partial)

- Add deliveryContext field to SystemEvent for routing preservation
- Wake interrupted session via heartbeat after restart (requestHeartbeatNow)
- Add retry logic for outbound delivery (2 attempts with 750ms delay)
- Preserve threadId routing through wake path
- Always enqueue wake even when delivery fails

Partial port of upstream 1c9f62f:
- Core: system-events deliveryContext, restart sentinel wake
- Deferred: heartbeat-runner turnSource integration, targets.ts routing updates
  (complex changes, requires more analysis)

Upstream commit:
- 1c9f62f: fix(gateway): restart sentinel wakes session after restart (openclaw#53940)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

- Add resolveFreshestSessionStoreMatchFromStoreKeys to prefer newest updatedAt
- Add resolveFreshestSessionEntryFromStoreKeys wrapper
- Use in sessions.preview for duplicate row handling

Handles case-insensitive and legacy alias keys by sorting duplicates
by updatedAt timestamp and returning the freshest entry.

Upstream commits:
- f48571b: fix: prefer freshest duplicate rows in session loads
- 40f820f: fix: prefer freshest duplicate session rows in reads

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Remove enqueueDelivery/ackDelivery/failDelivery usage from
server-restart-sentinel.ts as delivery-queue.ts is not yet ported
to our codebase. Keep retry logic but use agentCommand directly.

Resolves build failure.

Wrap individual channel startup in try-catch so one broken channel
(e.g. WhatsApp missing runtime) doesn't block other channels
(e.g. Discord) from starting.

Changes:
- Add try-catch in startChannels loop
- Log per-channel startup errors
- Continue starting remaining channels after failure

Upstream commit:
- 30e80fb: fix: isolate channel startup failures (openclaw#54215)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

- Implement runBeforeDispatch directly using getHooksForName
- Remove GroupName check (not in our MsgContext)
- Fix TypeScript build errors

Resolves build failures after before_dispatch hook port.

Scope rate-limit cooldowns per model so one 429 no longer blocks every
model on the same auth profile. Replace exponential 1min→1h escalation
with stepped 30s/1min/5min ladder.

Key changes:
- Add cooldownReason and cooldownModel to ProfileUsageStats type
- Implement stepped backoff: 30s → 1min → 5min (was: 1min → 5min → 25min → 60min)
- Add model-aware cooldown bypass in isProfileInCooldown()
- Track model scope when marking auth profile failures
- Pass modelId through markAuthProfileFailure and related calls
- Update isProfileInCooldown calls in model-fallback and pi-embedded-runner to pass forModel

Upstream ref: 8440122

Surface rate limit and overload errors that occur mid-turn (after tool
calls) instead of silently returning an empty response.

Only applies when the assistant produced no valid (non-error) reply text,
so tool-level rate-limit messages don't override a successful turn.

Changes:
- Add isReasoning field to EmbeddedPiRunResult payload type
- Detect mid-turn rate limits in agent-runner-execution.ts when there's
  no valid content (checking for text/media, excluding errors/reasoning)
- Import isRateLimitErrorMessage and isOverloadedErrorMessage
- Replace empty responses with user-facing rate limit message

Note: Skipped upstream's incomplete turn detection in run.ts (detecting
stopReason=toolUse with no payloads) as it requires deep understanding
of our specific agent loop structure and could cause false positives.
The agent-runner-execution.ts check catches the issue at final output.

Upstream ref: 4ae4d1f (partial port)

…artial)

Changes:
- Add hasInternalHookListeners() to check for listeners before cloning
- Add session:patch hook with structuredClone to isolate payload
- Add SessionPatchHookEvent and isSessionPatchEvent() type guard
- Only clone and dispatch when listeners are registered (performance)

Why structuredClone:
Fire-and-forget hooks cannot mutate objects used by the response path.
Deep cloning sessionEntry, patch, and cfg prevents plugin corruption.

Skip model default reasoning guards (6c04ce3, b91374e):
Those patches require agentEntry.reasoningDefault and
modelState.resolveDefaultReasoningLevel() which haven't been ported yet.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>