Skip to content

Garry's Opinionated OpenClaw/Hermes Agent Brain

TypeScript
9.3 KB
Raw Blame

Foundations 2 notes (capy/foundations-2, v0.60.53.0)

One integrated PR. Five lanes were built in parallel from master at 5bd9e8497 and merged here with merge commits in the order F (scale and operations), A (crash robot), E (keys and grants), C (write attribution), D (silent failures). Lane E's placeholder migrations became v202 legacy_token_grant_conversion and v203 oauth_client_grant_axes, the next numbers after master's v201 (engine graduation). Generated files (migration registry, CLI flag registry, schema blobs, llms bundles, goldens) were regenerated, never hand-merged; the module-size ratchet rows that two lanes both grew were merged to measured line counts, keeping every lane's rationale.

Headline

Outcome Before After Measured on
Crash runs per full persistence gate run (600 s budget) 0 PGLite 75 (Bun 1.4.0) and 69 (Bun 1.4.2); Postgres through PgBouncer 18 per Bun version; 11 crash points; 0 violations persistence-validation.yml run 37234740186 on this branch
Local robot long run n/a PGLite 300 s budget: 40 sequence x crash-point runs, 3 process faults, 0 violations; Postgres 600 s through PgBouncer: 12 runs, 4 process faults, 0 violations integrated head 8bd4e81f9
Past persistence fixes caught when reverted (12 that reverse-apply cleanly) 0 1 on each engine (6fe43935e, hung git child) Lane A branch, would-have-caught.ts
Facts from 10 meeting pages on a served PGLite brain, no command 0 after 22 min, 10 jobs waiting 49 facts from 10 pages at about 20 min ($0.084) Lane D branch, real model
Dashboard-minted keys with full access 20 of 20 0 of 20 serve --http fixture, integrated head
Legacy tokens authorized from the JSON grant after boot 50 of 50 0 of 50 (50 of 50 requests succeed) same fixture
Unattributed writer references 90 in 40 files 7 in 3 files test/write-attribution-legacy.test.ts inventory
Scale tier, 10,000 pages, enforced n/a PGLite 25 of 25 gates, Postgres 24 of 24 scale-tier.yml run 37234737519 on this branch
Scale tier, 20,000 pages nightly red (PGLite timed out at 50 min) green on both engines, about 8 min per job Lane F run 37220459411

Foreground MCP latency with and without the live facts drain (12 pages, real model): get_page p95 3.2 to 3.3 ms, list_pages 9.1 to 8.8 ms, search 439 to 389 ms.

Scale tier at 10k on this branch: MCP search p50 37.3 ms (PGLite) and 78.6 ms (Postgres); hybrid query 25 and 201 ms; find_orphans 41.4 and 31.9 ms; import 57 s and 102 s. The Postgres planner_stats gate is report-only and reports no pg_stats rows yet for three tables right after import (autovacuum owns statistics there). Ceiling enforcement stays report-only.

Items

Lane What shipped Priority
A Crash robot over the real operations (generator, reference model, SIGKILL at every crash seam, process faults, replay and shrink); crash-robot CI job per engine with PgBouncer; restarted PGLite owner releases its dead predecessor's claims; buildHistoryFixture; lock-order trace; would-have-caught runner P0
C Unmanaged writers through maintenanceTransaction, bounded per batch; attribution-backfill verified on a 10k managed history; calibration undo-wave on Postgres P1
D Automatic facts drain on PGLite (serve, serve --http, facts_drain cycle phase) with job deferral; #5299 cycle fence; repair take-supersession (#5886); doctor revision_backfill and the backfill on managed brains (#5216); chronicle event identity and invite projection P0 drain, P1 rest
E Dashboard key mint fix (read+write default, revoke by id); v202 bulk legacy grant conversion and column reads on both HTTP transports; v203 client --sources none and --takes-holders; doctor legacy_token_null_scope P0 mint fix, P2 rest
F find_orphans known answer on Postgres; PGLite planner stats in the timeline walk; autocommit WAL checkpoint guard; migrate refusals with why and fix; sources refresh inside a resident PGLite owner; projection-readiness cache; F4d ceilings in the scale tier; seat heartbeat and pattern-page seat follow-ups P0 scale fixes and refusal, P2 rest

Every P0 item shipped. Nothing was cut.

Integration fixes

  • 8bd4e81f9: Lane F's planner-stats refresh flushed the timeline batch before page N, so one 100-row maintenance transaction split in two and Lane C's test/write-attribution-timeline-10k.slow.test.ts saw 101 transactions for 10,000 rows. The refresh now runs after every N pages, where the batch is already flushed; Lane F's planner-stats test walks 301 pages so its check after page 300 still sees 600 rows.

