@blackbelt-technology/pi-dashboard-context-budget0.11.0

@blackbelt-technology/pi-dashboard-context-budget

Measure what a pi session actually sends the model, every turn, before any work happens.

The meter attaches to before_provider_request, which fires after the provider payload is built and before it is sent — so the numbers are the real request, not a reconstruction of one. ctx.getSystemPrompt() reports pi's system prompt string, which is not the same thing as the serialized payload.

Why

A context trim can apply cleanly, produce a valid config, and change nothing on the wire. That happened during this package's own development: a "skills": ["-…"] exclusion in project settings was accepted, looked correct, and was a silent no-op (package-provided skills must be excluded in the package entry). Only a before/after measurement caught it.

diff --expect-removed encodes that lesson: it exits non-zero when something you expected to remove is still on the wire.

Usage

# measure the current directory's session cost
npx context-budget measure -o before.json

# ...change settings / skills / tools...
npx context-budget measure -o after.json

# prove the change reached the wire
npx context-budget diff before.json after.json --expect-removed ctx-doctor,memory_remove

measure runs one headless pi turn (pi -e <meter> -p "reply with the single word: ok") in the current working directory, so it reflects that project's settings, skills and extensions.

To measure the effect of dropping tools without editing any config:

CONTEXT_BUDGET_DROP=memory_remove,recall npx context-budget measure -o trimmed.json

What it reports

• payload total, split into system / tools / messages
• system-prompt blocks — skills-catalogue, project-context, memory-policy, and other. The parts always sum to the whole: a block that stops matching shows up as other growing rather than as bytes disappearing.
• per-tool schema bytes, sorted — the cost of each registered tool
• per-skill catalogue bytes — what each advertised skill costs every turn

Library

import { analyzePayload, comparePayloads, checkBudget } from "@blackbelt-technology/pi-dashboard-context-budget";

const b = analyzePayload(capturedPayload);
const { ok, violations } = checkBudget(b, { maxPayloadBytes: 120_000 });

analyzePayload is pure — no pi, no network, no filesystem — so budget assertions can run in CI against a stored capture.

Env