alexgreensh/token-optimizer v5.8.0 on GitHub

Codex runtime-state intelligence + live quality refresh

Runtime-state reader — subagent token cost and leak detection, memory overhead, and goal-budget utilization, read directly (read-only) from Codex's SQLite state. New codex-state command and dashboard surfacing.
Modern session parsing — collab subagent events, compaction compression ratios, rate limits, per-turn effort, and tool durations.
SubagentStart/SubagentStop hooks — real-time subagent sprawl nudges (balanced/aggressive profiles; opt out with --no-subagent-hooks).
Doctor and status-line updates aligned to the current Codex surface.

Context quality now refreshes during long autonomous runs (throttled, on tool use), so the status-line score reflects live session state instead of the last prompt.
Per-session context-fill attribution keeps concurrent sessions independent.
The status line marks a score that isn't live.

Quality refreshes on tool execution, not just on prompts (token-optimizer-opencode 1.0.2).

Read-only, defense-in-depth SQLite access. Full feature parity preserved across Claude Code, Codex, and OpenCode.