fix(subagent): harden read-tool overflow guards and sticky reply threading (#19508)

* fix(gateway): avoid premature agent.wait completion on transient errors

* fix(agent): preemptively guard tool results against context overflow

* fix: harden tool-result context guard and add message_id metadata

* fix: use importOriginal in session-key mock to include DEFAULT_ACCOUNT_ID

The run.skill-filter test was mocking ../../routing/session-key.js with only
buildAgentMainSessionKey and normalizeAgentId, but the module also exports
DEFAULT_ACCOUNT_ID which is required transitively by src/web/auth-store.ts.

Switch to importOriginal pattern so all real exports are preserved alongside
the mocked functions.

* pi-runner: guard accumulated tool-result overflow in transformContext

* PI runner: compact overflowing tool-result context

* Subagent: harden tool-result context recovery

* Enhance tool-result context handling by adding support for legacy tool outputs and improving character estimation for message truncation. This includes a new function to create legacy tool results and updates to existing functions to better manage context overflow scenarios.

* Enhance iMessage handling by adding reply tag support in send functions and tests. This includes modifications to prepend or rewrite reply tags based on provided replyToId, ensuring proper message formatting for replies.

* Enhance message delivery across multiple channels by implementing sticky reply context for chunked messages. This includes preserving reply references in Discord, Telegram, and iMessage, ensuring that follow-up messages maintain their intended reply targets. Additionally, improve handling of reply tags in system prompts and tests to support consistent reply behavior.

* Enhance read tool functionality by implementing auto-paging across chunks when no explicit limit is provided, scaling output budget based on model context window. Additionally, add tests for adaptive reading behavior and capped continuation guidance for large outputs. Update related functions to support these features.

* Refine tool-result context management by stripping oversized read-tool details payloads during compaction, ensuring repeated read calls do not bypass context limits. Introduce new utility functions for handling truncation content and enhance character estimation for tool results. Add tests to validate the removal of excessive details in context overflow scenarios.

* Refine message delivery logic in Matrix and Telegram by introducing a flag to track if a text chunk was sent. This ensures that replies are only marked as delivered when a text chunk has been successfully sent, improving the accuracy of reply handling in both channels.

* fix: tighten reply threading coverage and prep fixes (#19508) (thanks @tyler6204)

CHANGELOG.md

@@ -8,6 +8,9 @@ Docs: https://docs.openclaw.ai
 
 - Agents/Subagents: add an accepted response note for `sessions_spawn` explaining polling subagents are disabled for one-off calls. Thanks @tyler6204.
 - Agents/Subagents: prefix spawned subagent task messages with context to preserve source information in downstream handling. Thanks @tyler6204.
+- iMessage: support `replyToId` on outbound text/media sends and normalize leading `[[reply_to:<id>]]` tags so replies target the intended iMessage. Thanks @tyler6204.
+- UI/Sessions: avoid duplicating typed session prefixes in display names (for example `Subagent Subagent ...`). Thanks @tyler6204.
+- Auto-reply/Prompts: include trusted inbound `message_id` in conversation metadata payloads for downstream targeting workflows. Thanks @tyler6204.
 - iOS/Talk: add a `Background Listening` toggle that keeps Talk Mode active while the app is backgrounded (off by default for battery safety). Thanks @zeulewan.
 - iOS/Talk: harden barge-in behavior by disabling interrupt-on-speech when output route is built-in speaker/receiver, reducing false interruptions from local TTS bleed-through. Thanks @zeulewan.
 - iOS/Talk: add a `Voice Directive Hint` toggle for Talk Mode prompts so users can disable ElevenLabs voice-switching instructions to save tokens when not needed. (#18250) Thanks @zeulewan.
@@ -42,6 +45,12 @@ Docs: https://docs.openclaw.ai
 
 ### Fixes
 
+- Agents/Subagents: preemptively guard accumulated tool-result context before model calls by truncating oversized outputs and compacting oldest tool-result messages to avoid context-window overflow crashes. Thanks @tyler6204.
+- Agents/Subagents: add explicit subagent guidance to recover from `[compacted: tool output removed to free context]` / `[truncated: output exceeded context limit]` markers by re-reading with smaller chunks instead of full-file `cat`. Thanks @tyler6204.
+- Agents/Tools: make `read` auto-page across chunks (when no explicit `limit` is provided) and scale its per-call output budget from model `contextWindow`, so larger contexts can read more before context guards kick in. Thanks @tyler6204.
+- Agents/Tools: strip duplicated `read` truncation payloads from tool-result `details` and make pre-call context guarding account for heavy tool-result metadata, so repeated `read` calls no longer bypass compaction and overflow model context windows. Thanks @tyler6204.
+- Reply threading: keep reply context sticky across streamed/split chunks and preserve `replyToId` on all chunk sends across shared and channel-specific delivery paths (including iMessage, BlueBubbles, Telegram, Discord, and Matrix), so follow-up bubbles stay attached to the same referenced message. Thanks @tyler6204.
+- Gateway/Agent: defer transient lifecycle `error` snapshots with a short grace window so `agent.wait` does not resolve early during retry/failover. Thanks @tyler6204.
 - iOS/Onboarding: stop auth Step 3 retry-loop churn by pausing reconnect attempts on unauthorized/missing-token gateway errors and keeping auth/pairing issue state sticky during manual retry. (#19153) Thanks @mbelinky.
 - Voice-call: auto-end calls when media streams disconnect to prevent stuck active calls. (#18435) Thanks @JayMishra-source.
 - Voice call/Gateway: prevent overlapping closed-loop turn races with per-call turn locking, route transcript dedupe via source-aware fingerprints with strict cache eviction bounds, and harden `voicecall latency` stats for large logs without spread-operator stack overflow. (#19140) Thanks @mbelinky.