agentxp: Skill for Claude Code

Skill: /analyze

Purpose

The analyze verb runs from a sealed brief through verdict + readout. Its first action is to verify the brief's three-part integrity lock (design_chain_hash + metric_snapshot + expected_shape). Any mismatch refuses entry; there is no --force.

On verify pass, the loop is fixed by the worldview:

1. SRM first (R2) — no metric value is read before the assignment ratio is checked
2. Stats only from the whitelist (R4) — every number comes from agentxp.stats.*
3. Verdict from the decision tree (R3) — walk_tree() is the only path
4. Narrator blind to hypothesis direction (R5) — AnalystNarratorBundle enforces this
5. Critic fires at every commit (R6)
6. Confidence labels computed, not chosen (R8)
7. Every readout claim cites an artifact (R7)

When to invoke

Direct: /analyze --brief experiments/<id>/brief.sealed.yaml

Plain-English routing:

Phrase	What to do
"Analyze experiment exp_001"	/analyze --brief experiments/exp_001/brief.sealed.yaml
"What does the data say" (sealed brief in context)	/analyze --brief <path>
"Run the analysis for the checkout test"	/analyze --brief <path>
"Analyze before sealing the brief"	REFUSE — R1 + R11 require a sealed pre-registration

Procedure (do these in order)

1. Verify the seal (R11 wall)

from pathlib import Path
from agentxp.workflows.analyze import verify_and_open
from agentxp.schemas.brief_seal import BriefSealMismatch

try:
    sealed = verify_and_open(Path(args.brief), Path.cwd())
except BriefSealMismatch as exc:
    print(f"REFUSED: {exc}")
    raise SystemExit(1)

If verify fails: print the specific reason (which lock component, current vs sealed hash), exit. The user resolves the drift (rev the brief, restore the metric, etc.) and re-invokes.

If verify passes: print the experiment id + the integrity-lock receipt (truncated hashes for display).

2. SRM check (R2 — always first)

Before any metric value is read, run SRM: