadFailure, isRekeyableUploadConflict, } from "@bli-cockpit/telemetry-core";
 import crypto from "node:crypto";
 import fs from "node:fs/promises";
+import { rekeyedEvidencePointer } from "./evidence-upload-rekey.js";
 import { describeError } from "./health-detail.js";
 const DEFAULT_MAX_ATTEMPTS = 3;
 const RETRY_DELAY_MS = 250;
@@ -182,10 +183,18 @@ export async function uploadRawEvidenceFilesChunked(options) {
                 resolvedEntries.add(entry);
                 continue;
             }
-            const outcome = await uploadOneObject(options, entry, disposition, chunkSizeBytes);
-            await reportAbandonedUpload(options, disposition, outcome);
+            // A conflict a different name can settle gets one, here, before the
+            // upload is attempted (BLI-3552). Anything else comes back unchanged and
+            // fails on its own reason inside `uploadOneObject`.
+            const resolved = await rekeyConflictedEntry(options, entry, disposition, chunkSizeBytes);
+            const outcome = await uploadOneObject(options, resolved.entry, resolved.disposition, chunkSizeBytes);
+            await reportAbandonedUpload(options, resolved.disposition, outcome);
             outcomes.push(outcome);
-            for (const duplicate of entry.duplicates) {
+            // The duplicates of a re-keyed primary were re-keyed with it: a duplicate
+            // is byte-identical and shared the primary's key, so it shares the new
+            // one too. Leaving them on the old key would point their evidence refs at
+            // somebody else's durable object.
+            for (const duplicate of resolved.entry.duplicates) {
                 outcomes.push(duplicateOutcome(outcome, duplicate));
             }
             resolvedEntries.add(entry);
@@ -281,6 +290,105 @@ function conformReasonLabel(reason) {
     const conformed = reason.replace(/[^a-z0-9_:.-]/gi, "_").slice(0, 120);
     return conformed.length > 0 ? conformed : "upload_failed_unlabelled";
 }
+/**
+ * Settle a re-keyable conflict by asking for a second name (BLI-3552).
+ *
+ * `hash_mismatch_committed_object` means the key already holds different bytes
+ * that are already durable. Re-sending is the loop; deleting the durable object
+ * to make room is destroying evidence. The third answer is to offer this
+ * content under a name derived from its own hash, which is what this does, once
+ * per object per sync.
+ *
+ * Every path that cannot get there returns the ORIGINAL pair, so the object
+ * fails on the conflict reason `begin` actually gave rather than on a label
+ * invented here. Nothing is written off permanently either way: the upload
+ * cursor only remembers successes, so the next eligible sync offers the bytes
+ * again — at delivery-backoff cadence now instead of every 15 minutes.
+ */
+async function rekeyConflictedEntry(options, entry, disposition, chunkSizeBytes) {
+    if (disposition.disposition !== "conflict" ||
+        !isRekeyableUploadConflict(disposition.reason)) {
+        return { entry, disposition };
+    }
+    const contentHashPrefix = (entry.file.pointer.content_hash_sha256 ?? "none").slice(0, 12);
+    const rekeyedPointer = rekeyedEvidencePointer(entry.file.pointer);
+    if (!rekeyedPointer) {
+        // The key already names this content and the server still says it holds
+        // something else. A rename cannot answer that; a person has to.
