Skip to content

v0.60.32.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 03 Oct 05:06
v0.60.32.0
48ed5e8
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Fix wave 7: automatic capture stops double-storing what you remember, the maintenance sweep reads whole transcripts, Gmail commitments and timelines work again on managed brains, gbrain upgrade refuses a Bun it can't run, the contradiction judge stops excusing undated conflicts, "who is waiting on me" ages and ranks requests honestly, and 62 community pull requests land.

When two notes give different numbers for the same thing and neither note has a date, nothing tells you which came first, so that is a conflict you need to resolve. The judge kept calling those pairs a change over time instead, guessing an order from the numbers themselves (the bigger headcount must be newer) or from the kind of note (a board deck must be newer than a memo). Prompt v4 tells it plainly not to do that, and that one note saying nothing about headcount does not contradict another note that gives one. On a fresh test world it now calls 149 of 150 planted same-time conflicts contradictions (was 131), all 50 undated ones included (was 37). The price is a few more false alarms: 14 of 820 unrelated pairs (was 11) and 5 of 50 compatible pairs (was 2).

gbrain waiting had four rough edges, all found by the gbrain-evals open-loops wave. A reply that only said "Thanks!" to someone's question closed the loop although the question was still open. An FYI whose only question mark sat inside a link opened a loop as if you had asked something. A first sync of an old inbox ranked a request that had waited 30 days as no older than one from yesterday, because a loop's age started when gbrain noticed it. And the ranking moved with the wall clock, so a published order could not be reproduced. All four are fixed; --as-of pins the clock.

After upgrading Before After
Undated same-fact conflicts called a contradiction (fresh N2 seed) 37/50 50/50
All same-time conflicts called a contradiction (fresh N2 seed) 131/150 149/150
False alarms on unrelated / compatible pairs (fresh N2 seed) 11/820, 2/50 14/820, 5/50
"Thanks!" in reply to a question closes the loop loop stays open
FYI whose only ? is inside a link opens a loop after 72 h opens nothing
A 30-day-old request found on first sync ranked as new ranked 30 days old
gbrain waiting order moves with the wall clock pinnable with --as-of
"Who works at Acme?" when "Acme Labs" also exists no seed, arm silent seeds from the page titled "Acme"
A meeting note's Participants: list mentions attended
segments.join("/") in a file that defines join resolved as a caller of join unresolved
Keyword-only answer to a question no page can answer graded moderate graded weak
check-backlinks fix on a managed brain one refusal per page; dry run claims success one clear refusal, dry run too
A fact remembered, then seen again by automatic capture stored twice stored once (corrections always kept)
Sweep extraction of a 120 KB session transcript first 8,000 characters whole transcript, in windows
Gmail commitments on a managed brain never extracted extracted every sweep
gbrain upgrade to a release needing a newer Bun installs, then fails to start refuses (exit 78); autopilot holds
Remote title search, rare term, 40k pages (Postgres) ~490 ms ~34 ms
Conversation-facts backlog with source-evidence pages marked conversation_parseable: false never reaches zero drains
chronicle-backfill --limit N, run twice the same N pages again the next N pages

These came from the gbrain-evals category waves A4 (abstention), N2 (contradictions), N7 (open loops), N9 (multi-hop), N12 (meeting formats) and N13 (code intelligence), plus open issues garrytan#5341, garrytan#5330 and garrytan#5329, and the GBRA-35 lane's fixes for garrytan#5888, garrytan#5887, garrytan#5889, garrytan#5890, garrytan#5892, garrytan#5855, garrytan#5902, garrytan#5875, garrytan#5867, garrytan#5856, garrytan#5877, garrytan#5854, garrytan#5868 and garrytan#5869.

Memory capture got two fixes that change what lands in your brain. A fact your agent saved with remember was often stored a second time by automatic capture a moment later; capture now skips an exact same-claim copy on the same entity, and never skips a correction. And the maintenance sweep only ever read the first 8,000 characters of each session transcript, so everything said after that never became a fact. It now reads whole transcripts in turn-aligned windows and remembers where it stopped. That costs more model calls on long sessions (about 15 for a typical 120 KB transcript, capped per sweep).

