rate-awareness

name	rate-awareness
description	Token-bucket advisory rate limiter per skill+session. Surfaces runaway tool-call loops in real time before pech/budget-watcher reports them post-hoc. Auto-fires on PreToolUse. Configurable per-skill limits via state/buckets.json. Do not use for cost reporting (see budget-watcher) or for blocking installs (see package-gate).
model	haiku
tools	["Read","Bash"]

rate-awareness

Token-bucket advisory limiter. Each tool call consumes one token from a per (session, skill) bucket. When the bucket is empty, an advisory is emitted to stderr — the tool still proceeds (advisory-only contract per ../vis/packages/core/conduct/hooks.md).

Preconditions

• state/buckets.json exists (opt-in gate — absence = silent no-op).
• The PreToolUse hook is registered via hooks/hooks.json.

Inputs

• Hook event: PreToolUse on every tool (matcher .*).
• Bucket config: state/buckets.json — global_default plus optional per_skill overrides.
• Session id: $CLAUDE_SESSION_ID if set, else SHA-1 hash of cwd:ppid truncated to 16 chars.
• Skill id: best-effort — most recent skills/<name>/ ancestor of cwd, else ungated.

Steps

1. Hook (hooks/pretooluse-rate.sh) fires; if state/buckets.json missing, exits 0 silently.
2. scripts/rate-check.py opens state/buckets.json r+, takes an exclusive lock (msvcrt on Windows, fcntl elsewhere).
3. Reads state. Resolves capacity + refill rate via per-skill override or global_default.
4. Refills the (session, skill) bucket: tokens = min(capacity, tokens + dt * refill).
5. If tokens >= 1 → consume 1, no advisory. Else → record empty, set advisory = true.
6. Write-back atomically (truncate + fsync) under the same lock.
7. If advisory: emit stderr message; always exit 0.

Success criterion: every PreToolUse decrements (or refills) exactly one bucket entry; the 61st call within a 1-second window for a bucket of capacity 60 emits the advisory; the tool call proceeds regardless.

Outputs

• Updated state/buckets.json (mutated _buckets map keyed by <session>::<skill>).
• Stderr advisory line on bucket-empty (visible to the developer; not injected into the model context).

Configuration

state/buckets.json schema:

{
  "global_default": {"capacity": 60, "refill_per_sec": 1.0},
  "per_skill": {
    "deep-research": {"capacity": 200, "refill_per_sec": 5.0}
  }
}

Add per-skill entries as needed. Delete the file entirely to disable the plugin (opt-in by file presence).

Handoff

Downstream: pech/budget-watcher still owns post-hoc cost ceilings. This plugin only surfaces velocity anomalies before the spend lands.

Failure modes

Code	Scenario	Counter
F09	Two PreToolUse hooks writing buckets.json concurrently	Exclusive flock/msvcrt lock around the read-modify-write
F05	Advisory spam on every call once bucket empties	Tokens stay at 0 until refill catches up; one stderr line per emptied call is the signal, not a bug
F13	Skill mis-identified as ungated when cwd is outside any skills/ tree	Default ungated bucket uses global_default; per-skill overrides only fire when detection succeeds
F08	Hook calling the script with wrong python	python3 per pech convention; matches budget-watcher hooks