+        console.error("[evidence-rekey] the key already carries this content hash; leaving the conflict for a person", JSON.stringify({
+            reason: disposition.reason,
+            kind: entry.file.kind ?? "unknown",
+            content_hash_prefix: contentHashPrefix,
+            byte_size: entry.bytes.byteLength,
+        }));
+        return { entry, disposition };
+    }
+    const rekeyedEntry = {
+        ...entry,
+        file: { ...entry.file, pointer: rekeyedPointer },
+        duplicates: entry.duplicates.map((duplicate) => ({
+            ...duplicate,
+            pointer: rekeyedEvidencePointer(duplicate.pointer) ?? duplicate.pointer,
+        })),
+    };
+    const rekeyedDisposition = await beginOneObject(options, rekeyedEntry, chunkSizeBytes);
+    if (!rekeyedDisposition) {
+        return { entry, disposition };
+    }
+    console.error("[evidence-rekey] committed object holds other content; offering these bytes under their own hash", JSON.stringify({
+        reason: disposition.reason,
+        kind: entry.file.kind ?? "unknown",
+        content_hash_prefix: contentHashPrefix,
+        byte_size: entry.bytes.byteLength,
+        chunk_count: entry.chunkCount,
+        rekeyed_disposition: rekeyedDisposition.disposition,
+        rekeyed_reason: rekeyedDisposition.reason ?? "none",
+        duplicate_count: entry.duplicates.length,
+    }));
+    return { entry: rekeyedEntry, disposition: rekeyedDisposition };
+}
+/**
+ * `begin` for a single object. Null whenever the answer cannot be trusted —
+ * transport failure, an unparseable body, a duplicate or missing key, or a
+ * pointer id that is not the one we sent — and the caller then keeps the
+ * original conflict rather than acting on a guess.
+ */
+async function beginOneObject(options, entry, chunkSizeBytes) {
+    const objectKey = entry.file.pointer.object_key ?? "";
+    const response = await requestJson(options, "/api/ambient/evidence/upload/begin", {
+        schema_version: "ambient-raw-evidence-upload-begin.v1",
+        generated_at: options.generatedAt,
+        provenance: options.provenance,
+        objects: [
+            {
+                pointer: entry.file.pointer,
+                chunk_size_bytes: chunkSizeBytes,
+                chunk_count: entry.chunkCount,
+            },
+        ],
+    });
+    if (!response.ok) {
+        console.error("[evidence-rekey] begin refused the re-keyed object; keeping the original conflict", JSON.stringify({
+            reason: "rekey_begin_rejected",
+            http_status: response.status,
+            server_reason: safeFailureDetail(response.body) ?? "none",
+            kind: entry.file.kind ?? "unknown",
+        }));
+        return null;
+    }
+    const parsed = RawEvidenceUploadBeginResponseSchema.safeParse(response.body);
+    if (!parsed.success)
+        return null;
+    const dispositions = readBeginDispositions(parsed.data);
+    const disposition = dispositions?.get(objectKey);
+    if (!disposition)
+        return null;
+    if (disposition.raw_evidence_pointer_id !==
+        entry.file.pointer.raw_evidence_pointer_id) {
+        return null;
+    }
+    return disposition;
+}
 async function uploadOneObject(options, entry, disposition, chunkSizeBytes) {
     const objectKey = entry.file.pointer.object_key ?? "";
     if (disposition.disposition === "already_committed") {

package/dist/evidence-upload-rekey.js

@@ -0,0 +1,40 @@
+/** How much of the content hash goes into the name, matching the pack keys. */
+const KEY_HASH_PREFIX_LENGTH = 16;
+/**
+ * The same object under a name that carries its own content hash.
+ *
+ * `null` when a re-key cannot help and must not be attempted:
+ * - there is no object key to rewrite;
+ * - the last segment already starts with this content's hash prefix, so the
+ *   re-keyed name would be the identical string and the second `begin` would
+ *   answer the identical conflict;
+ * - the key has no `/`, which the server's namespace rules make impossible, so
+ *   rewriting it would be inventing a key rather than deriving one.
+ *
+ * The new segment is `<hash16>-<old segment>`: still inside the operator's
+ * readable namespace (so `evidenceObjectKeyBelongsToWorkContext` still passes),
+ * still inside the key charset, and still legible to a person doing an incident
+ * walk — they can see which pack file it came from.
+ */
+export function rekeyedEvidencePointer(pointer) {
+    const objectKey = pointer.object_key;
+    if (!objectKey)
+        return null;
+    const cut = objectKey.lastIndexOf("/");
+    if (cut <= 0 || cut === objectKey.length - 1)
+        return null;
+    const prefix = objectKey.slice(0, cut);
+    const segment = objectKey.slice(cut + 1);
+    // A pointer with no content hash cannot name itself, and the upload routes
+    // reject it anyway; there is nothing to re-key it to.