Managed brains with Google connected had quietly stopped doing several jobs: Gmail commitments were never extracted, Gmail and Calendar pages got no links or timeline entries, quiet threads never opened their loops once the waiting window passed, and closing a commitment loop claimed it retired the fact when it hadn't. All four work again. gbrain upgrade now checks the target release's Bun requirement before swapping, so a host can no longer upgrade itself into a binary it can't start. An exact-title page ranks first in title search again, and remote title search is index-backed (a rare term on a 40k-page Postgres brain: about 490 ms to about 34 ms).

The release also absorbs 62 open community pull requests, each re-checked against current master: sync stops deleting pages it does not own and keeps skillpack files out of managed checkouts, backups and storage restores cover every source and every live page, Claude Sonnet 5.5 works and Sonnet 5 is priced at its standard rate, list_pages pages losslessly, a multi-line fact no longer blocks its page from being written, and two search paths get faster. Thank you to everyone listed below.

To take advantage of v0.60.32.0

gbrain upgrade installs the binary. There is no schema migration. Restart every gbrain serve, autopilot and worker afterwards so they run the new code.

  1. Meeting pages re-derive their attendees on their own. The link extractor version moved, so the next gbrain extract --stale (or the autopilot cycle) re-reads Participants: lists.
  2. Expect one paid re-judge of contradictions. The judge prompt version moved to 4, so the next gbrain eval suspected-contradictions re-judges every pair. It prints its cost estimate first, as on any prompt change.
  3. Existing loops get request-time ages when their thread is next checked. An open thread loop keeps the earlier of its stored age and the oldest unanswered message's time, so the next sync that touches a thread corrects an age set at detection time.
  4. Code call graphs pick up the member-call fix when a file re-imports. To refresh every code page now: gbrain reindex-code.
  5. Preview the two new explicit repairs, and apply only after you agree. gbrain doctor reports captured_facts_active (facts captured before v0.60.30.0 from gbrain's own claude-cli sessions or pasted text) and loop_facts_drift (facts of loops closed before this release): 
    gbrain repair captured-facts            # preview; apply with --apply --expect <hash>
gbrain repair loop-facts                # preview
    Guide: docs/guides/repair.md.
  6. Decide on connector atom extraction. Atom extraction of connector email and meeting pages is now opt-in on every brain. To keep it (the text of each thread or event goes to your extract_atoms model): gbrain config set cycle.extract_atoms.connector_pages true.
  7. Cached search results are invalidated once (KNOBS_HASH_VERSION 29 -> 30). If you changed your FTS language, run gbrain reindex-search-vector so remote title search matches it.
  8. Verify: 
    gbrain waiting --as-of 2026-10-01T12:00:00Z --json | grep as_of
gbrain sweep --once --json        # corpus_files[] shows windows_done / windows_remaining
gbrain doctor
  9. If any step fails, file an issue at https://github.com/garrytan/gbrain/issues with the output of gbrain doctor and ~/.gbrain/upgrade-errors.jsonl if it exists.

