Skip to content

feat(scripts): backfill_tasks — re-run only missing/failed tasks and merge into existing row#43

Merged
jonathanbbechtel merged 4 commits into
mainfrom
feat/backfill-missing-tasks
Jun 11, 2026
Merged

feat(scripts): backfill_tasks — re-run only missing/failed tasks and merge into existing row#43
jonathanbbechtel merged 4 commits into
mainfrom
feat/backfill-missing-tasks

Conversation

@jonathanbbechtel

Copy link
Copy Markdown
Collaborator

Reconstructs TaskRunResults from result.json + raw_outputs.jsonl, re-runs
only the named task IDs (fixtures resolve via explicit pack_id), re-aggregates
with summed judge cost, and rewrites the row in place with .bak backups.
Clears stale failures.json when the row becomes complete. Validated end-to-end
with the stub adapter (overwrite-merge and fresh-row paths).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com

JonathanBechtel and others added 2 commits June 10, 2026 20:22
…ces')

OpenRouter returns error objects with HTTP 200 for some failures (context
length exceeded, provider moderation). The adapter crashed with an opaque
KeyError('choices'); now it raises RuntimeError with the provider's actual
error message. Found live when gpt-oss-120b (131k context) hit the ~184k-token
operations tasks.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…merge into existing row

Reconstructs TaskRunResults from result.json + raw_outputs.jsonl, re-runs
only the named task IDs (fixtures resolve via explicit pack_id), re-aggregates
with summed judge cost, and rewrites the row in place with .bak backups.
Clears stale failures.json when the row becomes complete. Validated end-to-end
with the stub adapter (overwrite-merge and fresh-row paths).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@pearleng-atlantis

Copy link
Copy Markdown
Error: This repo is not allowlisted for Atlantis.

JonathanBechtel and others added 2 commits June 11, 2026 10:25
Mid-stream connection drops ('peer closed connection') killed whole tasks —
three leaderboard rows were damaged this way. Both HTTP call sites (candidate
run() and the shared post_chat_completion used by the judge) now retry up to
GRADE_OPENROUTER_RETRIES attempts (default 3) with exponential backoff on
httpx.TransportError and 429/500/502/503/504 responses. Timeouts still
propagate unretried (each attempt already waits the full configured timeout).
run()'s hardcoded 120s timeout now honors GRADE_OPENROUTER_TIMEOUT too.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@jonathanbbechtel jonathanbbechtel merged commit 81e1608 into main Jun 11, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants