Moved: this skill now lives in kimasplund/thinkbetter-skills, together with the rest of the reasoning suite. This repo is archived.
A reasoning suite for coding agents: nine reasoning methodologies as installable skills, two infrastructure skills that let them run in parallel and hand context to each other, and a router that picks which one a problem actually calls for — or tells you no framework is warranted at all.
Most reasoning failures aren't bad reasoning inside a method; they're the wrong method — exploring when you should eliminate, optimizing when you should synthesize, running a framework on a question that needed two sentences. The router makes that selection explicit, scored, and auditable; the patterns are the methods it routes to.
The router — skills/integrated-reasoning
Scores eleven problem characteristics (sequential dependencies, criteria clarity, evidence availability, novelty, time pressure, stakeholder complexity, …), computes weighted affinity scores for all nine patterns with validation rules (nothing to attack → adversarial scores zero; one decision-maker → negotiation scores zero), propagates uncertainty when a ±1 wobble would flip the winner, and orchestrates multi-pattern chains — sequential (explore → optimize → validate), parallel with explicit merge strategies, a 15-minute checkpoint loop, and bounded agreement-based confidence aggregation. A time-pressure fast-path skips straight to triage in emergencies, and a Direct Analysis threshold answers simple questions plainly instead of running the ceremony.
| Skill | The move | Reach for it when |
|---|---|---|
| tree-of-thoughts | Deep branching exploration with pruning | An optimization with clear criteria — find THE best |
| breadth-of-thought | Wide shallow sweep, conservative pruning | Unknown solution space — you can't afford to miss options |
| self-reflecting-chain | Stepwise reasoning with per-step validation | Proofs, traces, plans with dependent steps |
| hypothesis-elimination | Discriminating-evidence elimination | Diagnosis and debugging — find THE cause |
| adversarial-reasoning | Systematic attack on a candidate solution | Validation before committing — security, robustness |
| dialectical-reasoning | Thesis–antithesis–synthesis | Genuine trade-offs where both sides are right |
| analogical-transfer | Structured cross-domain borrowing | Novel problems with no in-domain precedent |
| rapid-triage-reasoning | Time-boxed good-enough decisions | Incidents and emergencies — minutes, not hours |
| negotiated-decision-framework | Multi-stakeholder alignment mechanics | Decisions that need actual buy-in, not just an answer |
| Skill | Role |
|---|---|
| parallel-execution | Fan-out/fan-in for running patterns or branches concurrently, with merge strategies (consensus / voting / aggregation / best-of-N) |
| reasoning-handover-protocol | The .reasoning/ directory protocol for chaining patterns — state, checkpoints, and context transfer |
git clone https://github.com/kimasplund/integrated-reasoning.git /tmp/integrated-reasoning
cp -r /tmp/integrated-reasoning/skills/* ~/.claude/skills/ # Claude Code; ~/.agents/skills/ for Kimi CodeEvery skill directory is self-contained — copy only what you want. The router functions standalone (its selection methodology and per-pattern guidance are inlined) but routes best with the library installed.
The router went through a blind, pre-registered with-skill vs no-skill round under skill-gauntlet (2026-08-18, frontier-model executors, 16 cells, sealed keys — full record in evolution/validation-round-1/). Honest results:
- NOT VALIDATED for decision-quality lift on a frontier model. Sealed-key detection and judged decision quality landed at exact parity — the base model already reasoned at ceiling on the test briefs.
- The restraint machinery passed both gates: the Direct Analysis threshold deflated a trivial question more crisply than the baseline, and the time-pressure fast-path opened an incident response with the rollback, not analysis.
- A machinery-leakage defect was found — and fixed in v2.2: the round-1 router narrated its scoring apparatus (affinity arithmetic, dimension tables) into user-facing answers. A suite-wide backstage clause + mechanical leak scan was then gauntleted as a candidate (evolution/validation-round-2/): it eliminated 100% of the leaks (0 vs 9 in the champion suite's main cells), lifted the leakage guard +1.50 (p=0.004), and recovered both guards to statistical parity with the no-skill baseline. The pre-registered gate still returned NO ADOPTION on one leg — the candidate's planted-key detection came in at 93.8% vs the champion's 100%, a 1.5-point gap made of three torn-edge-call thinness docks with no outright miss — and v2.2 ships by disclosed owner decision over that gap, with the residual wording improvement ("rewrite as content — never delete analysis") specified for a future round.
On strong models, treat the router today as reasoning hygiene and restraint rather than a quality lift. The routing-for-small-models claim (the original design target — the skill carries a measured 7.5B-model failure-mode guard) is untested; that round is the interesting next experiment.
The eleven library skills have not been individually gauntleted. The router round is the only blind validation any part of this suite has; the patterns' historical tests (in the old monorepo's .reasoning/ artifacts) predate the gauntlet protocol and are not blind or pre-registered. Until per-skill rounds run, treat each pattern as a documented methodology, not a validated one — and expect the router round's machinery-leakage finding to apply to patterns that instruct visible scoring apparatus.
- v1 router (evolution/) used an order-dependent decision tree over four patterns.
- v2 replaced it with weighted scoring over nine patterns plus Direct Analysis, uncertainty propagation, fast-paths, checkpoints, and statistically sane confidence aggregation. v2.1 fixed the DR formula and added validation rules. Renamed
integrated-reasoning-v2→integrated-reasoningat packaging (versions belong in changelogs, not names). - v2.2 (2026-08-18) added the backstage rule to the router and all nine patterns — the method never appears in the answer — plus a mechanical pre-delivery leak scan; adopted from validation round 2 by disclosed owner decision (see Validation status).
- Earlier copies of the whole suite lived in the claude_cognitive_reasoning monorepo; this repo supersedes them with the current, evolved versions and is the maintained home.