Behavior changes

  • Automatic capture skips an exact copy of a claim already stored in the same source on the same entity (or another entity the claim names), visible where the new fact would be, and written within 15 minutes of the turn or by capture in the same conversation. Reworded copies are never skipped, so corrections ("is not moving", a changed amount or place) are always kept. Explicit remember is never skipped (garrytan#5888).
  • The sweep extracts whole transcripts in turn-aligned windows of about 8,000 characters, with a <file>.progress sidecar, capped at 8 windows per file and 32 per sweep (GBRAIN_CORPUS_WINDOWS_PER_SWEEP). Transcripts marked done before this release are not re-read; only turns added after the upgrade are extracted (garrytan#5887).
  • gbrain upgrade and gbrain self-upgrade refuse a release whose Bun floor the host does not meet (exit 78, nothing changed); --no-bun-floor-check upgrades anyway when the floor cannot be read. Autopilot holds the upgrade and self_upgrade_health shows the hold and fix (garrytan#5855).
  • Remote title matching follows the language search_vector was built with (garrytan#5889).
  • A synthesis publish still pending after the 30 s writer wait is a phase warning (publish_pending), not a failure, and the next cycle finishes it (garrytan#5854).
  • Connector email and meeting atom extraction is opt-in on every brain (cycle.extract_atoms.connector_pages). The auto-drain's daily cap counts attempts at the per-attempt estimate, so 6 attempts a day by default (garrytan#5856).
  • loops_close returns { closed, id, status, fact_expired, retryable, reason? } and expires the fact and strikes its fence row in one coordinated write (garrytan#5869).
  • gbrain extract timeline --source db exits non-zero when a write is refused, naming the code and recovery (garrytan#5904 probe).
  • Short random passwords are redacted reliably. The high_entropy_assignment rule (retrieval redaction and memory relays) judges its 3.5 bits/char floor on bias-corrected entropy. Raw per-character entropy of a 12-16 character value is capped near log2 of its length, so the floor let through 68% of random 12-character secrets and 9% of 15-character ones; it now misses under 1%. A short weak password assigned to a password, secret or token key (password1234) now redacts too. All-digit values (counters, timestamps) never redact, as before.
  • Contradiction judge prompt v4. The time check comes first: the three temporal verdicts need two different dates (metadata or text) or explicit change wording, never an order guessed from the values, the document kind or which statement is listed first. A statement that does not mention an attribute does not conflict with one that states it. A negative claim about one party ("Fund A is not leading") and a positive claim about another ("Fund B is leading") agree. Matched A/B against prompt v3 on the fresh N2 development seed 20261002, one Haiku 4.5 judge sample per arm, 1,371 identical pairs: same-time conflicts caught 131/150 -> 149/150 (undated 37 -> 50, same-day 45 -> 49, mixed 49 -> 50), temporal recognition unchanged at 40/40, false contradictions on unplanted pairs 11/820 -> 14/820 and on compatible pairs 2/50 -> 5/50, judged-pair precision 91.0% -> 88.7%. A code rule that turned temporal verdicts without time evidence into contradictions was measured and rejected: on the counted seed it added about 65 false alarms, because Haiku uses a temporal label for unrelated pairs it misreads. The prompt version bump invalidates the judge cache (step 2).
  • An acknowledgement-only reply to a question does not close a thread loop. "Thanks!", "Got it, thanks." or "Thanks, Bob!" after their question leaves the reply-owed loop open and its clock running. "Will do" and any reply with content still count as the reply. An acknowledgement of a message that asked nothing still closes it.
  • A ? inside a link is not a question, so outbound FYIs that only share a link never open unanswered_outbound.
  • Loop age is request time. opened_at on a thread loop is the oldest message in the trailing run that carries the obligation, not the detection time.
  • open_loops takes as_of (ISO 8601, default now) for due-date proximity, loop age and rendered ages, and echoes it in the grouped result; gbrain waiting --as-of <iso> passes it through. An invalid value is refused with invalid_params.
  • Bare-name resolution prefers the exact title within one directory. When several pages share a bare name's prefix in one entity directory and exactly one is titled that name, it wins (Acme over Acme Labs). A collision across directories, or two pages with the same exact title, stays unresolved as before.
  • Participants is attendance evidence everywhere Attendees is: a Participants: line, **Participants:**, or a ## Participants link-list section.
  • OR-relaxed keyword rows are not "keyword_exact". A row from the keyword fallback that matched some query terms but not the query reads weak_semantic (create_safety: unknown), and CRAG grades a relaxed rank-1 weak with the new reason keyword_relaxed_top. A chunk that also appears in a strict list is not relaxed.
  • gbrain check-backlinks fix refuses a managed canonical worktree once, before scanning, with or without --dry-run. check still reports the gaps (garrytan#5341).
  • Conversation-facts eligibility is one rule. A page of a conversation type whose frontmatter says conversation_parseable: false (a single-message source-evidence page, say) is never claimed by extract-conversation-facts and never counted by doctor's conversation_facts_backlog or sampled by conversation_format_coverage, so the backlog can reach zero. Opt in to strict mode with gbrain config set cycle.conversation_facts_backfill.require_parseable_flag true to admit only type: conversation pages and pages marked conversation_parseable: true. The backlog now also counts granular collector types (slack-thread, email-digest, ...) the way the extractor already enumerated them (garrytan#5330).
  • chronicle_backfill moves forward on every run. Each chronicle_extract job is keyed by source, page and content hash, so a repeat run skips pages already enqueued or extracted for their current content (reported as already_enqueued) and enqueues up to limit new pages per type. An edited page is swept again (garrytan#5329, the backfill cursor).
  • Unpriced conversation models no longer stop fact extraction on a default cap. extract-conversation-facts, the conversation_facts_backfill phase and transcript-ingest facts drop only the default USD cap, with a warning, when the extraction model has no price. An explicit cap still refuses with no_pricing, and the stop now names the model (garrytan#5825).
  • list_pages takes updated_after_slug with updated_after to resume strictly after a (updated_at, slug) keyset, and returns updated_at_iso; the truncation hint teaches that recipe (garrytan#5641).
  • Claude 5 models get native request shapes. @ai-sdk/anthropic moves to 3.0.125 (and the @ai-sdk/provider-utils override to 4.0.56), so Claude 5 models are no longer sent temperature or emulated structured output; Sonnet 5 and Sonnet 5.5 price at $2/$10 per million tokens (garrytan#5688, garrytan#5690).
  • whoami and capabilities report the transport of the verified principal, not the caller's claim (garrytan#5899).