+    const hash = (pointer.content_hash_sha256 ?? "").slice(0, KEY_HASH_PREFIX_LENGTH);
+    if (!hash || segment.startsWith(`${hash}-`) || segment.startsWith(hash)) {
+        return null;
+    }
+    const rekeyed = `${prefix}/${hash}-${segment}`;
+    return {
+        ...pointer,
+        object_key: rekeyed,
+        raw_evidence_pointer_id: rekeyed,
+    };
+}

package/dist/log-rotation.js

@@ -0,0 +1,144 @@
+import fs from "node:fs/promises";
+import path from "node:path";
+/**
+ * BLI-3553. Measured on the reference Mac, 2026-09-04, not estimated:
+ * `sync.log` 219,866,695 bytes and `sync.err.log` 59,635,270 after two months
+ * of a 15-minute tick. The repo's logging contract says log the success branch
+ * too, which is correct and is exactly why the files grow: the cost of that
+ * rule is bounded here, not by logging less.
+ *
+ * A cap DID exist for sync.log — buried in `raw-evidence-gc.ts`, 50 MB,
+ * truncate-to-zero with no archive. It could not work: the GC is throttled to
+ * once a day and skipped outright by `COCKPIT_DISABLE_GC`, while the log grows
+ * about 8 MB an hour on an active machine, so the file spent almost all of its
+ * life several times over the cap (last GC 19 hours before the measurement
+ * above). sync.err.log had no cap at all. That copy is gone; this is the one
+ * owner, it runs every tick, and it keeps an archive instead of destroying the
+ * only record of what happened.
+ *
+ * WHY TRUNCATE-AFTER-COPY, NOT RENAME-AND-REOPEN
+ *
+ * launchd opens StandardOutPath/StandardErrorPath once and holds that fd for
+ * the life of the job (the paths are right there in the `launchctl print`
+ * capture pinned in autostart.test.ts). Renaming the file does not move the fd:
+ * the job would keep writing into the renamed — eventually unlinked — inode, so
+ * the disk would never be reclaimed and `sync.err.log` would stop receiving new
+ * lines entirely until the agent was reloaded. Only the tick can reopen it, and
+ * the tick is not the writer; launchd is.
+ *
+ * A dated StandardErrorPath plus a pruner was the other candidate and is worse:
+ * it changes the plist on a schedule, every change has to be re-registered
+ * (which on macOS means the BLI-2583 bootout/bootstrap hazard), and it makes
+ * the rendered-plist comparison in autostartStatus a moving target — the exact
+ * shape that made every Windows machine read "needs repair" in BLI-2541.
+ *
+ * So: copy the last `maxBytes` of the file aside, then `truncate(path, 0)`.
+ * Same inode, so launchd's fd stays valid and its next write lands at offset 0
+ * of a file operators can read. The tail is capped rather than copied whole, so
+ * the archives cannot themselves become the disk problem — worst case on disk
+ * is (keep + 1) x maxBytes per stream.
+ */
+export const DEFAULT_LOG_ROTATION_MAX_BYTES = 20 * 1024 * 1024;
+export const DEFAULT_LOG_ROTATION_KEEP = 3;
+/** The two files launchd writes for the scheduled tick. */
+export const ROTATED_LOG_NAMES = ["sync.log", "sync.err.log"];
+/**
+ * Caps the scheduled tick's own logs. Never throws: a rotation failure is data
+ * for the caller to log, never a reason a sync does not run.
