…ION_AGENT_SLUGS set [LOBE-9452]

… runs [LOBE-9449]

…LOBE-9451]

…traction [LOBE-9450]

…ath [LOBE-9455]

…c path [LOBE-9455]

…ation with execAgent path [LOBE-9454]

- Add `suppressSignal` and `sourceMessageId` to `ExecAgentAppContext`
  so background agent runs can flag themselves and skip
  `agent.user.message` re-emission in execAgent (#2).
- Fix character truncation in `ToolResultWithKind.data: unknown`
  (was bare ` nknown;`).
- Replace the invalid `chatConfig.enableAutoCreateTopic` field on the
  three self-iteration builtin agents (nightly-review / self-reflection /
  self-feedback-intent) with the required `plugins:
  ['agent-signal-self-iteration']` so execAgent actually injects the
  self-iteration tool manifest at runtime.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

The original PR commit (`feat(agent-signal): migrate skill management
action to execAgent async path [LOBE-9455]`) regressed
`runSkillDecisionAgentRuntime` to a `{ action: 'noop' }` stub and dropped
~1500 lines of supporting code (decision schemas, manifest, tool
executor, normalizers). This effectively disabled all skill-management
decisions.

Reverting the file to canary keeps the decision agent intact while we
redo the skill-write runner migration in a focused follow-up PR.
LOBE-9455 stays open for that work.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

execAgent returns immediately after queueing (`ExecAgentResult` does not
contain `finalState`). The previous implementation destructured
`finalState` as if execAgent were synchronous, making the
'did the memory write succeed?' branch unreachable and forcing every
memoryWriter run to fall through to skipped at runtime.

Changes:
- `runMemoryActionAgent` now enqueues via execAgent and returns
  `{operationId, status: 'applied'}` on successful queueing. Drop the
  inline finalState inspection helpers (hasSuccessful/FailedMemoryWrite)
  and the `resolveMemoryActionTargetFromState` projector — these belong
  in the completion policy once it grows a memory-writer handler.
- `runMemoryActionAgent` passes the writer-specific guidance via
  `instructions` (no `systemRoleOverride` exists on `ExecAgentParams`);
  a follow-up may introduce a `memory-writer` builtin agent so the
  systemRole replaces rather than appends.
- `userMemory.test.ts` drops the 300-line `resolveMemoryActionTargetFromState`
  describe block and the corresponding import (the function no longer
  exists). Remaining handler-level tests still run against the injected
  `memoryActionRunner` mock.
- `executeViaExecAgent` mirrors the same pattern: return `{operationId,
  enqueued}` instead of fake `finalState`. Drop the non-existent
  `systemRoleOverride` and the `createAgentSignalSelfIterationSystemRole(mode)`
  call that passed an argument to a no-arg function.
- Move `trigger: RequestTrigger.AgentSignal` from `appContext` to the
  top-level execAgent param (its actual location).
- `appContext.scope` is set to `'agent-signal'` (was `'chat'`) so
  cost/observability lanes are distinguishable.

NOTE: the post-completion bookkeeping (writing receipts, marking
idempotency only on real write success) is not yet wired up.  The
completion policy registered in this PR is the seam for that work but
no `onMemoryWriterCompleted` callback exists yet — tracked as a
follow-up under LOBE-9451 / LOBE-9455.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

The completion policy and the suppressSignal helper landed in the PR as
orphan code: `createCompletionPolicy` was never installed into the
default policy stack, and `shouldSuppressSignal` was never called from
execAgent's emit site. Both are wired up now.

- `completionPolicy.ts`: switch to the correct `defineSourceHandler`
  shape (the previous `{ handle, sourceType }` literal did not match
  `AgentSignalInstallableHandlerDefinition` and failed type-check).
  Drop `context.userId` — `RuntimeProcessorContext` has no such field;
  callers can look up userId via the operations table by operationId.
  Forward `topicId` from the source payload when present.
- `policies/index.ts`: register `createCompletionPolicy(options.completion)`
  in `DEFAULT_AGENT_SIGNAL_POLICY_FACTORIES`, and add a `completion`
  field to `CreateDefaultAgentSignalPoliciesOptions` for caller-provided
  callbacks.
- `aiAgent/index.ts`: guard the `agent.user.message` emit (lines
  1855-1874) with `if (!appContext?.suppressSignal)`. Builtin background
  agent runs (memory-writer, self-iteration reviewer, etc.) now skip the
  emit and avoid re-entering the analyzeIntent pipeline.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

The catch handler for the `agent.user.message` enqueue used the
`debug('lobe-server:...')` logger, which only surfaces output when the
`DEBUG` env var is set. An enqueue failure here means Agent Signal lost
a source event — that must be visible in production logs regardless of
debug flags.

Switch to `console.error` so the failure is always logged.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>