Skip to content

Codex Speed Playbook

Jump to bottom
Nick Hamze edited this page Jun 2, 2026 · 1 revision

Codex Speed Playbook

Codex Refit watches three broad kinds of local speed pressure: local state, runtime/config, and workflow context.

Local State

Refit checks:

  • active transcript size
  • media-heavy appshot/image sessions
  • archived pointers still sitting in active sessions
  • SQLite log and WAL size
  • crash dumps
  • rebuildable browser caches
  • stale active threads
  • oversized transcript counts
  • Codex-managed worktree pressure
  • Chronicle screen-cache and generated-memory load

Runtime And Config

Refit checks:

  • official Codex Doctor findings
  • CLI/app version drift
  • Codex launch latency
  • strict config validation
  • app-server transport health
  • SSH remote-host readiness
  • shell startup latency
  • CLI completion setup
  • background terminal commands
  • default model and reasoning fit
  • Plan Effort configuration
  • cloud handoff readiness
  • model-provider auth health
  • named speed-profile health
  • subagent fan-out settings
  • custom-agent load
  • MCP startup pressure
  • shell environment breadth
  • notification flow
  • history retention
  • storage paths
  • telemetry export
  • Codex exec automation hygiene
  • context budget and status-line settings
  • response verbosity and reasoning-summary shape
  • effective web-search mode
  • deprecated web-search toggles
  • network and sandbox friction
  • permissions profile shape
  • instruction-stack size
  • custom instruction overrides
  • command-rule pressure
  • lifecycle hook load
  • automatic approval review
  • memory-injection weight
  • deprecated custom prompts
  • skill-catalog metadata pressure
  • project local-environment setup/action readiness
  • file-based managed Codex config conflicts

Workflow Context

Refit samples active-thread metadata and markers for:

  • task clarity
  • done-when criteria
  • verification language
  • observed turn duration
  • first-token delay
  • rate-limit metadata
  • near-context-window token counts
  • approval and sandbox stop markers
  • empty or missing global/project AGENTS.md
  • oversized or capped durable guidance
  • missing .codex local setup/actions
  • current-project codex exec and GitHub Action patterns
  • large skill catalogs
  • broad concurrent thread settings
  • task/model mismatch

Plan Effort

Plan Effort is a dedicated Doctor and Speed Check signal.

It looks for plan_mode_reasoning_effort so ambiguous work can use deeper planning while everyday turns and named speed profiles stay lighter.

A good default shape is:

model_reasoning_effort = "medium"
plan_mode_reasoning_effort = "high"

For small, well-scoped tasks, a named speed profile can stay lighter:

model = "gpt-5.4-mini"
model_reasoning_effort = "low"
model_verbosity = "low"

Speed Check

Speed Check records a stable local readiness score using local data such as scan time, database latency, log pressure, transcript size, session-media pressure, task clarity, validation readiness, review readiness, Codex manual freshness, runtime load, and config shape.

It also records a live timing score that reflects current process load, background commands, helper pressure, and runtime health.

Speed Proof History

Speed Proof compares:

  • latest score
  • best score
  • trend
  • active-folder change
  • media-session change
  • task-clarity change
  • planning-template change
  • Plan Effort change
  • Goal Mode change
  • Validation Loop change
  • Review Lane change
  • Codex manual freshness change
  • turn-telemetry change
  • approval-friction change
  • permissions-profile change
  • Chronicle-load change
  • notification-flow change
  • history-retention change
  • storage-path change
  • telemetry-export change
  • status-line visibility change
  • CLI-completion change
  • strict-config change
  • Codex-launch change
  • automation change
  • app-server transport change
  • remote-host readiness change
  • shell-startup timing change
  • cloud-handoff change
  • live process-count change
  • worktree-pressure change
  • model-provider change
  • model-effort change
  • speed-profile health change
  • custom-agent change
  • response-shape change
  • effective web-search change
  • network-policy change
  • instruction-stack change
  • instruction-override change
  • command-rule change
  • custom-prompt change
  • skill-catalog changes
  • local-environment setup/action changes
  • managed-config changes

Practical Tuning Notes

  • Archive or split image-heavy active threads once the image context is no longer needed.
  • Use /compact on long threads before they turn into huge active transcripts.
  • Use /status to check context and rate-limit state.
  • Use /ps and /stop to inspect or stop background terminal work.
  • Keep Codex history retention capped if history.jsonl grows too large.
  • Keep SQLite and logs on local unsynced storage.
  • Turn off remote telemetry export outside tracing sessions.
  • Prefer clear AGENTS.md guidance over repeated prompt corrections.
  • Add a focused test/build/review proof path to projects Codex works on often.
  • Use /review before committing large local diffs.
  • Keep subagent fan-out conservative unless the task is deliberately parallel.
  • Review MCP servers with /mcp verbose when startup feels slow.
  • Use scoped permissions profiles instead of broad defaults.
  • Keep prompt shape simple: goal, context, constraints, done when.

Codex Refit

Clone this wiki locally