Fix malformed Face Mind retry routing
This commit is contained in:
parent
aa0b13ae1c
commit
72089c73e4
5 changed files with 78 additions and 14 deletions
|
|
@ -17,7 +17,7 @@ Git history and smoke artifacts carry routine proof. `state/evidence.jsonl` shou
|
|||
## Current Shape
|
||||
|
||||
- Discord bot and worker may be live, but the unattended repo Face / CTB swarm brake is active as of 2026-05-26.
|
||||
- The swarm was paused because live Face turns leaked owner-reply/Mind scaffolding into public Discord: Mind was being wrapped by the owner Discord reply prompt, then a worker fallback treated plain malformed Mind prose as public `SAY`. The fixed worker stack was restarted on 2026-05-26, typecheck/build pass, `agents:interpreter-loop-smoke` passes, and `agents:heartbeat-model-smoke -- --model gpt-5.4 --reasoning-effort none` passes. The Mind prompt has now been slimmed from incident-by-incident scar tissue into semantic route/retry/drop judgment plus a hard STATE NOTE/ARTICLE/SAY DSL contract. Keep `state/agent-swarm-paused.json` active until the operator explicitly asks to thaw; before thawing, run one controlled current-room turn and inspect model-output logs for `INTERPRETATION` plus explicit DSL actions only. Bifrost dispatch remains disabled by default behind `BIFROST_DISPATCH_ENABLED=false`.
|
||||
- The swarm is paused because live Face turns leaked owner-reply/Mind scaffolding into public Discord twice on 2026-05-26. The first failure wrapped parent Mind in the owner Discord reply prompt and let malformed plain Mind prose become public `SAY`. The second thaw attempt showed the deeper split: malformed parent Mind output caused the worker to ask the Face to revise, and that retry path turned owner-question boilerplate into Bifrost/Nibu `SAY`. The current worker retries malformed Mind output as Mind-owned interpretation, drops to diagnostic private summary instead of re-prompting the Face when Mind stays malformed, and never preserves raw dropped prose as the job summary. `scripts/inspect-model-output-logs.mjs` now flags repo-face parent interpreter records missing an `INTERPRETATION` decision. Keep `state/agent-swarm-paused.json` active until a new controlled thaw explicitly proves the live worker, parent Mind, model-output inspection, and Discord side effects all agree.
|
||||
- The teardown authority remains `notes/agent-swarm-teardown-plan-2026-05-21.md`, but it is now a restart checklist and smell ledger rather than a stop sign. Keep using it when a Face/Bifrost/CTB boundary smells like prompt/parser/transport Jenga.
|
||||
- The repo-controlled pause flag is `state/agent-swarm-paused.json`. While it is missing or has `"paused": false`, swarm runners may operate. While it has `"paused": true` or is malformed, the Face heartbeat runner skips and the orchestrator skips agent-swarm organs. This is the brake. Do not route around it.
|
||||
- Postgres owns jobs, audit events, interaction memory, and rate-limit state.
|
||||
|
|
|
|||
Loading…
Reference in a new issue