qs-plugins

A Claude Code plugin marketplace.

claude plugin marketplace add QSchlegel/qs-plugins

session-viz

Provides /qpact — analyse the current session, render an interactive HTML report into a pop-out window, and derive a /compact instruction tuned to what the session was actually about.

claude plugin install session-viz@qs-plugins

Plugin skills register at session start, so start a new session before /qpact resolves.

What it does

1. Extract — streams the session transcript into a compact turn spine. Human turns are only 2–4% of records, so a 42 MB transcript collapses to ~100 KB in under a second.
2. Score — deterministic prompting-quality metrics (see below).
3. Derive — Claude writes a TL;DR, an intent breakdown, and the /compact focus instructions.
4. Render — a self-contained HTML file opened in a real window, with the /compact line copy-pasteable at the top.

/compact is a built-in command that only the user can invoke, and PreCompact hooks cannot inject instructions into it — hence the copy-pasteable line rather than automatic compaction.

How the score works

Two tiers of evidence, weighted apart. Outcome penalties are things the transcript witnessed and dominate the score; form adjustments are guesses from the prompt text alone and move it only slightly.

Outcome signal	Points
re-sent verbatim (the first attempt did not land)	−30
interrupted mid-flight	−25
drew a correction on the next turn	−20
was itself a correction	−10
no tools ran and no question was asked	−8

Turns start at 72, not 100 — a prompt that merely did not fail is not the same as a good one. Terseness is penalised only for fresh prompts, never for mid-turn steering, which is short because the context is already loaded.

The session score is built from rates, not from a mean of turn scores: averaging hundreds of turns converges on the baseline by construction. Friction is counted per turn and deliberately not weighted by tokens — an interrupted turn is cheap precisely because it was interrupted, so cost-weighting erases the failures being measured. Wasted tokens are reported separately.

Below ~20 turns the score is reported with low confidence: the outcome signals have too little to witness for the number to mean much.

Using the scripts directly

node plugins/session-viz/scripts/extract.mjs --list
node plugins/session-viz/scripts/extract.mjs --project my-repo --json > spine.json
node plugins/session-viz/scripts/render.mjs spine.json --open

extract.mjs redacts API keys, tokens, JWTs and connection strings from prompt text by default. Transcripts contain whatever you have pasted into them; review before sharing a report.

Licence

MIT