study-long-lived

Metadata

Description

## Objective
Study and design a long-lived, periodically-waking "supervisor" (hard) agent inside the WG daemon — a Casa-style persistent persona (see docs/design-casa-wgfed-adapter.md) rather than an ephemeral per-task worker. It wakes on a tick (every N minutes), scans the task graph for tasks failed/stuck "for dumb reasons," and safely resets/requeues them. It keeps its own persistent memory of activity so it does not flap and can recognize recurrence.

## Research questions
- Wake cadence: tick interval vs event-driven vs hybrid; avoid thundering herd / double-acting with the existing coordinator reaper and `auto_rescue_on_eval_fail`.
- What counts as a "dumb failure" worth auto-resetting? Inventory the classes seen in this graph: `failed-pending-eval` / `pending-eval` (eval-pipeline exhaustion), `agent-exit-nonzero` crashes, dead-agent orphans, intentional-crash test tasks (e.g. `storm-source`), eval-satellite debris, respawn storms. For each: safe reset policy vs leave-for-human.
- Memory model: what activity history the supervisor keeps, where (graph task? sidecar file?), how it dedupes/recalls, and how it avoids reset loops (max-attempts, backoff, escalation to human).
- Composition with existing lifecycle: reaper, `wg recover`, `auto_rescue_on_eval_fail`, eval-lifecycle-reconcile. The supervisor must not fight them; define the boundary (what only the supervisor does).
- Failure mode of the supervisor itself: what if it hangs or mis-resets? Liveness, bounded blast radius, dry-run / audit-first mode.
- How detection results feed the rate-limit/parallelism controller (sibling studies).

## Inputs
- docs/design-casa-wgfed-adapter.md (long-lived persona pattern)
- src/ (coordinator, reaper, recover, eval-lifecycle) — survey current health/recovery code
- The hand-cleared incidents in this session (failed-pending-eval tar pits, storm-source loop) as worked examples

## Deliverable
docs/studies/supervisor-hard-agent-design.md

## Validation
- [ ] Design doc committed at docs/studies/supervisor-hard-agent-design.md
- [ ] Enumerates the dumb-failure classes present in this graph with a per-class reset/leave policy
- [ ] Defines wake cadence, memory model, loop-prevention, and the boundary vs existing reaper/recover/eval-reconcile
- [ ] Includes a fail-safe / dry-run / audit-first rollout plan and the supervisor's own failure handling
- [ ] Cites the specific source files/functions it would touch or add

Depends on

• (none)

Required by

• done .flip-study-long-lived
• done synthesis-roadmap-from

Log

• 2026-07-25T13:23:19.037793891+00:00 Task paused
• 2026-07-25T13:24:22.895142932+00:00 Task published
• 2026-07-25T14:08:29.537574911+00:00 Spawned by coordinator --executor pi --model zai:glm-5.2 --isolation required-worktree
• 2026-07-25T14:08:43.912193437+00:00 Starting: reading casa adapter doc + surveying reaper/recover/eval-lifecycle/coordinator code
• 2026-07-25T14:12:41.350200976+00:00 Survey complete: reaper(triage.rs), reconcile_durable_verdicts(eval_lifecycle.rs), respawn throttle, sweep, recover, coordinator loop (mod.rs). Writing design doc.
• 2026-07-25T14:16:45.570767942+00:00 Delivered docs/studies/supervisor-hard-agent-design.md. Validation: [x] doc committed [x] 8 dumb-failure classes C1-C8 with per-class reset/leave policy [x] wake cadence(180s slow tick)+memory model(graph-row+sidecar)+loop-prevention(cap/min-interval/per-pass)+boundary vs reaper/sweep/reconciler/recover [x] 4-stage dry-run rollout + supervisor self-failure handling(liveness/blast-radius/mis-reset/kill-switch) [x] cites triage.rs:233,coordinator.rs:4793/3990,eval_lifecycle.rs:2089/1590,sweep.rs:57,graph.rs:129/197,config.rs
• 2026-07-25T14:16:45.694250132+00:00 Committed: 0266c746 — pushed to remote (wg/agent-740/study-long-lived)
• 2026-07-25T14:17:18.171759657+00:00 Task marked as done; scheduled evaluator is advisory evidence only (execution is not a quality pass)