Skip to content

Dogfood Cases

RepoKeeper needs public proof, not synthetic examples. This page tracks real runs against this repository. Merged proof cases are separated from learning cases so the page does not overstate unmerged or failed runs.

Merged Proof Cases

Case Trigger Output Result Lesson
Code quality repair RepoKeeper branch repokeeper/ci-fix-1 PR #13 Merged, CI passed Larger autonomous repairs need strong verification evidence

Learning Log

Case Trigger Output Result Lesson
Architecture diagram Issue #26 + agent-todo PR #27 Generated, closed unmerged PR body proof format works; docs-only changes need clear merge workflow
Labeler exact-edit miss Issue #28 + agent-todo Error comment Failed Exact edits must fall back to patch or re-select files
Workflow guardrail Issue #32 + agent-todo Issue #33 Correctly refused Blocking workflow edits protects repository automation
Coverage scope boundary Issue #35 + agent-todo Issue #36 Failed with no changes Broad coverage goals need task decomposition before agent execution
Draft release generator Issue #22 + Pi backend PR #31 Open, under review Medium-risk generated PRs need human review before proof claims

Architecture Diagram

Field Link
Issue #26 Add a digram for RepoKeeper using mermaid to README.md
RepoKeeper PR #27 docs: add mermaid diagram to README.md
Outcome Closed without merge

Scope: Add a README architecture diagram. This was a low-risk documentation task and a good test of issue-to-PR generation.

RepoKeeper output:

  • Changed files: README.md
  • Verification: not run
  • LLM usage: 5,380 tokens, approximately $0.000825, deepseek-chat
  • Context: 1 file, approximately 2,166 context tokens

Result: RepoKeeper opened a structured PR with changed files, risk, context, and cost. The PR was closed instead of merged, so this is proof of PR generation, not proof of accepted output.

Code Quality Repair

Field Link
Pull request #13 fix: 10 code quality and safety improvements
Branch repokeeper/ci-fix-1
Outcome Merged on 2026-05-06

Scope: Apply a batch of code quality and safety fixes, including exception types, safer git helpers, logging, and tests.

RepoKeeper output:

  • Changed files: multiple source and test files
  • Verification: GitHub checks passed for Python 3.10, 3.11, 3.12, 3.13, 3.14
  • Human review notes: Codecov reported patch coverage gaps, but the test matrix passed and the PR was merged.

Result: This is the strongest merged proof case currently available. It should be replaced or complemented with smaller issue-linked merged cases as more RepoKeeper-generated PRs are accepted.

Labeler Exact-Edit Miss

Field Link
Issue #28 bug: remove repokeeper-labeler default label added by repokeeper labeler
Outcome Failed before PR creation

Scope: Remove an unwanted default label behavior from the auto-labeler.

RepoKeeper output:

  • It acknowledged the issue and began implementation.
  • It failed with Edit target not found in src/repokeeper/labeler.py.

Result: The failure was useful. Exact find/replace edits are safe, but brittle. This case supports improving fallback behavior: retry with a patch, refresh file context, or use the Pi backend for edits that drift from selected snippets.

Workflow Guardrail

Field Link
Issue #32 Revert commit 8c7e0a720b06cb860115e069f1aa212aeb93ee23
Follow-up issue #33 Project rules prohibit modifying files under .github/workflows/
Outcome Correctly refused

Scope: Revert a commit that only affected workflow files.

RepoKeeper output:

  • Pi analyzed the requested revert.
  • It refused because the only viable change modified .github/workflows/, a blocked path.

Result: This is a positive safety case. RepoKeeper avoided changing workflow automation with a write token, which matches the documented security boundary.

Coverage Scope Boundary

Field Link
Issue #35 Increase code coverage to >= 80%
Follow-up issue #36 RepoKeeper Pi finished but produced no file changes
Outcome Failed with no changes

Scope: Raise overall project coverage after Codecov reported project coverage below the configured threshold.

RepoKeeper output:

  • Pi ran but produced no file changes.
  • A follow-up issue captured the no-change failure.

Result: This is a task-shaping failure. Large coverage goals should be split into concrete test targets, for example one module or one missing branch at a time, before invoking the implementation agent.

Draft Release Generator

Field Link
Issue #22 Support draft release
RepoKeeper PR #31 feat: add draft release generator with AI-powered release notes
Outcome Open, under review

Scope: Add an AI-assisted draft release generator.

RepoKeeper output:

  • Changed files include CLI/profile integration, a new releaser module, tests, and unexpected files.
  • Verification was not run.
  • Risk was marked medium.

Result: This should not be marketed as a successful case yet. It is useful evidence that medium-risk generated PRs need stricter file controls and verification before they become public proof.

Criteria for Future Proof Cases

A case is strong enough for the README only when:

  • The original issue is public and clearly scoped.
  • RepoKeeper generated the PR or fix commit.
  • CI passed.
  • A maintainer merged the result.
  • The case records changed files, verification commands, model/cost, and human review notes.

Target next: five small merged cases covering docs, test additions, bug fixes, review feedback fixes, and CI repair.