O15 — Traces, health and cost¶
Status: design stub. Back to the component map.
Purpose¶
Explain what a session is doing and why it may be stuck, failing or expensive. Support the same useful computations in live views, historical inspection and optional host processors.
Existing foundation¶
Reuse MITM/observation boundaries, Codex findings and history analysis. Hosted exports and richer client reducers have separate implementation/qualification gates.
Proposed boundary¶
Capture and export must not block agent traffic. Keep native events, transport evidence and derived estimates distinct, deduplicate overlapping usage, and retain missing coverage. API-equivalent price estimates are not actual subscription bills. Findings are advisory; automated intervention is a later policy decision.
Inputs and outputs: Provider observations, histories, optional captures and price inputs → trace views, usage/context summaries and explainable health indicators.
O21 can select these processors and aggregate status. Place taps at the relevant session/relay boundary; a coordinator-owned relay or required MITM proxy is still on the live failure path. After a failure, preserve recorded coverage and mark missing tails, even when their exact size is unknown. This evidence supports agent-assisted recovery.
Candidate progression¶
- Mock: A tool loop, sandbox errors, pending user input, context pressure, a stalled connection and incomplete cost coverage.
- First usable slice: Show observed tool count, pending work, usage and provenance for one real session and its saved counterpart.
- Later: Optional MITM/OTLP/Logfire sinks, configurable diagnostics, fleet summaries and comparison of estimates across sources.
These are discussion candidates, not scheduled work or implemented capabilities.
Your additions¶
- Which conditions should draw attention automatically, and which should remain available only on inspection?
- Which trace/cost details help you decide what action to take?
Additional requirements, examples and decisions: TBD with you.