[dominno]

Summary

Applies the three HTDAG planning disciplines to this repo's meta-layer (the autonomous /loop) — building on the existing loop + run-log + memory wiki, without any new runtime (no pueue/hud/opencode analog). This is an additive discipline: it only ADDS constraints on top of the existing safety gates, never weakens one.

Scoping note: the product backlog (2 PRD stories, both done — deliberately a workflow test-stub) is the wrong place for a dependency DAG; the loop machinery is the layer with real typed activities, dependencies, and re-planning. So the DAG lives over the loop, not the stories.

What changed

The three disciplines

1. Persistent plan/DAG — new .claude/memory/loop-plan.md: typed nodes with id / description / after: edges / status / origin, a ready-node rule (deps done), and cascade + stale-recovery rules. The loop reads it at iteration start and rewrites it in the re-plan step. It is loop state (sibling to loop-run-log.md), not a knowledge topic.
2. Re-plan as a first-class step — loop.md step 9, positioned before dispatch so the loop always consults the fresh DAG (HTDAG's "consult the DAG before every dispatch"). Records a re-plan: delta in the run-log.
3. No action without a node — every L2/L3 dispatch (loop-iteration.js args.items) must be a ready plan node; node creation is the only exempt action (the sanctioned bootstrap), so an empty/first-iteration plan populates and dispatches in the same iteration rather than deadlocking.

Safety posture

• No .js engine touched. The discipline is prose + a lightweight plan schema.
• Readiness never bypasses the denylist (incl. .claude//CLAUDE.md) or trust levels — those gates still apply on top at dispatch.
• Trust-level defaults (L1 when unstated), maker/checker split, and confirmation gates are unchanged.

Consistency

loop-checklist.md §5, topics/workflow.md, the memory index.md, skills-index.md, CLAUDE.md, and the loop-iteration diagram label all updated to reference the plan/DAG and stay in sync.

Maker/checker — verified by our own gate

Ran /improve-skills (24 agents) as the independent checker, explicitly told to flag any weakened safety gate as a blocker:

• confirmedBlockers: 0 — no gate weakened (the core acceptance criterion).
• 14 important, all real+safe — genuine DAG-contract completeness gaps the checker found in the first draft, all applied:
  • in-progress was not excluded from the ready formula → a node orphaned by a crash/timeout could be re-read as ready and double-dispatched. Fixed + added stale-in-progress recovery.
  • No cascade when an after: dependency became dropped/escalated → dependents stranded blocked forever. Fixed (drop/re-point for dropped; escalate-with-note for escalated).
  • node/step terminology collision, an append-only-vs-prune ambiguity in the checklist, and stale CLAUDE.md/skills-index.md pointers — all corrected.

Verification

• /improve-skills gate: 0 blockers, 14 important (applied), 6 nice-to-have, 4 refuted.
• No .js changed; denylist + trust defaults + maker/checker confirmed intact by grep and by the gate.
• Ready-formula and terminology confirmed consistent across all seven surfaces.

🤖 Generated with Claude Code

---

Generated by Claude Code