Re-engineer the maister-copilot generator into a runnable Copilot CLI plugin

[robmar-net]

Summary

Re-engineers the maister-copilot generator (platforms/copilot-cli/build.sh + Makefile + CI) so the generated plugin is a runnable, coherent GitHub Copilot CLI plugin — not just an installable one. All changes are generator-side; the Claude plugins/maister/ source is 100% untouched.

Why

An initial documentation-based analysis (see #8) concluded the port "was not faithfully runnable and needed re-engineering of the orchestration layer." Live testing against a real GitHub Copilot CLI 1.0.73 overturned that: maister-copilot is runnable, and maister's orchestration maps onto Copilot's primitives almost 1:1:

maister (Claude)	Copilot CLI	Verified live
Task(subagent_type: "maister:X")	task(agent_type: "maister-copilot:X")	✅
Skill("X")	skill("X")	✅
parallel waves	task mode:"background" + read_agent/list_agents	✅
AskUserQuestion	ask_user	✅
hooks/hooks.json (${CLAUDE_PLUGIN_ROOT})	plugin hooks.json, same format	✅
.mcp.json, agents/*.md, skills/*/SKILL.md	same (agents namespaced maister-copilot:<name>)	✅

So the real defects were narrow and all generator-side: wrong reference naming and the generator deleting hooks/.

What changed

• Reference naming (core): the flat s/maister:/maister-/g — which mangled every reference — is replaced by a registry-driven, kind-aware rewrite: agent refs → maister-copilot:<name> (plugin-id namespaced), skill/command refs → bare <name>. Name-set registries are built from the source tree with a pairwise-disjointness assertion, and the pass fails loud on any unknown token. Illustrative compound tokens (maister:x:y, e.g. product-design's /maister:feature:new "does-not-exist" example) are handled in-generator, so the Claude source needs no edit.
• Keep hooks: removed rm -rf "$OUT/hooks". The Claude-format hooks.json (incl. ${CLAUDE_PLUGIN_ROOT}, SessionStart + PreToolUse) fires unchanged on Copilot — dropping it was a needless functional + safety regression (it includes the destructive-command guard).
• Manifest & branding: plugin.json description → "…for GitHub Copilot CLI" (targeted string edit, no JSON round-trip); fixed the "Use ask_user instead of ask_user" ordering tautology; hardened the .claude/CLAUDE.md path rewrite; scrubbed "Claude Code" branding/URLs.
• make validate hardening: new static checks (no wrong maister-<word> refs, string argument-hint, hooks/ present, no Claude-Code residue) + a check-deterministic target (byte-identical rebuild).
• WS7 compatibility harness (platforms/copilot-cli/compat-tests/, make test-copilot): loads the freshly-built plugin into a live Copilot CLI and asserts 7 runtime contracts (plugin/skills/agents register, task+skill delegation, hooks fire, MCP loads); emits a version-stamped report; side-effect-free; supports an auth-free subset.
• CI: a compat job runs the harness when a Copilot seat is provisioned (fails on any red contract), else skips with build+validate still gating. Purely additive — the existing build job is unchanged.

How it was tested

On GitHub Copilot CLI 1.0.73:

• make build ✅ · make validate ✅ (existing 6 + new checks) · make check-deterministic ✅ (byte-identical rebuild) · provenance: only build.sh / Makefile / build-copilot.yml + new compat-tests/ are source edits.
• make test-copilot → GREEN, 7 PASS · 0 FAIL (plugin loads; 17 skills register, 0 failures; 25 agents resolve as maister-copilot:<name>; task(agent_type) delegation executes; skill(<name>) executes; SessionStart+PreToolUse hooks fire with ${CLAUDE_PLUGIN_ROOT}; .mcp.json loads).

Notes

• Generator-only & deterministic: plugins/maister/ (the Claude plugin) has zero changes; rebuilds are byte-identical, so the existing CI auto-commit stays a no-op.
• Regenerated plugins/maister-copilot/** is included so the branch is self-consistent (matches the repo's "Rebuild Copilot CLI variant" convention).
• Copilot delegation is model-driven (it decides when to call task/skill); capability and addressability are proven, determinism is inherently softer than Claude's Task tool.

Refs #8, #9

🤖 Generated with Claude Code

[robmar-net]

Update — three commits added to this PR (all still Copilot-only; the Claude plugins/maister/ source stays untouched, and rebuilds remain byte-identical):

1. feat(copilot-gen): adapt destructive-command guard to 'ask' — live hook testing found the destructive-command guard is a silent no-op on Copilot: it keys on an agent_type that Copilot's PreToolUse payload never carries (verified on 1.0.73 — not in stdin, not in any env var). A generator override (build.sh WS2b) replaces only the output guard with a variant that asks the user to confirm any destructive command (permissionDecision: "ask"), which Copilot honors and does not bypass under --allow-all-tools (held in headless = fail-closed). Rationale is documented in the hook's header comment; make validate WS5.10 guards it.
2. test(copilot-gen): L1 hook-effect checks (make test-hooks) — a committed harness that verifies each hook's effect (not just that it fires) on a live Copilot CLI: skill-invocation reminder reaches the model; post-compact reminder's env dep is satisfied (Copilot sets $CLAUDE_PROJECT_DIR); and the destructive guard now gates (a subagent's rm -rf is held). make test-hooks → 5 PASS · 0 FAIL. See compat-tests/L1-FINDINGS.md.
3. ci: pin actions/upload-artifact to a commit SHA — supply-chain hardening so the workflow survives sha_pinning_required.

All verified against Copilot CLI 1.0.73: make build · make validate (incl. the new WS5.9/WS5.10 checks) · make check-deterministic (byte-identical) · make test-copilot (7 runtime contracts) · make test-hooks (hook effects).

Refs #8, #9