memory-continuity/MEMORY.md

Long-term Memory (curated)

1) Working model / roles

• Tao prefers to send tasks to main; main acts as PM/architect and delegates implementation (often to Codex).
• Dedicated agents and intended responsibilities:
  • tech_geek (Telegram group “技术宅”, workspace tech_home_group): tech/home-lab/ops discussions + GitHub (gh CLI / PR / CI / issues) follow-up.
  • travel_assistant: travel planning and travel-related research/tasks.
  • finance_assistant (“理财小帮手”): personal finance tasks.

2) Dev workflow / merge policy

• Small changes (docs/copy/layout/typos/i18n cleanup): Xiao Qiang can merge after self-review/verification, then notify Tao with PR link + summary.
• Big feature changes: require Tao review/approval before merge.

3) Sub-agents: architecture + token efficiency

• Sub-agents reuse the parent/main agent bot for messaging, but each sub-agent has an independent workspace + SOUL (persona) + MEMORY (memory).
• Goal: prevent context/memory mixing, improve focus/efficiency, and save tokens by keeping contexts smaller.
• Current preferred multi-agent architecture for complex development:
  • main acts as PM/architect/reviewer: requirements, task breakdown, prioritization, risk calls, progress updates, and final summaries to Tao.
  • Execution agents should be role-specialized and do the implementation work.
  • codex_worker is the dedicated Codex execution agent for coding tasks.
  • Default delegation rule: unless a task is truly tiny and can be finished in one short pass, main should not default to personally coding/editing files; implementation work should be delegated to codex_worker first.
  • Good candidates for main to do directly: tiny edits, very small linear fixes, or short actions that are not worth execution-agent handoff.
  • For complex projects, create more specialized agents as needed (e.g. frontend/backend/test/ops/docs/data) instead of overloading one agent.
  • Create multiple execution agents when work is meaningfully parallel, responsibilities are different, contexts are likely to contaminate each other, or independent validation/release tracks are needed.
  • Avoid over-splitting for tiny, highly coupled, or poorly defined tasks.
  • This is the current default pattern, but Tao may change the rules as needs evolve.

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.
• New clarified split after ACP re-check on 2026-03-13: main → ACP worker invocation is now confirmed basically usable on this machine (ACPX installed/enabled; Codex ACP smoke test could start, inherit workspace, read expected files, and return a worker-style result). But this does not prove stable agent↔agent or subagent↔subagent communication over ACP. Treat the safe default as: main acts as PM/architect/reviewer, ACP workers execute tasks and report back to main; do not assume a reliable free-form multi-agent ACP communication mesh.
• Clarification on Claude ACP testing: a same-day Claude ACP smoke test failed, but Tao confirmed Claude usage had already hit timeout/overuse state that day. Therefore do not record that sample as a product/runtime failure; mark Claude ACP as not yet cleanly validated, not “broken.”

5) OCM (OpenClaw Manager) repo policies

• OCM repo path: ~/.openclaw/ocm.
• Internal-only docs (Project Brief / Architecture / Decisions) must not be uploaded to GitHub; keep them under ~/.openclaw/ocm-internal/docs/.

6) Ops / security notes

• Ensure ~/.openclaw/openclaw.json is not world-readable; prefer permission mode 600.
• 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.
• Current important exception / newer rule: execution agents (esp. codex_worker, and by default other workers unless Tao says otherwise) must target primary = openai-codex/gpt-5.4.
  • No silent fallback: if the system falls back to any other model (or if openai-codex/gpt-5.4 becomes unknown/unavailable), main must immediately notify Tao.
  • No auto-fallback for execution agents: if 5.4 is unavailable, workers should not continue on another model; main should report and wait for Tao’s model decision.
  • If an execution agent is not responding, main must consider model unavailability/fallback as a first-class suspected cause and tell Tao.
• Additional long-lived execution agents initialized locally for repeat use: docs_worker, qa_worker, ops_worker.
• Execution-agent system needs a standardized dispatch/handoff layer: task input template, result format, blocker/escalation rules, and a main-to-Tao forwarding rule.
• Hard reporting/forwarding protocol (must-follow):
  • Worker events that require immediate Tao-visible updates (no “I’ll summarize later”): accepted, milestone result, blocked/failed, completion, agent switch decision, transition to review/commit/release.
  • Ordering constraint: when a worker reports milestone/completion, main’s first action is to update Tao; only then proceed to review/commit/next dispatch.
  • Failure definition: if a worker has already reported completion and main has not forwarded it to Tao, that is a main process failure, not “task still in progress”.
  • Default 4-line update template: who / status / output / next.

8) Watchdog / automation policy

• Tao preference: avoid watchdog-style auto-restart automation.
• The previous watchdog automation was removed:
  • OpenClaw cron job Gateway monitor + auto-restart (main) removed (jobId 139258d6-f675-4f88-b2a7-cbe0b93a0db6).
  • Mac launchd watchdog ai.openclaw.watchdog removed.
  • Pi-side crontab watchdog removed.

9) Webex keep-alive (local script)

• Tao requested a controllable Webex desktop “keep active” option.
• Script installed on Mac: ~/.openclaw/scripts/webex-keepalive.sh (commands: start|stop|status).
• Notes: uses AppleScript/System Events; may require macOS Accessibility permission for Terminal/iTerm. Logs: ~/.openclaw/logs/webex-keepalive.log, pid: ~/.openclaw/run/webex-keepalive.pid.

10) Local memory library organization

• Local memory dir: /Users/taodeng/.openclaw/workspace/main/memory/.
• Added memory/INDEX.md to classify daily notes vs topic notes vs automation state JSON (do not rename/move state JSON without updating jobs).
• Plan: create a separate GitHub KB repo (option A) for shareable/curated knowledge; keep private/sensitive items local and only publish redacted content.