evaluate

name	evaluate
description	Adversarial evaluation of code changes — runs build, tests, plan compliance, and security review. Use after implementing changes or before creating a PR. Usage: /evaluate [story-id] [--quick]

Triggers: "evaluate this", "check my changes", "review before PR", "run evaluator", "quality check"

---

You are the evaluation orchestrator. You launch the evaluator agent and present its findings.

Parse arguments from: $ARGUMENTS

• First argument: story ID or branch name (if omitted, use current branch)
• --quick flag: if present, only run hard gates (build + tests), skip plan compliance and adversarial review

---

Phase marker

Before spawning the evaluator agent (Step 2), write tasks/stories/<story-id>/phase.md per rules/phase-markers.md — overwrite it in full with the six plain key: value lines (schemaVersion: 1, phase: reviewing (the reviewing phase, displayed as Warden), role: reviewer, updated: <ISO-8601 UTC now>, skill: evaluate, detail: Step 2 — evaluator-agent). role is always reviewer for /evaluate. /evaluate writes this marker only when it has a story id / story workspace to write into — if there is no story ID and no tasks/stories/<id>/ workspace (e.g. running against a bare branch name with no plan file), skip the marker entirely.

---

Step 1 — Determine context

Run:

cd YOUR_PROJECT_ROOT && git branch --show-current && git diff --stat HEAD~1..HEAD

Determine:

• Branch name — from git output
• Story ID — from arguments, or extract from branch name (e.g., sprint3/4567-add-login-page → 4567)
• Scope — "quick" if --quick flag present, "full" otherwise

Check if a plan file exists:

ls YOUR_PROJECT_ROOT/tasks/stories/[story-id]/plan.md 2>/dev/null || echo "no plan"

If no plan file exists and no story ID was provided, set scope to "quick" automatically (no plan to check against).

---

Step 2 — Launch evaluator

Write the phase marker (per rules/phase-markers.md — the full six-key marker, see "Phase marker" above) before spawning, if a story workspace exists.

Spawn an evaluator-agent (foreground) with:

Story ID: [story-id or branch name] Plan path: [path to plan.md, or "none"] Scope: [full or quick]

Wait for it to return the full evaluation report.

---

Step 3 — Present findings

Output the evaluator's full report.

Then based on the verdict:

If ❌ NO (hard gates failed):

---

EVALUATION BLOCKED — build or tests are failing.

Fix the failures above, then run /evaluate again.

---

If ⚠️ WITH CAVEATS (high-confidence findings):

---

EVALUATION PASSED WITH CAVEATS — [N] findings need review.

Review the findings above. For each one, say "fix" (I'll help fix it) or "skip" (acceptable risk). Or say "proceed" to continue to PR with all findings as-is.

Before presenting a finding as skippable, apply the ship test (rules/deferrals.md) to it: with this left undone, does the change behave incorrectly for the project's real configured inputs — not for the values its tests use? Mark any finding that fails the test as BLOCKER (not deferrable) in the list, whatever severity the evaluator assigned it, and say why. A finding the human then skips must be registered as a tracker item, not left as prose.

---

If ✅ YES (all clear):

---

EVALUATION PASSED — no blockers, no high-confidence findings.

Ready to proceed to PR.

---

Hard rules

• Never skip the evaluator agent — always spawn it, even if changes look trivial
• Never override a hard gate failure — if build/tests fail, the verdict is always ❌ NO
• Present the full report — don't summarize or filter findings
• Apply the ship test (rules/deferrals.md) to every finding before presenting it as skippable — the evaluator's severity label is an input, not a verdict; it doesn't know what the project has configured
• If the user says "fix" for a finding, help implement the fix, then re-run /evaluate