maister-copilot? could I touch it perhaps (the scripts)?

[robmar-net]

⚠️ Update (2026-07-22): the analysis below is superseded by live testing on GitHub Copilot CLI 1.0.73 — see the correction comment. Key reversal: maister-copilot is runnable — .md agents register (as maister-copilot:<name>), the skill + task(agent_type) tools map maister's orchestration, and Claude-format hooks work. The real fix is generator-side reference-naming + not deleting hooks, not re-engineering.

---

Got this analysis, as I sometimes use maister-copilot, perhaps I could try to fix it a little? What do you think?

It's 100% auto-generated by platforms/copilot-cli/build.sh (a 9-step cp/rm/sed transform) from plugins/maister/, and CI (build-copilot.yml) rebuilds+auto-commits it on every source change. I confirmed empirically: the committed output is byte-identical to a fresh build (100 files) and passes all 6 make validate checks. So generation is deterministic, fresh, and self-consistent by its own contract.

① Functionality parity — near-total

┌─────────────────┬─────────┬─────────┬──────────────────┐
│ Category │ maister │ copilot │ Δ │
├─────────────────┼─────────┼─────────┼──────────────────┤
│ Skills │ 17 │ 17 │ ✓ │
├─────────────────┼─────────┼─────────┼──────────────────┤
│ Agents │ 25 │ 25 │ ✓ │
├─────────────────┼─────────┼─────────┼──────────────────┤
│ Commands │ 6 │ 6 │ ✓ │
├─────────────────┼─────────┼─────────┼──────────────────┤
│ Orchestrators │ 5 │ 5 │ ✓ │
├─────────────────┼─────────┼─────────┼──────────────────┤
│ MCP (.mcp.json) │ 1 │ 1 │ ✓ byte-identical │
├─────────────────┼─────────┼─────────┼──────────────────┤
│ Hooks │ 4 │ 0 │ −4 (dropped) │
└─────────────────┴─────────┴─────────┴──────────────────┘

104 → 100 files; the entire delta is the deleted hooks/ dir — the only functional loss. 63/100 common files are byte-identical; 37 differ only textually.

② Correctness & consistency — self-consistent, but NOT a faithful runnable port

The validator passes, but it only checks .md/command surface conventions and misses every semantically material gap. 9 deduplicated defects remain — 100% generator-side (build.sh); the source maister plugin is valid, zero source-side defects. Top ones:

• 🔴 D-ORCH (HIGH) — the one real execution blocker: maister orchestrates via a Task/subagent_type API + programmatic Skill tool; Copilot has capable subagents (task//fleet) and skills but not that API, and agents ship as .md not the required .agent.md → they don't register. Workflows don't run end-to-end as written.
• 🔴 D-HOOKS (HIGH) — hooks/ deleted even though Copilot fully supports plugin hooks; drops a PreToolUse destructive-command safety guard, yet CLAUDE.md still documents a ## Hooks section and an agent still claims that guard exists.
• 🔴 D-INSTR (HIGH) — split-brain: skills point to .github/copilot-instructions.md (35+ refs) that doesn't ship, while the plugin still ships an unrenamed CLAUDE.md and root prose still says "CLAUDE.md".
• 🟠 D-REFS (HIGH/MED) — the maister:→maister- rewrite mints ~107 maister-development / subagent_type: maister-gap-analyzer refs that resolve to nothing (Copilot invokes flat: /development), contradicting the plugin's own "no prefix" note.
• 🟡 MED — plugin.json description still says "for Claude Code"; a whole ## Claude Code Documentation section survives; commands/ may not even load (undeclared path).
• ⚪ LOW — the platform note literally reads "Use ask_user instead of ask_user" (a transform-ordering bug).

Verification caught a significant error: the platform gatherer first claimed ask_user doesn't exist ("interaction is dead"). I verified against first-party GitHub docs — ask_user is a real native Copilot tool (--no-ask-user flag; Issue #2109 is only about the ACP bridge). I overturned that verdict, so the port has one real blocker (orchestration), not two.

