Skip to content

v0.5.1 — the health release

Choose a tag to compare

@borges-funes borges-funes released this 11 Nov 02:27
· 2311 commits to main since this release

The health release. An agent that keeps a memory for months needs a way to ask whether that memory is still in good shape. v0.5.1 rebuilds the health score around that question: the raw counts, the signals computed from them, and the curves that grade each signal are now declared as data, a build refuses to score a graph it does not match, and a small workbench shows what a curve change would do before you make it. Also in this release: generalization edges (a memory can point up to the principle it exemplifies, shown when it is recalled), a fixed budget on everything the hooks inject into context, and a telemetry store for measuring hook reliability.

The curve shapes shipped here are a first judgement made by one team on its own graphs; expect them to move. This version went through five release candidates in six days, and each soak found something. The last finding, one wrong sentence in a nap advisory, is fixed on the development branch and ships in the next release.

Upgrading: ask your agent to upgrade ENGRAM. It will walk the full upgrade routine — which is more than the mechanical commands — and it can answer any questions you have about what changed and whether anything affects you.


[v0.5.1] - 2026-09-14

rc5 fix round — 2026-09-19 (found on the rc4 soak; tracking #2941)

  • engram_nap renamed to engram_checkpoint; engram_nap kept as a
    deprecated alias for one release.
    The tool name collided with the
    engram-nap SKILL name (the eight-step nap routine, of which calling the
    tool is only one step), and drowsiness/nap-warning nudges that said "Run
    engram_nap" read ambiguously as either "call the tool" or "run the skill"
    — an agent that took it literally called only the tool and skipped the
    skill's other steps (notably Step 5b, the warm-briefing rotation). The two
    hook nudges (context_tracker.py, engram-surface-hook.py) now say to
    load and run the engram-nap skill instead of naming the tool directly.
    engram_nap delegates to engram_checkpoint with identical behavior and
    output shape (#3018, closes #3015).
  • Nap advisory re-anchored on the last checkpoint, not session start.
    written_this_session (renamed written_since_last_nap) satisfied itself
    on one early write of warm-briefing.md and stayed silent for the rest of
    a long session — across many subsequent naps and a compaction where Step
    5b was actually skipped, measured live on two seats (a section stale by 4
    naps / 10 hours, and a post-compaction section, both read as non-stale).
    Every nap/sleep already appends a timestamped header to session_log.md;
    the advisory now compares against the most recent one instead of the
    per-session marker, re-arming after every nap and naturally covering the
    compaction case (#3017, closes #3014).
  • Boot charge's monitor-liveness check reads the monitor's stdout target,
    not just process presence.
    pgrep certifies a monitor process alive
    even when its stdout was armed to a dead end (a detached nohup … > file
    instead of through the Monitor tool) — measured live: nine real wake
    events detected, none reached the session, silently, for six hours. The
    charge now reads readlink /proc/<pid>/fd/1 (Linux, a socket: target is
    healthy) or lsof -a -p <pid> -d 1's TYPE column (Darwin, no /proc)
    (#3008, closes #3007).
  • Dream-record filename pinned to the NYC-local write date.
    ~/.engram/history/dream/YYYY-MM-DD.md never specified whether its date
    is the day the record was written or the day it consolidates; the
    sibling dream-report filename in the same skill was already pinned to
    NYC-local, making this an oversight rather than deliberate latitude
    (#3010, closes #2177).

rc4 fix round — 2026-09-18 (found on the rc3 soak and the 09-18 nightly; tracking #2941)

