todo2code

Participant: Codex (AI agent)

• Ticket: ticket-002
• Status: COMPLETE
• Workflow state: DONE

Understanding of the task

The objective is not merely to prove that todo2code completes on other repositories. The work must establish whether its semantic conclusions remain useful outside its own codebase, identify recurring causes of weak coverage or false diagnostics, and improve the library only where repeated measurements justify the change.

Included scope

1. Create isolated detached worktrees for the recorded external commits.
2. Run one normalized offline pipeline and reality report per repository.
3. Persist a compact machine-readable baseline and a reviewed Markdown report under this ticket.
4. Compare relation classes, diagnostics, unsupported languages, topic status and coverage rather than relying on record count alone.
5. Review representative false positives and false negatives.
6. Select the highest-impact shared defect that can be fixed without accepting ungrounded evidence.
7. Add gold/unit coverage, implement one correction and rerun the same corpus.
8. Record the delta and either retain or reject the correction.

Excluded scope

• Mutating, committing or cleaning external repositories.
• Reading private or untracked external inputs.
• Tuning a threshold only to improve headline coverage.
• Provider-dependent LLM calls in the primary baseline.
• Adding a new dependency without a separate license and security review.
• Implementing several semantic heuristics in one unmeasurable batch.

Execution plan

Phase 1 — reproducible baseline

1. Verify stable todo2code and Docker validation commands.
2. Define the shared document/task/communication policy and explicit repository exceptions.
3. Analyze the seven verified repositories at recorded detached commits.
4. Store per-repository JSON metrics, warnings and sampled diagnostic evidence.

Phase 2 — evidence review

1. Rank recurring gaps by frequency, severity and affected repositories.
2. Separate extractor, target-resolution, linker, diagnostics and unsupported-language failures.
3. Choose one defect with evidence in at least two repositories.

Phase 3 — one controlled improvement

1. Add a gold or focused unit regression, including a nearby negative.
2. Implement the smallest deterministic correction.
3. Run gold v2, focused tests and the unchanged external corpus.
4. Keep the change only if the target metric improves without a measured precision regression.

Phase 4 — validation and conclusions

1. Run the complete stable validation matrix and Docker checks.
2. Update ticket evidence, changelog, acceptance criteria and readiness conclusions.
3. Present the next ranked improvement as a separate continuation decision.

Candidate hypotheses, not decisions

• PL documentation to EN identifiers is still a measured knownGap.
• Changelog claims may lack implementation evidence because topic matching intentionally excludes changelog records.
• Configuration-only evidence may overstate aligned.
• Unsupported PHP and other languages may dominate reality gaps in some repositories.

The baseline decides which hypothesis is addressed first.

Approval gate

Approved by the user’s kontynuuj message on 2026-07-31 under P-CORE-008. Execution may proceed within the recorded scope.

Actual changes

• Initialized the standard ticket structure and project-level TODO entry.
• Verified Docker availability and the seven candidate repositories.
• Verified ticket formatting, absence of local absolute paths and compatibility with the generated-analysis guard.
• Ran the normalized deterministic pipeline successfully on all seven detached, tracked-only external worktrees.
• Preserved the complete baseline in baseline.json and its reviewed summary in baseline.md.
• Selected non-actionable changelog mechanics as the first controlled defect: it repeats across the corpus, but can be corrected without pretending that ungrounded release claims have implementation evidence.
• Added a focused red/green regression and a narrow changelog-signal classifier.
• Evaluated only this patch on the unchanged external corpus: graph fingerprints remained stable, gold v2 stayed perfect, and false review-required findings fell by 1,024 across five repositories.
• Added an independent red/green correction for generated-analysis verification: tracked audit quotations no longer masquerade as private input consumption, while newly introduced untracked references remain blocked.

Unfinished items and blockers

• No blocker inside ticket scope. Remaining library gaps are listed in docs/READINESS.md; they require separate controlled iterations.

This site is open source. Improve this page.