grahama1970/assess

STOP. READ THIS ENTIRE SKILL.MD BEFORE CALLING ANY ENDPOINT.

assess 是做什么的？

Step back and critically reassess project state. Provides interactive guidance (human-in-the-loop) and programmatic analysis (automated pipelines). Specialized for doc pruning and doc-code alignment.

CLI Usage

# Structured JSON assessment for pipelines
.pi/skills/assess/assess.py run . --output assessment.json

# Generate all figures from assessment
create-figure from-assess --input assessment.json --output-dir ./figures/

Assessment Flow

1. Scope the Assessment

Clear scope ("assess the auth module") → proceed directly. Open scope ("step back") → ask: concerns? code quality vs docs? known issues to skip? Doc pruning ("prune documentation") → ask: deprecated features? missing docs? alignment?

2. Find Project Root & Detect Ecosystem

Detect via marker files (pyproject.toml, package.json, Cargo.toml, go.mod). Read ecosystem-specific metadata, then universal docs (README.md, CONTEXT.md, AGENTS.md).

3. Load Skill Manifest (Route-Gated)

Load .pi/skills-manifest.json for assessments that could lead to "we should build X" recommendations. Skip for status checks and doc-only reviews.

cat .pi/skills-manifest.json | python3 -c "
import json, sys
data = json.load(sys.stdin)
print(f'Skills: {data[\"skill_count\"]} (generated: {data[\"generated\"]})')
"

Warn if manifest >7 days stale. Fallback: ls .pi/skills/ | wc -l.

4. Quick Scan → Check In → Deep Dive

1. Read project metadata, README, glob for structure
2. Note 2-3 initial observations, check in with user
3. Deep dive based on user guidance — for each finding: blocking? known? fix complexity?

5. Collaborative Report

Present findings and ask which to fix. Offer /create-figure for visualization.

Assessment Categories

1. Doc-Code Alignment

• README claims vs implementation
• Cross-reference validation (internal links, code examples)
• Placeholder markers and stale information
• Deprecation: docs for removed features

2. Aspirational vs Implemented

• Stubs (pass, raise NotImplementedError), placeholder notes
• Features in code structure but no logic, unused dependencies

3. Brittle Code

• Hardcoded values, missing error handling, fragile regex/parsing

4. Non-Working Code

• Dead code paths, silent exception swallowing, broken integrations

5. Over-Engineered Code

• Abstractions with single implementation, config for hypothetical flexibility

6. Test Coverage (Non-Negotiable)

Check each feature/module for test files, passing tests, edge cases. Flag pytest.mark.skip without reason, "0 tests collected", tests that don't assert.

Task file audit: Verify each task has Definition of Done with specific test/assertion. Block execution if implementation tasks lack Definition of Done.

7. Working Well

Acknowledge solid code to calibrate the assessment.

Output Format

# Assessment: <project-name>

## Summary
<2-3 sentences: overall health, key findings>

## Findings

### Doc-Code Alignment
| Claim | Reality | Action? |

### Issues Found
1. **file.py:123** - [Issue] — Severity: H/M/L — Suggested fix: [brief]

### Test Coverage
| Feature | Test Status | Action Needed |

### Working Well
- [Solid code to acknowledge]

## Recommended Next Steps
1-4 prioritized actions. Ask which to start.

Policies

• Read-only by default — never write files without explicit consent
• Print first, write with consent — show proposed content, then offer to write
• Collaborate — assessment is dialogue, not monologue
• Ask before running — get permission before tests, builds, paid services
• Escalate wisely — suggest /review-code for complex/critical issues
• Be specific — file paths and line numbers
• Be actionable — each finding suggests next steps
• Test coverage is non-negotiable — flag missing tests as blockers

External Research

Available when needed (ask before paid services):

• /context7 — library docs (free)
• /brave-search — general web (free)
• /perplexity — deep research (paid)

Common Mistakes

# WRONG: Treat all assessment findings as blocking issues
# → Agent auto-adds docstrings to 12 private helpers. Unnecessary bloat.
# RIGHT: Review findings with user. Escalate complex issues to /review-code.

# WRONG: Skip /assess after major changes
# → README examples use old function signature. Users hit errors.
# RIGHT: Run assess --focus "doc-code-alignment" after editing core APIs

# WRONG: Run /assess and act on stale results
# → Assessment was from 3 days ago. Codebase changed since.
# RIGHT: Always re-run assess; never cache results across sessions