Repository navigation
Prism runtime on T3 only: delete legacy path, models and limits from the T3 snapshot, drop the skill #116
Description
Activity
Evidence from the live-proof job router-80-release-t3 (2026-09-25), for stream B: the first Muse worker implemented #80 correctly and opened PR #112 within 11 minutes (commit f7cc65c). The job then looped for 1.5 hours and 3 more workers because (1) its proof failed twice only on the known flaky test_supply52 (#108), counted as the worker's failure; (2) after a Kimi turn passed the proof with no new changes (the fix was already committed), the dispatcher judged "no changes, no PR" although PR #112 existed on the same branch. Fixes for this stream: run the proof once on the base commit and ignore failures that already fail there (already in scope), rerun a single failing test once before counting a proof failure, and give the dispatcher the job's branch commits and existing PR so it does not mistake finished work for no work.
Contract update for stream B: stream A is changing role preferences to per role and per lane (easy, medium, hard; ordered {model, effort} lists, first primary, rest fallbacks). Read that shape from the snapshot once A pushes the updated contracts commit on feat/19-prism-toolkit. Details on #115.
- added a commit that references this issue
on Sep 25, 2026
Part of #115 (the spec; read it first). Stream B: router runtime.
Scope
resetsAt, 80 % degraded, missing reset re-checked on next refresh; Muse free keeps its own mark from the error'snextin epoch ms). Delete the hard-coded route inventory, Go cost sums, probe readers, reset parsing andprobe_due_routes. Build against a fake T3 server first so this does not wait for stream A.Proof
python3 scripts/test.py; fake T3 server tests for eligibility and limits. Do not release or install while jobrouter-80-release-t3is running.