cad-decision-review

Files (1)

• SKILL.md

SKILL.md (first lines)

<objective>
Stress-test ONE decision the user names, rather than take it on faith. The
review capability lives entirely in cadence-core/workflows/decision-review.md:
`cad-reviewer` (and, when configured, a cross-model provider) refutes the
decision through the review subsystem's reviewer resolution; the main model
then grounds each objection against documentation (library/API claims:
Context7 when its tools are available, else the installed source, lockfile
or vendored docs, with any claim none could check listed as unchecked) and
the real codebase (factual claims) and rules it `survives | partial | refuted`
with a concrete amendment where the decision needs one. When refutation
returns nothing, the grounding retargets onto the decision's own load-bearing
claims rather than lapsing, so a clean pass still rests on checked facts.

This is a manual, on-demand entry point only - it never auto-fires and there
is no review-trigger wiring for it. Durability (cad-context's `## Durable
decisions` filter) names candidates worth this deeper pass; picking one and
invoking this skill is the user's call, not a mechanical handoff.
</objective>

<execution_context>
@${CLAUDE_PLUGIN_ROOT}/cadence-core/workflows/decision-review.md
</execution_context>

<process>
1. **Resolve the target** from `$ARGUMENTS` - a path to a CONTEXT.md (naming
   which `- D-NN (...)` line) or a PROJECT.md (naming which Key Decisions
   row). If `$ARGUMENTS` is empty or ambiguous, ask (ask-user seam) for the
   path and the specific decision.

2. **Run the workflow** end-to-end: refute (the review subsystem's reviewer
   resolution), then adjudicate (docs grounding through Context7 when its
   tools are available, else installed source, lockfile or vendored docs,
   plus codebase grounding, per-objection ruling and amendments), then the qualitative cost report.

3. **Present** the per-objection rulings, groundings, and amendments - or, on
   a clean pass, the grounded load-bearing claims, never a bare "no findings"
   - plus which reviewers ran, and the tier/effort that reached the
   cross-model arm (the `cad-reviewer` arm resolves neither: no routing cell
   gives it a model, and its effort is frontmatter-frozen). Do NOT auto-apply any
   amendment - this is a review, not an edit; the user decides what to change
   and does it themselves.
</process>

From crenshawdev/cadence, MIT.

More from crenshawdev/cadence