Artanis fleet-overseer automation: autonomous control loop orchestrating stress + healing + scaling + external-yield (on artanis-administrator-tick, approval-gated)

[AtlantisPleb]

Owner-directed strict-bug-override: GitHub issues in this repo are normally strict, reproducible bugs only (apps/openagents.com/AGENTS.md). This issue is filed under explicit owner direction. Discussion/commentary belongs in the Product Promises Forum.

Child of #6316.

Why

The continuous stress harness, external-wins admission, reliability hardening, and throughput tuning all need a continuous operator. The owner wants Artanis to be that autonomous overseer -- watching fleet health + throughput + external demand, orchestrating the stress load (start/scale/back-off), triggering heal/scale/quarantine within approval-gated authority, and reporting -- as an extension of the existing artanis-administrator-tick + artanis-approval-gates, not a new system. Today Artanis runs a bounded live autonomous loop fenced by typed schemas and approval gates, but it does not read GPU/replica/throughput/demand; the GPU sensor (glm-pool-heartbeat.ts) runs as a separate tick. This wires Artanis onto that sensor as the orchestrator.

Scope

• New artanis-fleet-overseer-tick.ts exporting runArtanisFleetOverseerTick(db, deps) + a ...Scheduled Effect wrapper, mirroring runArtanisAdminTick/runArtanisAdminTickScheduled exactly: env-gated by ARTANIS_FLEET_OVERSEER_ENABLED, self-bounded cadence, every outcome a D1 row in artanis_fleet_overseer_decisions, schema-invalid -> blocked. Registered as one new observedEffect('ArtanisFleet.tick', ...) in the Worker scheduled Promise.all beside ArtanisAdmin.tick.
• Watch (assembleContext): read the existing glmPoolHeartbeatRoutingStateOracle (per-replica health/warm/draining + the new live-headroom read surface), aggregate throughput/goodput from telemetry, and external demand from the token-usage ledger. Do NOT build a new fleet collector -- the GLM heartbeat stays the sensor.
• Decide: reuse artanisMindComplete with a bounded S.Union action vocabulary validated by a typed schema.
• Act -- authority split:
  • Autonomous (no spend, non-destructive): start/scale/back-off the internal stress load (yield to external demand using the same external-wins signal); re-admit recovered / warm idle owned replicas; emit public-safe health/throughput reports.
  • Owner-gated (pending ArtanisApprovalGateRecord, effective only with operator approval + receipt): request paid scale-out (reuse provider_call/deployment); quarantine a replica (add a new fleet_mutation risky kind to ArtanisRiskyActionKind + ARTANIS_RISKY_ACTION_KINDS + rollbackRequiredKinds, with test + INVARIANTS.md update in the same change); any wallet spend/settlement (already gated, never widened).
• Report: fold a public-safe fleet health/throughput/goodput signal into the existing artanis-health.ts snapshot (new ArtanisHealthSignalKind) so stale/blocked fleet health structurally blocks overclaiming.

Acceptance (measurable)

• The overseer tick runs on the scheduled handler, env-gated, self-bounded, with every outcome a D1 row; a schema-invalid mind proposal yields a blocked row and dispatches nothing.
• Autonomously starts/scales/backs-off the stress harness keyed on live external demand, with 0 external-request failures during a demand spike.
• A replica-quarantine or paid-scale-out proposal emits a pending approval gate and does NOT execute until artanisApprovalGateEffective (operator approval + receipt); wallet spend stays gated.
• A fleet health/throughput signal appears in the Artanis health snapshot and a stale/blocked signal blocks overclaiming.
• Public-safe: no raw origin URLs, IPs, bearer material, prompts, or wallet material in any projection.

Refs

• docs/inference/2026-06-25-glm-fleet-max-throughput-stress-and-artanis-overseer.md (§5)
• apps/openagents.com/workers/api/src/artanis-administrator-tick.ts, artanis-approval-gates.ts, artanis-scheduled-runner.ts, artanis-health.ts, artanis-spend.ts, artanis-mind.ts, inference/glm-pool-heartbeat.ts
• The continuous stress harness, external-wins admission, and reliability hardening issues; MASTER: Maximize GLM-5.2-REAP usage in Khala (full pool, durability, throughput, quality) #6316

Strict-bug-override: filed as direction work under owner mandate, not a strict bug.

[AtlantisPleb]

#6321 safe-prep slice landed on main: 396ce3e5384f1daa7ab1a61d782805b5ce75b2ab (Gate Artanis fleet mutations behind approvals).

What changed:

• Added fleet_mutation to ArtanisRiskyActionKind / ARTANIS_RISKY_ACTION_KINDS.
• Made fleet_mutation rollback-required, so replica quarantine/drain-style fleet mutations cannot be treated as effective without an operator-approved gate plus rollback refs.
• Added regression coverage proving a pending fleet mutation is ineffective and public projections redact rollback/operator/private refs.
• Updated apps/openagents.com/INVARIANTS.md because adding a risky action kind is a policy-surface change.
• Updated the GLM/Artanis overseer design doc and Khala roadmap to state this is authority scaffolding only: no fleet-overseer scheduled tick, stress traffic, scale-out, quarantine, or spend is armed.

Pylon/Codex delegation proof:

• assignment: assignment.public.khala_coding.chatcmpl_91fef54d53024affbe6f932e1fc3fb3a
• durable request: chatcmpl_3cd73fdbe42e4c24abca3b12f6c881d4
• closeout: assignment.closeout.361664b11c56442300f87421
• exact tokens: 5,988,903 (5,959,054 input, 29,849 output, 6,586 reasoning, 5,641,344 cache-read)
• traces: 148 owner-only ATIF traces
• raw events: one owner-only raw Codex archive, 258 SDK events / 4,798,646 bytes
• public counter baseline/post: 463,848,927 -> 470,504,517; assignment proof remains the exact attribution source because the public counter is global and closeout-based.

Verification:

• bun run --cwd apps/openagents.com/workers/api test -- src/artanis-scheduled-runner.test.ts src/artanis-health.test.ts src/artanis-approval-gates.test.ts src/inference/glm-pool-heartbeat.test.ts src/inference/benchmark/stress-saturation-plan.test.ts (41 tests)
• bun run --cwd apps/openagents.com typecheck:api
• bun run --cwd apps/openagents.com check:deploy (also re-run by pre-push hook; green)

Still open: the actual env-gated overseer tick, health integration, stress orchestration, and live external-yield proof after #6317/#6318. This patch only adds the destructive-action approval guard needed before that future loop can safely exist.

[AtlantisPleb]

Resolved by #6597.

[AtlantisPleb]

Reopening: #6597 was merged but reverted (1a6def5) to keep main green. artanis-fleet-overseer-tick.ts adds 3 Effect.runPromise calls outside the named temporary bridge allowlist, failing the zero-debt architecture check. Needs a compliant redo (use the allowlisted bridge or restructure to stay in Effect).