External capacity snapshots

Frost reads one neutral usage snapshot and never runs a producer. This directory is an example producer for CodexBar 0.60.2 that writes the snapshot Frost's capacity_file points at. Any script, usage tool, or hand-edited file that emits the same shape works too; the contract is in the slice 2 plan.

Requirements

Install CodexBar and jq and ensure both are on PATH before running the wrapper. Configure the provider access CodexBar needs and edit the bindings for your accounts.

Files

• codexbar-to-frost.jq: converts CodexBar usage --json rows into the snapshot. Account emails, credit inventory, tokens, and raw error bodies are dropped.
• bindings.json: which CodexBar provider (and source) feeds which Frost pool, and which CodexBar window IDs to read. Copy it to ~/.config/frost/capacity-bindings.json and edit; it is the operator's document.
• refresh.sh: runs CodexBar per provider, converts, and atomically replaces the output. A provider that fails becomes a source_status: error pool; if every provider fails the previous file stays. Afterwards it runs frost capacity check when frost is on PATH.
• install-launchd.sh: schedules refresh.sh every five minutes as the user LaunchAgent net.vorwaller.frost-capacity, logging to ~/Library/Logs/frost/capacity.log. --uninstall removes it.

cp examples/capacity/bindings.json ~/.config/frost/capacity-bindings.json
examples/capacity/refresh.sh --bindings ~/.config/frost/capacity-bindings.json --dry-run | head
examples/capacity/install-launchd.sh
frost capacity check ~/.config/frost/capacity.json

What the snapshot can and cannot say

The operator config owns topology: which pools exist, which windows each requires, which windows earn the expiry preference, and which profiles draw from which pools. The snapshot only supplies observations keyed by those IDs; unknown pools or windows are validation errors. generated_at never freshens an observation; observed_at comes from CodexBar's updatedAt. A null window stays unknown, never 100%. After a reset boundary passes, Frost treats the old value as unknown until the next refresh.

Measurements follow CodexBar's dataConfidence: exact and percentOnly (Codex OAuth, Claude web) become exact, estimated (OpenCode Go) stays estimated, and a row with no confidence takes the binding's measurement field, else unknown. Estimated values count only when the policy opts in.

Verified and unverified bindings

Pool	CodexBar windows	Status on September 16, 2026
codex-main	secondary (weekly)	Codex is a flat weekly limit; the five-hour primary is null and not declared.
claude-main	primary (five-hour), secondary (weekly)	Claude web source reports both.
claude-fable	claude-weekly-scoped-fable	A Fable-only weekly allowance CodexBar exposes as an extra window; bound to the Fable profile alongside claude-main. CodexBar omits it on some runs, and then the pool is unknown until the next refresh, which leaves Fable unknown rather than excluded.
opencode-go	primary, secondary, tertiary	Estimated five-hour, weekly, monthly.
codex-spark	codex-spark, codex-spark-weekly	Which native models draw from Spark is unverified; mapping_verified = false keeps it unknown.

Antigravity and Grok pools wait for confirmed window IDs. Reset credits are inventory, not capacity, and are never emitted.