Skip to content

v2.11.0

Choose a tag to compare

@github-actions github-actions released this 14 Nov 02:49
· 10 commits to master since this release

One release for the whole 13-11-2026 audit: faster and steadier startup, search
that finds Spanish words, smaller answers, safer data and secrets, a unified
installer and four new commands. Still 15 tools; every field the zero-crbro card
reads keeps its name and meaning. The search index rebuilds once on first start.

Startup and daemon

  • Daemon mode: the proxy answers initialize, ping and tools/list itself
    at once, whatever the daemon is doing, from a snapshot of the real server
    written at build time (dist/mcp-surface.json): initialize in ~80–95 ms
    instead of waiting for a daemon to be found or started (603 → 94 ms when it
    has to start one). This is what made Codex give up on a slow start.
  • A client that waits before sending initialize is no longer cut every 5 s:
    the daemon's handshake clock stops once the proxy has authenticated.
  • A proxy that had to serve in-process keeps looking for a daemon and moves
    back once one answers and no call is running; three lost daemons in a minute
    now mean a pause, not in-process for good. Getting a daemon is bounded by one
    12 s budget. After any move, the proxy repeats the conversation's last
    crbro_boot (answer kept to itself) so its project binding survives.
  • One daemon per release: the build id is a hash of the compiled code
    (dist/build-id.json, written by npm run build), so two installs of the
    same version share one daemon.
  • daemon stop / off let running calls finish (up to 10 s); a call the
    daemon had not started is sent to the next backend, and a call cut by a crash
    says it may or may not have taken effect.
  • .daemon is cleaned on every daemon start and by daemon status|stop|off:
    descriptors whose daemon does not answer at its endpoint (never judged by PID,
    which Windows reuses) and stale temporary, lock and blocked files.
  • A daemon that cannot start says why in .daemon/daemon.log and the waiting
    client stops waiting at once.
  • CRBRO_STARTUP_TRACE=1 prints the startup phases with timings to stderr and
    <brain>/.daemon/startup-trace.log.
  • A single MCP message over 32 MB gets -32600 for its id instead of being
    split into two broken messages.

Search and recall

  • Words with á, í, ú, ñ or ü are now found with or without the accent, and
    hyphenated or slashed compounds by any of their parts (memory finds
    crbro-memory). The index rebuilds once (format 8).
  • crbro_recall says no_answer: true when nothing stored clearly matches and
    returns at most three weak rows without also_matched, instead of five rows
    that read like answers. CRBRO_NO_ANSWER=0 restores the old behaviour.
  • A semantic-only match is strong above this brain's own noise line (measured
    in the background, stored with the index) instead of a fixed 0.86.
  • With several queries, the row shown is one phrasing's own row, so its
    confidence and scores agree; also_matched never repeats an entry.
  • One large neuron no longer wins by size; the bound project's neurons get a
    ×1.25 boost instead of being placed above better global matches.
  • Entries over 1,200 characters are indexed in overlapping parts and recall
    returns the part that matched (matched_part).
  • A search index that is only behind the files is refreshed, not rebuilt (boot
    with a stale index on a 500-neuron brain: 1.9 s → 0.4 s). A running server
    sees what other processes saved before every search and never overwrites a
    newer index with an older view. Classic stdio mode writes the index when the
    client leaves (at most 2 s).
  • Smaller index and memory: only searched fields are indexed, no sort data.
  • Semantic vectors are added and saved in slices while a large pass runs (the
    upgrade itself splits long entries into parts that need new vectors), so a
    process that leaves halfway keeps what it did. The pass pauses while a search
    runs, and while it lasts recall answers keyword-only with
    semantic: "warming". Measured on a 563-neuron brain right after upgrading:
    recall 0.05–0.4 s during the ~4-minute pass, instead of 1.5–3 s with 20 s
    stalls.
  • A running server's once-a-minute check of the brain's files reads their
    dates in parallel: ~0.5 s instead of ~20 s on a busy machine.
  • Semantic layer: recall waits at most 1.5 s for the model
    (CRBRO_SEMANTIC_WAIT_MS), then answers keyword-only with
    semantic: "warming"; only semantic install downloads; missing vectors are
    re-embedded and orphans pruned at start; the runtime is pinned to
    @huggingface/transformers 3.8.1 and the model files to SHA-256 hashes.
  • New Spanish retrieval benchmark gating npm run bench; the old gate's floor
    rises.

