ce-work's codex route hard-codes reasoning effort; the review peer's route_effort() takes an override

Open
#1,569 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

  • #1616 by @ucirello — closed without merging

Assessment

Difficulty
3/5
Estimated time
1-2 days
Newbie friendliness
35/100
Issue type
Feature
Clarity
Mostly clear
Activity status
Stale
Tech stack
shell, typescript
Domain
tooling

Research direction

Start by reading skills/ce-work/scripts/cross-model-work.sh, especially adapter_argv(), then compare it with route_effort() in skills/ce-code-review/scripts/cross-model-adversarial-review.sh. Decide whether the work route should accept an effort override while preserving its current default; the issue asks for that hook, not a default change. The linked pull request #1616 is closed and unmerged, so check its context before choosing a next step.

Written by the indexing model from the issue text.

Description

Summary

On 3.22.4 the two codex adapters resolve reasoning effort by opposite rules. ce-code-review goes through route_effort(), which reads CROSS_MODEL_EFFORT_OVERRIDE and falls back to a per-route editorial default. ce-work writes the value inline in adapter_argv() and offers no override path.

Same CLI, opposite policy. I searched open and closed issues and found nobody asking about the difference, so I am asking rather than sending a patch.

I filed #1565 this morning against the review path's effort whitelist. This is the neighbouring question on the work path, one layer down: #1565 is a stale value list inside a knob that exists, and this is the absence of the knob.

The two paths

skills/ce-work/scripts/cross-model-work.sh, adapter_argv():

codex)
  # --ignore-user-config drops the user's model_reasoning_effort, so pin the
  # editorial tier explicitly, matching the claude/grok routes' --effort high.
  printf '%s\0' codex exec --ignore-user-config --ignore-rules --ephemeral \
    -s workspace-write -C "$WORKSPACE" --json -o "$RAW_RESULT" \
    -c model_reasoning_effort=high

skills/ce-code-review/scripts/cross-model-adversarial-review.sh:

route_effort() {   # <route> -> requested effort: the override where the route takes one, else editorial
  if [ -n "${CROSS_MODEL_EFFORT_OVERRIDE:-}" ]; then
    case "$1" in codex|claude|grok-cli) printf '%s' "$CROSS_MODEL_EFFORT_OVERRIDE"; return 0 ;; esac
  fi
  case "$1" in
    codex) printf 'xhigh' ;;
    claude|grok-cli) printf 'high' ;;

The review script also pins M_CODEX="gpt-5.6-luna" at line 80. So the plugin already runs Luna, and runs it above high.

Why #1289 and #1314 do not cover this

Both reported that the codex work route ran at the CLI default while claude and grok already carried --effort high. The ask was parity, not selectability. #1289's author proposed pinning to high in his own report and shipped #1290 two hours later. #1314 was pointed at that merge and closed.

The question of whether the value should be choosable has never been put, which is why I am opening this instead of commenting on either.

Why a user cannot route around it

The three obstacles #1289 listed still stand. --ignore-user-config drops ~/.codex/config.toml, correctly. The engine config contract forbids CLI flags in work_engine_preferences. The codex CLI carries effort as a separate key rather than an effort-suffixed slug, so a model pin cannot smuggle one.

What makes the gap hard to read from outside

The same model writes code at high and reviews it at xhigh. You closed #1261 by explaining that the luna/xhigh mapping was deliberate, adopted in #1208 after the peer benchmark was re-run with cost priced in API dollars, and written up in docs/solutions/skill-design/benchmark-review-peer-model-and-reasoning-tier.md. That reasoning is legible. The work route's high has a comment that explains what it matches, not what it was measured against.

Luna is also where the dial moves most. Artificial Analysis puts it at 50.7 on the Intelligence Index at medium effort and 71.4 at max, and max still costs $0.21 per task. Whether high is right for implementation I have no data on. My point is narrower: on this model the value is worth choosing, and right now it cannot be.

The question

Is high on the work route a measured editorial decision like the review peer's xhigh, or is it the parity default from #1290 that has not been revisited since?

If the first, a pointer to the write-up closes this and I will stop asking.

If the second, I would like to know whether an override hook on the work adapter is welcome, shaped like the one route_effort() already implements.

#1567 is open and unreviewed at the time of writing, so I am not leaning on it. If it lands, max and none become legal on the codex route and a work-side hook inherits a value space that is already settled. If it does not, the question here stands anyway: the review path takes an override today and the work path does not.

What this is not

Not a request to change the default. high can stay the editorial tier for every route. I am asking about the hook, not the number.

Environment

  • compound-engineering-plugin 3.22.4
  • codex-cli 0.150.1
  • macOS 26.5.2

Observed argv from bash skills/ce-work/scripts/cross-model-work.sh --emit-adapter codex:

codex exec --ignore-user-config --ignore-rules --ephemeral -s workspace-write -C <workspace> --json -o <raw-result> -c model_reasoning_effort=high -
Dominant language
TypeScript
Stars
25.6k
Forks
2.1k
Avg merge
3h 57m
Merged PRs (30d)
170

Getting set up

This project ships no dev container, Dockerfile or contributing guide, so setting up is up to you: start from its README, and see our first-contribution guide for the general steps.

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from EveryInc/compound-engineering-plugin

All issues in EveryInc/compound-engineering-plugin

Similar issues

More TypeScript issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.