Move GIGAFACTORY_ROADMAP.md and GIGAFACTORY_SYSTEM_OVERVIEW.md under agent-queue/docs/gigafactory/ so the scattered top-level docs are easy to discover. Update the README links, the overview code-map, and all phase job-spec source-of-truth paths to the new location. Pure docs move; no behavior change. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
9.5 KiB
| engine | cwd | yolo | lock | timeout |
|---|---|---|---|---|
| devin | /Users/sd9235/code/mygh/learning_ai_devops_tools | true | agent-queue | 4h |
ROLE: Senior bash + distributed-systems engineer. Implement PHASE 2 SLICE 3 —
FACTORY-AGENT INTEGRATION: make the single-host agent-queue.sh runner act as a
"factory" that registers / heartbeats / claims / reports against the already-merged
fleet coordinator in platform-service, behind a feature flag, while keeping
the existing offline git-queue path 100% intact when the flag is off.
NON-NEGOTIABLE DESIGN RULE (prevents merge churn + regressions):
- Put ALL coordinator-client logic in a NEW separate file
agent-queue/lib/fleet-client.shthatagent-queue.shsources. Touchagent-queue.shonly at a few well-defined hook points (claim source, stage-transition reporting, dispatch/help). The offline git-queue code path MUST be byte-for-byte behaviorally unchanged whenAQ_FLEETis unset/0. - Gate every coordinator interaction on
AQ_FLEET=1. Default (unset) = today's offline behavior. All 53 existing selftest checks MUST still pass unchanged.
READ FIRST (verify the real contract — do not guess):
- agent-queue/agent-queue.sh — the runner. Study: the manifest/lifecycle stages
(queued→assigned→building→review→testing→shipped + blocked/failed/dead_letter),
run_worker/cmd_run/ship/promote, the Slice-4tracker_apicurl wrapper +_api_call+ awk JSON helpers (REUSE these patterns — POSIX awk, curl-only, no jq), and the Slice-4 auto-echo hooks. Mirror that style exactly. - agent-queue/selftest.sh — how stub-driven HTTP tests work (the tracker stub overrides the curl wrapper). Build the fleet stub the same way.
- THE COORDINATOR CONTRACT (read-only, in the sibling repo
../learning_ai_common_plat/services/platform-service/src/modules/fleet/routes.ts):
all routes are registered under the
/apiprefix. Exact endpoints: POST /api/fleet/factories/heartbeat {factoryId, capabilities[], health, load} POST /api/fleet/claim {factoryId, capabilities[]} -> job + leaseEpoch + lease expiry (or empty) GET /api/fleet/jobs/:id PATCH /api/fleet/jobs/:id fenced stage transition: {stage, checkpoint?, leaseEpoch} POST /api/fleet/jobs/:id/lease/renew {leaseEpoch} POST /api/fleet/jobs/:id/lease/release {leaseEpoch} GET /api/fleet/jobs/:id/runs GET /api/fleet/jobs/:id/events Note: there is NO client-side "register factory" or "append event" endpoint — registration is the heartbeat upsert, andfleet_eventsare written SERVER-SIDE by the coordinator on each PATCH/claim. The coordinator ownsleaseEpochfencing: a PATCH/renew carrying a stale epoch is rejected (409/conflict). - ../learning_ai_devops_tools/agent-queue/docs/gigafactory/GIGAFACTORY_ROADMAP.md §7 (claim loop), §8 (factory/heartbeat/claim/report/drain), §9 (split-brain/offline-degrade), §18 (fencing).
PREREQUISITE / BRANCHING:
- Branch off CURRENT
main(Phase 1 complete; foundation + hardening merged). New branch: feat/gigafactory-p2-slice3. Commit in logical steps. Push + open a PR. DO NOT merge.
CONFIG BLOCK (env, in fleet-client.sh; document in README):
- AQ_FLEET (0/1, default 0 — master switch; 0 = pure offline git-queue)
- AQ_FLEET_API (default http://localhost:4003/api)
- AQ_FLEET_TOKEN (bearer; never hardcode)
- AQ_PRODUCT_ID (reuse the Slice-4 var; X-Product-Id header)
- AQ_FACTORY_ID (default: hostname + short rand; stable per process)
- AQ_FLEET_LEASE_RENEW_SEC (default 300), AQ_FLEET_CAPS (auto-detected caps override)
DELIVERABLES
-
agent-queue/lib/fleet-client.sh(new) — a sourced library, curl-only + POSIX awk (reuse Slice-4 helpers; do not add deps):fleet_enabled— returns true iff AQ_FLEET=1 (guard for every other fn).fleet_api METHOD PATH [json]— curl wrapper adding bearer + X-Product-Id; returns body; captures HTTP code; non-2xx is logged and surfaced (never crashes the runner).fleet_detect_caps— reuse the runner's existing capability auto-detection (os, engines, tools) to build the capabilities array.fleet_heartbeat— POST factories/heartbeat (registration == first heartbeat); call at loop start + every AQ_FLEET_LEASE_RENEW_SEC during long runs.fleet_claim— POST /fleet/claim with caps; parse job id + bodyMd + leaseEpoch + lease expiry; materialize a transient local job file (reuse the Slice-4 from-tracker materialization) so the existing runner executes it unchanged. Store leaseEpoch in the job meta.fleet_report STAGE [checkpoint]— PATCH /fleet/jobs/:id with {stage, checkpoint?, leaseEpoch}. Fencing-aware: if the coordinator returns conflict/409 (stale epoch), the worker MUST self-abort the job (stop work, do NOT ship/merge) and log a fenced-abort event — a reclaimed/zombie worker can never corrupt coordinator state.fleet_lease_renew/fleet_lease_release— fenced; renew on a timer while building; release on terminal stages.fleet_checkpoint— capture {wipBranch, wipCommit} and send via fleet_report so a reclaim can resume (durability, §25).
-
Wire
agent-queue.shat MINIMAL hook points (all guarded byfleet_enabled):- source
lib/fleet-client.shnear the top. - claim: when AQ_FLEET=1 and the local inbox is empty, try
fleet_claimbefore idling (coordinator jobs interleave with local.mdfiles; local files still work). - stage transitions (building/review/testing/shipped/failed): call
fleet_report+ checkpoint — REPLACE the meaning of the Slice-4 direct tracker echo when AQ_FLEET=1 (the coordinator recordsfleet_events, becoming the audit source of truth → "tracker echo routed through fleet_events"); keep the direct tracker echo as the offline path. - heartbeat timer in the run loop; lease renew while a fleet job is building; release on done.
- new subcommands:
aq fleet-status(heartbeat + show claimable count) and surface factoryId/leaseEpoch instatus; add to dispatch + help.
- source
-
OFFLINE-DEGRADE + SPLIT-BRAIN (§9/§18): if the coordinator is unreachable mid-job, the runner finishes the in-flight job locally and reconciles on the next reachable call; on reconnect it presents its leaseEpoch — if the coordinator reports it stale (reclaimed), the local result is quarantined (marked, NOT auto-shipped) and surfaced for human triage.
TESTS — extend agent-queue/selftest.sh (stub the fleet API exactly like the tracker stub;
tests are sacred, all 53 prior checks stay green):
- flag off (default): AQ_FLEET unset → ZERO fleet API calls; existing offline flow identical (re-assert a couple of the offline cases under flag-off).
- heartbeat/register: AQ_FLEET=1 loop start → stub receives POST factories/heartbeat with caps.
- claim: stub returns a job → runner materializes a local job (bodyMd + leaseEpoch in meta) and executes it to review/.
- report + checkpoint: building/review/testing → stub receives PATCH /fleet/jobs/:id with the correct stage + leaseEpoch (+ checkpoint on building).
- FENCING: stub returns conflict on PATCH (stale epoch) → worker self-aborts, job NOT shipped, a fenced-abort is logged/surfaced.
- lease renew: long-running stub → at least one renew call with current leaseEpoch.
- offline-degrade: stub returns connection error mid-job → job still completes locally; on next call presenting a now-stale epoch → result quarantined (not auto-shipped).
- no-leak: assert the prompt/bodyMd + token are never sent in a report/comment payload they shouldn't be (reuse the Slice-4 sentinel check).
VERIFY GATE (must all pass):
- bash agent-queue/selftest.sh (all prior 53 + new fleet cases green; none weakened)
- bash -n agent-queue/agent-queue.sh && bash -n agent-queue/lib/fleet-client.sh
- node --check agent-queue/dashboard.mjs (if present/unchanged)
- shellcheck --severity=error agent-queue/agent-queue.sh agent-queue/lib/fleet-client.sh
DOCS:
- README: a "Fleet integration (Phase 2)" section — the AQ_FLEET flag, env table, the claim/heartbeat/report/fence/renew protocol, offline-degrade + quarantine behavior, and a one-paragraph "offline vs fleet mode" explainer.
- Tick the relevant §8/§9/§14 Phase-2 boxes in GIGAFACTORY_ROADMAP.md with a P2-S3 slice note.
CONSTRAINTS: bash + curl + POSIX awk only (no jq, no new deps); reuse Slice-4 helpers; never hardcode tokens/secrets; offline path unchanged when AQ_FLEET unset; conventional commits (feat(agent-queue): ...); never weaken a test; do not edit the sibling common-plat repo.
FINAL OUTPUT — print the report in EXACTLY this format:
Implementation Report — Phase 2 Slice 3 (factory-agent integration)
Branch & commits / PR
Files changed
- :
What was implemented
- fleet-client.sh: <functions + flag gating>
- agent-queue.sh hook points: <the few places touched + why minimal>
- fencing + offline-degrade + quarantine:
- tracker echo via fleet_events:
Tests added
- : (esp. flag-off no-op, claim, fenced self-abort, offline quarantine)
- selftest summary: <N checks = 53 prior + M new>
Verify gate results
- selftest / bash -n / node --check / shellcheck:
Deviations / assumptions
- <claim/lease contract details, anything stubbed, how registration maps to heartbeat>
Suggested next slice
- Phase 2 remaining: scheduler/router wiring, factory enrollment + scoped tokens, feature-flag shadow/dual-run, and the two-factory parallel demo (Phase 2 exit criteria).