todo2code

Participant: Codex (AI agent)

• Ticket: ticket-005
• Status: DONE
• Workflow state: DONE

Understanding

Ticket-004 proved that multilingual similarity is useful for ordering candidates but unsafe as relation evidence. The next candidate therefore separates recall from acceptance: retrieval finds a small shortlist, while an audited reranker must explain an accepted module using repository-owned evidence or abstain.

Before introducing another semantic stage, the current communication boundary must be measured. The governance standard names participants through user-<identity> and ai-<provider> files; those records must remain distinct from ticket specifications and must produce an actionable response owner when human and agent intent diverge.

Execution plan

1. Audit user-*/ai-* extraction and communication analysis on current todo2code tickets.
2. Add red regressions for participant filename recognition, evidence-file exclusion and response ownership.
3. Implement the minimal deterministic communication correction.
4. Re-run the corrected analysis on todo2code and external tracked projects.
5. Specify the candidate, decision, provenance and abstention contracts.
6. Add red contract tests and cross-language gold projection fixtures.
7. Implement the optional orchestration boundary outside the deterministic linker.
8. Evaluate a constrained reranker on the six gold positives and negatives.
9. Run tracked A/B on todo2code, subactor/platform and one additional repository selected from the existing seven-repository corpus.
10. Manually review every newly proposed relation.
11. Retain the implementation only if every precision and coverage criterion passes; otherwise remove it and retain the evidence.
12. Run the full release validation and update readiness documentation.

Planned code locations

• src/: public contracts and optional orchestration.
• test/: contract, hard-negative and integration tests.
• evaluation/gold/: versioned evaluation fixtures if the schema requires it.
• scripts/research/: optional manually invoked reproducer only.
• project/ticket-005/: specifications, logs, captured results and decisions only.

Risks

• A reranker may restate semantic similarity without adding evidence.
• Candidate text may bias a model into selecting a module instead of abstaining.
• Multi-module requirements may be incorrectly collapsed to one module.
• Provider-dependent evaluation may be nondeterministic or unavailable.
• Curated gold projections may overfit six examples without improving a real repository.

Guardrails

• No relation from retrieval score alone.
• No silent fallback from an unavailable reranker to raw embeddings.
• No network-dependent default or offline-CI requirement.
• No external untracked content.
• No executable files under the ticket directory.

Actual changes

• Initialized the reviewable plan only.
• No linker behavior has changed.
• Owner approved execution and added the user-*/ai-* divergence audit.
• Added section-aware conversion in src/extractors/communication.ts for governance participant files and excluded ticket evidence plus raw ai-*-logs.txt from the participant channel.
• Added explicit response ownership in src/communication/analyzer.ts to every communication issue and a separate issue for an agent claim about an unconfirmed human decision.
• Added migration warnings for unstructured participant files in src/extractors/communication.ts, normalized filename identities, ignored numeric Markdown markers and recognized bare filenames as repository paths in src/core/text.ts.
• Prevented opposite statements about two explicit, different files from becoming a false intent conflict.
• Tested historical wellmanifest/new-project prompts and agent analyses in a read-only migration captured by project/ticket-005/audit.md. Correct request/message typing produced zero issues for Opus; GPT retained three unanswered prompt fragments and no false file conflict.
• Focused communication, NL, pipeline and task-synthesis tests pass.
• Added versioned, bounded candidate and reranker result contracts in src/semantic/reranker.ts. Retrieval alone cannot mutate a graph; an accepted result must cite exact repository-owned evidence, and ambiguity or multi-module scope abstains.
• Added a strict tracked-snapshot network boundary and a research reproducer under scripts/research/; no executable source was added to the ticket.
• Added captured gold reranking fixtures to evaluation/gold/v2/dataset.json: 6/6 expected cross-language relations, 0/6 forbidden violations and one hard-negative abstention.
• Ran three live attempts on clean subactor/platform commit 3e96573; provider output violated the structured contract each time, so no relation or coverage change was accepted.
• Removed reranker exports from the public package in src/index.ts. The deterministic linker, CLI, MCP and A2A remain unchanged.

Blockers

• The evaluated provider/model does not reliably honor the structured result contract, and no real-repository coverage improvement was demonstrated. This blocks production retention but does not block closing the rejected experiment.

Conclusion

Retain the communication correction and offline evidence contracts. Reject the live semantic production path until a provider-pinned candidate passes the same real-repository boundary.

This site is open source. Improve this page.