evaluate

name	evaluate
description	Evaluate codebase quality across four dimensions (correctness, maintainability, performance, testability) plus the principal-engineer lens. Produces a scorecard usable as a quality gate.
user_invocable	true

Evaluate

Comprehensive quality evaluation of the Synto codebase across four dimensions — Correctness, Maintainability, Performance, Testability — plus the principal-engineer lens.

The entire multi-round fan-out runs in the evaluate workflow (.claude/workflows/evaluate.js). This skill invokes it and presents the result. Moving the orchestration into a workflow flattens what used to be a deep team of sub-agents into one deterministic script driving leaf agents — no nested-agent wall, and the run is resumable.

Run it

Call the Workflow tool:

Workflow({ name: 'evaluate' })

The GitHub issue reconciliation is off by default (skipSync defaults to true) — a plain run is findings-only and never mutates the repo. Opt in with Workflow({ name: 'evaluate', args: { skipSync: false } }) once you actually want findings synced to evaluation-labelled GitHub issues (the first such run creates the evaluation label and the issues).

What the workflow does

1. Baseline — an agent runs dotnet build --no-restore -c Debug, dotnet test --no-build -c Debug, and dotnet format --verify-no-changes; build failure is a critical correctness finding, failing tests critical/high. (Analyzer warnings surfaced by the build are findings, not hard green-gate failures — TreatWarningsAsErrors is not set in Synto.) Results feed every reviewer.
2. Triage — maps each project to its most-relevant dimensions (from each project's README.md and .csproj) so reviewers read high-relevance projects in depth and skim the rest.
3. Round 1 (discovery) — 5 reviewers in parallel (4 dimensions + the principal-engineer lens), each reading .claude/playbook/architecture.md, principles.md, standards.md, and their own checklist/focus areas. Evidence (a 3–10 line snippet) is required for every critical/high.
4. Round 2 (verification + cross-pollination) — 5 fresh reviewers verify their own criticals/highs against the code, challenge false positives, reinforce agreements, discover cross-dimensional findings, and adjust severities. The principal-engineer also receives mediums so they can promote convergent risks and challenge workarounds.
5. Synthesis — the team-lead agent resolves conflicts, correlates root causes, builds the prioritized action plan, and writes the full report.
6. Sync (only when skipSync: false) — reconciles findings with evaluation-labelled GitHub issues: stale review (close fixed / keep valid / respect deferred), then create / update / re-confirm. Returns a sync summary.

Scoring model

The reviewers and synthesis follow .claude/playbook/standards.md — severity definitions, per-dimension scores (Excellent / Good enough / Needs improvement / Broken), and the quality-gate tiers (Correctness is the hard gate; Maintainability/Performance must be Good enough+; Testability must not be Broken), plus the principal-engineer (consequences) cross-cutting lens.

Present the result

The workflow returns { report, gate, syncSummary, counts }. Print report verbatim (it is already formatted as the Quality Scorecard → Findings → Domain Expert Assessment → Root Causes → Prioritized Action Plan → Strengths → Verification Notes), then append the syncSummary. Lead with the quality gate: PASS / FAIL.

Notes

• Live cross-expert consultation (reviewers messaging each other mid-round) is not part of the workflow model — leaf agents aren't addressable to one another. The two-round structure (Round 2 reviewers see Round 1's findings) preserves the cross-dimensional discovery that consultation provided.
• The principal-engineer's findings are attributed to the dimension they most affect and counted in that dimension's score (no separate scorecard column), per standards.md.