You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
office-hours-spec-review E2E deterministically hits error_max_turns (9 turns) on main #2473
CI eval-store comparison: FAIL → FAIL vs the previous baseline run (same 9-turn shape)
Local run on pure origin/main @ a325940: error_max_turns, 9 turns, 33s
The deterministic 9-turn ceiling on every run suggests the scenario's turn budget no longer fits the flow (a skill prose change or CC version shift added turns), rather than flakiness. Candidate next steps: dump the failing transcript (the harness drops reasoning/tool traces on failure — see the E2E observability TODO), check what the agent spends turns 1-9 on, and either raise maxTurns or trim the scenario's prompt scope.
Not caused by PR #2472 (receipts above — identical failure on main without the branch).
Root-caused and fixed in PR #2472 (commits 7d1b9e9, c7bf129, 85804bb). Three distinct budget problems, all verified fail-first and green on CI run 3:
office-hours-spec-review: the Spec Review Loop content is carved into office-hours/sections/, so the agent needs discovery hops before writing — 8 tool turns + closing text = 9 > the pre-carve 8-turn budget. maxTurns 8 → 12.
review-dashboard-via: session startup starves under the file's concurrent sessions on a loaded runner; a clean run measures 231-270s against a 180s budget. Inner 180s → 300s, outer 240s → 360s.
retro-base-branch: clean pass measures 225-239s against a 240s budget — one second of headroom (local receipt: PASS at exactly 239s). Inner 240s → 360s, outer 300s → 480s.
Activity
fix(e2e): office-hours-spec-review turn budget fits the carved skill …
fix(e2e): review-dashboard-via session budget survives runner content…
fix(e2e): retro-base-branch session budget survives runner contention (…
garrytan commentedon Aug 7, 2026
Root-caused and fixed in PR #2472 (commits 7d1b9e9, c7bf129, 85804bb). Three distinct budget problems, all verified fail-first and green on CI run 3:
office-hours-spec-review: the Spec Review Loop content is carved intooffice-hours/sections/, so the agent needs discovery hops before writing — 8 tool turns + closing text = 9 > the pre-carve 8-turn budget. maxTurns 8 → 12.review-dashboard-via: session startup starves under the file's concurrent sessions on a loaded runner; a clean run measures 231-270s against a 180s budget. Inner 180s → 300s, outer 240s → 360s.retro-base-branch: clean pass measures 225-239s against a 240s budget — one second of headroom (local receipt: PASS at exactly 239s). Inner 240s → 360s, outer 300s → 480s.Closes with #2472's merge via the PR reference.
v1.61.0.0 fix wave: guards failing open / silent failures (9 fixes, 4…