Local coding agents (Codex / Claude Code)

codex and claude are agent CLIs on the host. Treat a request to them like passing a message to a colleague: send it, wait, relay the answer. Handing off one task is a single step — don't announce a plan, just run it.

Host execution only works when the agent owner is talking to you. If exec says host execution is restricted, or the CLI is missing / not logged in, say so plainly instead of retrying variations.

Use the CLI the user named; if they didn't name one, use codex. Always call it as command codex / command claude (bypasses shell aliases that add unsafe flags).

Default: one call, then reply

For anything that isn't a change to an existing repository — images, questions, explanations, small standalone scripts — run the CLI once in your Working Directory (shown in Runtime info) with a foreground exec and timeout: 600. Pass the user's request through faithfully; add only context the CLI can't know.

Codex

cd "<Working Directory>" && command codex exec -s workspace-write --skip-git-repo-check \
  --json -o .codex-last.md "<request>" > .codex-events.jsonl 2> .codex-stderr.log
echo "exit=$?"; cat .codex-last.md
grep -m1 '"thread.started"' .codex-events.jsonl

• Codex can generate images itself. They are saved under ${CODEX_HOME:-$HOME/.codex}/generated_images/<thread_id>/, where thread_id comes from the thread.started event above. Copy them into your Working Directory with a meaningful name and show them in your reply:

  cp "${CODEX_HOME:-$HOME/.codex}/generated_images/<thread_id>/"*.png ./cat.png

  then reply with ![cat](cat.png). Any other file Codex wrote in the Working Directory (an SVG, a script) is delivered the same way, by relative path.

• If the exit code is non-zero, show the user the last lines of .codex-stderr.log.

Claude Code

cd "<Working Directory>" && command claude -p "<request>" \
  --permission-mode acceptEdits --output-format json > .claude-result.json 2> .claude-stderr.log
echo "exit=$?"; cat .claude-result.json

• The reply is in result. If is_error is true, result holds the reason (e.g. "Credit balance is too low", not logged in) — tell the user as-is.
• Claude Code cannot generate raster images; for a picture it can write an SVG/HTML file, which you deliver by relative path.

Then answer the user in your own words, say which CLI did the work, and include any images/files. Keep the reply short.

Follow-ups in the same thread ("make the cat orange"): resume instead of starting over — command codex exec resume <thread_id> "<follow-up>" (same flags), or command claude -p "<follow-up>" --resume <session_id> using session_id from .claude-result.json.

Changing code in a repository

Only when the user wants an existing repo modified. Ask which repo if unclear.

1. Isolate the work so the user's checkout is untouched:

   REPO=/path/to/repo; TASK=fc-$(date +%Y%m%d-%H%M%S)
   WT="$REPO/../.fc-worktrees/$TASK"
   git -C "$REPO" worktree add -b "$TASK" "$WT" HEAD

2. Run the CLI inside $WT with a self-contained brief (goal, constraints, "run the tests; do not commit"). Repo changes can take many minutes, so use run_in_background: true and check with bash_output until it exits.

   • Codex: same command as above with cd "$WT". Network is off in workspace-write; add -c sandbox_workspace_write.network_access=true only if it must install dependencies.
   • Claude Code: add --allowedTools "Bash" so it can run builds and tests.

3. Verify yourself: git -C "$WT" status --short && git -C "$WT" diff --stat, and run the tests if it claims they pass. If it's wrong, resume the same session with concrete feedback (at most two rounds).

4. Deliver: commit, push the branch and open a PR (gh pr create --fill) if the user wants one; otherwise summarise the diff and name the branch. Send the PR link with a 2–4 line summary.

Never use --dangerously-bypass-approvals-and-sandbox (Codex) or bypassPermissions (Claude Code). For read-only questions about a repo, run in the repo with -s read-only (Codex) or --permission-mode plan (Claude Code) — no worktree needed.