coding-agent

Files (1)

• SKILL.md

SKILL.md (first lines)

# Coding Agent

Use for background feature builds, PR reviews, large refactors, and issue-to-PR loops. Do not use for simple edits, read-only lookup, ACP thread-bound work, or any run inside `~/.openclaw`, `$OPENCLAW_STATE_DIR`, or active OpenClaw state dirs.

## Hard rules

- Always launch with `background:true`.
- Codex and OpenCode: use `pty:true`.
- Codex: never inherit ambient `CODEX_HOME` or the default `~/.codex`. Use a
  separately authenticated coding-agent home and scope it to each Codex command.
- Claude Code: no PTY; use `claude --permission-mode bypassPermissions --print`.
- Capture a real notification route before spawning.
- Worker must send completion/failure via `openclaw message send`.
- Do not rely on heartbeat, system events, or notify-on-exit.
- Monitor with `process`; do not kill slow workers without cause.
- If user asked for a specific agent, use that agent.
- If worker fails/hangs, respawn or ask; do not silently hand-code instead.
- Never checkout branches or run background coding agents in `~/Projects/openclaw`; use an isolated checkout.
- Classify the source ref as trusted or untrusted before any checkout or worktree creation. Never materialize a contributor-controlled ref outside the repository's approved untrusted-PR sandbox/review workflow, and never launch a permission-bypassed worker in it.
- For tasks that modify a Git-backed project, prepare and verify the Git worktree before launch, then include the exact Git preparation block below in the worker prompt.

## Mandatory Git preparation

Before launching Codex, Claude Code, or OpenCode for work that modifies a Git-backed project:

1. Establish the intended target repository, then select its canonical remote. Prefer `upstream` when it exists and matches that target; otherwise verify `origin`. Resolve the selected remote's default branch dynamically. Determine the target base from an explicit task branch or authoritative existing-PR metadata; for other shared branches, prove the configured/tracked base or ask. Use the canonical default only for new work with no other specified base. Stop if the repository, remote, or target base cannot be proven.
2. Classify the source ref as trusted or untrusted before any checkout or worktree creation. For contributor-controlled refs, use the repository's approved untrusted-PR sandbox/review workflow, which must own ref materialization inside the sandbox, or stop. The remaining steps and launch forms are for trusted refs only.
3. For trusted new work, run `git fetch --prune <canonical>` immediately before creating a new isolated worktree and branch from `<canonical>/<targetBaseBranch>`.
4. For trusted new work, verify the worktree's initial `HEAD` equals the fetched target-base SHA. Record the canonical remote, canonical default branch, target base branch, base SHA, worktree path, and branch.
5. For a trusted existing PR or shared branch, fetch the canonical target base and source branch immediately before creating an isolated worktree from the fetched source branch. Record that source ref and starting SHA, report its divergence from the refreshed target base, and do not automatically rebase, merge, reset, force-push, or otherwise rewrite shared history.
6. Launch the worker in the isolated worktree, never the primary checkout. For OpenClaw, the primary checkout under `~/Projects/openclaw` remains forbidden.

For tasks that modify a Git-backed project, append this block to the worker prompt with real values:

```text
Git preparation (mandatory before edits):
- canonical remote: <canonicalRemote>
- canonical default branch: <canonicalDefaultBranch>
- target base branch: <targetBaseBranch>
- fetched target base SHA: <targetBaseSha>
- preparation mode: <new work | existing PR/shared branch>
- checkout trust: trusted
- prepared source ref: <canonicalRemote/targetBaseBranch | fetched trusted source ref>
- prepared start SHA: <preparedStartSha>
- isolated worktree: <worktreePath>
- working branch: <branch>
- preparation receipt: <new work: `git fetch --prune <canonicalRemote>` ran immediately before creation from `<canonicalRemote>/<targetBaseBranch>` | existing branch: the canonical target base and trusted source ref were fetched immediately before the worktree was created from `<preparedSourceRef>` at `<preparedStartSha>`>

Before editing, verify the current directory is the isolated worktree and its initial HEAD equals <preparedStartSha>. For new work, that SHA must equal <targetBaseSha>. Never edit the primary checkout. For existing PR/shared-branch work, report divergence and do not rebase, merge, reset, force-push, or otherwise rewrite shared history unless explicitly asked.
Immediately before the final push or PR for newly authored work, run `git fetch --prune <canonicalRemote>` and `git merge-base --is-ancestor <canonicalRemote>/<targetBaseBranch> HEAD`. If the ancestry check fails, update the new branch onto the latest target base, rerun the relevant proof, and only then push without force. For existing PR/shared-branch work, report a failed ancestry check and follow the repository workflow without rewriting the branch.
```

For trusted refs, the launcher must create and verify the worktree before starting the editing worker; do not delegate worktree creation to that worker. The approved untrusted-PR workflow must instead own checkout and worktree materialization inside its sandbox. Never start a worker in `~/Projects/openclaw`. Read-only tasks and non-project scratch work do not require the Git preparation block.

## Notification block

Append this shape to every worker prompt with real values:

```text
Notification route:
- channel: <notifyChannel>
- target: <notifyTarget>
- account: <notifyAccount or omit>
- reply_to: <notifyReplyTo or omit>
- thread_id: <notifyThreadId or omit>

When finished, send exactly one completion or failure message using:
openclaw message send --channel <channel> --target '<target>' --message '<brief result>'
Add --account, --reply-to, or --thread-id only when present above.
Do not use openclaw system event or heartbeat.
```

If no trustworthy route exists, say completion auto-notify is unavailable.

## Launch forms

Write the worker prompt to a temp file first. This avoids shell quoting bugs when the required notification block contains quotes or newlines.

```bash
PROMPT=$(mktemp -t openclaw-worker-prompt.XXXXXX)
cat >"$PROMPT" <<'EOF'
Task.
<mandatory Git preparation block>
<notification block>
EOF
printf 'prompt file: %s\n' "$PROMPT"
```

Use `$PROMPT` when launching from the same shell/session. If using a separate tool call, substitute the printed path. The launch forms below are for trusted checkouts only; untrusted contributor refs require the repository's approved sandbox/review workflow.

Before the first Codex worker on a host, prepare a dedicated auth home in the
foreground. Do not copy `auth.json` or other credentials from ambient
`~/.codex`; authorize this home separately. Codex scopes both file and keyring
credentials by `CODEX_HOME`.

```bash
CODEX_WORKER_HOME="$HOME/.codex-coding-agent"
mkdir -p "$CODEX_WORKER_HOME"
if ! env -u CODEX_API_KEY -u CODEX_ACCESS_TOKEN -u OPENAI_API_KEY \
  CODEX_HOME="$CODEX_WORKER_HOME" codex login status >/dev/null 2>&1; then
  printf 'Codex coding-agent login required for %s\n' "$CODEX_WORKER_HOME"
  env -u CODEX_API_KEY -u CODEX_ACCESS_TOKEN -u OPENAI_API_KEY \
    CODEX_HOME="$CODEX_WORKER_HOME" codex login --device-auth
fi
printf 'Codex worker home: %s\n' "$CODEX_WORKER_HOME"
```

The login is an interactive foreground setup step, not a background worker.
The launch command repeats the fixed, quoted home and removes ambient Codex and
OpenAI auth overrides. Never export the worker home into the OpenClaw Gateway
environment.

Codex:

```bash
bash pty:true background:true workdir:/path/isolated-worktree command:"env -u CODEX_API_KEY -u CODEX_ACCESS_TOKEN -u OPENAI_API_KEY CODEX_HOME=\"$HOME/.codex-coding-agent\" codex exec - < \"$PROMPT\""
```

Claude Code:

From openclaw/openclaw, MIT.

More from openclaw/openclaw