重构主控编排与运行时预热链路，统一工作区提示词/专家调度协议并补齐 wiki 记忆注入与写回闭环。

同时收敛启动与运维脚本默认行为（含 wiki worker）、更新 Admin 可观测性与相关测试，降低首轮时延并提高运行稳定性。

Made-with: Cursor

openclaw/skills/taskflow/SKILL.md

@@ -0,0 +1,149 @@
+---
+name: taskflow
+description: Use when work should span one or more detached tasks but still behave like one job with a single owner context. TaskFlow is the durable flow substrate under authoring layers like Lobster, ACPX, plugins, or plain code. Keep conditional logic in the caller; use TaskFlow for flow identity, child-task linkage, waiting state, revision-checked mutations, and user-facing emergence.
+metadata: { "openclaw": { "emoji": "🪝" } }
+---
+
+# TaskFlow
+
+Use TaskFlow when a job needs to outlive one prompt or one detached run, but you still want one owner session, one return context, and one place to inspect or resume the work.
+
+## When to use it
+
+- Multi-step background work with one owner
+- Work that waits on detached ACP or subagent tasks
+- Jobs that may need to emit one clear update back to the owner
+- Jobs that need small persisted state between steps
+- Plugin or tool work that must survive restarts and revision conflicts cleanly
+
+## What TaskFlow owns
+
+- flow identity
+- owner session and requester origin
+- `currentStep`, `stateJson`, and `waitJson`
+- linked child tasks and their parent flow id
+- finish, fail, cancel, waiting, and blocked state
+- revision tracking for conflict-safe mutations
+
+It does **not** own branching or business logic. Put that in Lobster, acpx, or the calling code.
+
+## Current runtime shape
+
+Canonical plugin/runtime entrypoint:
+
+- `api.runtime.tasks.flow`
+- `api.runtime.taskFlow` still exists as an alias, but `api.runtime.tasks.flow` is the canonical shape
+
+Binding:
+
+- `api.runtime.tasks.flow.fromToolContext(ctx)` when you already have trusted tool context with `sessionKey`
+- `api.runtime.tasks.flow.bindSession({ sessionKey, requesterOrigin })` when your binding layer already resolved the session and delivery context
+
+Managed-flow lifecycle:
+
+1. `createManaged(...)`
+2. `runTask(...)`
+3. `setWaiting(...)` when waiting on a person or an external system
+4. `resume(...)` when work can continue
+5. `finish(...)` or `fail(...)`
+6. `requestCancel(...)` or `cancel(...)` when the whole job should stop
+
+## Design constraints
+
+- Use **managed** TaskFlows when your code owns the orchestration.
+- One-task **mirrored** flows are created by core runtime for detached ACP/subagent work; this skill is mainly about managed flows.
+- Treat `stateJson` as the persisted state bag. There is no separate `setFlowOutput` or `appendFlowOutput` API.
+- Every mutating method after creation is revision-checked. Carry forward the latest `flow.revision` after each successful mutation.
+- `runTask(...)` links the child task to the flow. Use it instead of manually creating detached tasks when you want parent orchestration.
+
+## Example shape
+
+```ts
+const taskFlow = api.runtime.tasks.flow.fromToolContext(ctx);
+
+const created = taskFlow.createManaged({
+  controllerId: "my-plugin/inbox-triage",
+  goal: "triage inbox",
+  currentStep: "classify",
+  stateJson: {
+    businessThreads: [],
+    personalItems: [],
+    eodSummary: [],
+  },
+});
+
+const classify = taskFlow.runTask({
+  flowId: created.flowId,
+  runtime: "acp",
+  childSessionKey: "agent:main:subagent:classifier",
+  runId: "inbox-classify-1",
+  task: "Classify inbox messages",
+  status: "running",
+  startedAt: Date.now(),
+  lastEventAt: Date.now(),
+});
+
+if (!classify.created) {
+  throw new Error(classify.reason);
+}
+
+const waiting = taskFlow.setWaiting({
+  flowId: created.flowId,
+  expectedRevision: created.revision,
+  currentStep: "await_business_reply",
+  stateJson: {
+    businessThreads: ["slack:thread-1"],
+    personalItems: [],
+    eodSummary: [],
+  },
+  waitJson: {
+    kind: "reply",
+    channel: "slack",
+    threadKey: "slack:thread-1",
+  },
+});
+
+if (!waiting.applied) {
+  throw new Error(waiting.code);
+}
+
+const resumed = taskFlow.resume({
+  flowId: waiting.flow.flowId,
+  expectedRevision: waiting.flow.revision,
+  status: "running",
+  currentStep: "finalize",
+  stateJson: waiting.flow.stateJson,
+});
+
+if (!resumed.applied) {
+  throw new Error(resumed.code);
+}
+
+taskFlow.finish({
+  flowId: resumed.flow.flowId,
+  expectedRevision: resumed.flow.revision,
+  stateJson: resumed.flow.stateJson,
+});
+```
+
+## Keep conditionals above the runtime
+
+Use the flow runtime for state and task linkage. Keep decisions in the authoring layer:
+
+- `business` → post to Slack and wait
+- `personal` → notify the owner now
+- `later` → append to an end-of-day summary bucket
+
+## Operational pattern
+
+- Store only the minimum state needed to resume.
+- Put human-readable wait reasons in `blockedSummary` or structured wait metadata in `waitJson`.
+- Use `getTaskSummary(flowId)` when the orchestrator needs a compact health view of child work.
+- Use `requestCancel(...)` when a caller wants the flow to stop scheduling immediately.
+- Use `cancel(...)` when you also want active linked child tasks cancelled.
+
+## Examples
+
+- See `skills/taskflow/examples/inbox-triage.lobster`
+- See `skills/taskflow/examples/pr-intake.lobster`
+- See `skills/taskflow-inbox-triage/SKILL.md` for a concrete routing pattern