+ */
+export async function rotateCollectorLogs(paths, options = {}) {
+    const maxBytes = options.maxBytes ?? DEFAULT_LOG_ROTATION_MAX_BYTES;
+    const keep = options.keep ?? DEFAULT_LOG_ROTATION_KEEP;
+    const names = options.names ?? ROTATED_LOG_NAMES;
+    const result = { rotated: [], failures: [] };
+    for (const name of names) {
+        const filePath = path.join(paths.state_dir, name);
+        const size = await fs
+            .stat(filePath)
+            .then((info) => (info.isFile() ? info.size : null))
+            .catch(() => null);
+        if (size === null || size <= maxBytes)
+            continue;
+        try {
+            const kept = await rotateOne(filePath, maxBytes, keep);
+            result.rotated.push({
+                name,
+                bytes_before: size,
+                bytes_kept: kept,
+                archives: keep,
+            });
+        }
+        catch (error) {
+            result.failures.push({
+                name,
+                reason: error instanceof Error && error.message
+                    ? `rotation_failed:${error.code ?? "unknown"}`
+                    : "rotation_failed:unknown",
+            });
+        }
+    }
+    return result;
+}
+async function rotateOne(filePath, maxBytes, keep) {
+    // Shift the archives down first: .2 -> .3, .1 -> .2, and drop whatever fell
+    // off the end. Done before the copy so a crash mid-rotation loses an old
+    // archive, never the live log.
+    await fs.rm(`${filePath}.${keep}`, { force: true });
+    for (let index = keep - 1; index >= 1; index -= 1) {
+        await fs
+            .rename(`${filePath}.${index}`, `${filePath}.${index + 1}`)
+            .catch(() => undefined);
+    }
+    const handle = await fs.open(filePath, "r+");
+    try {
+        const info = await handle.stat();
+        const start = Math.max(0, info.size - maxBytes);
+        const buffer = Buffer.allocUnsafe(info.size - start);
+        await handle.read(buffer, 0, buffer.length, start);
+        // Drop the partial first line so the archive never opens mid-JSON.
+        const newline = start > 0 ? buffer.indexOf(0x0a) : -1;
+        const tail = newline >= 0 ? buffer.subarray(newline + 1) : buffer;
+        await fs.writeFile(`${filePath}.1`, tail);
+        // The load-bearing line: same inode, so launchd's held fd keeps working.
+        await handle.truncate(0);
+        return tail.length;
+    }
+    finally {
+        await handle.close();
+    }
+}
+/**
+ * Rotates and says what happened, on both branches. Safe to call every tick:
+ * the size check is one stat per file and a rotation is rare.
+ */
+export async function rotateCollectorLogsBestEffort(paths, options = {}) {
+    let result = { rotated: [], failures: [] };
+    try {
+        result = await rotateCollectorLogs(paths, options);
+    }
+    catch (error) {
+        result = {
+            rotated: [],
+            failures: [
+                {
+                    name: "*",
+                    reason: `rotation_threw:${error?.code ?? "unknown"}`,
+                },
+            ],
+        };
+    }
+    // Metadata only: file names (fixed strings) and byte counts, never a path
+    // and never a line of content.
+    for (const rotated of result.rotated) {
+        console.error("[log-rotation] capped a scheduled-tick log", JSON.stringify({
+            file: rotated.name,
+            bytes_before: rotated.bytes_before,
+            bytes_kept: rotated.bytes_kept,
+            archives_kept: rotated.archives,
+        }));
+    }
+    for (const failure of result.failures) {
+        console.error("[log-rotation] could not cap a scheduled-tick log", JSON.stringify({ file: failure.name, reason: failure.reason }));
+    }
+    return result;
+}

package/dist/onboarding-roots.js

@@ -4,6 +4,23 @@ import os from "node:os";
 import path from "node:path";
 import { isSamePath, normalizeCollectionRoots, } from "./root-normalization.js";
 export const COLLECTION_ROOT_REQUIRED = "collection_root_required";
+/**
+ * No approved collection root, so there is nothing this machine may look at.
+ *
+ * A type rather than a message prefix (BLI-3551): the health-receipt classifier
+ * used to recognise this by searching the error text, which is the same habit
+ * that filed `session_report_unposted` as an auth failure. The message is
+ * unchanged — `collection_root_required: <what to do about it>` — so every
+ * operator-facing string and every test that reads one still matches; what
+ * changed is that the classifier reads the type.
