code-agent-orchestration（代码 Agent 编排）

Code Agent Orchestration

Use openclaw-code-agent to run Claude Code or Codex sessions as background coding jobs from chat.

Launch

• Do not pass channel manually. Routing comes from agentChannels, the current chat context, and fallbackChannel.
• Sessions are multi-turn. Continue existing work with agent_respond or agent_launch(..., resume_session_id=...); do not start a fresh session for the same task.
• Always set a short kebab-case name when you care about later follow-up.
• Set workdir to the target repo.
• Use permission_mode: "plan" when the user wants a real review gate before implementation.
• Use permission_mode: "bypassPermissions" only for autonomous execution.
• defaultWorktreeStrategy now defaults to off. Opt into a worktree strategy explicitly when you want branch isolation.
• In plan mode, the plan belongs in normal session output. Do not ask the coding agent to write plan docs or transcript artifacts unless the user explicitly asked for a file.

Example:

agent_launch(
  prompt: "Fix the auth middleware bug and add tests",
  name: "fix-auth",
  workdir: "/home/user/projects/my-app"
)

Resume, Don't Respawn

When a session already exists for the task, keep using it.

• Waiting for plan approval: agent_respond(session, message, approve=true) or agent_request_plan_approval(...) if delegated approval must escalate to the user
• Waiting for a question answer: agent_respond(session, message)
• Killed/stopped by restart: agent_respond(session, message)
• Completed but needs follow-up: agent_launch(resume_session_id=session_id, prompt="...")
• Fresh agent_launch is only for genuinely independent work

Do not launch a new coding session from a wake event for the same task.

State and Monitoring

Use:

agent_sessions()
agent_output(session: "fix-auth", lines: 100)
agent_output(session: "fix-auth", full: true)

For worktree follow-through, inspect:

agent_worktree_status()
agent_worktree_status(session: "fix-auth")

Treat that tool's lifecycle, derived state, cleanup disposition, and retained reasons as authoritative. Do not infer cleanup safety from a transcript summary or from branch names alone.

Treat these wake fields as authoritative state when present:

• requestedPermissionMode
• effectivePermissionMode / currentPermissionMode
• approvalExecutionState

Use those deterministic fields instead of inferring behavior from transcript fragments.

Approval/execution meanings:

• approved_then_implemented: normal approved execution
• implemented_without_required_approval: actual approval bypass
• awaiting_approval: still stopped at the approval gate
• not_plan_gated: no plan gate applied

Completion ownership:

• The plugin sends the canonical completion notification.
• The plugin owns the canonical completion status line; the orchestrator owns any additional plain-text follow-up.
• After a coding-agent session completes, the orchestrator should usually add at least a short human-useful summary of what changed, what was done, or the concrete outcome.
• That expectation applies to ordinary terminal/manual completions, manual no-change completions, and delegated worktree completions alike.
• Treat the plugin's canonical ✅ as the status signal and your follow-up as the factual outcome summary that should usually come right after it.
• That summary can be brief; one sentence is often enough.
• Extra synthesis, risk framing, and next-step guidance are optional. Add them when useful; do not force them every time.
• Do not generate your own heuristic completion summary from transcript tail lines. Base any summary on reliable result data such as agent_output(..., full=true), diff context, or deterministic tool state.
• Skip the summary only in narrow cases:

- no user-facing follow-up will be sent at all because the orchestrator is silently continuing an internal multi-phase pipeline - the completion produced no meaningful outcome to report, or the reliable result data is still too incomplete to support even a short factual summary

Respond Rules

Auto-respond immediately only for:

• permission requests for file reads, writes, or shell commands
• explicit continuation prompts such as "Should I continue?"

Forward everything else to the user:

• architecture or design choices
• destructive operations
• scope changes
• credentials or production questions
• ambiguous requirements

When forwarding, quote the session's exact question. Do not add commentary.