  • macOS install unblocked. install-local-marketplace.sh carried a heredoc
    apostrophe that Apple's bash 3.2 mis-parses, so a fresh rc2/rc3 install on a
    Mac whose PATH resolves bash to /bin/bash failed; fixed with a regression
    test and a macos-bash32-syntax CI job that runs the installer under bash 3.2
    syntax rules (#3000, closes #2979).
  • Nap advisory for a skipped warm-briefing rotation. engram_nap returns a
    warm_briefing block stating whether the "From this session" section moved
    since the session started (optional session_id, session_started_at as the
    comparison anchor), so a nap whose Step 5b rotation did not run is visible in
    the return the agent already reads instead of only in the file's mtime
    (#2998, closes #2997).
  • Nightly no longer fails on its own fix. The #2989 erase-safety control
    read its "pre-fix template" from the live dev ref, so it was correct while
    the PR was open and stale the moment it merged; the control is now a frozen
    literal fixture (#2996, closes #2995).
  • persistent: true is conditional on the Monitor schema. The flag was
    removed from Claude Code's Monitor tool between 2.1.267 and 2.1.273 and is
    accepted-and-ignored on later versions; the collaborating-loop skill now
    says to pass it where the schema lists it and otherwise to arm at the
    1800000 ms effective cap and re-arm on expiry, reading the arm line rather
    than the success response (#3006, part of #3005).
  • CLAUDE.md gh-stack note: approval dismissal on a stacked merge comes from
    the base branch's dismiss_stale_reviews_on_push setting, not from the
    auto-rebase itself, with five dated cases (#2999, closes #2838); the
    generic-404 back-reference dropped by the earlier consistency fix is restored
    (#3004).

rc3 fix round — 2026-09-17 (found on the rc2 soak and the 09-16 dream cycle; tracking #2941)

  • Warm-briefing erase landmine removed. The template's protected region
    carried two literal ## From this session occurrences, so a sleep erase could
    land inside the user's letter; the section now ships as an empty placeholder
    as the file's final line with the structural-invariant comment above it, and
    the sleep skill specifies the erase mechanically (last line-anchored match to
    EOF, re-append the bare heading, post-condition: the file's last line is the
    bare heading). The in-place-erase and remove-entirely dispositions are retired
    (#2989).
  • Dream-master routing keys on category, not suggestion prose. Category 7/8/9
    findings were routed by regex over the free-text suggestion field and fell
    silently to the log-only bucket when phrased naturally (#2988).
  • Dream projection staleness is now a signal. A long-lived MCP server started
    before the 0.5.0 metric consolidation kept emitting the old metric set with
    nothing to say so; the dream-master runs an engine-version staleness check, the
    repair-record validator covers the full row, and the nightly reports the
    backfill-cap total it matched (#2985).
  • Dream-fairy stale-scan spec: the cornerstone exemption keys on the node's
    own type instead of focus-list membership, which no longer holds stable
    principles (#2984).
  • CLAUDE.md templates: the MCP-schema load-time parenthetical is deleted
    rather than reverted, per the rule that CLAUDE.md carries key rules and their
    explanation, never technical details that drift (#2991).
  • CI: the team-practice drift gate checkout no longer persists credentials
    and its advisory message is fork-agnostic (#2975).

rc2 fix round — 2026-09-16 (found on the rc1 soak; tracking #2941)