Itemized changes

Contradictions (N2)

  • Undated same-time conflicts: prompt v4 (above). Measured on the counted N2 seed and on a fresh development seed against prompt v3; numbers in behavior changes.

Open loops (N7)

  • N7-2 acknowledgement-only replies (isAcknowledgementOnly), N7-5 pinnable ranking clock, N7-6 request-time opened_at (OpenLoopUpsert.openedAt; an open row keeps the earlier value), N7-7 link question marks (asksQuestion).

Graph and retrieval (N9, N12, A4)

  • N9-5 exact-title tiebreak in bare-name prefix expansion, so the relational arm fires for colliding names.
  • N12-6 Participants attendance evidence; LINK_EXTRACTOR_VERSION_TS moves to 2026-10-02T00:00:00Z.
  • A4-2 relaxed keyword tops grade weak.

Code intelligence (N13)

  • N13-8 a call through a receiver the extractor cannot type (obj.m(), emitted as bare m) is flagged memberCall and stored as edge_metadata.member_call; the within-file resolver never resolves it to a same-file top-level function. Bare-name code_callers matching is unchanged.

Maintenance

  • garrytan#5341 check-backlinks fix managed-worktree refusal.
  • garrytan#5330 shared conversation-facts eligibility (isConversationFactsEligiblePage and its SQL twin in src/core/facts/conversation-types.ts).
  • garrytan#5329 chronicle_backfill idempotency keys and paging, no schema change. The issue's other parts (a cadence gate, extraction idempotency for direct chronicle_extract runs, the no-sub-events page, the judge output cap) remain open.

Memory capture (garrytan#5888, garrytan#5887)

  • What the agent saves with remember is no longer stored a second time by automatic capture (garrytan#5888). The writeback hook, the compaction harvest and the corpus sweep skip a fact only on an exact normalized match (rules above). Capture counts same-entity near duplicates (cosine 0.92 or higher) without dropping them, so a threshold can be measured first. Hot memory and the context_pack/delta facts show one line per duplicate group. Each skip is logged on stderr and in the hook heartbeat, and gbrain doctor (memory_writeback) reports cross_lane_duplicates_7d and near_duplicates_shadow_7d.
  • The maintenance sweep reads whole session transcripts, and a resumed session no longer re-extracts the same head (garrytan#5887). Windows break at turn boundaries and carry their [user] or [assistant] label, pasted blocks are removed first, and the progress sidecar makes appends and compaction rewrites cost only their new turns. gbrain sweep --once --json reports corpus_files[] with windows_done and windows_remaining. Ported from garrytan#5897. Contributed by @javieraldape.
  • A capture or atom drain right after another managed write no longer refuses as "writer busy". The managed-facts preflight, the atom-maintenance session and the connector-sync preflight probed the worktree lock with no wait, while this process's own persistence consumer was still finishing the previous write (its receipt already read committed). A remember followed by automatic capture could lose that capture for a whole sweep cycle, and an atom drain burned one of its capped attempts. All three now share one bounded probe that waits up to 1 s (MANAGED_WRITER_PROBE_WAIT_MS); a writer still busy after that refuses with writer_lock_unavailable as before. The connector-sync probe only exists to re-stamp a moved checkout's device, so a busy result there still lets the sync continue.
  • New explicit repair gbrain repair captured-facts expires facts captured before v0.60.30.0 from gbrain's own claude-cli sessions or from pasted text (doctor captured_facts_active; the post-upgrade banner names the preview). Self-capture facts with evidence expire with --apply --expect <hash>; paste candidates (a word-overlap heuristic against the retained session corpus) only with --include-ambiguous. A claim with a legitimate active copy is kept. Facts are expired, never withdrawn, so remember can save the claim again.