+ */
+export class CollectionRootRequiredError extends Error {
+    reason = COLLECTION_ROOT_REQUIRED;
+    constructor(detail) {
+        super(`${COLLECTION_ROOT_REQUIRED}: ${detail}`);
+        this.name = "CollectionRootRequiredError";
+    }
+}
 export function homeRootConsentPrompt(homeDirInput) {
     const homeDir = path.resolve(homeDirInput ?? os.homedir());
     return `You're in your home folder (${homeDir}).\n  Sync ALL projects on this machine? Every git repo under here gets captured now and in the future.\n  This is a work machine — exclude personal projects yourself if needed.\n  Press Enter to sync everything, or answer n to name one folder instead.  [Y/n]: `;
@@ -22,7 +39,7 @@ export async function resolveOnboardingRoots(options) {
     }
     if (hasRootInput(explicitInput) && explicit.rejected.length > 0) {
         if (!options.interactive) {
-            throw new Error(`${COLLECTION_ROOT_REQUIRED}: ${rootRejectionExplanation(explicit.rejected[0], options)}`);
+            throw new CollectionRootRequiredError(rootRejectionExplanation(explicit.rejected[0], options));
         }
         explainRejectedRoots(options, withoutHomeRejections(explicit.rejected));
         const homeOptIn = await promptForHomeRootOptIn(options, explicit.rejected);
@@ -54,7 +71,7 @@ export async function resolveOnboardingRoots(options) {
             return resolvedRoots(options, [homeDir], "cwd_likely_root", true);
         }
         if (!options.interactive) {
-            throw new Error(`${COLLECTION_ROOT_REQUIRED}: ${homeRootTutorial(homeDir)}`);
+            throw new CollectionRootRequiredError(homeRootTutorial(homeDir));
         }
         const homeOptIn = await promptForHomeRootOptIn(options, [
             { input: homeDir, reason: "home_dir" },
@@ -63,7 +80,7 @@ export async function resolveOnboardingRoots(options) {
             return homeOptIn;
     }
     if (!options.interactive) {
-        throw new Error(`${COLLECTION_ROOT_REQUIRED}: ${missingCollectionRootMessage(options)}`);
+        throw new CollectionRootRequiredError(missingCollectionRootMessage(options));
     }
     const cwdRoot = likelyBliRootFromCwd(cwd);
     if (cwdRoot) {
@@ -160,7 +177,7 @@ async function promptForRoots(options, message) {
             prompt.message?.(missingCollectionRootMessage(options));
         }
     }
-    throw new Error(`${COLLECTION_ROOT_REQUIRED}: ${missingCollectionRootMessage(options)}`);
+    throw new CollectionRootRequiredError(missingCollectionRootMessage(options));
 }
 function savedRootsPrompt(roots) {
     if (roots.length === 1) {
@@ -174,7 +191,7 @@ function savedRootsPrompt(roots) {
 }
 function requirePrompt(options) {
     if (!options.prompt) {
-        throw new Error(`${COLLECTION_ROOT_REQUIRED}: interactive prompt unavailable.`);
+        throw new CollectionRootRequiredError(`interactive prompt unavailable.`);
     }
     return options.prompt;
 }
@@ -269,7 +286,7 @@ async function promptForDeclinedHomeRoot(options, homeDir) {
         break;
     }
     prompt.message?.(HOME_ROOT_NO_FOLDER_CHOSEN_MESSAGE);
-    throw new Error(`${COLLECTION_ROOT_REQUIRED}: ${HOME_ROOT_NO_FOLDER_CHOSEN_MESSAGE}`);
+    throw new CollectionRootRequiredError(HOME_ROOT_NO_FOLDER_CHOSEN_MESSAGE);
 }
 function resolveRootInputPath(input, homeDir, cwd) {
     const expanded = input === "~" || /^~[\\/]/u.test(input)