  • Health score is single-sourced on health-v2. The pre-v2 formula that
    engram_stats/diagnose/the session banner still reported (95.0 beside the
    v2 panel's 100) is retired (#2957).
  • Installer hardening. The jq dependency is gone; config is read with
    python3 (#2958). Malformed config.json now fails loud on the venv-provisioning
    reads instead of silently resolving tier/multi-agent to defaults (#2964), and a
    well-formed config with a wrong-typed install_tier/multi_agent value is
    rejected the same way (#2973). A missing src/forum/requirements.txt on a
    multi-agent seat warns instead of silently skipping (#2964).
  • Forum package requirements are installed into the multi-agent venv, so a
    fresh dev venv can collect the forum tests (#2949); the test-id gate now scans
    src/forum/tests (#2948) and enforces scan-roots ↔ tests.yml sync (#2959).
  • dag_conformance no longer reports a phantom one-node cycle for
    supersedes/retracts edges (#2947).
  • Viz server: gzip responses, ThreadingHTTPServer, and an in-flight guard,
    with config writes serialized under one lock (#2926).
  • Skills: the health-exam skill matches the rc1 runtime and is portable
    (#2946); two skills name the live engram-viz-user.service unit and
    tools/operator-setup-viz.sh as the canonical restart path (#2971); the
    upgrade skill documents the Codex 0.154 live-refresh boundary (#2954).
  • Baton gates: the REST flip endpoint now runs the same CI-green /
    approval-staleness / colleague-approval gate as the CLI, fail-closed when the
    shared gate module cannot load, with an author: field and a legacy-author
    fallback (#2966); cmd_merge's Gate 2 accepts a maintainer-role holder, not
    only the literal pool sentinel (#2969).
  • Team practice: the "which colleague takes the seat" block of CLAUDE.md
    is a generated region checked by a CI drift gate against the private
    team-practice repo (#2963); engram_add_observation's docstring is trimmed to
    its contract with the WHY moved to the engram-observe/engram-standpoint
    skills (#2962); the interim soma review-convention exception is retired (#2956).
  • CI: both nightly workflows record the DST-changeover weighing decision
    beside their cron line (#2970).

A scoring-system release. 204 commits since the v0.5.0-rc3 cut
(2026-08-30 through 2026-09-14). The headline arc is the Health Score v2.0
curve system
, with the new Curve Workbench tool built to tune it —
together the single largest body of work in this window. A second arc, the
hook injection budget, bounds every hook's context injection for the
first time. A third, daemon liveness / socket-guard / WAL-guard
hardening
, is a cluster of reliability fixes to ENGRAM's background
processes and its crash-safety instrumentation.

Added — Health Score v2.0: curves declared as data (epic #1991, #2518, #2537, #2563, #2575, #2905)

  • The health-scoring engine has been rebuilt around a declarative
    measurement → signal → curve pipeline
    , replacing the earlier hardcoded
    scoring arithmetic. Raw measurements (L0) are collected purely additively;
    L1 signals are now declared as data (SIGNAL_DEFS) rather than scattered
    arithmetic; the scoring curves that turn a signal into a 0–1 contribution
    are themselves declared as data (CurveSpec.knots, a versioned knots file)
    instead of hardcoded functions. A composed score, hold-out/renormalize
    behavior at both the signal and curve level, and a dream-side projection
    round out the model (#2187, #2189, #2205, #2494, #2520, #2523, #2556,
    #2577, #2578, #2582).
  • A version gate now prevents a stale build from scoring against fresh
    data (or vice versa).
    The gate diffs the code's declared L0/L1/curve set
    against the max deployable tag and fails closed (ILLEGAL) or denies
    (over-severity DENY) on mismatch, rather than silently scoring with a
    mismatched engine (#2563, #2588, #2725).
  • Two more version-gate hardening fixes: the build-identity check no
    longer falls back to the undiscriminated repo resolver, which could
    silently record a data-repo sha as engine_build (#2650); and the gate's
    tie-break now uses tag creation time, so its DENY message no longer
    conflates "unavailable" with "unchanged" — a case that was producing
    wrong DENY verdicts on otherwise-fine PRs (#2805).
  • The health snapshot store gained a historical series — declarations are
    replayed as-of any stored point in time, with a git-replay driver, a
    backfill/verify equality control, and export via a new
    /api/health/v2/series endpoint, so trends over time are queryable, not
    just the latest snapshot (#2575: #2593, #2599, #2603, #2605, #2607, #2608,
    #2616, #2869/#2891).
  • The v2 curve system shipped in two steps: v1.1 introduced curve knots
    as a versioned data file with a generic placeholder curve set (#2788/#2794),
    then v1.2 created the two named curves badness_pending_premise and
    goodness_coverage and rebound the relevant metrics onto them (#2905/#2908).
  • A MAJOR version bump unifies two of the L0 measurement set's counters
    into one.
    taint_rate (retract-only) and stale_derivation_rate
    (supersede-only, derivations-only) were two separate signals; they're now
    one pending_premise_rate signal backed by a single new L0 measurement,
    claim_bearing_pending_marker — a DISTINCT count over claim-bearing nodes
    carrying either a pending retract or a pending supersede cascade marker
    (a node carrying both counts once, not twice). The L0 set's own identity
    changing (two measurements retired, one added) is exactly the
    add/remove/redefine class the health-version gate exists to track, hence
    MAJOR rather than MINOR — MINOR resets to 0 on the bump (#2905/#2922).
  • Data-integrity hardening for the underlying FTS5 index: a
    content/index mismatch detector that distinguishes an unknown "partial"
    state from a confirmed "mismatch", plus an AFTER UPDATE trigger that
    resyncs FTS on a plain content rewrite so the index can no longer silently
    drift from the content table (#2678, #2731, #2742/#2745, #2744).
  • Snapshot access hardening: explicit file permissions on health
    snapshots, no more default VIZ_OPERATOR, and a fleet repair tool for
    existing installs whose snapshots were written under the old defaults
    (#2720/#2734).
  • A composite index + PRAGMA optimize keeps the query planner's stats
    current as the series grows (#2559/#2685).

