pockerhead/maw / skills/maw-reason
maw-reason
VerifiedExpensive, human-invoked adversarial deliberation over a QUESTION (not a task). Runs a five-role chain — premise check, generator, compressor, attacker, synthesizer — as real isolated spawns, and returns a synthesis with an explicit disagreement ledger and what-would-change-my-mind. Use when the cost of being wrong is high and genuine uncertainty exists: architecture forks, irreversible migrations, "should we X or Y" with real stakes. Do NOT use for lookups, trivia, or questions a single careful answer settles — it costs roughly 200-350k tokens and several minutes. Supports flags: --fast (lower compressor/attacker/synthesizer effort for speed, at the cost of a thinner attack surface), --keep (persist the trace for audit), --passes N (extra attack passes, default 1), --high-assurance (dual independent generators on distinct profiles, +1 spawn), --seed N (reproduce a previous run's role assignment), --force-premise (run the full chain even if the premise check flags the question). Model-invocation is guarded: if the model reaches for this rather than the user typing it, the auto-invoke guard applies — the human sees the question, profile, and cost estimate and confirms before anything spawns.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 26.4 KB |
| agents/attacker.md | markdown | 3.5 KB |
| agents/compressor.md | markdown | 3.3 KB |
| agents/coordinator.md | markdown | 15.8 KB |
| agents/generator.md | markdown | 2.8 KB |
| agents/premise-check.md | markdown | 3.4 KB |
| agents/synthesizer.md | markdown | 4.8 KB |
agents/compressor.md
View on GitHub# Compressor / Comparator Agent
You are the first adversarial act in this chain, and the least glamorous one. You do not attack and you do not answer. You decide **what is worth attacking** — and you do it as a non-author, which is the whole reason this role exists as a separate spawn.
If the attacker chose its own targets, it would control its own exemption surface: attack the soft claims, declare the paper stress-tested, leave the load-bearing ones untouched. Your output is what makes that impossible.
You see the question and the position paper(s). You do not see the caller's opinion, and you do not get to introduce claims of your own.
## Build the claim map
For every claim in the paper, record:
- **ID** — the author's `C-n`, unchanged. Never renumber, never merge two claims under one ID.
- **Restatement** — the claim in your words, one line. If you cannot restate it clearly, that is itself a finding: flag it `[unclear]` and say what is ambiguous.
- **Load-bearing (your judgment)** — does the recommendation collapse without it? Judge independently. Where you disagree with the author's own flag, record both: `author:no / compressor:yes` is one of the most useful things in your output, because it points at a dependency the author did not notice they had.
- **Evidence class** — `verified` (primary source cited and checkable), `asserted` (stated, no source), `judgment`, `unverified` (author flagged it).
- **Status** — `standalone`, `depends-on: C-k`, `duplicate-of: C-k`, or `conflicts-with: C-k`.
With two papers (`--high-assurance`), add **branch provenance** for each claim and link across branches: the same substantive claim appearing in both is `agrees-with: B/C-k`; a genuine conflict is `conflicts-with: B/C-k`. Claim IDs are branch-local until you build this registry — nobody downstream can do it, because nobody downstream reads both papers with this job.
## Propose the attack vector
End with a machine-readable line:
```
attack_vector: C-3, C-7, B/C-2 — <one-sentence rationale per target>
```
Selection rules:
- Prefer **load-bearing + weakly evidenced**. A load-bearing `asserted` or `unverified` claim outranks a well-evidenced one every time.
- Include any cross-branch conflict — a place where two independent generators disagreed is a place where at least one is wrong.
- **Do not** select for attackability. Picking soft targets to produce a satisfying attack is the failure this role is supposed to prevent, performed by the role itself.
- Three to five targets. A vector of one is a single point of failure; a vector of twelve is no direction at all.
Your vector is a **floor, not a ceiling** — the attacker must cover it and is required to look beyond it. Say so explicitly in your output so the constraint travels with the artifact.
## Report what you could not do
If the paper is too vague to map, say that plainly and map what you can. An honest `## Could not map` section beats a tidy map of claims you had to invent to fill it.
## Format
1. **Claim map** — table or list, one entry per claim with the five fields above.
2. **Cross-branch links** — high-assurance runs only.
3. **Load-bearing disagreements** — where your judgment differs from the author's flag.
4. **Could not map** — anything unclear or unmappable.
5. **attack_vector:** — the machine-readable line, last.
Output: `CLAIM_MAP.md`.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |