Record the architecture for moving code implementation AFK, decided in a design/grilling session. The owner wants the human-in-the-loop boundary to stop at design + spec: once an issue is triaged ready-for-agent, an agent should implement it test-first, push it, and see it to a healthy deploy on its own, escalating only when it can't proceed. Decisions captured: - claude-agent-service is the control plane (poller + watcher + safety); a dedicated in-cluster T3 Code instance is the executor + cockpit, because T3 can only show sessions it launched itself -> we dispatch into it (ADR 0003). - AFK code pushes straight to master; on a broken deploy it fix-forwards then freezes the broken state for forensics rather than reverting (ADR 0002). - Implementation agents use persistent per-repo checkouts + git worktrees on SSD-NFS for warm caches, reversing the throwaway-clone rule for this path because concurrency is serial-within-repo (ADR 0004). Pilot-gated: five integration unknowns must be validated against a dedicated T3 instance before the poller is wired. No code yet. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
13 KiB
AFK implementation pipeline — design
Date: 2026-06-14
Status: proposed — pilot pending (see "Pilot" below; no code yet)
Scope: A new autonomous path that turns a triaged ready-for-agent issue
into tested, deployed code with no human at the keyboard. claude-agent-service
becomes the control plane; a dedicated in-cluster T3 Code instance
becomes the executor + cockpit. Touches: claude-agent-service (new poller
- dispatch + watcher), a new T3 stack in
infra/, a shared SSD-NFS volume, and the per-repo issue trackers.
Provenance: this design is the output of a long grilling session (2026-06-14). It records the decisions and the alternatives that were considered and dropped, so the reasoning survives. The three hardest-to-reverse calls are split into ADRs 0002–0004.
Problem
Today the development flow is grill-with-docs → to-prd → to-issues → triage →
implement, and every stage is human-in-the-loop (HITL), including
implementation. The owner wants the HITL boundary to stop at design + spec:
once an issue is triaged ready-for-agent, an agent should pick it up and
implement it AFK (away from keyboard) — write it test-first, push it, and
see it through to a healthy deploy — escalating to a human only when it genuinely
can't proceed.
Two gaps block this today:
- The only existing issue→agent automation is the infra
issue-responder, which fires onuser-report/feature-requestlabels on theinfrarepo only — not onready-for-agent, not on the other sub-project repos that the general design flow produces. claude-agent-serviceonly ever clonesinfra, runs one-shot fire-and-forgetclaude -pjobs (no session, no live stream, no attach), and has no multi-repo checkout. The owner wants to watch and steer in-flight work, which the batch model can't offer.
Goal
- HITL covers design + spec only. Publishing
ready-for-agentissues is the release signal (theto-issuesquiz is the review gate). - An autonomous loop picks up unblocked
ready-for-agentissues from enrolled repos, implements them test-first, and lands them — pushing straight tomasterso CI deploys them (see ADR 0002 for the risk posture). - The owner can see all in-flight workers and converse with any of them from one UI — the T3 cockpit (see ADR 0003).
- Reuse before building: lean on the existing CI/CD chain, the design skills, T3 Code's multi-agent cockpit, and the persistence/worktree machinery — rather than hand-building a session console and a bespoke runtime.
Design
Roles: control plane vs executor + cockpit
| Concern | Owner |
|---|---|
| When to start, which issue, the prompt, the safety envelope | claude-agent-service (control plane) — poller + watcher |
| Running the agent (Claude Agent SDK), the worktree, the fleet UI | T3 Code (executor + cockpit) — one dedicated in-cluster instance |
| Build → image → deploy → rollout | existing CI/CD (GHA → ghcr → Woodpecker → Keel) |
| Issue queue + state | the per-repo GitHub issue trackers |
The pivotal constraint that forces this split: T3 can only display sessions it
launched itself — it has no command to adopt an externally-started session. So
"viewable in T3" ⟺ "launched by T3". To keep claude-agent-service in charge
and get the fleet view, the control plane dispatches into T3 rather than
running claude itself. See ADR 0003.
End-to-end flow
HUMAN (interactive session)
/grill-with-docs → /to-prd → /to-issues → /triage
└ produces ready-for-agent issues (dependency-ordered), labeled by a
trusted collaborator. Publishing them = the release signal.
