Workflows

The repo ships seven GitHub Actions workflows under .github/workflows/. Most read model assignments from auto_tune_config.yml.

claude.yml — interactive agent

Triggered when @claude is mentioned in an issue or PR comment, or when an issue is assigned/labelled accordingly. Runs the anthropics/claude-code-action@v1 action with the model configured as models.claude_code.

Use it to ask Claude to answer questions, review changes, or implement small-to-medium tasks directly from GitHub.

claude-code-review.yml — automated review

Runs on pull requests and asks Claude (using models.code_review) to review the diff. The review focuses on correctness, readability, and security, as laid out in CLAUDE.md.

auto-improve-discover.yml + auto-improve-verify.yml — self-tuning loop

The self-tuning loop is split into two workflows that share the auto-improve label namespace but own separate halves of the lifecycle.

auto-improve-discover.yml runs on the cron schedule defined in auto_improve.schedule (and can be dispatched manually). It inspects recent workflow runs and Claude Code session transcripts, parses them via the scripts in scripts/, and maintains a long-term issue tracker for each improvement subject. When it creates a new issue it writes a ## Baseline (before fix) section into the body — a frozen snapshot of the signal counts and excerpts as they were at the moment the problem was promoted to an issue. When it can fix a problem automatically, it opens a focused PR that references the issue via Refs #<num> (never Fixes #, Closes #, or Resolves # — GitHub’s auto-close behaviour is intentionally disabled so that only the verify workflow can close). The discover workflow only advances issues along raised → pr-open.

auto-improve-verify.yml runs on the cron schedule defined in auto_improve_verify.schedule (daily by default, and dispatchable with an issue_number input to verify one issue at a time). For each target issue it resolves a WINDOW_START timestamp — the linked PR’s mergedAt if the fix PR has merged, otherwise the issue creation time — and passes it to the extractor subagent as a hard filter. The subagent then only parses workflow logs and session transcripts from GitHub Actions runs created at or after that timestamp, so conversations and logs appended before the fix cannot enter the “after” snapshot. The subagent is also scoped to the issue’s fingerprint key so it only returns matching signals. Verify compares the in-window counts to the frozen ## Baseline (before fix) section, appends a ## Verification history entry, and makes the lifecycle decision. It is the only workflow allowed to advance issues to auto-improve:merged, close them as auto-improve:solved, or reopen them on regression. There is no verification threshold to tune: a single clean verify run with zero in-window matches is enough to close, and regressions after closure are caught by the daily cron.

Session transcripts analyzed per run are capped by auto_improve.default_conversation_limit; workflow logs are always parsed in full. When the local-transcript lane is enabled (hub.local_transcripts.enabled: true) and the HUB_TOKEN secret is provisioned, both workflows fetch session transcripts published from local Docker runs via the shared hub repo before invoking Claude. The workflow-insights-extractor subagent then folds those local-run signals into the same clustering pass alongside CI transcripts. Forks without the secret provisioned skip this step silently — the auto-improve loop works exactly as before.

The goal is to let the workspace grow incrementally: when a pattern of failure or friction repeats, the loop captures the lesson rather than repeating the mistake, and a fix is only considered “done” once a dedicated per-issue run has confirmed the problem is actually gone.

hub-daily-sweep.yml — cross-workspace improvement sharing

Runs daily at 03:00 UTC (and on manual dispatch with a configurable lookback window). Phase 1 of the hub protocol: scans PRs merged into main in the last 24 hours, uses Claude to judge which are generalizable, and opens proposal issues in the configured hub repo (hub.repo in auto_tune_config.yml). This workflow is publish-only — it never adopts proposals from other forks. Gated by the hub.enabled master switch.

config-sanity.yml — dead-key guard

Runs on every PR targeting main, on pushes to main, and daily at 06:00 UTC. Executes scripts/check-config-keys.py, which enumerates every leaf key in auto_tune_config.yml and fails if any key has no reader in scripts/, .github/workflows/, or run.sh. This catches the class of bug where a config entry sits dead because the consumer was deleted or hardcodes a literal fallback.

Docs-sync agent

A daily docs-sync agent keeps pages under docs/ aligned with what landed on main in the last 24 hours. It is deliberately narrow: it only edits docs/, never proposes code changes, and uses docs/.docsrules to route changed source files to the doc page most likely to need updating. Its full instructions live in scripts/docs-sync-prompt.md.

Scripts

Supporting scripts live in scripts/: