context-control

context-control

A local UI for testing the selection half of ghost's loop. steering-control measures what a ghost package buys downstream of generation; context-control measures whether the right guidance even makes it into context. Selection is the unit. No generation, ever.

ghost gather does no filtering: the menu is always the whole catalog, and selection happens in the model's head against each node's for payload. So what this bench actually tests is whether for payloads are good enough retrieval payloads — a node with a vague for is invisible at selection time no matter how good its body is.

Run

pnpm --filter @design-intelligence/ghost build   # bench shells the real CLI
tmp=$(mktemp -d)
node packages/ghost/dist/bin.js init --package "$tmp/.ghost"
node packages/context-control/cli.mjs \
  --package "$tmp/.ghost" \
  --asks packages/context-control/demo/asks.md
# → http://127.0.0.1:4114

Options: --package <dir> (default ./.ghost), --asks <file>, --port <n> (default 4114), --ghost <bin> (default: repo dist build, else ghost on PATH).

Screens

package — the catalog behind the selection surface: id, kind, for payload, material count, and coverage. Click an item to see its real ghost pull output in a drawer. Review for payloads as retrieval payloads, not file contents; an item with no for is flagged as invisible.

bench — type an ask (or run the whole asks suite), fire N single-shot selection trials, and read the heatmap: nodes × asks, each cell the fraction of trials that selected the node. Solid column = confident for payload. Speckled = coin-flip. Empty row = dead node. Blue outline = the ask's expected set. Scores above the map: consistency (mean pairwise Jaccard), mean per-trial precision and recall, poison-selection rate, unknown ids, and nodes ever selected.

Selection runs against the exact agent-facing Markdown from ghost gather <ask>. The system prompt only requests applicable IDs; it does not reconstruct the menu or repeat Ghost's selection mechanics. One caveat remains: a live agent also carries task context (open files, prior turns) that the single-shot selector lacks.

replay — the real .ghost/.events tape grouped into sessions: each gather with its ask, the pulls that followed, re-gathers, and pull misses. Ground truth from the field keeps the bench honest. Events stay local; nothing leaves the machine.

Asks format

Use the same ask blocks as steering-control. expect: lists nodes a good selector should pull. poison: lists nodes whose condition does not apply. discount: remains available to steering-control and is ignored here.

## Ask 1 — notification settings

Build a dense settings screen for notification preferences.

expect: foundation.composition, foundation.controls, foundation.layout
poison: context.conversation

Asks without expect: still score consistency. Expected and poison sets add precision, recall, and poison rate. List only selectable menu ids. Do not list the manifest cover: gather excludes it from selection and every pull includes it automatically.

Model adapters

The contract is one function: select({ ask, menu, trial }) -> ids. Two adapters ship, selectable in the bench UI:

• fake-lexical — deterministic lexical-overlap stub with per-trial jitter near the decision boundary. No network, instant. Use it to exercise the UI loop and as a dumb-retrieval baseline.
• openai-compatible — a real LLM behind any OpenAI-compatible chat API. Configure CONTEXT_CONTROL_BASE_URL, CONTEXT_CONTROL_API_KEY, and CONTEXT_CONTROL_MODEL in an untracked .env or .env.local at the working directory. The CLI loads it on startup and makes this adapter the default when all three values are present. Trials run concurrently and sample at the endpoint's default temperature because trial-to-trial variance is the signal being measured.

Add providers to MODEL_ADAPTERS in lib/model.mjs.

Layout

cli.mjs          # context-control → serves the UI
lib/ghost.mjs    # shells exact Markdown for trials; JSON for inspection and pull
lib/model.mjs    # model adapters; real models receive literal gather Markdown
lib/bench.mjs    # trial runner + asks.md parser
lib/score.mjs    # jaccard, consistency, precision/recall, rates, coverage
lib/tape.mjs     # .ghost/.events parser + session grouping
lib/server.mjs   # node:http JSON endpoints + static UI
ui/index.html    # single-file UI, no build step
demo/asks.md     # demo suite against the default ghost init skeleton