Smaller, clearer answers

  • crbro_boot is about half the size: protocols whole, then counts, open items
    (due first, then newest; 180 characters; age_days; review: true past 30
    days without a due date), closed items as titles, hot topics, session
    headlines and a shorter memory_discipline. New pending block whose
    numbers always match message, and a more line naming what was left out.
    open_items_total, recently_closed_total, active_protocols and
    sessions_note are gone; retired_tools moved to crbro_inspect view=status.
  • crbro_boot no longer fails when project_neurons names a neuron that is
    gone (listed in unknown_project_neurons). Over its budget it drops session
    headlines and hot topics before open items.
  • crbro_recall rows leave out fields at their default (temporal,
    verification, freshness, provenance), as the description says.
    relevance is {score}; relevance_score, matched_terms and query_terms
    left the MCP rows. Previews are capped at 160 characters. The "say it may be
    out of date" instruction appears once, in stale_warning. A corrupt neuron no
    longer fails the whole recall (unreadable_neurons).
  • crbro_context writes return what changed and the counts
    (return_context: true for the old reply): 22,381 → 427 characters.
  • Not-found and bad-parameter answers carry isError: true in revise, forget,
    map, space, share and learn. crbro_forget entire tokens cover the whole
    neuron and are valid only in the process that issued them.
  • crbro_learn warns (never refuses) on facts over ~600 characters and neurons
    over 300 entries or 300 KB.
  • Tool definitions (tools/list): 44,434 → under 30,000 characters.

Data integrity

  • backup --keep must be a whole number ≥ 1 ("siete" or a missing value used
    to delete every copy). Each new copy is flushed and read back before any
    rotation and is never rotated itself; a copy dated in the future no longer
    counts as the newest.
  • Tiered backup retention: the newest N copies (--keep, default 7) plus one a
    day for N days, one a week for 4 weeks and one a month for 6 months.
  • A neuron file with a byte-order mark reads normally. A file that does not
    parse no longer breaks recall, crbro_inspect, crbro_maintenance,
    crbro_consolidate or the boot's protocols: it is skipped and named with
    where a good copy is, and writes to it are refused.
  • Similar topic names no longer merge ("Sprint B" stays out of "Sprint A");
    only the same name with or without underscores, or a plural, resolves to an
    existing neuron (resolved_by: "variant", ambiguous when several match).
    "C", "C++" and "C#" are three neurons; non-Latin names get their own neuron;
    project_x no longer becomes project_project_x; types are inferred from
    whole words.
  • crbro_learn with a neuron_id that does not exist is an error instead of
    writing to whatever the topic matched. supersedes also applies when the new
    fact already exists, and the retirement reaches shared spaces.
  • Entries that differ only in whitespace or case are one entry.
  • Synapse writes take the same lock as every other write; a write never brings
    back a neuron deleted meanwhile. Decay follows elapsed time (decayed_at),
    pruned synapses get a quarantine copy, and archiving never takes protocols or
    shared neurons and keeps an archived neuron's links.
  • Quarantine retention: CRBRO_QUARANTINE_DAYS (default 180; 0 keeps copies
    forever); crbro_consolidate and crbro_maintenance prune older copies and
    report quarantine_pruned. Forgetting some entries keeps a copy of just those
    entries; repair and purge_boilerplate also keep what they remove.
  • Opt-in encryption at rest of backups and quarantine copies
    (CRBRO_ENCRYPT_AT_REST=1, AES-256-GCM, 32-byte key CRBRO_AT_REST_KEY in the
    OS keychain). Without the key nothing is written in clear and nothing deleted.
  • Heat is kept in prefrontal/heat_state.json: consolidate re-reads only
    changed neurons and never rewrites a neuron for heat (first consolidate on a
    500-neuron brain: 1.5 s → 0.36 s).
  • crbro_maintenance stale_list pages through every stale entry and
    stale_review verifies or retires many at once; the dry run adds a read-only
    hygiene report (split_partitions, merge_candidates, summary_missing).
  • The shelf-life detector no longer marks Spanish thousands ("1.200.000"),
    monthly counts or the name "Max" as volatile.

Secrets, keychain and team spaces

  • Breaking: crbro_secret get no longer returns arbitrary environment
    variables of the server process. An environment value counts only as
    CRBRO_SECRET_<NAME> or when <NAME> is listed in CRBRO_SECRET_ENV; answers
    say source: "env" | "keychain". Names must be SCREAMING_SNAKE_CASE.
  • Through MCP, overwriting a secret with a different value and removing one are
    two-step (confirm_token, then confirm); the first call writes nothing.
  • Windows keychain: a missing keys.dpapi is an error with recovery steps when a
    store existed, never an empty store (secret set NAME --force starts one on
    purpose). Every write keeps rotated sealed copies (CRBRO_KEYS_COPIES,
    default 5); secret copies lists them and secret restore <id> --yes puts one
    back after validating it.
  • Keychain and git calls run asynchronously in the server and the store probe
    is remembered: a credential read no longer freezes a shared daemon (865 ms →
    24 ms) and a repeated read takes ~0 ms instead of ~0.43 s. macOS and Linux
    failures are never read as "empty"; Linux lists names without pulling values.
  • Team spaces: names must be simple slugs inside shared/; remotes with
    embedded credentials, a leading - or control characters are refused; a sync
    never merges into, or marks as shared, a local neuron you did not share there
    (left_alone); incoming notes are validated and redacted (ops_dropped).
    crbro_share also checks rationales, keywords, revision notes, names,
    summaries and tags for credentials.
  • Secret filter: private keys are redacted whole, and "PASS value" /
    "contraseña value" forms without a colon are caught; two false positives seen
    on a real brain are gone. Rationales, keywords and revision notes pass through
    the filter and crbro_audit counts them. Benchmark: 24/24 caught, 0/28 false
    positives, measured per value.
  • Locks: a dead owner's lock is reclaimed even when Windows reused its PID;
    release retries a transient EPERM/EBUSY; a timeout names the holder.
  • Miner scheduler: the crontab is written through stdin (existing lines kept
    byte for byte, the previous one saved); the Windows task runs the absolute
    npx.cmd with correct quoting.