Search (garrytan#5889, garrytan#5890, garrytan#5892)

  • An exact-title page ranks first again in title search, for CLI and MCP callers on both engines (garrytan#5889). Title ties broke by page id, so a page titled exactly "Acme" lost to dozens of "Acme weekly sync N" pages. Remote title search matches pages.search_vector with title-weight lexemes, with no migration: on a 40k-page Postgres brain, rare term 490.9 -> 33.6 ms, 2,000 matches 545.1 -> 453.3 ms, OR fallback 1142.6 -> 498.0 ms (median of 9). Diagnosis credit: @morven-ai (garrytan#5853).
  • think and synthesis see the full evidence set again with search.adaptive_return on (garrytan#5890). Gather was cut to 2 or 6 pages before synthesis; gather, brainstorm's close set, grade-takes evidence, whoknows and enrich now opt out of both reader-facing trims (6 of 12 matching pages -> 12 of 12 on the reproduction).
  • --json output stays valid JSON when a claude-cli model runs (garrytan#5892). AI SDK warnings go to stderr; a user-set AI_SDK_LOG_WARNINGS is kept.

Upgrades and operations (garrytan#5855, garrytan#5902, garrytan#5854)

  • gbrain upgrade no longer installs a release the host's Bun cannot run (garrytan#5855). A clone fetches and checks package.json at the exact commit it then fast-forwards to. The post-swap smoke runs gbrain --help and prints per-method recovery when it fails. Clone installs should upgrade with gbrain upgrade, not git pull, so the check runs. Ported from garrytan#5860. Contributed by @andreineacsu.
  • An unpaced embed backfill no longer takes every database connection (garrytan#5902). Without an explicit concurrency, embed --stale and the embed-backfill job use at most half the engine's real pool (20 with no pool); an explicit GBRAIN_EMBED_CONCURRENCY keeps the full-pool clamp.
  • Dream synthesis no longer fails its phase when the canonical writer is busy (garrytan#5854). Maintenance publishes wait up to 30 s (one wait per job, bounded by the job deadline); a still-pending publish is deferred without stamping the cooldown.

Managed brains and Google connectors (garrytan#5875, garrytan#5867, garrytan#5868, garrytan#5869, garrytan#5856, garrytan#5877)