══════════════════════ HANDOFF ══════════════════════
CONTROL PLANE (claude-agent-service, in-cluster)
poller CronJob (every few min):
for repo in allowlist:
skip repo if it already has an agent-in-progress issue (per-repo lock)
pick highest-priority ready-for-agent issue where:
• all "Blocked by" closed • labeled by a trusted collaborator
→ stamp agent-in-progress
→ POST /api/orchestration/dispatch (thread.turn.start + bootstrap:
create thread, prepare worktree, run setup, deliver the prompt)
EXECUTOR + COCKPIT (dedicated T3 instance, in-cluster)
runs the issue-implementer agent (our prompt) in the worktree:
read issue + AGENT-BRIEF + repo CONTEXT.md/ADRs → TDD red-green-refactor
→ commit (paraphrase issue, "Closes #N", AFK trailer) → push master
watcher (control plane) polls GET /api/orchestration/snapshot + CI:
├─ healthy ──────► comment + close issue, drop lock, notify ✅
├─ pre-push block ► do NOT push, relabel ready-for-human, escalate
└─ post-push red ► fix-forward (≤5 attempts / 60 min)
├─ recovers ► healthy
└─ exhausts ► FREEZE broken (preserve forensics),
relabel ready-for-human, hard page
Trigger & dispatch predicate
A poller CronJob (mirrors the existing beads-dispatcher pattern; stays
in-cluster because neither the service nor T3 has public ingress). It dispatches
issue I in repo R iff all hold:
Ris in the allowlist ConfigMap, and the kill switch is off;Ihas labelready-for-agent, applied by a trusted collaborator (the trust gate — on private repos only collaborators can label, so the label is the authorization; external/bot issues never auto-run);- every issue in
I's "Blocked by" is closed; Rhas no issue currently labeledagent-in-progress(the per-repo lock).
On dispatch it stamps agent-in-progress; on any terminal outcome it removes it.
Concurrency & locking
Parallel across repos, serial within a repo. Multiple repos progress at
once; at most one agent per repo (two agents in one repo would collide on the
working tree). Enforced by the agent-in-progress label as a per-repo lock.
Starting value; raise later.
Merge & failure posture — see ADR 0002
- Always push to master (no PR gate). Tests-green is the merge gate; CI + rollback are the safety net, matching the human allow-then-audit model.
- Pre-push failure (can't get green / blocked / would need a disallowed op):
do not push; relabel
ready-for-human; comment what was tried; page. - Post-push failure (CI build or rollout red): fix-forward up to 5
attempts or 60 minutes, then if still red freeze in the broken state
(preserve forensics — do not auto-revert), relabel
ready-for-human, hard page. The owner explicitly chose debuggability over availability here. - Budget:
max_budget_usd = 100per issue (time/attempt caps usually bite first).
Build/test environment & worktrees — see ADR 0004
The agent must run the target repo's test suite (TDD gate) before pushing. Therefore:
- Local toolchains scoped to the allowlist — the executor image carries only the enrolled repos' runtimes; the toolchain set grows in lockstep with the allowlist.
- Persistent per-repo checkout +
git worktreeper issue on a shared SSD-NFS volume, so git objects, installed deps, and package-manager caches stay warm across jobs. This supersedes the throwawaygit clone --localmodel from2026-06-02-parallel-execution-design.md; that rejection was correct for concurrent same-repo jobs, but the serial-within-repo choice here removes the.gitcontention it guarded against (ADR 0004). It pays off precisely becauseto-issuesclusters many slices in one repo, processed serially — slice N reuses the warm checkout slice 1 paid for.
T3 integration: thin dispatch — see ADR 0003
The control plane holds a capability-scoped orchestration:operate bearer
token (minted via t3 auth, stored in Vault, refreshed for the 1-hour expiry)
and calls T3's HTTP API:
POST /api/orchestration/dispatch→thread.turn.startwith abootstrapthat creates the thread, prepares the worktree, optionally runs a setup script, and delivers the prompt — one call spawns a worktree-isolated worker.GET /api/orchestration/snapshot→ the full fleet read-model (per-threadrunning/idle/error,hasPendingUserInput,hasPendingApprovals,branch,worktreePath). T3 has no outbound webhooks, so the watcher polls this to drive CI-watch, freeze, and label transitions.
The AFK behavior and safety (issue-implementer prompt, guardrails, always-push, fix-forward/freeze, issue integration) live in our thin layer, so T3 is a swappable, version-pinned backend — never Keel-auto-upgraded, reversible to a self-hosted runtime if it goes sideways.