Plan Approval

Use permission_mode: "plan" whenever the user wants a real planning checkpoint.

planApproval: "ask"

• Approval belongs to the user.
• The plugin sends the canonical Approve / Revise / Reject prompt directly to the user.
• If the user requests changes, wait for the revised plan from that same session; the revised submission becomes the latest actionable review version automatically.
• Wait for the user's answer, then forward it with agent_respond(...).
• Do not send a duplicate approval recap or second approval prompt.

planApproval: "delegate"

• Approval belongs to the orchestrator first.
• This is wake-first: the plugin wakes the orchestrator without user buttons.
• Before deciding, read the full plan with agent_output(session, full=true); do not rely on the truncated preview.
• Approve directly with agent_respond(..., approve=true) only when the latest actionable plan version is clearly in-bounds and low risk.
• When approving directly, pass a structured rationale with approval_rationale, for example: agent_respond(session='...', message='Approved. Go ahead.', approve=true, approval_rationale='Scope matches the request and the changes are low risk.')
• After approving directly, send the user a short plain-text follow-up explaining what was approved and why. The plugin's 👍 Plan approved line is only a fallback signal, not the full explanation.
• If a prior version had changes_requested, that stale state should not block approval of the latest revised plan version.
• If escalation is needed, call agent_request_plan_approval(session='...', summary='...') exactly once so the plugin sends the single canonical user approval prompt.
• That escalation summary must concisely explain why you are escalating, plus risk/scope notes the user needs to decide.
• After that canonical prompt exists, wait for the user's decision; do not send a second plain-text approval summary.

planApproval: "approve"

• Auto-approve only after verification per the session policy.

Worktree Decisions

Treat worktrees as temporary task sandboxes, not as generic branch inventory.

Lifecycle meanings:

• pending_decision: still waiting for merge / PR / dismiss follow-through
• pr_open: PR exists; preserve the sandbox
• merged: normal ancestry merge landed
• released: content already landed on the base branch even though SHAs differ after rebase, squash, or cherry-pick
• dismissed: sandbox intentionally discarded
• no_change: no committed delta

If agent_worktree_status reports released, treat that sandbox as already landed. Do not narrate it as “still unmerged” just because the branch appears ahead.

off

• No worktree. The session runs in the main checkout.

ask

• The plugin owns the user-facing completion/decision message and button UI.
• Do not call agent_merge or agent_pr unless the user explicitly asks after that.
• A completed ask-session worktree may later resolve as released if its content already landed on base through another path. Confirm that with agent_worktree_status(...) before deciding what follow-up is still needed.

delegate

• The plugin wakes the orchestrator with diff context and no automatic user buttons.
• Read the diff context and decide whether a local merge is clearly safe.
• agent_merge is acceptable for low-risk, clearly scoped changes that match the task.
• Never call agent_pr() autonomously in delegate flows. Escalate PR decisions to the user.
• If the wake already says the plugin sent the canonical completion notification, do not repeat that status line, but you should still usually add a short summary of the completed outcome.

manual

• Wait for an explicit user request before calling agent_merge or agent_pr.

Cleanup

• Use agent_worktree_cleanup(mode: "preview_safe") to review what Clean all safe would remove.
• Use agent_worktree_cleanup(mode: "clean_safe") only when the user asked to clean up safe sandboxes.
• Use agent_worktree_cleanup(mode: "preview_all") when you need both safe candidates and retained reasons.
• Respect retained reasons from agent_worktree_status / agent_worktree_cleanup; they are the lifecycle model, not advisory prose.

Never

• Never use raw git merge or raw PR commands in place of plugin tools.
• Never invent your own workaround for a pending worktree decision; use agent_worktree_cleanup(session: "...", dismiss_session: true) to dismiss permanently.
• Never use agent_worktree_cleanup to force-delete unresolved worktrees. The supported bulk action is "clean all safe": omit session and let the plugin remove only lifecycle-safe worktrees while preserving anything active, pending, dirty, or PR-open.
• Never merge or PR an ask worktree behind the user's back.

File Artifact Policy

• Do not ask the coding agent to write planning documents, investigation notes, or analysis artifacts as files unless the user explicitly requested a file.
• Do not commit planning documents, investigation notes, or transcript-summary artifacts to the branch.
• Commit only actual code, configuration, tests, and explicitly requested documentation.

Anti-Patterns

• Do not pass multi_turn or multi_turn_disabled; all sessions are multi-turn.
• Do not pass channel manually unless you are debugging routing.
• Do not auto-answer design or scope questions.
• Do not infer approval/completion ownership from old transcript snippets when deterministic fields are present.
• Do not post duplicate completion or approval recaps when the plugin already sent the canonical message.