⚠️ If you run your own ENGRAM instance: shipping this on dev/in a
release is not the same as your instance running it.
Your server and
hooks need their own rebuild + restart to pick up the new health-scoring
version — until then, the health score your instance reports can lag what
is actually in the code you pulled (it may still say health-v1.0 /
v1.1 after upgrading). The version gate above exists specifically to
catch this class of mismatch and refuse to score with a stale/fresh
mismatch rather than silently mis-scoring; run your normal upgrade routine
(tools/install-local-marketplace.sh + restart) and check the reported
health-engine version if this surprises you.

Added — Curve Workbench: a new dev-tier tool for tuning health curves (#2727, #2738, #2748, #2754, #2758, #2870, #2881, #2894, #2896)

  • A new dev-tier web page for picking curve knots and importances against
    real history
    — value-checked knot edits with a hold-out lock, real-time
    history + focus readout, per-knot add, set-as-reference, and
    unavailable-bands rendering (#2727/#2730, #2738/#2740, #2748/#2750,
    #2870/#2875, #2881/#2882, #2896/#2903).
  • Moved off the main health dashboard's :5001 to its own :5021 entry
    point, and gained a systemd user unit so it can run persistently
    (#2754/#2762, #2894/#2907, #2909/#2917).
  • The main health dashboard's Detail tab is a read-only fork of the
    Workbench
    — a three-column-grid history/distribution view over
    /api/health/v2/series, alongside a rounded-to-2-decimals Normal-view
    score panel (#2518, #2754/#2759, #2867/#2871, #2868/#2872/#2873). The
    Normal-view Score panel itself now renders the v2 composed score

    (previously v1) with a coverage caveat and a details link into the
    Detail tab — the number users see there changed meaning (#2761/#2757).
  • The Workbench's own startup gate now covers all four dev-tier files
    (previously missed some), with a reason-string stage flag replacing a
    locals() guess (#2804).

Added — Generalization edges: a new general→specific abstraction relation (epic #2753)

  • A new generalizes edge type connects a general claim to the specific
    claims it abstracts, with the substrate for the edge itself (WS1, #2812),
    rendering of generalization chains at the tripwire and surface hooks (WS2,
    #2814), a read-only migration-proposal script to surface candidate edges
    in existing graphs (WS3-A, #2862), and a dream-fairy scan category that
    proposes new generalization candidates during sleep (WS3-B/C, #2865).
    Landed incrementally across the window; check the epic for current status
    before assuming it's fully closed out.

