Jev Codex Pilot

The 1 to 5 score selects the default model and reasoning route from configuration and Codex can step up reasoning mid-ticket. Scores guide routing only and never block execution.

No. Decision questions stay provider-neutral and gateway specifics live in one module, so a direct TypeSafe adapter can be added without duplicating the ticket logic.

Each project runs Codex concurrently in its own home with separate sessions and state, reusing your Codex auth. Account limits stay shared, and launching a sequence drains pending tickets in queue order.

The Git workspace shows the branch, unpushed commits, and per-commit diffs, with optional per-ticket Conventional Commit commits. Nothing pushes without using the explicit push action, and no branch is created automatically.

Level 1 is reserved for reading or explaining documentation, project information or code, plus limited edits to README.md or AGENTS.md. Level 2 covers simple implementation, running tests, installation commands, and tiny cosmetic visual tweaks that do not change layout, accessibility or interaction. Any UI/UX change starts at level 3, complex UI/UX or 3D work can be level 4, and critical or long-horizon work can be level 5. The shipped defaults are GPT-6 Luna Medium at level 1, GPT-6 Luna High at level 2, and GPT-6.1 Sol Low, Medium or High for levels 3 to 5, editable in config/model.toml.

Two ways. When another ticket is queued, JEV reviews whether it relates to the completed work, clears unrelated work, and compacts a related or uncertain thread once context passes 100,000 tokens. Before native Codex compaction it replays the active transcript, pairs completed tool calls with their results, and decides per pair whether to keep it verbatim, keep the call with a bounded result, or drop the pair. The newest items are pinned, malformed decisions fall back to native compaction, and a SessionStart hook restores missing high-priority content from a secrets-redacted checkpoint.

No, and the README is explicit that it is an estimate from reported input tokens at $0.04 per million, not a billing receipt. It covers ticket evaluation and re-evaluation, continuity reviews, and the native hooks that run during implementation and compaction, and it stays attached to the ticket across retries, escalation and quota resumes. Codex token totals are reported from completed turns, while account usage is labelled separately when available, and missing telemetry shows as unavailable or incomplete rather than as zero.