deadfish-docs

deadfish-docs — Track-Boundary Reconciliation

7 Living Docs (in docs/living/)

Doc	Budget (chars)	Content
TECH_STACK.md	3200	Languages, frameworks, deps, versions
PATTERNS.md	3200	Architecture patterns, conventions, idioms
PITFALLS.md	2800	Known gotchas, footguns, anti-patterns
RISKS.md	2000	Security, operational, business risks
PRODUCT.md	2800	Features, API surface, user-facing behavior
WORKFLOW.md	2800	CI/CD, scripts, deployment, dev workflow
GLOSSARY.md	2000	Domain terms, abbreviations, naming
Total	~18800

Trigger and Scope

• Reconciliation starts only when .deadfish/reconcile/<track_id>.trigger exists.
• This skill is track-boundary only. No per-task DOCSYNC, no scratch-buffer lifecycle.
• Default cadence: exactly one reconciliation run per completed track.

Role Wiring (Required)

• Doc-keeper (Haiku): Phase 1 proposal generation only.
• Reviewer A (Conductor / Opus 4.6): independent review verdicts.
• Reviewer B (Planner / GPT-5.2 high via mcp__codex-planner__codex): independent review verdicts.
• Integrator (Sonnet): applies approved diffs, writes reconciliation record, commits once, deletes trigger.

Phase 1 — Proposal (Doc-keeper)

For each of the seven living docs:

1. Read tracks/<track_id>/SPEC.md.
2. Read full track diff (base_commit..HEAD) and relevant commits.
3. Compare implementation evidence to current doc content.
4. Produce one proposal record per doc:

doc: TECH_STACK.md
action: UPDATE # UPDATE or NOP
proposed_diff: |
  + - **PyYAML 6.0**: YAML parsing for sentinel blocks.
rationale: Track added parse-blocks.py importing yaml.
evidence:
  - file: bin/parse-blocks.py
  - file: requirements.txt

Phase 1 output is proposals only. Doc-keeper does not apply edits.

Phase 2 — Multi-Model Debate

Run for each proposal where action=UPDATE.

Round 1:

• Reviewer A and Reviewer B independently return:
  • verdict: APPROVE | REJECT | MODIFY
  • reasoning
  • concerns
  • modified_diff (required for MODIFY)

Decision table:

• Both APPROVE -> auto_approved (apply original proposal)
• Both REJECT -> auto_rejected (skip apply)
• Both MODIFY with semantically equivalent diff -> auto_approved_modified
• Any disagreement -> continue to Round 2

Round 2:

• Each reviewer sees the other review and re-evaluates.

Round 3:

• Each reviewer sees full debate history and emits final verdict.
• If still unresolved after Round 3 -> escalated_to_human.

Phase 3 — Apply, Record, Cleanup (Integrator)

Integrator actions:

1. Apply all auto-approved diffs to docs/living/*.
2. Escalate sustained disagreements with both reviewer perspectives.
3. Commit once:
   • docs(deadfish): Reconcile docs for track '<track_id>'
4. Write .deadfish/reconcile/<track_id>.yaml.
5. Delete .deadfish/reconcile/<track_id>.trigger.

Reconciliation record schema:

track_id: v32-realignment
timestamp: 2026-02-09T14:30:00Z
trigger: .deadfish/reconcile/v32-realignment.trigger
docs_evaluated: 7
proposals:
  TECH_STACK:
    action: UPDATE
    debate_rounds: 1
    outcome: auto_approved
    reviewers: {opus: APPROVE, gpt52: APPROVE}
docs_changed: [TECH_STACK]
commit_sha: abc123
debate_log:
  - doc: TECH_STACK
    round: 1
    reviewer: opus
    verdict: APPROVE
    reasoning: Dependency manifest changed with matching implementation use.

Required top-level keys:

• track_id, timestamp, trigger, docs_evaluated, proposals, docs_changed, commit_sha

Docsync Sentinel

Doc-keeper emits reconciliation-only sentinel:

action: RECONCILE
track_id: <track_id>
trigger: .deadfish/reconcile/<track_id>.trigger
phase: proposal|debate|apply
status: READY_FOR_DEBATE|AUTO_APPLY|ESCALATE_HUMAN|COMPLETED
summary: <concise track-boundary reconciliation status>

Budget Enforcement

If a doc exceeds 80% of budget, prefer compression edits before expansion.