ce-work cross-model adapters: codex effort unset, grok writes always cancelled (v3.21.0)

[t851029]

Two issues in skills/ce-work/scripts/cross-model-work.sh on v3.21.0, macOS. Both make a configured work_engine_preferences route behave differently than the equivalent review route.

1. Codex work route never sets reasoning effort

adapter_argv() builds the codex route as:

codex exec --ignore-user-config --ignore-rules --ephemeral \
  -s workspace-write -C "$WORKSPACE" --json -o "$RAW_RESULT"

--ignore-user-config means ~/.codex/config.toml is not read, and no -c model_reasoning_effort is passed, so implementation runs at the model's built-in default. For gpt-5.6-luna that is medium (per codex debug models: "default_reasoning_level": "medium").

The review adapters on the same CLI do pass it — skills/ce-code-review/scripts/cross-model-adversarial-review.sh line 209:

-o "$RAW_OUT" -m "$(route_model codex)" -c 'model_reasoning_effort="xhigh"' -c 'hide_agent_reasoning=false'

So reviews get xhigh while implementation silently gets medium, and there is no supported override: work_engine_preferences exposes only harness and model, and the docs state "Do not put CLI commands or flags in configuration."

Adding -c 'model_reasoning_effort="high"' to the codex work branch fixes it locally.

2. Grok work route produces no edits at all

The grok branch uses --permission-mode acceptEdits. In a headless --prompt-file run this does not auto-approve write-class tools. Every attempt returns:

User cancelled the execution for tool `write`
"stopReason":"cancelled"

The model reasons correctly and emits the right diff, then the write is cancelled and the unit completes having changed nothing.

Reproduced 5 times on grok 0.2.118 with the adapter's exact flags, varying one thing at a time:

Variation	Result
--permission-mode acceptEdits (adapter default)	cancelled (write)
--permission-mode dontAsk	cancelled (write)
no --tools allowlist	cancelled (search_replace)
workspace added to ~/.grok/trusted_folders.toml	cancelled (write)
workspace under the real runs root /tmp/compound-engineering-$UID/ce-work/...	cancelled (write)

Not a trust, tool-allowlist, or permission-mode issue. grok --help lists --always-approve ("Auto-approve all tool executions") as the only flag that auto-approves; the adapter never passes it. I could not verify that it fixes the run in my environment, so treat it as the likely fix rather than a confirmed one.

Worth noting the review script comments explicitly forbid emitting --always-approve — correct for a read-only review, but the work route needs some equivalent, or a preflight that fails loudly instead of returning a successful unit with an empty diff.

Impact

A user who configures grok as a work engine gets units that report success and change nothing. A user who configures codex gets a quieter surprise: implementation at medium while their reviews run at xhigh.

Environment

• compound-engineering 3.21.0
• codex-cli 0.146.0, grok 0.2.118
• macOS 15 (Darwin 25.3.0)

[t851029]

Confirmed the fix and the mechanism for issue 2.

--always-approve does resolve it. Same workspace, same argv, only the approval flag swapped:

grok --prompt-file prompt.txt --cwd "$WS" --effort medium --always-approve \
  --tools Read,Write,Edit --disable-web-search --no-memory --no-subagents \
  --no-plan --max-turns 50 --output-format streaming-json --verbatim --model grok-4.5

"stopReason":"end_turn"
PASS - file written:
print("Hello from grok-4.5")

Why the permission-mode aliases can't work here. Per the permissions docs, grok has exactly three modes:

• ask — prompts for anything not already approved
• auto — classifier auto-approves safe tools, still prompts for dangerous ones
• always-approve — approves without prompting

The CLI's Claude-compatible aliases (acceptEdits, dontAsk, bypassPermissions, plan) are import shims that collapse onto those three; acceptEdits and dontAsk both land on auto. That is why every mode I tried behaved identically — they were the same mode. Under auto a write-class tool hits the classifier, the classifier wants a prompt, and a headless --prompt-file run has no one to answer, so it resolves to cancelled.

That makes this a mode-selection bug rather than an environment quirk: the grok work route cannot write files under any --permission-mode value, on any machine.

Suggested change in adapter_argv():

       printf '%s\0' grok --prompt-file "$PROMPT_FILE" --cwd "$WORKSPACE" \
-        --effort high --permission-mode acceptEdits \
+        --effort high --always-approve \

I recognize the review scripts deliberately forbid emitting --always-approve, which is right for a read-only review. The work route is the opposite case: it runs in a controller-owned per-unit workspace under -s workspace-write semantics, where writing is the entire point. If that is still unwanted, the alternative is a preflight that fails the route loudly rather than returning a successful unit with an empty diff — the current silent-no-op is the worst of the three outcomes.

[moorecatherine06]

状态管理别过度抽象就好 ⭐

[dunncharles0112]

🤝🔥 Mock 和服务虚拟化怎么选

[harrisonmichael809]

回滚预案写在 runbook 里了吗 ✨📦

[harrisonmichael809]

🧠👍 token 成本有没有做过粗算

[blackwellmiranda8]

🤝⚙️🚀 这个方向很实用，收藏学习了

[hansenisaac8461]

GPU 显存不够时有量化方案吗 🧠😄

[harveysamuel99]

🧪 flaky test 怎么治理的

[barrettchelsea0]

跨平台兼容性测过 Windows 吗 🔥🙌