session-analysis

Session Analysis Skill (1b)

You are producing the judgment half of session drift-detection: given a session's deterministic ask-vs-done extract (from 1a), assess whether the agent stayed on task and what it acknowledged leaving undone — then persist a combined record. This is interpretive work; a deterministic parser cannot do it, which is exactly why it costs LLM tokens and is opt-in.

Prerequisites

This skill drives the claude_prospector CLI. The package must be installed in the environment Claude Code uses — see the README install steps.

Step 1 — Identify the target session

The session-id (or transcript path) is `` if provided.

• If `` is a session-id, use it directly.
• If `` is empty, find the most recent transcript for the current project under ~/.claude/projects/<encoded-cwd>/*.jsonl (newest mtime) and confirm the session-id with the user before spending tokens — variance analysis is opt-in, don't guess silently.

Step 2 — Load the deterministic extract (free, 1a)

python -m claude_prospector session-audit --session-id <id> --format json

This returns {original_ask, prior_asks, actions}. original_ask is the authoritative first ask; prior_asks are later distinct asks in the session; actions are the Edit/Write/NotebookEdit file paths. Reason over this — it is your ground truth for "what was asked" and "what was changed".

For richer context (reasoning, tool failures, what the agent said it was doing), also read the transcript itself at the resolved ~/.claude/projects/<…>/<id>.jsonl. Use it to judge intent, not to recompute the deterministic fields.

Step 3 — Form the judgment

Assess two fields against original_ask (+ prior_asks for multi-task sessions):

• variance — did the agent stay on the original ask? Note scope creep, approach pivots, or drift onto a later ask at the expense of the first. Cite specifics (a file in actions unrelated to the ask; a pivot point in the transcript). If it stayed on task, say so plainly — "no variance" is a valid, useful finding.
• not_done — what did the agent acknowledge skipping or defer? Prefer the agent's own admissions in the transcript over your speculation. If nothing was skipped, say so.

Optionally assign severity (integer 0-3): 0 = on task, 1 = minor drift, 2 = notable unrequested scope or skipped ask, 3 = the session largely did not do what was asked. Omit (null) if you can't justify a number.

Be evidence-bound: every claim cites a file path, a prior_asks entry, or a transcript moment. Do not invent drift to fill the field.

Step 4 — Persist the combined record

Write the judgment to a temp JSON file (prose is multi-line; don't fight shell escaping), then call variance-save. It re-loads 1a internally and writes the combined {1a fields + your judgment} to <data>/variance/<id>.json:

# judgment.json: {"variance": "...", "not_done": "...", "severity": <int|null>}
python -m claude_prospector variance-save --session-id <id> --judgment-file judgment.json

variance-save finds the transcript under ~/.claude and writes output under the plugin data dir by default — no extra flags needed. It prints the written path; surface that to the user.

Step 5 — Report

Give the user a short variance report (NOT a JSON dump):

Session <id> — variance: <one-line verdict, severity if assigned>
  Asked:    <original_ask, trimmed>
  Variance: <the judgment, with its citation>
  Skipped:  <not_done, with its citation>
  Saved:    <path printed by variance-save>

When to run (and not)

• Run it on a session you suspect drifted, a long multi-task session, or one flagged by prior_asks.length > 0. Selective use is the whole cost argument — 1a+1b beats the abandoned always-on hook only when 1b runs on a minority of sessions.
• Don't auto-run it on every session — that re-introduces the per-session cost this design exists to avoid.