* fix: stabilize telegram draft stream message boundaries
* fix: suppress NO_REPLY lead-fragment leaks
* fix: keep underscore guard for non-NO_REPLY prefixes
* fix: skip assistant-start rotation only after real lane rotation
* fix: preserve finalized state when pre-rotation does not force
* fix: reset finalized preview state on message-start boundary
* fix: document Telegram draft boundary + NO_REPLY reliability updates (#33169) (thanks @obviyus)
* feat(hooks): add trigger and channelId to plugin hook agent context
Adds `trigger` and `channelId` fields to `PluginHookAgentContext` so
plugins can determine what initiated the agent run and which channel
it originated from, without session-key parsing or Redis bridging.
trigger values: "user", "heartbeat", "cron", "memory"
channelId values: "telegram", "discord", "whatsapp", etc.
Both fields are threaded through run.ts and attempt.ts hookCtx so all
hook phases receive them (before_model_resolve, before_prompt_build,
before_agent_start, llm_input, llm_output, agent_end).
channelId falls back from messageChannel to messageProvider when the
former is not set. followup-runner passes originatingChannel so queued
followup runs also carry channel context.
* docs(changelog): note hook context parity fix for #28623
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
The `forceFlushTranscriptBytes` path (introduced in d729ab21) bypasses the
`memoryFlushCompactionCount` guard that prevents repeated flushes within the
same compaction cycle. Once the session transcript exceeds 2 MB, memory flush
fires on every single message — even when token count is well under the
compaction threshold.
Extract `hasAlreadyFlushedForCurrentCompaction()` from the inline guard in
`shouldRunMemoryFlush` and apply it to both the token-based and the
transcript-size trigger paths.
Fixes#32317
Signed-off-by: HCL <chenglunhu@gmail.com>
Address Greptile review: when sessionKey is undefined the fallback
matched any enabled cron job, which could silently suppress the guard
note due to jobs from unrelated sessions. Return false instead so the
note always appears when session scoping is not possible.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Before appending the "I did not schedule a reminder" guard note, check the
cron store for enabled jobs matching the current session key. This prevents
false positives when the agent references an existing cron created in a
prior turn (e.g. "I'll ping you when it's done" while a monitoring cron is
already running).
The check only fires on the rare path where the text matches commitment
patterns AND no cron was added in the current turn, so the added I/O is
negligible.
Closes#32228
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The followup runner (which processes queued messages) was calling
runEmbeddedPiAgent without currentChannelId or currentThreadTs.
This meant the message tool's toolContext had no channel routing
info, causing reactions (and other target-inferred actions) to
fail with 'Action react requires a target' on queued messages.
Pass originatingTo as currentChannelId so the message tool can
infer the reaction target from context, matching the behavior
of the initial (non-queued) agent run.
The /\s+/g whitespace normalizer collapsed newlines along with spaces/tabs,
destroying paragraph structure in multi-line messages before they reached
the LLM. Use /[^\S\n]+/g to only collapse horizontal whitespace while
preserving line breaks.
Closes#32216
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds group context fields to MessageSentHookContext so hooks can
correlate sent events with received events for the same conversation.
Previously, message:received included isGroup/groupId but message:sent
did not, forcing hooks to use mismatched identifiers (e.g. groupId vs
numeric chat ID) when tracking conversations.
Fields are derived from MsgContext in dispatch-from-config and threaded
through route-reply and deliver via the mirror parameter.
Addresses feedback from matskevich (production user, 550+ events)
reported on PR #6797.
Adds two new internal hook events that fire after media/link processing:
- message:transcribed: fires when audio has been transcribed, providing
the transcript text alongside the original body and media metadata.
Useful for logging, analytics, or routing based on spoken content.
- message:preprocessed: fires for every message after all media + link
understanding completes. Gives hooks access to the fully enriched body
(transcripts, image descriptions, link summaries) before the agent sees it.
Both hooks are added in get-reply.ts, after applyMediaUnderstanding and
applyLinkUnderstanding. message:received and message:sent are already
in upstream (f07bb8e8) and are not duplicated here.
Typed contexts (MessageTranscribedHookContext, MessagePreprocessedHookContext)
and type guards (isMessageTranscribedEvent, isMessagePreprocessedEvent) added
to internal-hooks.ts alongside the existing received/sent types.
Test coverage in src/hooks/message-hooks.test.ts.
After a drain loop empties the queue it deletes the key from
FOLLOWUP_QUEUES. If a new message arrives at that moment
enqueueFollowupRun creates a fresh queue object with draining:false
but never starts a drain, leaving the message stranded until the
next run completes and calls finalizeWithFollowup.
Fix: persist the most recent runFollowup callback per queue key in
FOLLOWUP_RUN_CALLBACKS (drain.ts). enqueueFollowupRun now calls
kickFollowupDrainIfIdle after a successful push; if a cached
callback exists and no drain is running it calls scheduleFollowupDrain
to restart immediately. clearSessionQueues cleans up the callback
cache alongside the queue state.