Added — Telemetry store: a JSONL/records.db mechanism for measuring hook-emitter reliability (epic #2574)

  • A never-raise telemetry writer + aggregator, so ENGRAM's own hook
    system can be measured (did an emitter site actually fire, was its output
    attended to?) without that measurement ever risking a hook's own
    operation: a write-time JSONL append (WS1, #2847) with an atexit
    rename-on-exit and two-mechanism drop-counting; a manifest telemetry key
    • gate declaring which sites participate (WS2, #2846); hook-side site
      attestation with try-narrowing and reached-tracking (WS3, #2853); a
      consumer-side projection registry that surfaces unprojected verdict rows
      rather than silently dropping them (WS4, #2850); and server-side emitter
      initialization + narrow-try protection at every call site (WS5, #2854).
      Convenience-tier internal instrumentation — no new user-facing CLI or MCP
      surface — but a real shipped mechanism, not a design doc (#2816, #2848).

Added — Hook injection budget: every hook's context injection is now bounded (#2863)

  • New shared renderer emit_context() (plus a per-harness budget table)
    replaces hand-rolled additionalContext construction across all hooks —
    every SHIPPED hook injection is now bounded under a named per-harness
    ceiling, never grows unpredictably, and carries its highest-value content
    first, so a harness that silently truncates the tail of an oversized
    injection cannot silently drop the important part (#2874, #2876, #2886,
    #2887, #2888, #2889, #2892/#2893/#2898). A CI gate (the "violator gate")
    now catches new hand-rolled injections that don't go through the shared
    renderer, tuned to exclude non-emitter false positives (#2910) — but
    #2863 itself stays open: the completeness-guaranteeing enforcement gate
    (WS3) is still being hardened, so the claim above covers every hook
    shipped so far, not a closed guarantee against a future one. A related
    fix caps the lesson-tripwire nudge block to a self-measured character
    ceiling (#2858/#2860), and the daemon-supervision/persist-threshold
    ordering within the shared renderer was corrected so a must-not-miss line
    can't be pushed past the cut (#2859/#2925).

Added — Evidence capture & write-time quality advisories

  • Evidence snapshots now record an archive kind and retain the original
    URL
    even after a web/PDF page has been archived, with evidence_type
    stamped at capture time (#2606 facet 1, #2634).
  • New hard-token anchoring measurement + write-time advisory: when an
    observation isn't anchored by verbatim tokens from its cited source, a new
    measurement now surfaces this as an advisory nudge at write time
    (#2602/#2627).

Added — New operational tools

  • A fairy-report extraction tool recovers dispatched-fairy reports after
    a session reboot, for when the dispatching session ends before reading a
    fairy's final report (#2760/#2766).
  • A fleet-wide per-seat supervision probe for checking process/cgroup
    health across a multi-seat deployment, with a follow-up fix so it gates
    cgroup-emptiness on real content rather than a lying stat size (#2749/
    #2758, #2790/#2822).

Changed — Baton / forum CLI (multi-agent coordination tooling)

  • baton init now requires --repo or refuses — it can no longer mint
    a bare PR anchor with no repo attached (#1924/#2915).
  • cmd_flip's CI/approval/colleague gate now fires on any maintainer-role
    flip
    , not only the literal pool-sentinel name — closing a gap where a
    differently-named maintainer flip bypassed the gate (#2912/#2914).
  • baton gained a set-colleague verb with a persisted colleague field
    (#2764/#2771), and cmd_flip now refuses an unqualified PR anchor on a
    pool-sentinel flip (#2511/#2512).
  • Step 1 of 2 retiring the single/multi-agent mode gate — not yet
    complete.
    Step 1 shipped baton, inter-agent letters, and forum
    unconditionally to everyone (#2857/#2864). Step 2 disposed of the
    remaining 6-item residue — 3 dead files plus their symlink mirrors
    deleted, 2 files reclassified into tiers.json's excluded[] (#2929) —
    but the gate itself is not fully retired: the --multi-agent build flag,
    is_multi_agent_mode(), one multi_agent: true tier entry
    (tools/collab-baton-monitor.sh), and template.CLAUDE.multi-agent.md
    all still select on it. #2857 stays open until the flag itself is
    removed.
  • Forum thread-read header now carries the post count (#2477/#2478).

