Clarify execution-agent-dispatch scope and harden gateway rules

2026-07-22 05:25:07 +00:00 · 2026-03-12 19:33:24 +10:00
parent 6ba8353506
commit 22e6292aba
5 changed files with 31 additions and 1 deletions
@@ -145,6 +145,7 @@ Ordering rule:
 - Don't exfiltrate private data. Ever.
 - Don't run destructive commands without asking.
 - `trash` > `rm` (recoverable beats gone forever)
 - On Tao's machine: gateway lifecycle is a high-risk operation. Never run `openclaw gateway stop`, and do not run restart-style gateway lifecycle commands on your own. Always check `openclaw gateway status` first, then ask before any disruptive gateway action.
 - When in doubt, ask.
 ## External vs Internal
@@ -27,6 +27,7 @@
 ## 4) ACP/Codex delegation constraints
 - Telegram channel plugin currently does **not** support `subagent_spawning` hooks → cannot bind persistent subagent sessions with `thread=true`; use `sessions_spawn(mode="run")` as workaround.
 - Current durable conclusion on `execution-agent-dispatch`: it is a **process/protocol skill**, not a fix for OpenClaw/ACP runtime communication. We previously tested parent↔child / agent↔agent flows and did **not** get stable bidirectional communication. Default behavior is still closer to spawn + announce than reliable free-form agent-to-agent conversation. Keep the skill frozen as a workflow aid only; wait for OpenClaw ACP/runtime support to become stable before resuming development aimed at true inter-agent communication.
 ## 5) OCM (OpenClaw Manager) repo policies
 - OCM repo path: `~/.openclaw/ocm`.
@@ -34,7 +35,10 @@
 ## 6) Ops / security notes
 - Ensure `~/.openclaw/openclaw.json` is not world-readable; prefer permission mode **600**.
- **Gateway control rule (important):** on Tao’s machine, do **not** run `openclaw gateway stop` as part of recovery/reload work. Stopping the gateway can cut off the agent’s own control path, and the service may then require Tao to manually run `openclaw gateway install` to restore it. Preferred action is `openclaw gateway restart` only when truly needed, and first check `openclaw gateway status` before touching the service.
+- **Gateway control rule (critical hard ban):** on Tao’s machine, do **not** run `openclaw gateway stop` under any circumstance during normal assistance/recovery/reload work.
 - **Related hard ban:** do **not** run any gateway command that may internally perform stop→start semantics unless Tao explicitly asks for that exact action and accepts the risk. Treat `openclaw gateway restart` as unsafe-by-default as well, because in practice it may still tear down the active service/control path.
 - Stopping or restart-style control can cut off the agent’s own control path, and the service may then require Tao to manually run `openclaw gateway install` to restore it.
 - Required sequence before any gateway intervention: (1) run `openclaw gateway status`; (2) report findings to Tao; (3) prefer non-disruptive diagnosis first; (4) only touch gateway lifecycle if Tao explicitly approves the exact command.
 ## 7) Model policy (current preference)
 - Historical temporary preference once was: all subagents **primary** = `openai-codex/gpt-5.2`, **fallback** = `github-copilot/claude-opus-4.6`.
@@ -0,0 +1,6 @@
 # 2026-03-12 execution-agent-dispatch status
 - Tao confirmed the remembered conclusion: `execution-agent-dispatch` was tested before and did not solve the real inter-agent communication problem.
 - Durable conclusion: the skill only helps with dispatch structure, worker reply format, escalation rules, and main-to-Tao forwarding discipline.
 - It does **not** provide stable bidirectional communication between OpenClaw agents/subagents/ACP sessions.
 - Current policy: freeze further development of this skill as a communication-layer solution; wait until OpenClaw ACP/runtime support is stable before resuming work in that direction.
@@ -0,0 +1,5 @@
 # 2026-03-12
 - Tao explicitly reinforced a hard rule after gateway disruption: on Tao's machine, do not run `openclaw gateway stop`, and do not assume `openclaw gateway restart` is safe. Treat restart-style lifecycle control as disruptive and approval-required.
 - Durable lesson: before any gateway intervention, first run `openclaw gateway status`, report findings, prefer non-disruptive diagnosis, and get Tao approval for the exact lifecycle command if one is still necessary.
 - Trigger for this rule: attempted gateway restart led to service becoming unavailable and Tao needing to manually reinstall/restore gateway service again. This is a high-friction failure and must not be repeated.
@@ -7,6 +7,20 @@ description: Standardize how main dispatches work to execution agents and how ex
 Use this skill after execution architecture is decided.
 ## Status / scope boundary
 This skill is a **workflow / protocol skill**.
 It standardizes dispatch packets, worker replies, escalation points, and main-to-Tao forwarding discipline.
 It is **not** a runtime transport fix.
 It does **not** create stable bidirectional communication between OpenClaw agents, subagents, or ACP sessions.
 It does **not** guarantee parent-first completion routing, persistent child↔parent conversation, or reliable agent-to-agent backchannels.
 Current project status:
 - treat this skill as useful only for **process discipline**
 - do **not** treat it as a solution to ACP / subagent communication instability
 - if stable agent-to-agent communication is needed, wait for OpenClaw / ACP runtime support to mature first
 ## Goal
 Turn vague delegation into a clear handoff.