Observability & interaction
The "active sessions layer" and the "attach and converse" surface converge
into one screen — the T3 cockpit: a live list of all worker threads grouped by
project; click one to stream its transcript and send it a turn. This dissolves
the earlier intermediate ideas of a generalized-breakglass console and a
raw-tmux hybrid attach — T3 provides converse / approve / resume natively
(thread.user-input.respond, thread.approval.respond).
Cross-system, durable signals the control plane still emits:
- Phase-checklist comment on the issue, edited in place as phases complete (worktree → tests-red → green → pushed → CI → deployed). Durable, low-noise, lives on the issue, doubles as audit trail.
- Loki logs labeled
{repo, issue}for deep-dive. - Presence claim per running session (
repo:<name>, purposeAFK #N), heartbeated — so AFK work shows up next to human sessions in the layer the prompt hook already injects. - Doorbell: Slack / ntfy ping on terminal states, deep-linking into the T3 thread. Notify, not control — the dedicated-Slack-control-plane idea is dropped in favour of the T3 cockpit.
Safety envelope
- Trust gate — only collaborator-labeled
ready-for-agentissues run. - Allowlist — a repo is untouchable until enrolled (prereqs: tests + GHA CI
CONTEXT.md). Start with 1–2 repos; expand deliberately.
- Kill switch — one ConfigMap flag pauses all pickup (the Keel scale-to-0 reflex, built in from day one).
- Per-repo lock — ≤1 agent per repo.
- Guardrails (reused from
issue-responder) — no PVC/PV deletes, no direct Vault edits, no force-push to master, infra changes Terraform-only, never[ci skip]. - Identity & audit — shared service identity; each commit body paraphrases
the issue and carries
Closes #N+ an AFK-agent trailer, so the commit message stays the audit trail.
Parameters (chosen starting values — all tunable)
| Knob | Value |
|---|---|
| Merge gate | always push to master |
| Post-push failure | fix-forward, then freeze-broken |
| Fix-forward cap | 5 attempts or 60 minutes |
| Per-issue budget | max_budget_usd = 100 |
| Concurrency | parallel across repos, serial within a repo |
| Repo scope | opt-in allowlist, start small |
| Progress detail | phase-checklist on issue + Loki logs |
| Alert channel | Slack (+ ntfy), as a doorbell into T3 |
| Executor | dedicated in-cluster T3 (thin dispatch), version-pinned |
Pilot — validate before wiring the poller
The thin model rests on five unknowns. Stand up the dedicated T3 instance and drive a couple of allowlist-repo issues by hand via the dispatch API to confirm each, before building the poller and committing the architecture:
- Per-thread custom agent + skip-permissions — can a dispatched thread
carry our
issue-implementersystem prompt and run unattended without stalling on T3's approval gating? (biggest unknown) - Dispatch auth — mint
orchestration:operate, store in Vault, refresh the 1-hour token. - Status/completion — drive CI-watch/freeze/labels purely from polling
GET /api/orchestration/snapshot. - Worktree reconciliation — T3's native
prepareWorktreevs our persistent-checkout-with-warm-caches; pick one or make them cooperate on the volume. - The in-cluster T3 pod — headless
t3 serve --no-browser, version-pinned and Keel-excluded, internal ingress + Authentik, with tokens / toolchains / SSD volume /claude authprovisioned.
Relationship to prior decisions
- Supersedes the worktree rejection in
2026-06-02-parallel-execution-design.md(contextualized, not contradicted — ADR 0004). - Drops two intermediate ideas explored and rejected this session:
evolving
claude-agent-serviceinto its own session/tmux/worktree runtime, and building a bespoke breakglass-generalized console — both replaced by T3. - Reuses the
issue-responderguardrails, the CI/CD chain, thebeads-dispatcherCronJob pattern, presence, Loki, and the design skills.
Out of scope / open questions
- Raw-terminal "take-over" of a worker (T3 is a GUI cockpit, not a terminal); if ever needed, that's a separate add-on.
- Multi-tenant T3 (it is single-operator by design — fine, it matches the shared service identity).
- Cross-repo dependency orchestration beyond per-issue "Blocked by".
- T3 Code is pre-1.0 (~v0.0.x) and churny; the version-pin + Keel-exclude + swappable-backend discipline (ADR 0003) is the mitigation.