fix: flywheel kill outcome stamping#21
Open
n1arash wants to merge 3 commits into
Open
Conversation
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…urns (VisionForge-OU#1) Turn-budget exhaustion was the dominant failure in the dogfood soak test (49% of runs killed_turns) because the per-run turn cap had no relationship to the model running the work — a 30-turn cap starves a small/cheap model. - turns.py: pure policy. classify_model (tier by family substring; unknown -> small/generous), effective_turns (exact per-model pin, else floor = max(configured, tier_floor x phase_factor)), resolve_budget. - config.py: turn_budget_by_model / turn_tiers / phase_turn_factors; extension_wall_min (30) + extension_cost_usd (3.0); max_turn_extensions default 2 -> 6 (now a backstop). Validation + round-trip + installer YAML. - runner.should_extend: cumulative wall + cost ceiling is the primary stop; count is a backstop (0 disables it). New run_duration_min helper. - Wired resolve_budget + chain wall/cost tracking into all 3 extension loops (pipeline._spawn, issue_run worker, scheduler _run_agent_with_extensions) and the init spec. - Loud killed_turns: TUI worker_finished flags a non-clean terminal_reason; BuildReport gains a "Turn-killed runs" section populated from usage records. Tests: +38 (test_turns, config, budget_policy, controller, pipeline wiring, scheduler report). Full suite 406 passed. Existing turn-kill tests neutralize the new floor (they target the extension loop, not budget sizing). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The dogfood campaign drafted 0 retro proposals while 49% of runs ended killed_turns. Intermediate turn-extension runs (and phase agents) were persisted with a blank outcome, the read side maps blank -> legacy, and cluster_failures skips legacy — so the dominant failure was invisible and retro correctly-but-uselessly found nothing. - metrics: add completed / killed_* / error as first-class outcomes, KILL_OUTCOMES, and terminal_outcome() - state.write_run_record: default a blank outcome from terminal_reason at the one choke point every run flows through (worker intermediates, phase agents, bounced attempts) — no terminal run is ever persisted blank again - retro._signature: cluster the kill reasons (surfacing killed_turns); exclude the non-failures completed / killed_user - retro.propose_for_clusters: a high kill rate (>=20% of runs, >=2 runs) deterministically drafts a concrete, gated fix — killed_turns proposes the turn-budget fix directly - driver.analyze: prepend the deterministic kill-rate proposals (deduped) so a silent analysis agent no longer means "found nothing" Full suite: 417 passed. 11 new tests cover every acceptance criterion, including a regression: a fixture of killed runs yields a non-empty proposal set. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
#2