  • Master merge (v0.60.52.0: engine graduation, auto_chronicle date quality, contributor fixes): runMigrateEngine runs graduation's Windows platform check before Lane F's source checks; the legacy copier's PGLite refusal, which now only fires with graduation turned off, asks the user to turn graduation back on and preview the read-only plan, or to stay on PGLite and share it with gbrain mcp expose; chronicle publication applies the invite marking to the events the date screen keeps, so an ended invite is projected (no judge call) and still dated by its end; auth rescope --client takes --operations all and --sources none / --takes-holders together; core-services-1.md was split under the 60 KB cap.

  • test/e2e/graduation-clients.test.ts failed on master once v0.60.52.0 was tagged (its "older releases" included the graduation release); #6031 fixes it on master and this branch takes that version.

  • consumer.ts composes Lane A's consumer:prepared seam and first-tick PGLite releaseAbandonedClaims with Lane F's refresh-fence predicate in the idle probe; released claims are still fenced when they are claimed again.

  • gbrain sources add on a 20,000-file checkout failed intermittently on 4-vCPU runners (also on branches without this wave): the background git gc that the fixture's large commit starts pruned .git/objects/<xx> while the physical-root overlap scan walked the checkout, and the scan's readdirSync threw ENOENT. The scan now skips a subdirectory that vanished mid-walk (test/persistence-physical-root.test.ts races it against a churning directory). Reproduced 4 of 5 times on a standard-4 VM before the fix.

  • test/sweep.test.ts counted 4 raw queries instead of 3 when an earlier file in the same process had registered a local CLI writer in the shared test home: snapshot brains share one brain_id, so the maintenance principal lookup found and verified that registration. The test now runs with its own GBRAIN_HOME.

  • test/decide/retrieval-think.serial.test.ts (from #5797): the late-answer test now waits for both of its intent receipts before returning, so the late fallback_regex receipt no longer lands in the next test's count.

  • PR-budget coverage cut: the 150 s pull-request robot run on Postgres skips the seams and the fault that can wait out a dead owner's 2-minute claim lease (effect:*, consumer:prepared, publication:prepared, publication:before_publication, publication:after_publication, and the pooler disconnect fault). Before the cut that run took 669 s locally (three publication seams at about 128 s each, the pooler disconnect at 125 s). The manifest lists the cut as lease_bound_seams_skipped and lease_bound_faults_skipped; 600 s runs on pushes, schedules and manual dispatch still crash every seam and run every fault.

Gates

  • bun run verify (69 checks) and bun run typecheck.
  • bun run ci:ubicloud on 8bd4e81f9: 3,345 items, 0 failed (gitleaks 1, unit 2,468, serial 433, slow 30, E2E 412 against Postgres and PgBouncer, verify 1), 8 min 27 s wall.
  • Lane focused tests after the merge: the 65 unit, serial and slow files green locally; the 11 E2E files green in ci:ubicloud.
  • persistence-validation.yml (manual run, every Bun version): all jobs green.
  • OpenClaw native context-engine startup test against pgvector pg16 with openclaw 2026.9.4: 1 of 1 pass.
  • uvx semgrep scan --config p/default --config p/typescript --error --baseline-commit origin/master: 0 findings on 177 changed files.
  • bun run wave-security-scan origin/master..HEAD: no obfuscation or eval; admin/dist changed and rebuilds byte-identical from admin/src; one gitleaks hit, a function-body sha256 in the migrations golden, which .gitleaks.toml allowlists.
  • bun run check:agent-contract: OK.
  • gstack credential pre-push guard on every push: no HIGH findings (MEDIUM hits are minified constants in admin/dist and a version-shaped string).

Deferred (TODOS)

P1: a held sync lock reaches agents as internal_error (skipped test; the agent-output contract owns it); ENOSPC and lost-exit process faults; a seeded mutation pass; source and writer-lifecycle ops in the robot. P2: a restarted Postgres owner waits out its predecessor's claims; a CI check that every mutating operation is registered with the robot; pre-activation claim lock order; the three link-extraction writers' attribution; one write path for unmanaged brains; the permissions mirror after 2026-11-04; Postgres MCP search at 20k pages; chronicle occurrence status, keyless invite projection and invite-aware cost accounting. P3: forget as backfill proof, register-client --takes-holders, embed --stale budget overshoot, findOrphans paging.