feat(mattermost): add ambient session ingestion mode

[And1rew132]

Summary

Add support for ambient session ingestion in the Mattermost connector — a new channel mode where messages are silently stored into Hermes session history without triggering an LLM response.

Also includes the changes from #20874 (Mattermost thread-mode fixes), merged in.

Problem

The existing MATTERMOST_FREE_RESPONSE_CHANNELS bypasses mention gating and immediately invokes handle_message() for every message. In multi-agent and large-team environments this causes:

• Excessive token usage
• Response loops between agents
• Noisy behavior in busy channels
• No way to build passive/ambient memory

Solution

Introduce a new env var:

MATTERMOST_AMBIENT_CHANNELS=channel_id_1,channel_id_2

Behavior matrix:

Channel type	Message	Result
Normal	no mention	ignored
Normal	@mention	LLM runs
Free-response	any	LLM runs immediately
Ambient	no mention	stored silently, no LLM
Ambient	@mention	LLM runs with accumulated context

Ambient messages are stored with sender attribution ([username]: text) so the agent has full context on who said what when eventually triggered.

Implementation

gateway/platforms/base.py

• Add trigger_llm: bool = True field to MessageEvent
• BasePlatformAdapter.handle_message() short-circuits when trigger_llm=False:
  • Calls session_store.get_or_create_session() + append_to_transcript() with sender attribution
  • Returns without invoking _message_handler or spawning background tasks

gateway/platforms/mattermost.py

• Parse MATTERMOST_AMBIENT_CHANNELS (also readable from config.yaml via ambient_channels key)
• In _handle_ws_event(): detect ambient channels, set trigger_llm=False when no @mention present, trigger_llm=True on @mention
• Pass trigger_llm to MessageEvent constructor

Use cases

• Multi-agent orchestration (prevent response loops)
• Passive memory accumulation
• Background context awareness
• Enterprise chat environments
• Mattermost-based operational workflows

[alt-glitch]

This appears to be a resubmission of closed #26663 (same author, same title). Also includes changes from #20874. Reviewers: please check what changed vs the previous submission.

[teknium1]

Thanks for the ambient-ingestion proposal. The use case is valid, but the current implementation has blocking correctness and boundary issues.

Problems

• gateway/platforms/base.py:2834 writes and returns before _message_handler. On current main that handler is GatewayRunner._handle_message (gateway/run.py:7063), where authorization occurs at gateway/run.py:8958-8998; unapproved senders can therefore inject transcript content.
• The proposed write uses event.source (gateway/platforms/base.py:2837), while Mattermost sources retain each sender ID (gateway/platforms/mattermost.py:932-946). Current session keys isolate normal group/channel sessions per user by default (gateway/session.py:892-958, gateway/config.py:700), so it does not produce the claimed shared accumulated channel context.
• .github/workflows/build-cr-image.yml:6-55 is unrelated branch-specific publishing automation and must not be part of this feature.

Suggested changes

• Rework against the current plugin adapter at plugins/platforms/mattermost/adapter.py (migrated in af973e407).
• Preserve the gateway authorization boundary and define shared-context semantics explicitly. The documented pre_gateway_dispatch hook already supports silent ingestion/buffering (website/docs/user-guide/features/hooks.md:999-1057).
• Keep behavior configuration in config.yaml and remove the unrelated workflow.

Automated hermes-sweeper review.

[teknium1]

This returns before the GatewayRunner handler, which performs the normal authorization/pairing gate. An unapproved participant can therefore write [sender]: text into a session that an authorized user later loads. Preserve authorization before any ambient transcript write.

[teknium1]

This derives the transcript from the individual sender's event.source. With the default per-user group-session isolation, ambient posts from different channel members land in different sessions, so the later mention cannot receive the claimed channel-wide accumulated context.

[teknium1]

Remove this branch-specific publishing workflow from the feature PR. It is unrelated to Mattermost ingestion and grants package-publishing capability for an external Coding-Reality image.

[GottZ]

This was generated by AI during triage.

Summary

Two PRs address the same Mattermost ambient-ingestion use case: both add a trigger_llm path that writes unmentioned channel messages directly to session transcripts, while #26901 resubmits #26663 and adds unrelated thread-mode and Codex-stream changes. The shared implementation does not yet fix the reported cause safely because it bypasses authorization and, under default per-user session isolation, does not create the claimed shared channel context.

Related pull requests

• #26663 [closed] duplicate — (+771/-45) — superseded: Closed #26663 remains relevant as the original implementation, but its ambient-ingestion diff is carried forward substantially unchanged in #26901; it also mixes in Mattermost thread fixes and branch-specific image-publishing automation.
• #26901 related — (+954/-60) — keep open for substantial rework, not merge-ready: The contributor keep_open review on #26901 is supported by the diff: silent writes return before the gateway authorization handler, use sender-specific sources that preserve default per-user session isolation instead of shared channel context, target the pre-migration Mattermost adapter, and include unrelated publishing, thread-mode, and Codex-stream changes.

Duplicates

#26663 and #26901 implement substantially the same ambient-ingestion change; #26901 is the superseding resubmission and contains additional unrelated changes.

Suggested consolidation

Do not merge #26901 yet; retain it as the single open consolidation point only if the author rebases the feature onto plugins/platforms/mattermost/adapter.py, moves ingestion behind authorization, defines session keying that actually provides the intended accumulated channel context, and removes the unrelated workflow, thread-mode, and Codex-stream changes. This follows the contributor keep_open review on #26901 while treating its listed issues as blocking; #26663 is already closed and can remain closed as superseded by #26901.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    subgraph Dup26663 ["PRs duplicating each other"]
        P26663["PR #26663 (closed)"]
        P26901["PR #26901 (open)"]
    end
    class P26663 closed
    class P26901 open
    class P26901 target
    click P26663 "https://github.com/NousResearch/hermes-agent/pull/26663"
    click P26901 "https://github.com/NousResearch/hermes-agent/pull/26901"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed (state tag in the node label).

Cross-PR triage: Reviewed 2 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 104 kB of PR diffs, 4 kB of issue/PR text, 2 kB of discussion (3 comments), 0 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.