openclaw/skills/taskflow/examples/inbox-triage.lobster

@@ -0,0 +1,33 @@
+# Illustrative Lobster authoring example for a TaskFlow-style inbox triage job.
+# Swap the placeholder commands for your own tools or scripts.
+
+name: inbox-triage
+steps:
+  - id: fetch
+    command: gog.gmail.search --query 'newer_than:1d' --max 20
+
+  - id: classify
+    command: >-
+      openclaw.invoke --tool llm-task --action json --args-json
+      '{"prompt":"Classify each inbox item as business, personal, or later. Return one JSON object per item with route and summary.","thinking":"low","schema":{"type":"object","properties":{"items":{"type":"array"}},"required":["items"],"additionalProperties":false}}'
+    stdin: $fetch.stdout
+
+  - id: post_business
+    command: slack-route --bucket business
+    stdin: $classify.stdout
+    condition: $classify.json.items[0].route == "business"
+
+  - id: wait_for_business_reply
+    command: echo '{"status":"waiting","reason":"slack_reply"}'
+    condition: $classify.json.items[0].route == "business"
+
+  - id: notify_personal
+    command: >-
+      openclaw.invoke --tool message --action send --args-json
+      '{"provider":"telegram","to":"owner-thread","content":"Personal inbox item needs attention."}'
+    condition: $classify.json.items[0].route == "personal"
+
+  - id: stash_for_eod
+    command: summary-append --bucket eod
+    stdin: $classify.stdout
+    condition: $classify.json.items[0].route == "later"

openclaw/skills/taskflow/examples/pr-intake.lobster

@@ -0,0 +1,32 @@
+# Illustrative Lobster authoring example for a TaskFlow-style PR intake lane.
+# Replace the placeholder commands with repo-specific tooling.
+
+name: pr-intake
+steps:
+  - id: fetch
+    command: gh pr list --repo owner/repo --state open --json number,title,body,headRefName
+
+  - id: classify
+    command: >-
+      openclaw.invoke --tool llm-task --action json --args-json
+      '{"prompt":"Classify each PR as close, request_changes, refactor, or maintainer_review. Return intent and recommended next action.","thinking":"low","schema":{"type":"object","properties":{"items":{"type":"array"}},"required":["items"],"additionalProperties":false}}'
+    stdin: $fetch.stdout
+
+  - id: close_low_signal
+    command: pr-close-low-signal
+    stdin: $classify.stdout
+    condition: $classify.json.items[0].nextAction == "close"
+
+  - id: request_changes
+    command: pr-request-changes
+    stdin: $classify.stdout
+    condition: $classify.json.items[0].nextAction == "request_changes"
+
+  - id: refactor_branch
+    command: pr-refactor-branch
+    stdin: $classify.stdout
+    condition: $classify.json.items[0].nextAction == "refactor"
+
+  - id: escalate
+    command: echo '{"status":"notify","target":"maintainer"}'
+    condition: $classify.json.items[0].nextAction == "maintainer_review"