Changed — Sleep cycle & lesson system

  • Blocking prevention_type lessons gained a stateless override +
    auto-demotion
    — a lesson that keeps blocking without being overridden
    eventually auto-demotes rather than blocking forever (#1944), with a
    durable write for the read-time auto-demotion lease (#2544/#2621) and an
    OSError-tolerant reconciliation walk that also closes a NULL-status gap
    (#2646, #2647/#2670).
  • Recall-summary cohort selection is uncapped for fresh items, with
    backfill capped at 60, changing how much gets summarized in a single sleep
    pass (#2819).
  • Dream-lottery's cascade-count + urgency-interrupt mechanism was retired
    now that the cascade machine's own frontier/resolve flow handles that
    work (#2399/#2823).

Fixed — Daemon liveness & socket-guard hardening (#2652)

  • A cluster of reliability fixes to start-engram-daemon.sh and the socket
    guard: the script is now an explicit four-state supervisor
    (assert-liveness-and-report) with ENGRAM_HOME normalization
    (#2700); a socket is only unlinked by the process that actually bound it,
    never a stale guess (#2713/#2723); _resolve_engram_home() now returns an
    absolute path, closing a fail-open in the socket guard under a relative
    ENGRAM_HOME (#2695/#2697); .service/.plist launch templates now bind
    their argv to the reap pattern's own ENGRAM_HOME scoping (#2686); a
    socket the reap pass didn't match is no longer force-removed (#2687); the
    liveness probe now logs which exception it caught and which branch fired,
    instead of failing silently (#2676/#2682, #2667); and a bare except: in
    the probe that was swallowing its own intentional SystemExit(0) is fixed
    (#2666/#2669).

Fixed — WAL-guard data-integrity hardening

  • The write-ahead-log guard now distinguishes "detection unavailable"
    from "confirmed healthy"
    — previously an unavailable check could read as
    a clean bill of health (#2651/#2675). Forensic capture (locks + fd→inode
    mapping) now fires on shm-displacement detection (#2635/#2653). A
    position-based fd-scan cap that was silently dropping evidence under high
    ambient fd counts is fixed (#2705/#2712), and _LOCK_ROWS_LIMIT is now
    resolved at call time rather than at function-definition time (#2714/
    #2747).

Fixed — Install / upgrade flow

  • Untagged deploys are now refused by default, with an --allow-untagged
    escape hatch and an alpha_tag marker recorded on deploy (#2586/#2595);
    the deploy-tag exception is now checked before the branch-name guard,
    fixing an ordering bug that could refuse a legitimate tagged deploy
    (#2711/#2724).
  • rc-upgrade-verify now recognizes an already-migrated seat as CLEAN
    instead of misreporting it (#2411/#2461).
  • A declared-but-never-copied build-manifest entry now fails the build
    loudly
    instead of silently shipping an incomplete install (#1962), and
    subdirectory manifest entries are no longer silently dropped by the
    manifest-driven copy loop (part of #2537).
  • Upgrade-runbook doc fixes: Step 8's interpreter-pattern matching widened
    (it missed 3 of 5 real shapes, #2797/#2799); Step 8's engram_stats check
    no longer implies it proves the server actually reconnected (#2786/#2791);
    the hooks/tools symlink note is now install-path-dependent guidance
    instead of a single wrong answer (#2560).
  • A 4-state viz-server probe now runs on SessionStart so a broken
    dashboard is surfaced at session start instead of discovered later
    (#2783/#2795).
  • Autostart heartbeat guidance is now selected per harness (Claude
    Code vs. Codex, etc.) instead of one-size-fits-all (#2508/#2517).
  • engram-nap's backup-health check, which was constant-FAILED regardless
    of actual backup state, is fixed (#2680/#2729).

