docs: extracting-plan-dag skill + gc/non-gc orchestration strategy

[scarson]

Summary

Codifies the DAG-extraction methodology that emerged across the prior conversation about coordinating multi-agent plan execution. Three artifacts:

1. .claude/skills/extracting-plan-dag/SKILL.md — production-quality skill that chains after plan-review-cycle when the execution model warrants it. Detects gc / non-gc projects and adjusts mandatoriness of the gate (Phase 1) and tracker sync (Phase 8) accordingly. Bridges the phase-granularity Living Document Contract banners to task-granularity DAG nodes.

2. dev/research-findings/dag-extraction-and-orchestration.md — strategy + context doc explaining how the Living Document Contract, DAG extraction, and Beads-backed trackers (Gas City) layer; who has authority for what; and how the workflow stays the same on gc and non-gc projects with a small mechanical delta (skip the in-progress banner; force DAG extraction + tracker sync).

3. dev/plans/2026-03-10-phase9-health-review-remediation-dag.md — retrofit of the existing phase 9 DAG with header pointers to the methodology and an Adversarial Review section documenting the nine rounds the DAG actually went through during its initial production.

Strategy in one paragraph

Three tools, three layers, one direction of authority: plan → DAG → tracker. Banners stay; Beads adds atomic claim and queryable runtime state when the project's execution model warrants it. The gc / non-gc split is small and mechanical: detect once, drop the 🚧 banner row, force DAG extraction and Phase 8 sync. Everything else is the same workflow on both kinds of project.

Adversarial review of the new skill

Four rounds run against extracting-plan-dag per its own Phase 7 process. Findings applied between rounds; final loop pass produced zero material findings.

Round	Lens	Notable findings applied
1	Naive fresh agent	Vocabulary normalized between Phase 1 gate and Phase 8 table; idempotency-verification clarified; rebuttal for "Beads is overkill" softened to match table semantics
2	Coverage auditor	Granularity bridge added (banners are phase-level, DAG nodes are task-level — parent_phase + parent_phase_banner capture both); plan-revision-after-review now requires fresh plan-review-cycle pass; banner-state internal-inconsistency row added to Phase 9
3	Inference-discipline auditor	Strong claims softened (e.g. "single most common failure mode" → "a common failure mode"); 3-round canonical structure justified by mapping each round to a specific failure mode
4	Tool-coupling auditor (skill-specific)	Format pluralism added (Mermaid default; DOT and structured-text valid); Beads-specific examples in social proof flagged as exemplary not canonical

Test plan

• Skill renders correctly in the skill list (description visible, frontmatter parses)
• Strategy doc renders cleanly on GitHub (Mermaid/tables format correctly)
• DAG doc retrofit reads clearly and the new Adversarial review section is accurate to the conversation arc
• Cross-references between the three artifacts resolve to real paths
• No accidental commits to main; PR base is dev per worktree-flow convention

Open questions for review

1. writing-plans-enhanced integration. The new skill chains after plan-review-cycle but writing-plans-enhanced doesn't yet call it. A future change to writing-plans-enhanced adding a Step 6 ("Run extracting-plan-dag if Phase 1 gate triggers") would make this automatic. Out of scope for this PR; flagged for follow-up.

2. gc-mode wrapper vs. inline conditional. The skill handles gc / non-gc inline rather than via a writing-plans-gc wrapper. Strategy doc tracks this as a maintenance concern: if the gc-specific delta grows past ~5 items across multiple skills, a single skill with a mode: gc parameter would be cleaner than a wrapper fork.

3. Beads detection markers. The skill lists plausible markers (.gc/, .beads/, gas-city config, CLAUDE.md setting) but doesn't pick a canonical one. Worth nailing down before the first gc-project plan extraction.

https://claude.ai/code/session_01BWBzqwVmxrSJmpgoSgkx6J

---

Generated by Claude Code