fix(telegram): reply to latest queued follow-up

[Hilo-Hilo]

Bug Description

When multiple Telegram follow-ups arrive while a session is busy—or are coalesced by Telegram ingress batching—Hermes can combine their text but visibly reply to an older message or inject stale quoted-message context into the model.

Root Cause

Merged events appended newer content while retaining parts of the first event's trigger tuple. The production queue-mode debounce path, Telegram text batching, photo batching, and album/media-group batching could bypass or partially apply the original metadata refresh.

That produced mixed events such as a newest message_id paired with an older reply_to_text, author, raw update, timestamp, or trust provenance. An internal event could also leave a merged user event marked internal.

Fix

• Centralize merged-event trigger/reply/provenance refresh.
• Apply it to direct text/media merges, queue-mode debounce and flush, Telegram ingress text batching, photo batching, and media groups.
• Advance message ID, reply ID/text/author/ownership, raw message, platform update ID, timestamp, channel context, and metadata coherently to the newest event.
• Merge metadata with newest values winning.
• Keep a merged event internal=True only when every constituent event was internal, preventing privilege escalation.
• Preserve the last concrete platform message ID for intentionally ID-less synthetic inputs while replacing stale quoted-message context.

Test Plan

• Regressions for newest-message anchoring, clearing stale quotes, and ID-less synthetic inputs.
• End-to-end queue debounce → flush regression verifies coherent newest reply metadata.
• Telegram ingress text-batching regression verifies newest trigger/reply/provenance metadata.
• Photo and album/media-group batching regression verifies newest reply targets.
• Mixed internal/user merge regression verifies latest raw/update/channel metadata and no privilege escalation.
• Rebased onto current origin/main (7cb2d2cd4).
• Focused queue/debounce/batching/pending surface: 54 passed.
• Combined current-upstream gateway suite with PR fix(telegram): prevent duplicate finals after failed edit #66026: 9462 passed, 13 skipped, 15 failed.
• Untouched current origin/main under the same environment: 9449 passed, 13 skipped, 15 failed.
• The 15 failing node IDs are identical between baseline and combined trees; the combined tree adds 13 passing regressions and introduces no new failures.
• Fresh isolated adversarial regressions: 5 passed.
• git diff --check and Ruff pass.

Risk Assessment

Low-to-moderate. Accumulated content and media behavior are unchanged; only the identity, reply context, provenance, and trust state of an already-merged event are made coherent with its newest constituent.

[alt-glitch]

This was generated by AI during triage.

Related to #43787 (Telegram batching) and #64752 (queued-turn return anchoring). This fixes the shared pending-event merge path, so it is related rather than a duplicate.

[tonydwb]

Code Review Summary

Verdict: Comment

Small fix (64 lines). Telegram queued reply target fix.

Checked diff — no security concerns, no hardcoded secrets.

Looks good. No blocking issues.

---

Reviewed by Hermes Agent

[tonydwb]

Code Review Summary (Re-review)

Verdict: LGTM / Comment

Fix ensures Telegram bot replies to the latest queued follow-up rather than an earlier one. Prior COMMENT review confirmed. No issues found.

[teknium1]

Thanks for tracing the queued-event and Telegram ingress paths. The premise is confirmed on current main: gateway/platforms/base.py:2157-2164 appends merged text without updating event identity or reply context, while gateway/run.py:11485-11499 injects that reply context into the model prompt. The proposed helper updates the relevant trigger tuple and preserves the existing message ID for ID-less synthetic inputs; the Telegram batching call sites and regressions cover the stated scope.

Suggested changes

• Consider a follow-up audit of other native text batchers. plugins/platforms/discord/adapter.py:6877-6882, plugins/platforms/matrix/adapter.py:3588-3595, and plugins/platforms/whatsapp/adapter.py:1297-1303 use the same append-without-metadata-refresh pattern. This PR is correctly scoped to Telegram, but those paths may need the same treatment if they rely on MessageEvent.message_id or reply context.

Automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Four PRs were reviewed: #46459 and #63298 directly address #45560, while #31496 fixes a separate three-sender delivery-loss path and #66027 fixes stale Telegram trigger/reply metadata after merging. #46459 adds visible boundaries while still collapsing canonical turns; #63298 preserves FIFO, history, and SessionDB boundaries and merges only the provider copy.

Related pull requests

• fix(gateway): prevent third-sender message drop in queue-mode debounce #31496 related — (+67/-8) — n/a: Prevents a third sender's text from being dropped in a queue-mode debounce collision, but the diff unconditionally merges C into A's event rather than preserving separate user turns. The contributor keep-open review blocks that implementation because it loses source attribution and requests separate FIFO entries plus a drain-path regression.
• fix(agent): preserve later user message boundaries #46459 partial — (+51/-26) — n/a: Adds a [Later user message] marker to runtime, Anthropic, and gateway merges, but repair_message_sequence still collapses canonical user objects and the TUI busy-queue merge remains outside its scope. Despite the contributor keep-open review, its repair_message_sequence aggressively merges separate user messages (v0.16.0+) #45560 work is superseded by fix: preserve queued prompt boundaries end to end #63298's canonical-boundary and FIFO implementation.
• fix: preserve queued prompt boundaries end to end #63298 best fix — (+2338/-261) — n/a: The recorded best fix replaces the TUI merged slot with an identity-preserving FIFO, keeps adjacent user messages separate in canonical history and SessionDB, and merges only the transient provider copy with [Next user message]. The maintainer-bot keep-open verdict identified lost steer behavior; the shown later changes and contributor evidence restore accepted steering and rejected/unavailable-steer fallback queueing.
• fix(telegram): reply to latest queued follow-up #66027 related — (+346/-8) — n/a: Refreshes merged-event trigger, reply, provenance, and timestamp metadata across the shared pending-event path and Telegram batching so the newest follow-up controls the reply target. As its maintainer-bot keep-open verdict reflects, this is a salvageable but distinct Telegram anchoring fix, not canonical user-turn preservation for repair_message_sequence aggressively merges separate user messages (v0.16.0+) #45560.

Duplicates

#46459 and #63298 overlap on #45560 and explicit provider-visible boundary markers; #46459 is the narrower implementation and can be closed as a duplicate superseded by #63298. #31496 and #66027 address separate gateway defects and are not duplicates of them.

Suggested consolidation

Keep #63298 open with the salvage path centered on its FIFO, canonical-history/SessionDB boundary preservation, stable source IDs, and wire-only alternation repair; require the current head to retain the restored steer contract. Close #46459 as duplicate of #63298 despite its keep-open review because its diff still collapses canonical turns and omits the TUI queue root cause; keep #66027 open as a separate Telegram metadata fix, and require author action on #31496 to rebase or split out a separate-event FIFO solution that resolves the contributor's source-attribution blocker.

Cross-PR triage: Reviewed 4 pull requests and 1 issue in this complex. Each diff was read against this issue; Assessment working set: 186 kB of PR diffs, 10 kB of issue/PR text, 12 kB of discussion (16 comments), 3 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.