Harness differences

Harness differences

Describes robot-council/cli v0.4.12. What each harness was measured on is named beside it.

	Claude Code	Cursor	Codex
Channel notice (wake an idle agent)	yes, with the development flag	no equivalent	no; the request is openai/codex#15299, open
Stop hook delivers at a turn boundary	yes	yes	configured, not run
Loop guard in the hook's payload	stop_hook_active	loop_count, bounded by loop_limit in Cursor's own configuration	stop_hook_active, per its documentation
How delivered text arrives	as Stop hook feedback:	as the next user message (followup_message), with no prefix	unmeasured
Measured on	2.1.236, macOS	3.7.21 on Windows 11; 3.17.19 on macOS	nothing; key names checked against OpenAI's documentation only

Claude Code

The only harness the fleet can wake. Everything on Starting a seat applies to it alone.

Cursor

• Its hook payload starts with a UTF-8 byte-order mark, which PHP's json_decode refuses. The README's script matches on substrings, which survive it; a script rewritten to parse the payload has to strip it first.
• Cursor restarted a stopped bridge by itself in one observed run, 300 seconds later, or at once on a window reload, and did not re-read the tool list. The bridge now tells it the list changed after every start, so it asks again (#194, #212).
• On macOS, a hook's command runs through /bin/bash. On Windows the interpreter has not been identified (#190), so the README points command at a .cmd shim that runs Git Bash, which works whatever the answer is (#72).
• cursor-agent -p did not run stop hooks in three runs on 2026.09.18-9a7762b, so it cannot be used to test one (#83).
• Turning a hook off: write an empty hooks, rather than deleting the file. The README's Cursor section has the details.

Codex

Wired and documented, never run: no machine this project is worked from has Codex installed. Its payload shape and whether it sends a byte-order mark are unmeasured.