Install, CLI, doctor and hooks

  • New npx crbro-memory install [--clients …] [--dry-run] [--engine-dir DIR]:
    copies the engine to ~/.crbro-engine/<version> and registers it in Claude
    Code, Claude Desktop, Codex and Cursor (no npx at client start; one build, one
    daemon). New uninstall removes only what CRBRO added and never the brain.
  • Codex: install-boot registers the server in config.toml when missing and
    adds startup_timeout_sec = 60 / tool_timeout_sec = 120 only when absent; a
    packaged boot hook runs from a stable folder, never from an npx cache.
  • Every installer write keeps a timestamped backup, is refused if the file
    changed since it was read, and keeps line endings, indentation and links.
    CLAUDE_CONFIG_DIR and CODEX_HOME are honoured.
  • Installed hook copies refresh themselves when a newer release finds them
    unmodified; edited copies are kept and reported once (hooks_notice).
    CRBRO_HOOK_REFRESH=0 turns it off.
  • CLI: unknown commands and options exit 2 instead of starting the server;
    --version; <command> --help runs nothing; failures exit 1.
  • doctor checks hooks, daemon leftovers, quarantine, the guard index, client
    registrations, Codex timeouts and backup age, and can now say ok.
  • Guard: the current project's lessons first, others only inside their folder;
    team and miner lessons stay out unless CRBRO_GUARD_ORIGINS=all; at most 6
    per session.
  • Miner: no more checklist lines, bare links or headings stored as facts;
    crbro_maintenance purge_boilerplate removes the ones already stored.

New commands

  • crbro-memory export [--out DIR] [--format md|json] [--include-retired]:
    the brain as an Obsidian-compatible Markdown vault (a note per neuron with
    frontmatter, entries by kind, synapses as [[links]], domain index notes,
    day logs, open items) or one crbro-export.json that import reads back.
    Read-only and safe while a client runs; credentials redacted; never writes
    inside the brain or into a non-empty folder.
  • crbro-memory import PATH [--apply] [--topic NAME] [--type …]: learns what a
    CLAUDE.md / AGENTS.md, a Claude Code memory folder or a CRBRO export says.
    Dry run by default; writes go through the MCP tools, so validation, redaction
    and de-duplication apply, and importing twice adds nothing.
  • crbro-memory view [--port N] [--no-open]: a local page (127.0.0.1, one-time
    token, strict CSP) to browse neurons and entries with their date, status,
    verification and freshness, search, resolve or discard open items, verify or
    retire stale entries and remove one entry after confirmation.
  • crbro-memory serve --http: MCP over Streamable HTTP at
    http://127.0.0.1:8787/mcp with a bearer token (CRBRO_HTTP_TOKEN or the
    keychain; --new-token creates one without printing it). Loopback unless
    --allow-remote; no TLS and no OAuth; crbro_secret is left out unless
    --allow-secrets.

Fixed

  • A neuron reference with a path separator (../x) was used as a file name, so
    crbro_inspect and crbro_learn could read or write JSON files outside the
    brain. It is now always a name.
  • crbro_inspect view=status reports the folder the server reads as
    brain_path, not the path the manifest recorded when the brain was created
    (a copied or moved brain showed its old location).

Packaging and releases

  • Requires Node.js 20 or later. npm publish runs the tests first. The package
    no longer ships sourcemaps or declaration files (205 → 61 files).
  • Releases are published from a version tag by CI (release.yml) with npm
    provenance; CI runs on Linux, Windows and macOS with Node 20, 22 and 24,
    starts the real server over stdio and builds the .mcpb and Docker image.
  • README rewritten in three layers; long sections moved to docs/.
  • Dependencies moved to patched versions: zod 4.6.7 (the minimum is now
    ^4.6.7), and through the MCP SDK body-parser 2.4.0, hono 4.13.18,
    @hono/node-server 2.1.8, fast-uri 3.1.9 and qs 6.17.1. npm audit
    reports no advisories at release time (not a guarantee about later ones).