session-analysislisted

Use when the user wants a judgment-level read on whether a Claude Code session stayed on task — did the agent do what was originally asked, and what did it acknowledge skipping. This is the interpretive (LLM) complement to the deterministic `session-audit` CLI: 1a extracts ask-vs-done for free; this skill adds the `Variance` and `What was NOT done` judgment that a parser can't compute, then persists a combined record for drift analysis. Cost: ~1-3k tokens of the current session, paid only when invoked. Opt-in by design — run it selectively, not on every session. Best run in a Sonnet (or stronger) session; the judgment quality depends on it. Use `session-audit` (CLI) for the free deterministic ask/actions extract. Use `usage-analysis` for token-spend insights. Use `claude-audit` for agent/skill config overlap. Use THIS skill for "did this session drift from what I asked". Trigger phrases: "/session-analysis", "did this session stay on task", "analyze session drift", "did the agent do what I asked", "what did t
glitchwerks/claude-prospector · ★ 1 · AI & Automation · score 67

Install: claude install-skill glitchwerks/claude-prospector

# 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](https://github.com/glitchwerks/claude-prospector#install-as-a-claude-code-plugin). ## 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) ```bash 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 co