Skip to content

Harden execution with durable self-correction - #102

Open
sfw wants to merge 5 commits into
openai-codexfrom
codex/harden-self-correction
Open

Harden execution with durable self-correction#102
sfw wants to merge 5 commits into
openai-codexfrom
codex/harden-self-correction

Conversation

@sfw

@sfw sfw commented Jul 28, 2026

Copy link
Copy Markdown
Owner

Summary

  • add a durable correction controller with typed blockers, repairability policy, deterministic handlers, structured progress detection, and restart-stable cycles
  • route verifier, tool/source, artifact-context, placeholder, remediation, iteration-gate, and replan failures through the unified correction lifecycle
  • give CSV schema mismatches a dedicated automatic repair lane with exact file targets, an LLM-generated in-place repair plan, deterministic re-verification, and one bounded attempt beyond the ordinary retry ceiling
  • stop repeat fetches of login-walled, access-denied, anti-bot, or missing URLs and redirect research toward search/alternate public sources
  • normalize equivalent tool-budget blockers into checkpoint continuations, count shrinking target sets as progress, and allow one bounded completion pass for a progressing near-complete subtask
  • make long-run failure analysis use the newest event window so recovered early 403/404 failures cannot mask the terminal cause
  • harden runner context compaction, malformed tool transcript recovery, fact-checking concurrency/timeouts, method-resilient verification, and targeted artifact refresh
  • add SQLite migration 20260728_008 for correction cycles, attempts, and idempotent actions
  • preserve fail-closed behavior for safety, path, artifact-integrity, canonical-write, authentication, and destructive-action blockers
  • replace captured execution fixtures with synthetic organizations, queries, URLs, paths, and identifiers

Validation

  • uv run ruff check .
  • uv run python scripts/check_db_migration_policy.py
  • uv run pytest -q
  • 3217 passed, 6 skipped
  • schema-repair focused suite: 112 passed
  • verifier/runner/API regression suite: 320 passed
  • privacy-focused Python fixtures: 165 passed
  • privacy-sanitized desktop fixtures: 94 passed

Historical replay coverage

  • login-walled 403/404 source fallback and repeated-URL suppression
  • CSV row-width mismatch recovery after ordinary retries are exhausted
  • 1,005-event long-run terminal-cause selection
  • tool-budget reason aliasing and checkpoint continuity
  • bounded progress extension at the ordinary retry ceiling

Notes

  • leaves the unrelated modified desktop icon out of the commits
  • existing databases migrate in place during normal startup

@sfw
sfw force-pushed the codex/harden-self-correction branch from 1555dd2 to d27e6d7 Compare July 28, 2026 19:48
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.

1 participant