Bottom line

Installable and partially functional, but not faithfully run-as-written. A sed-only transform can't bridge maister's execution model. A viable port needs real re-engineering (rename agents → .agent.md, remap the Task/Skill orchestration onto Copilot's task//fleet/skills, convert—not delete—hooks, fix manifest/branding, and extend make validate to catch these classes). All recommendations target build.sh/source — never the generated files.

[robmar-net]

Update — corrected after live testing on GitHub Copilot CLI 1.0.73

The analysis in the description above was documentation-based, and it turned out to be too pessimistic on every load-bearing point. I loaded maister-copilot into a real, working Copilot CLI (1.0.73) and tested each non-obvious compatibility against actual runtime behavior (debug logs, not model self-report).

Bottom line: maister-copilot is runnable, not "needs real re-engineering." maister's orchestration model maps onto Copilot's primitives almost 1:1. The remaining work is a tractable build.sh change, not a re-architecture.

What the live tests showed (✅ = works)

Compatibility	Description claimed	Live result
Install from nested .claude-plugin/plugin.json	(implied problematic)	✅ installs & loads (the already-installed copy uses exactly this)
Agents shipped as .md	"must be .agent.md → don't register"	✅ register, namespaced maister-copilot:<name>
Programmatic Skill tool	"not that API"	✅ a skill tool exists and runs
Task / subagent_type routing	"no equivalent → workflows don't run"	✅ task(agent_type:"maister-copilot:<agent>") executes the named agent; the task tool's own prompt enumerates all 24 maister agents as valid types
Parallel waves	(part of the blocker)	✅ task mode:"background" + read_agent / list_agents
Hooks	"Copilot supports them" (correct)	✅ Claude-format hooks.json fires as-is, incl. ${CLAUDE_PLUGIN_ROOT} (no version:1 needed) → the fix is simply don't delete them
Plugin .mcp.json (playwright)	byte-identical	✅ loads + initializes
ask_user	real (correct)	✅ exists (--no-ask-user flag)
Reads CLAUDE.md / AGENTS.md / copilot-instructions.md	—	✅ all three
Plugin commands/*.md as slash commands	"may not load"	❌ confirmed: they don't register — the one genuine runtime gap (low impact; entry points are skills)

Evidence: forcing task(agent_type:"…:probe-md-agent") produced SessionAgentExecutor.execute() called for "…" and returned the delegated reply; the --agent/task surfaces list every maister-copilot:* agent. Full write-up + reproduction steps are in the local maister-tests/COMPAT-TEST-REPORT.md.

Corrections to specific claims above

• D-ORCH is not a blocker. Task(subagent_type) → task(agent_type), Skill() → skill(), waves → background tasks + read_agent. Agents do not need renaming to .agent.md.
• D-REFS is real but different. The maister:→maister- rewrite is wrong, but the fix isn't "flat /development": agent refs need maister-copilot:<name> (colon), skill refs need the bare name. So it's a position-aware rewrite, not a global sed.
• D-HOOKS — correct that Copilot supports hooks; live-confirmed the Claude format works unchanged, so the fix is removing the rm -rf "$OUT/hooks" line (this restores the PreToolUse destructive-command guard too).
• init skill — Copilot enforces a frontmatter schema (argument-hint must be a string). Already fixed in v2.2.2 by quoting it.

Actual remaining work (all generator-side, in build.sh)

1. Reference naming (core): agents → maister-copilot:<name>; skills → bare <name> — a structured/position-aware rewrite, not s/maister:/maister-/g.
2. Stop deleting hooks/ (remove one line).
3. Commands: declare a commands path or convert to skills (optional, low impact).
4. Cosmetics: plugin.json description ("for Claude Code"), residual "Claude Code" prose, the CLAUDE.md/copilot-instructions.md split, the "Use ask_user instead of ask_user" ordering bug.
5. make validate: forbid maister-<agent> refs (require maister-copilot: for agents), require string argument-hint, and assert hooks/ is present.

So: yes — worth doing, and it's a tractable generator change, not a re-architecture.

Caveat: Copilot's delegation is model-driven (it decides when to call task/skill per the orchestrator instructions) — the capability and addressability are proven; determinism is inherently softer than Claude's Task tool. Live agent runs consumed a few AI credits.