Audit Pi Session

Analyze the structured session embedded in a Pi HTML export. Prefer decoded session data over scraping rendered HTML.

Workflow

1. Confirm the target .html file. Treat it as sensitive: exports can contain source code, prompts, local paths, images, credentials, and tool output.
2. Resolve this skill's directory and run the bundled extractor locally:

uv run <skill-dir>/scripts/extract_pi_session.py <export.html> summary

3. Read references/session-schema.md when interpreting Pi entry types, branches, or extractor output.
4. Choose the smallest useful evidence view:

# Entries on the active branch; thinking text is omitted by default
uv run <skill-dir>/scripts/extract_pi_session.py <export.html> active-path

# Branch leaves and their sizes
uv run <skill-dir>/scripts/extract_pi_session.py <export.html> branches

# Tool calls correlated with results across the full tree
uv run <skill-dir>/scripts/extract_pi_session.py <export.html> tools

# Inspect one abandoned or selected branch
uv run <skill-dir>/scripts/extract_pi_session.py <export.html> active-path --leaf-id <entry-id>
uv run <skill-dir>/scripts/extract_pi_session.py <export.html> tools --leaf-id <entry-id>

5. Increase --max-text-chars only when truncated evidence matters. Use raw only for fields unavailable in the bounded views; it emits the complete decoded payload.
6. Read references/audit-rubric.md, apply only dimensions relevant to the user's question, and cite entry IDs, tool-call IDs, timestamps, or exact bounded excerpts.
7. Separate direct observation, reasoned inference, and facts requiring repository inspection.

Scope Rules

• Use active-path metrics for the work Pi would continue from. Do not describe global tree totals as active-branch behavior.
• Inspect abandoned branches only when they explain retries, alternatives, useful discoveries, or wasted effort.
• Correlate assistant tool calls with toolResult entries before judging success. A call without a result is pending or interrupted, not successful.
• Treat command output and exit status as evidence of execution, not proof that the resulting code is correct.
• Treat final assistant claims as claims until supported by tool output or repository evidence.
• Do not grade or reproduce private reasoning. Evaluate observable prompts, responses, tool use, branch choices, and validation. The extractor reports thinking size but omits thinking text unless --include-thinking is explicitly requested.
• Preserve unknown entry types as evidence rather than assuming they are irrelevant.

Reporting

Default to a concise report:

# Pi Session Audit

## Executive summary
## Requested and claimed outcome
## Material findings
- Severity: high | medium | low
  Evidence: entry/tool ID and timestamp
  Observation: directly recorded behavior
  Impact: why it matters
  Improvement: concrete alternative
## Validation evidence and gaps
## Better workflow
## Metrics

Report only material findings. Avoid generic praise, checklist dumping, unsupported intent attribution, and precise duration claims unless timestamps establish them.

Privacy

Keep extraction local. Do not upload or web-search export contents unless the user explicitly requests it. Never repeat suspected secrets in the report; identify their location by entry ID and recommend rotation or redaction instead.