feat(cell): uptake cell-harness 0.3.0, agent-protocol 0.2.0, celld v0.3.0

[coccyx]

Takes into APM the framework changes that landed while Kidder was being built: the two npm packages, the celld 0.3.0 runtime, and the infra + docs changes needed to consume them.

Stacked on #155 (refactor/cell-infra-fleet-module) — the diff here shows only the uptake chunk. Merge #155 first.

Commits

• 5dea567 — chore(deps): cell-harness 0.3.0, agent-protocol 0.2.0
• cd73be7 — feat(cell/infra): celld v0.3.0, pinned module ref, optional secrets
• 839ec0e — docs(cell): DEPLOY/README for celld 0.3.0 + harness 0.3.0 vars
• a session log commit

1. Deps — every change is additive, so no APM source edits

@criblio/agent-protocol  ^0.1.1 → ^0.2.0
@criblio/cell-harness    ^0.1.1 → ^0.3.0

Verified rather than assumed:

Framework change	Why no APM edit
Env extends CellEnv	CellEnv is a name-identical superset of what APM declared — the node's existing /etc/celld/env still binds. No SSM/Terraform change.
createTools(env, sql?)	Optional param; APM's payload is API-backed and ignores sql.
imageCount? on tool frames	Optional, and the UI imports wire types from @criblio/agent-protocol rather than mirroring them, so the type widens for free.
agent_images table	CREATE TABLE IF NOT EXISTS in the session DO constructor ⇒ existing DOs gain it on rehydration. No migration.
DO class names	CoordinatorDO / InvestigationDO unchanged ⇒ durable state stays addressable across the deploy.

The headline behavior change is the bounded watchdog: MAX_WATCHDOG_ATTEMPTS = 3, counted in durable storage under watchdogAttempts, written before the turn (the isolate dying is the event being counted). Previously unbounded — a turn whose tool overran celld's handler budget killed the process and got retried forever, restarting the node every ~240s.

2. Infra

• celld v0.2.0 → v0.3.0.
• CELLD_DURABILITY pinned to fleet. v0.3.0 added a replicated write-behind log and moved the default write-ack posture bucket → fleet. That reads like a regression on a single-node fleet but isn't: with no peers celld behaves exactly like sync-to-bucket (bucket-proven acks) and starts using peers the moment one joins. Pinned rather than left to the default so a future celld default change can't move this node's durability without showing up in a plan.
• CELLD_HANDLER_BUDGET_S pinned at the default, 300. Exceeding it kills the celld process, not just the isolate, so every session on the node dies — this is the budget the bounded watchdog stops retrying against. Pinned for visibility.
• Module source is a pinned git ref, not a sibling checkout. The app stopped consuming the framework from file:../ when it moved to GitHub Packages, so nothing guarantees a checkout sits beside this one and terraform init on a fresh machine has to resolve without one. The repo is public, so https needs no credentials.
• required_secret_keys = [UI_BEARER, TICKET_SECRET, LLM_API_KEY] — only these three abort the boot. WEBHOOK_BEARER only guards POST /alerts/fire, and the coordinator pulls firing alerts from $vt_results now, so autonomous investigation survives its absence; CRIBL_CLIENT_ID/SECRET are checked at use. Skipped keys land in /etc/celld/missing-secrets.
• Staging config moved from -var flags into variables.tf defaults. All of it renders into user_data under user_data_replace_on_change = true, so an apply that forgot a flag would replace the node into stub mode (empty llm_base_url ⇒ no agent loop) with no error. bucket_name is the only flag now.
• eslint.config.js ignores **/.terraform/ — the pinned git module means terraform init drops a full framework clone there and lint failed on vendored source. It's gitignored, but eslint doesn't read .gitignore.

Test plan

Local (all green on the tip):

npm run lint      # 0 errors
npm test          # 32 files, 440 tests passed
npx tsc --noEmit  # rc=0
cd cell && npx tsc --noEmit  # rc=0

Validate on staging — already done for the bundle

The new bundle is deployed and verified on the live node; sequenced deliberately one variable at a time so a bad bundle would still have a working node to roll back on.

• celld deploy → version 9d5358e345bf0c41 (was 7bd185e1a91dd9d7), with deploy/current.json confirmed pointing at it before the restart. celld version IDs are content hashes and current.json is a separate pointer, so the bucket can hold a newer bundle than the pointer selects — that exact trap left the node on old code all of the previous day.
• Bundle diff proves the new harness is in the artifact: watchdogAttempts present in the new index.js, absent from the old.
• Restarted via SSM (graceful SIGTERM). No state lost: 30 investigations before → 30 after, 0 lost, 0 new, 0 title/status drift; all DOs restored fresh=false; journalctl -p warning → No entries; Started celld = 1 (no crash-loop).
• Real end-to-end investigation on the new bundle (inv-ed5faf29-…): 69 frames, 19 real tool calls against live staging, found 393 error oteldemo.PaymentService/Charge spans at sub-ms latency.

Anyone can re-check right now:

curl -s https://54-71-34-177.sslip.io/healthz     # {"ok":true,"disabled":false}

Known limitations

• The terraform apply has NOT been run. The celld 0.2.0 → 0.3.0 binary upgrade replaces the instance, so it needs a human at the confirmation prompt. Plan verified, blast radius from the plan JSON: module.fleet.aws_instance.cell (replace on user_data) + module.fleet.aws_eip_association.cell (replace on instance_id), 10 no-ops — 2 to add, 0 to change, 2 to destroy. The EIP is a separate resource so cell_url stays https://54-71-34-177.sslip.io and durable state survives in the bucket. Pre-flight done: user_data rendered through templatefile and bash -n'd, and the celld v0.3.0 aarch64 asset returns 200 (a 404 would brick the replacement node). The live node therefore still runs celld 0.2.0 with neither new env var; the bundle is already in place, so that apply is the only remaining step.
• The verification investigation ended failed on "hit the 12-turn cap without concluding". Not a regression — the identical string is in the old bundle, and autonomous mode is hard-capped at MAX_TURNS = 12 regardless of TURN_BUDGET (sessionDO.js:832: interactive ? row.turn_budget : MAX_TURNS). Worth revisiting on its own.
• Downgrading off celld v0.3.0 is one-way-ish: do not start a v0.2.x binary on a node that has run v0.3.0 unless its shutdown log shows node-log close: sealed epoch. Documented in variables.tf and cell/infra/README.md.

Session log: docs/sessions/2026-08-22-uptake-cell-harness-0.3.0.md

🤖 Generated with Claude Code

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	react-router-dom@​7.18.1 ⏵ 7.18.2				-1
	@​criblio/​agent-protocol@​0.2.0	N/A	N/A	N/A	N/A	N/A
	@​criblio/​app-tooling@​0.2.0	N/A	N/A	N/A	N/A	N/A
	@​criblio/​app-utils@​0.5.1	N/A	N/A	N/A	N/A	N/A
	@​criblio/​cell-harness@​0.3.0	N/A	N/A	N/A	N/A	N/A
	@​criblio/​cell-workspace@​0.1.1	N/A	N/A	N/A	N/A	N/A

View full report