TWO FINDINGS, one from cleaning up and one from refusing to keep guessing.
THE LEAK. `./scripts/test.sh` left three containers running every time — 289 had
accumulated. The cause was a comment that lied: `warm_pool.rs` said "Shutdown
destroys assigned AND pooled sandboxes", while `SandboxManager::shutdown` drains
the POOL only. Its own doc says why — assigned sandboxes persist deliberately so
a redeploy can reuse them, and production reaps the strays with
`reconcile_orphans` at boot. A test has no next boot, so each one that assigned a
sandbox simply left it running. The three tests now call the `release_agent` that
already existed, and the comment says what the code does. Verified: 0 leaked,
where the same run leaked 3 before.
THE EVAL. The independent judge failed the same correct phase FOUR times, each
time citing a different invented requirement. I blamed the condition's wording
twice and rewrote it twice — the second rewrite made it worse, by naming a
command a tool-using judge then ran in its own container. Then a control showed
the same model answering MET to the same question asked directly, and a third
wording test showed a STRICTER phrasing scoring UNMET. Prose wording was not the
variable. Continuing to iterate would have been fitting the fixture to noise.
`scripts/judge-eval.sh` measures the thing instead: five cases drawn from real
incidents, each with an answer a careful human would agree with. This project has
557 tests and had zero evals, which is backwards — a test pins OUR code, an eval
pins the MODEL, and the model changes without us touching anything.
The result is why it was worth building:
glm-4.7 4/5 — wrong on kernel-ok: says UNMET when MET
kimi-for-coding 4/5 — wrong on goodhart: says MET when UNMET
Identical scores, opposite failure modes. GLM fails good work; KIMI passes work
where 14 assertions were deleted and the failing module removed to make a suite
"pass" — the exact incident the verifying judge was built after. Swapping the
validator to Kimi because it passes our failing case would have installed a
rubber stamp. Keep GLM: a judge that is too strict costs a re-run, a judge that
is too lenient costs the guarantee.
The eval also caught a bug in itself before I trusted it: Kimi answers with a
`thinking` block first, and a 160-token budget was consumed entirely by it, which
the harness scored as NO-ANSWER. An eval that misreads a model is worse than no
eval, so it now reads thinking blocks as a fallback and has room to answer.
557 tests pass, clippy clean.
Co-Authored-By: Claude Opus 5 <[email protected]>
Root cause of the "reuse must not drain the pool" flake: the test called
`manager.exec(AgentId::new(), ...)` with a random UUID that had no agents
row. `agent_containers::upsert` uses `INSERT ... FROM agents WHERE a.id = $1`,
which silently inserts zero rows when no agent matches — so the "assigned"
sandbox never persisted to the DB. The next exec's reuse lookup returned
None, fell through to the provision branch, and popped from the warm pool
instead of reusing the assigned sandbox. When the warmer hadn't refilled by
the time we asserted, pool_size == 1 instead of 2.
Seed a workspace + owner + agent up front (mirrors soak.rs's setup). Now
upsert commits a real row, the second exec hits the reuse branch, and the
pool stays whole — the test asserts what its name claims.
The tolerant-health-check change in cm-runtime (15cffba) stays as a
defensive improvement for prod under docker daemon load, but the real fix
for the test is here.
Decouples "many users" + "many server replicas" from "many machines" so the
platform is tenant-isolated and horizontally safe on the current single node.
- Per-signup workspaces (cm-auth): a new hosted-identity sign-in provisions and
owns its own workspace instead of joining the first. Config-gated by
auth.per_signup_workspace (default off); concurrent first-logins serialized by
a per-subject advisory lock so no duplicate workspaces.
- Terminal tickets in Postgres (migration 0016, hashed, single-use): any replica
can redeem a ticket minted by another. Drops the in-process ticket map.
- Container registry in Postgres (migration 0017, agent_containers): Terminal
and Sandbox managers resolve an agent's container through a shared registry,
so a 2nd replica reuses it instead of spawning a duplicate. node_id recorded
as 'local' (Phase 2 hook). Boot reconcile removes only true orphans, so
terminals now survive a redeploy (tmux sessions resume).
- Per-workspace quotas (cm-api/quota.rs): plan-tier caps on agents + live
containers, enforced at agent create + terminal spin-up (reconnects allowed),
returned as HTTP 402. New GET /api/quota surfaces usage vs limits.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Full-depth rename per the approved plan; the 'claw' product vocabulary
(claws, /claws routes, clawId, Claw Chat) stays — it is now the brand.
- Display brand: Clawmates (manifest, titles, hero, login/rail logo
'clawmates'); default host app.clawmates.work; registry
ghcr.io/clawmates
- Crates tc-* -> cm-* (16 crates + all imports); binaries
clawmates-server/broker/bundler; images clawmates/*; env prefix
CLAWMATES_* (+ CM_TEST_DATABASE_URL / CM_LIVE_LLM); config
clawmates.toml; helm chart deploy/helm/clawmates with clawmates-*
resources; db names clawmates*; sockets /run/clawmates; cookie
cm_session; kind cluster clawmates-test; seccomp node profile
clawmates-agent-profile.json
- All 9 Playwright brand assertions updated in lockstep; historical
spec document left untouched as the only remaining 'TeamClaw'
- Local env migrated: dev pg clawmates-dev-pg/clawmates_dev, shared
test server clawmates-test-pg, kind cluster recreated with image +
profile, compose images rebuilt under clawmates/*
Verified end to end: 161 Rust + 68 frontend tests, 29 Playwright
journeys, 4 live kind tests, helm/install/LOC/placeholder gates, and
the clean-room install rehearsal serving the clawmates login page from
a signed bundle of the rebuilt images.
Co-Authored-By: Claude Fable 5 <[email protected]>
