Codex Rollout Insights

A local-first VS Code extension for observing raw Codex rollout token telemetry. It is deliberately not a ChatGPT quota, credits, or billing calculator.

Status: functional V1 alpha with a packaged VSIX. See V1 readiness for the remaining Marketplace-release work.

Current implementation

The first milestone is a streaming rollout analyzer. It reads Codex session JSONL without modifying it, normalizes the evolving raw schema, and reports:

• latest turn and cumulative thread usage when Codex supplies them;
• input, cached-input, cache-write, output, reasoning-output, and total tokens;
• model metadata, inference-call count, and context-window capacity;
• the source and confidence of each usage total.
• a native session tree that separates prompt turns, background activity, model inferences, and tool call/result pairs.

Run it against the auto-discovered Codex sessions directory, or provide a rollout file explicitly:

npm install
npm run analyze -- "C:\path\to\rollout.jsonl"

Run the fixture tests with npm test.

No prompt text, source code, tool output, credentials, or telemetry leave the machine. The extension does not proxy traffic, attach to Codex's app-server connection, or modify rollout files.

What is and is not feasible

The local rollout format currently includes token_usage_record entries with per-inference, cumulative-turn, and cumulative-thread token telemetry. This makes the raw usage view feasible without an API key. token_count snapshots remain a compatible fallback and expose context-window capacity.

Rollout telemetry alone cannot authoritatively establish ChatGPT plan quota, credits, billing cost, or rate limits. It also cannot always prove attribution across compaction, retries, background work, or sub-agent threads. Those values must remain unavailable or explicitly estimated until a supported authoritative Codex surface provides them.

model_context_window reports capacity, not current occupancy. The monitor will never represent cumulative session usage as context-window occupancy.

See architecture for the local session/turn model and privacy for its data-handling guarantees.

Multiple sessions on one machine

Each Codex rollout is treated as an independent observed session. Open the native Sessions tree to select one and pin it for the current VS Code workspace; the live observer then tails that exact rollout. Expand the selected session to inspect its prompt turns, model inferences, and tool-call/result pairs without exposing their contents.

This explicit choice is intentional: local rollout files do not reliably state which external Codex chat or tab a person considers current when multiple sessions are active at once. The session catalog scans only the configured number of recent sessions (20 by default), hides sub-agent/internal threads by default, and stores no prompt or tool-output contents. Local session names and prompt previews are opt-in settings because either may contain prompt-derived text.