Active

- **Date**: 2026-05-10
- **Agent**: claude-opus-4-7[1m] main session; no subagent dispatches
- **Task**: Added `component-design-with-tdad` skill (issue #313) — methodology guidance loadable by `spec-writer`, `tdd-agent`, or human brainstorming when designing a new plugin component. Names the five design questions implied by the four-layer TDAD architecture: component type, layer targeting, scenario `Then` shape, new-vs-modification, scenario-vs-finding. Loadable rather than dispatchable — no new agent, no new dispatch surface.
- **Surprise**: The user's question started as "Consider whether we could do with an agent..." and the honest answer was *no*. Adding a new agent would directly contradict cartograph story #3 of PR #309 (the same-day decision: single tdd-agent + branch over a separate tdad-agent, citing the architectural failure mode of "two agents that share a charter"). The cartograph's promoted convention from yesterday — name the design choices openly so future contributors see what the project decided — surfaced this contradiction immediately. What was worth recording is *how cleanly* the prior cartograph stories provided the right reference frame: the user's framing pointed at the right gap (design intelligence + TDAD discipline are coupled), but the wrong shape (a new agent). The right shape — a skill — was the answer cartograph #3 was pre-loaded to suggest. The mechanism worked: yesterday's adjudication surfaced today's decision.
- **Proposal**: None for AGENTS.md or HARNESS.md. The skill itself captures the design intelligence; the cartograph story #3 already provides the architectural rationale for "skill not agent." The pattern is now visible in the corpus: when in doubt about adding an agent vs a skill, default to skill unless the work has a genuinely new dispatch surface that cannot be loaded into context.
- **Improvement**: One improvement worth surfacing: when an existing cartograph story's pattern applies to a new design question, citing the story's ID directly in the new design discussion is faster than re-deriving the rationale. The cartograph corpus is becoming a reference frame for future design work, not just an audit trail of past decisions. Consider, at a future curation pass, whether to surface the load-bearing cartograph stories (especially #1 self-exemption clause and #3 single-agent-charter) into AGENTS.md ARCH_DECISIONS where they are loaded by every session. They currently live only in the per-spec story records — discoverable but not loaded by default.
- **Signal**: workflow
- **Constraint**: none — methodology packaging is judgemental craft, not falsifiable enforcement. The HARNESS constraints from PR #309 already cover the *outputs* of TDAD-aware design (scenario presence, structural correctness); the design intelligence itself is loaded into context, not gated.
- **Session metadata**:
  - Duration: ~30 min (user question → exploratory recommendation → AskUserQuestion → skill authoring → brief reflection → PR ceremony)
  - Model tiers used: claude-opus-4-7[1m] throughout; no subagent dispatches.
  - Pipeline stages completed: brainstorm-by-question (4-option AskUserQuestion) → skill authoring (single SKILL.md, ~270 lines) → reflection authoring → PR with chore exemption per AGENTS.md STYLE conditions (issue #313 + reflection capture + additive + honest version bump).
  - Agent delegation: manual — direct edits, no subagent dispatches. The skill is loadable by future agents; this PR ships the loadable surface.