Session Intelligence

Session Intelligence

Improve the user's AI-assisted work from local behavioral evidence while keeping transcript text behind explicit consent.

Non-negotiable boundaries

• Never read raw transcript files into model context. Use scripts/si.py only.
• doctor and diagnose may run without excerpt consent because their output contains no transcript text.
• Never run evidence without explicit consent for one named finding and bounded release.
• Never write history or an artifact without explicit approval.
• Never treat cross-tool occurrence as an automatic confidence boost.
• Never compare incompatible provider token totals or invent exact cost.
• Never recommend from low-confidence, single-session, or structurally unreliable evidence.
• Hook, MCP, and direct configuration mutations are out of scope in v1. Represent them as reviewed note recommendations only.

Select mode and scope

Modes are arguments to this one skill: full (default), rules, docs, waste, prompts, success, or verify. The focused /si-rules, /si-docs, /si-waste, and /si-prompts entrypoints are intentionally retained in commands/; they share this staged engine but each has one narrow goal.

Scope is global or project. If ambiguous, ask one short choice. Project scope requires the selected project root; descendants belong to that project. Documentation extraction is project-only.

Default window is 30 days. The CLI expands to 90 days only when fewer than ten usable sessions exist and reports the expansion.

Default workflow

1. Read references/running.md, then run scripts/si.py run for the selected scope. This performs doctor followed by diagnose. Do not inspect transcript sources directly.
2. Check coverage and data quality first. If evidence is insufficient or structurally unreliable, report that and stop; silence is valid.
3. Apply the selected lens from references/capabilities.md. Confirm candidates with references/detection-playbook.md.
4. Present two strengths worth preserving and at most three ranked opportunities. For each include support, evidence quality, recommendation, expected benefit, trade-off, and next measurement.
5. Ask whether the user consents to bounded excerpts for a specific finding. Refusal ends excerpt work without penalty; structural recommendations may still be discussed.
6. After consent, run evidence only for that finding and stay within approved turn and character limits. Mention release counts and redactions.
7. Before drafting an artifact, read the relevant section of references/improving-ai-workflows.md. Use only rule, skill, slash-command, agent, doc, note, stat, or none.
8. Run route preview, show the exact diff and proposal hash, and wait for explicit approval. Run route apply --approved only for that unchanged hash.
9. Offer opt-in baseline recording. Later verification needs at least five comparable sessions in each window and must say association, never causation.

Report contract

Lead with any coverage warning. Then provide two supported strengths and no more than three findings ranked by support, detector quality, completeness, recency, and actionability. Each finding states logical session count and sources, evidence kind, confidence, plausible alternatives, benefit, trade-off, and next measurement. Cross-tool spread is descriptive only. Do not quote transcripts before consent.

Use stable finding IDs from si.diagnostic/v1. Do not expose opaque pointers unless needed for evidence selection.

References

• running.md: CLI, staged workflow, and routing
• measurement.md: definitions and causal limits
• privacy.md: release and state boundaries
• contracts.md: public schemas
• capabilities.md: analysis lenses
• detection-playbook.md: confirmation and alternatives
• improving-ai-workflows.md: artifact craft
• sources.md: adapters and degradation

Validation

PYTHONPATH=scripts python3 -m unittest discover -s tests -v
python3 scripts/si.py diagnose --scope project --path /Users/dev/acme-web --root evals/fixtures --window-days 90

The fixture run must report 8 project sessions across all three sources, contain no transcript text, and produce no persistent write.