  • Gmail and Calendar pages get links and timeline entries again (garrytan#5875). Cycles with no on-disk checkout (connector sources, Postgres brains without --dir) now run the bounded, source-scoped database-only stale drain each cycle.
  • Managed brains extract commitments from Gmail again (garrytan#5867). Every sweep queues loops_extract on managed and unmanaged brains, the sync result carries loops_enqueue (enqueued, deferred, skipped_reason), and every skip is logged. A one-time catch-up on managed sources analyzes email threads from the last 30 days that were never extracted.
  • Quiet Gmail threads open their loops when the waiting window ends (garrytan#5868). The sweep records a grace hold and re-evaluates the thread when due, from the stored page when nothing changed; the first sweep after upgrading seeds holds for threads active in the last 14 days, with no Gmail calls. Grace holds are not held items.
  • Closing a commitment loop retires its fact, and says so honestly (garrytan#5869). On a managed brain the expiry was refused, the error swallowed, and loops_close still answered fact_expired: true; the fence row was never struck on any brain. A fact another open loop still uses is kept, and a close whose retirement did not commit is retryable. New explicit repair gbrain repair loop-facts (doctor loop_facts_drift) retires the facts of loops closed before this release.
  • Automatic atom drains stop dead-lettering on connector sources (garrytan#5856). When opted in, each attempt runs under one budget across its batches; with a model the tracker cannot price, only the attempt count bounds spend. A job refused before any model call dead-letters once and does not count; checkout and connector sources take turns. Off switch: gbrain config set autopilot.auto_drain.enabled false. The connector_database write target is ported from garrytan#5865. Contributed by @andreineacsu.
  • gbrain doctor's orphan_ratio fix command runs, and connector mail no longer inflates the ratio (garrytan#5877). The printed command includes --source db (and --source-id when scoped), and orphaned email/meeting renders of connector sources are reported apart as connector_renders_excluded.
  • gbrain extract timeline --source db writes on fresh and managed brains (garrytan#5904 probe). It publishes per page through the coordinator instead of inserting raw rows every managed brain refuses. garrytan#5904's library-level entry point stays open.

Issue reports with reproductions for this group: @andreineacsu (garrytan#5888, garrytan#5887, garrytan#5889, garrytan#5890, garrytan#5892, garrytan#5855, garrytan#5875, garrytan#5867, garrytan#5856, garrytan#5877, garrytan#5868, garrytan#5869), @nezovskii (garrytan#5902, garrytan#5904), @ywlee7922 (garrytan#5854).

Security

  • Short random secrets clear the entropy gate (v0.60.31.0 follow-up). A password= value of 15 random characters with one letter repeated three times read 3.46 bits/char and stayed plaintext; the gate now uses the Miller-Madow estimate (correctedEntropy) and requires a non-digit, and the A5 test that flaked on such values runs on fixed seeded values plus that case.

Known limits

  • A remember saved after automatic capture already stored the same claim leaves two active rows, and two capture writers racing on the same claim can both insert it. Reworded duplicates still appear twice until a measured cosine threshold ships: single-word corrections ("Lisbon" -> "Porto") embed at 0.95 or higher, which no token guard catches (garrytan#5888).
  • Hosts on v0.60.31.0 or older run their own upgrade code, which has no Bun floor check (garrytan#5855).
  • With an unpriced extract_atoms model the auto-drain's dollar limit is not enforced (garrytan#5856).
  • garrytan#5902 is mitigated, not reproduced end to end; garrytan#5854 fixes the wait and phase failure, not the reported duplicate child re-runs, which were not proven.

Community fixes

Sync and persistence:

Backup, storage, migrate and doctor:

Search, graph and timeline:

Facts, models and extraction:

CLI, hooks, connectors and loops:

For contributors

  • KNOBS_HASH_VERSION is 30.
  • The managed connector job contract harness (test/helpers/managed-connector-job-contract.ts, PGLite and Postgres) drives every automatic connector-source job on a managed brain through a real queue and worker and counts every managed_writer_guard refusal.
  • src/commands/extract-timeline-db.ts is the shared coordinated --source db timeline path.
  • PROMPT_VERSION is 4.
  • isManagedFilesystemPath(path) in src/core/persistence/filesystem-guard.ts is the predicate assertManagedFilesystemWrite uses.
  • ExtractedEdge.memberCall and CodeEdgeInput.edge_metadata.member_call.
  • ThreadLoopSpec.openedMs, OpenLoopUpsert.openedAt; rankGroups and renderText take the reference time.
  • ConfidenceGrade.reason gains keyword_relaxed_top.
  • ALLOWED_TYPE_ALIASES and pageTypesForAllowed moved to src/core/facts/conversation-types.ts (re-exported from extract-conversation-facts.ts), next to isConversationFactsEligiblePage, conversationFactsEligibleSql and requireParseableConversationFlag.

Contributors

h6y3, andreineacsu, and 21 other contributors
Assets 4
Loading