Agent Delegate: catch headless agent runs that did nothing
Run codex exec or any CLI agent on a spec in its own git worktree; get exit 65 when the run changed nothing.
Free
Works with the AI tools you already use
Claude Code
Cursor
Codex CLI
Muse
Dots+22 moreClaude CodeCursorCodex CLIMuseDots
OpenClaw
Manus
Grok
Grok Bot+18 agentsAgent Delegate: catch headless agent runs that did nothing
Example session with this skill installed
Delegate docs/specs/01-json-flag.md to the headless agent in a fresh worktree and tell me whether it actually changed anything.
- Read your context and instructions
- Compiled the agent delegate
Summary
The agent exited 0 but changed nothing, so delegate.sh returned
65. This is the real output of delegate.sh against a stand-in agent that only prints a plan and asks "Shall I proceed?" (macOS, bash):
spec: ~/myproj / docs/specs/01-json-flag.md
agent: ~/fake-agent
log: ~/myproj/.delegate-runs/20261004-163927-01-json-flag.log
--- effective settings ---
model: demo-agent reasoning effort: high
--------------------------
exit=0 / log 2 lines
!! nothing changed - the agent did no work.
Check the end of the log; it most likely asked for approval or
reported a blocker: tail -30 ~/myproj/.delegate-runs/20261004-163927-01-json-flag.log
echo $? after set -o pipefail prints 65.
| Check | Result |
|---|---|
| Agent exit code | 0 |
| git status / HEAD moved | no |
| delegate.sh exit code | 65 (ran, nothing changed) |
| Effective settings echoed | yes, from the agent's own header |
Next steps
- Read the last lines of the log. Here the agent stopped at "Shall I proceed?".
- Add "do not ask for approval, execute the spec" to the spec or extra instructions and run again.
- After a run that does change files,
./review-status.sh <project>lists the commits still waiting for review.
Connects securely to your tools. The creator never sees your data.
About this skill
Headless coding agents can end a run with exit 0 and pass every test while changing nothing. In one workspace this happened when the agent wrote a plan, asked "shall I proceed?", and exited, because nobody was there to answer. The log looked complete. This skill is three bash scripts (68 to 207 lines, bash + git) and a spec template.
- delegate.sh runs the agent on a spec with stdin closed (from /dev/null), echoes the agent's own config header so you see the effective reasoning effort, compares git status and HEAD before and after, and returns 65 if neither moved. worktree.sh makes one git worktree per task, sharing node_modules. review-status.sh lists agent commits that do not have a review tag yet. SPEC-TEMPLATE.md has a file ownership table, an example table, and runnable acceptance commands.
Exit codes: 0 changes made, 64 bad usage, 65 ran but nothing changed, 66 project or spec not found, anything else is the agent's own exit code. test.sh reproduces 10 cases with a fake agent (passed 10, failed 0). What it does not do: it does not review the agent's code and it cannot fix a bad spec. It makes a run that did nothing, hung, or lost its exit code visible instead of green.
How to install
Works the same in every agent - Claude, Cursor, Codex, Copilot and 20+ more.
- 1
Download the ZIP
Free skills download straight away. Paid skills unlock right after purchase.
- 2
Unzip into your skills folder
Every agent reads skills from one folder on your machine. Drop the unzipped folder in there.
- 3
Ask your agent to use it
Restart the agent if it was already running. It picks the skill up automatically - no config needed.
Skills folder by agent
Click the path to copy it. Create the folder if it does not exist yet.
Reviews
2 people have installed this skill.
Trust & safety
Security scanned
Verified clean today
- Free to download with an account
Needs access to
Gumroad
Github
Blogspot