Factor AGENTS.md procedures into agent skills (local dev, board/issue workflow, CI, publishing, …)

[cliffhall]

Problem

AGENTS.md is now the single, very large contract for changing this codebase — the issue/board workflow, the priority rubric, the branch and label rules, the TypeScript/Mantine conventions, the test tiers and coverage gate, the pre-push gate, and the release procedure. It works, but it is loaded in full, on every turn, for every task, whether the task is a one-line CSS fix or a release cut.

That has three costs:

1. Context. The whole contract is resident even when almost none of it applies.
2. Discoverability. A procedure like "cut a release" or "triage new issues" is a multi-step recipe buried in prose among a dozen unrelated recipes; it is found by reading, not by asking for it.
3. Drift. The recipes carry live IDs (board field/option IDs, project node IDs) and exact command sequences. When they go stale there is nothing that runs them, so staleness is discovered by a failed command mid-task.

Agent skills are the mechanism designed for exactly this shape: a named, on-demand bundle of instructions loaded only when the task matches, invokable by name (/<skill>).

Proposal

Add a .claude/skills/ set that factors the procedural parts of AGENTS.md into invokable skills, leaving AGENTS.md as the always-on rules (conventions a reviewer cites against a diff) and moving the procedures (multi-step recipes with commands) into skills.

Candidate skills, roughly one per procedure that today lives as prose:

Skill	Covers
local-dev	The non-workspace install cascade, per-client npm install, clients/web && npm run dev vs the built launcher, which script to run for which client, the @inspector/core alias, dependency placement rules (root-only SDK packages, bundler external lists, overrides)
issue-create	The full create flow: version label + type label + milestone + board card + Status + Priority — all five steps, with the current board IDs
issue-triage	Pass 1 sweep (the unboarded diff against both boards), the priority rubric with the score comment, and the board audit query
board-ops	The gh project recipes for #28 and #11: move a card, re-score, delete a card, plus the option-deletion hazard, the snapshot rule, and the recovery recipe
pre-push-gate	npm run format → npm run ci, what each tier covers, the DCO signoff rule and its repair, and what to do when a specific gate fails
testing	Test placement (side-by-side vs src/test/ vs __tests__/), the tiers, the ≥90 per-file gate, the v8 ignore policy, renderWithMantine
pr-flow	Branch naming, Closes #N, screenshots into the gitignored pr-screenshots/, requesting Copilot review, responding to review rounds, and closing out (manual close + Done)
release	Cutting a release: bump on v2/main first, milestone merge, tag origin/main bare x.y.z, pack:verify, the Docker image
test-servers	Picking and running a showcase config, ensure-test-servers, the stale-build hazard

Not exhaustive — the list should be settled while doing the work, and some of the above may merge.

Constraints

• AGENTS.md stays the source of truth for rules. A skill must not become a second, divergent statement of a convention. Where a skill needs a rule, it should state the procedure and point at AGENTS.md for the rule — the same relationship .github/copilot-instructions.md has, and with the same failure mode (silent drift) to guard against.
• Anything moved out of AGENTS.md must be removed from it, not duplicated. Two copies of the board IDs is strictly worse than one.
• Non-Claude agents read AGENTS.md. Skills are a Claude Code mechanism; other agents (and humans) get only the file. So the split has to leave AGENTS.md still complete enough to work from — which argues for moving recipes (which a human can find via a pointer) rather than rules (which a reviewer must have resident).
• Mirror any review-relevant change into .github/copilot-instructions.md, per the existing rule.

Open questions

• Is the right split rules-vs-procedures, or task-vs-task? The table above mixes both.
• Should AGENTS.md keep a short index of the skills so a human/non-Claude agent knows what exists and where to look?
• Do any of these warrant being project skills committed to the repo (so every contributor gets them) vs personal? Committed, presumably — but that makes them a reviewed surface with its own maintenance cost.
• Is there a way to make a skill's recipe executable (a script in the skill dir) rather than prose, so the board IDs live in one place that fails loudly when stale?