Automatic observations and control boundaries

What the MCP now does

The remote plan_context_budget is still an explicit-number calculator. Its purpose is reproducible arithmetic and structured results, not private telemetry.

The separate **local stdio** inspect_local_context removes manual transcription of available Codex metrics. It reads the exact bound root and workspace, reports compatible usage/model/effort/cache/throughput/account observations, recalculates the reference budget and suggests a task size/output reserve. A watch command refreshes these local recommendations automatically. Unknown/stale data stays unknown and cannot justify a current action. These are estimates, not exact active context or monetary budgets.

The local tool never appears in remote HTTP discovery or the Docker image. It has no account, credential or paid model API call. Its result omits message text, chat/turn IDs and paths. It does not modify Codex configuration or hooks.

Model, effort and chat changes

These three controls are **disabled in open-source 1.0.0**. A read-only transcript does not grant control of the existing Desktop chat. Starting another app-server does not establish ownership of that chat. The source contains existing guarded continuation work, but its current Desktop owner/complete-state gates have not been verified. This release does not promote it to automatic mode.

For a future controlled integration, require a supported connection to the exact owner, correct root/turn binding, complete active-task/background state, explicit allowed models and effort ranges, a cost/plan policy and normal hook review. Check request receipts before retrying or creating another chat. Preserve policy and task state during handover. Do not treat an interrupt reply as proof of a complete stop or skip guards because a read-only estimate looks near a threshold.

App Server documentation describes model/effort overrides on user turns and thread lifecycle primitives. Using those in a harness you own is different from controlling the Desktop chat. A successful isolated fixture or local TLS test does not verify the latter.

Until the supported owner path and runtime policy are available, show a concrete recommendation and say the control was not executed. Do not change global config, silently start model turns, spend credits, bypass /hooks, manipulate UI settings or invent that a chat was moved. Missing cost/entitlement information remains unknown; changing a model/effort is not a guaranteed saving.

Source: official App Server documentation.