Mola-maker/mstack / skills/m-team
m-team
VerifiedEvolving agent team for open-ended, large-budget goals; a supervisor redesigns roles each round. Use for 组队/一直优化/卡在局部最优.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 11.4 KB |
| references/dsh-team.md | markdown | 3.5 KB |
| references/role-catalog.md | markdown | 5.1 KB |
| references/supervisor-checklist.md | markdown | 4.6 KB |
| references/team-config-template.md | markdown | 2.9 KB |
references/role-catalog.md
View on GitHub# Role catalog (m-team)
Roles are invented per round to fit the current bottleneck — this catalog is a starting menu, not a fixed roster. Mix, rename, split, or invent new roles; retire any role whose job is done. Every role contract has four parts: **duty**, **may change**, **must not touch**, **delivers** (evidence, as files — never a prose summary alone). Every role reads `SYNTHESIS.md` before starting.
## Focused Debugger
- **Duty:** reproduce and root-cause one named failure.
- **May change:** only its isolated branch (`agent/debug-<slug>`) and its scratch files.
- **Must not touch:** the main artifact and any other role's branch.
- **Delivers:** a minimal failing case, the root cause with `file:line`, a proposed fix as a diff against the isolated branch.
- **Add when:** the same failure has survived several rounds, or debugging is churning the main artifact (supervisor pattern 3).
## Disciplinarian
- **Duty:** keep the team on the chosen architecture and working method.
- **May change:** nothing directly — it reviews and rejects.
- **Must not touch:** implementation code.
- **Delivers:** a per-commit verdict: architecture respected? one change per commit? measured before editing? Rejects with reasons.
- **Add when:** rounds drift toward whatever is easy to implement, or commits mix several unmeasured changes (diagnosis: discipline problem).
## Knowledge Integrator
- **Duty:** merge findings from prior rounds into one actionable document.
- **May change:** `SYNTHESIS.md` and `LEARNINGS.md` only.
- **Must not touch:** the artifact under optimization, branches, configs.
- **Delivers:** an updated `SYNTHESIS.md` (current best design, known dead ends, open questions) and deduped `LEARNINGS.md` entries.
- **Add when:** a conclusion from an earlier round was lost, re-derived, or contradicted (supervisor pattern 2).
## Challenger
- **Duty:** attack the claim that the metric has reached its ceiling.
- **May change:** its own branch and benchmark scripts.
- **Must not touch:** the main branch.
- **Delivers:** either a counter-example (a measurement beating the "ceiling") or a written argument with data for why the ceiling is real.
- **Add when:** the metric is flat and the team's only explanation is "we're at the limit" (diagnosis: exhausted search).
## Auditor
- **Duty:** independently re-check delivered evidence before the supervisor scores it.
- **May change:** nothing.
- **Must not touch:** everything — read-only.
- **Delivers:** byte-level comparison of claimed vs actual artifacts: does the diff exist, do the numbers reproduce, do the gates actually run.
- **Add when:** any role's report and its artifacts disagreed, or scoring decisions felt shaky last round.
## Profile-Driven Optimizer
- **Duty:** improve the metric, but only where fresh measurements point.
- **May change:** the main artifact on its assigned branch.
- **Must not touch:** anything outside the profiled hot path.
- **Delivers:** a fresh profile taken this round, each change tied to a hot spot, before/after measurements for every commit.
- **Add when:** tuning decisions are citing measurements from before the last structural change (supervisor pattern 4).
## Dead-End Revisitor
- **Duty:** re-open one hard direction that an earlier round abandoned.
- **May change:** its own branch; may reuse prior failed attempts as a starting point.
- **Must not touch:** the current best candidate.
- **Delivers:** what was tried before (from `history.md`), what is different now, and either a working spike or a sharper explanation of the blocker.
- **Add when:** `history.md` shows an abandoned direction whose blocker may no longer hold (new knowledge, new tools, changed architecture).
## Research
- **Duty:** fill a named knowledge gap from external sources.
- **May change:** only its notes file under the round's knowledge placement.
- **Must not touch:** code, configs, benchmarks.
- **Delivers:** sourced findings (multi-channel, cross-checked per `m-research`), each marked verified or single-source.
- **Add when:** the diagnosis is knowledge gap — the team needs facts it doesn't have (hardware behavior, library semantics, prior art).
## Architect
- **Duty:** redesign the structure when incremental work cannot reach the target.
- **May change:** a design doc and a spike branch.
- **Must not touch:** the main branch directly.
- **Delivers:** a concrete redesign (what changes, in what order, what it enables), plus a spike proving the riskiest part.
- **Add when:** several rounds of optimization produced only marginal gains and the bottleneck is structural (diagnosis: exhausted search).
## Composing a round
- One role can be the whole team (coordination shape: solo) when the diagnosis points at a single bottleneck. That solo-shaped round is already the baseline: do not add a second solo agent. A pipeline or parallel round still runs one extra solo agent at the team's shared budget; that agent is not a catalog role.
- Pipeline (Research → Architect → Profile-Driven Optimizer) when knowledge must precede structure which must precede tuning.
- Parallel branches (Focused Debugger isolated + main-line work continuing) when a bug blocks progress but the metric work shouldn't stop.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |