Platform
A limit that fits how you actually work.
Caps are read by the guard on every tool call. Everything is off until you switch it on, and every limit is a number you set — or one of three presets sized from this machine’s own median session and median day. The fourth setting is Off.
Big reads ask first
A read with no offset or limit, on a file whose size implies more than the threshold, pauses and asks. A paged or searched read goes straight through.
Three budgets
A session budget, a daily one, and a tighter unattended one. The meter is the session's own transcript, parsed incrementally from a stored offset.
Unattended means nobody is there
When you have not touched the keyboard for a few minutes, nobody can answer a prompt. That work gets the tighter cap, because a loop at 3am is what empties a budget.
This layer fails open, on purpose
Any error inside the cap logic — an unreadable policy, a half-written transcript, a corrupt tally — lets the call through. The checks above it, for originals and destructive verbs, stay fail-closed.
That asymmetry is deliberate and it is the design's most important line. A safety control that fails open is not a control. A cost control that fails closed stops your work because a JSON file was half-written, and a cost control that stops your work gets uninstalled by the end of the week.
A cap that trips says what it is about
It uses the same ask-and-wait protocol as a doctrine block — the hook writes a request, the app raises the prompt, you answer — and the prompt says in as many words that this one is about cost rather than safety. Confusing the two teaches people to dismiss both.
Presets, measured rather than invented
Careful, everyday and lean are derived from your own transcripts: the median session and the median day on this machine. Off is the fourth setting and is derived from nothing — it is the state everything starts in. A limit copied from someone else’s working pattern is a limit you will spend a week fighting.
What is deliberately not built, in the cost layer
This list is about spend and budgets. The Agent Roster reads eight tools’ configs, and fixes and drift alerts exist elsewhere in the product — they are not part of this.
- No OpenTelemetry receiver. Cost is rebuilt from Claude Code and Codex transcripts; spend through any other agent is invisible to it.
- No remediation scripts for spend, and no alerting on a budget.
- No way for Mockingbyrd to change an agent’s effort level.