Fixed — Security

  • Evidence-snapshot capture now enforces http/https and blocks internal
    hosts across redirects
    — an SSRF-hardening fix closing a path where a
    redirect chain could reach an internal host the original URL check would
    have refused (#2639, #2715/#2691).
  • The viz dashboard's seat table now escapes a.name in its row rendering,
    closing an injection path via an untrusted seat name (#2751/#2802).

Fixed — Core / MCP tool behavior

  • engram_derive's retract path now requires replacement_id to name a
    pre-created node
    , retiring the old mint-inside-retraction path
    (#2612/#2614) — a behavior change to the retract/replace flow.

  • Server now fails loud at startup when the embedding model is unavailable,
    instead of degrading silently (part of #345, #2485).

  • A local embedding-model directory now works when the HuggingFace Hub is
    unreachable
    (#2835). If your network blocks the Hub, the documented
    workaround is to download the model elsewhere, put it on disk, and point
    embedding.model at that directory. That did not work: the server refused
    to start and printed a HuggingFace HFValidationError traceback instead of
    the message that tells you what to fix. It now starts.

    If you use a local model directory, two things to check on upgrade:

    • Give the absolute path. ~ is not expanded — ~/models/my-model
      will still be refused. Use /home/you/models/my-model.
    • The directory must contain modules.json (a sentence-transformers
      export) or config.json (a plain transformer directory). A directory
      with neither is refused on purpose, because starting without a usable
      model would file nodes with no embedding at all, and those are invisible
      to semantic recall afterwards with no way to tell.

    When something is wrong, the startup message now names the directory and
    says which of those files it looked for, instead of showing a stack trace.

    Two related cases are fixed by the same change: a model name that is one of
    the older no-organization names (e.g. bert-base-uncased) is no longer
    reported as uncached when it is in fact cached, and a malformed path is
    refused cleanly rather than by raising.

  • DAG-conformance cycle checking now resolves superseded endpoints through
    the supersedes chain before walking for cycles, fixing a class of false
    cycle detections (#2813/#2827).

  • PDF-quote verification errors now name the actual format gap instead of
    accusing a correctly-quoted passage of paraphrasing (#2765/#2768).

  • _generate_snapshot's "Derives from:" line now includes
    acknowledged_derives_from premises, which were previously missing from
    the rendered snapshot (#2657/#2658).

  • Session-log re-parsing in the indexer is now gated on file size rather
    than mtime, fixing a class of missed/duplicate re-parses (#2785/#2792).

  • The health repair pass's stale .error sentinel check is now turn-based,
    so a sentinel left over from a previous night no longer skips the whole
    pass forever (4c666542, closes #2525).

  • Three per-prompt hooks' candidate-path resolution is now OSError-tolerant
    instead of raising (#2671/#2673); the surface hook's self-anchoring gate
    now excludes <loop-meta> lines (#2482/#2483); a Codex-specific bell
    notification bug (polling the wrong stream) is fixed (#2490/#2492).

Fixed — Viz dashboard (general, non-health)

  • Stats tab no longer drops the connection on a cross-user WAL read
    (#2571/#2573). health_score now fails honest on ImportError instead of
    fabricating a 0.0 score (#2524/#2763). Frozen-header typography,
    consolidated nav CSS, and --access-log flag parity between
    viz.server and viz.workbench round out this window's dashboard
    polish (#2555, #2767/#2769).

Docs

  • engram_list's claim key carrying recall_summary is now documented
    (#2694/#2721).
  • The Model Identity Gate section of the shipped identity template
    (template.CLAUDE.md) now states its actual purpose — self-recognition,
    not lineage hygiene — for anyone setting up a fresh agent identity
    (#2488).
  • README now links engram-agents.org from the public-facing page (#2153).
  • CLAUDE.md now states the review-convergence rule Lei ruled on 2026-09-13:
    reviewer-fairy APPROVE is the author's own pre-peer self-check, not a
    re-run gate on every push; the merge gate is a non-author peer APPROVE on
    the current tip (#2930).