feat(slack): backfill thread history on mid-thread mention by Aaronontheweb · Pull Request #576 · netclaw-dev/netclaw

Aaronontheweb · 2026-04-09T19:25:12Z

Summary

Adds Slack thread catch-up when the bot is @-mentioned in an existing thread, so Netclaw can rebuild context before responding.
Uses an atomic ingress model: historical gap + live triggering event are merged into a single inbound user turn before enqueue.
Persists a per-thread cursor inside SlackThreadBindingActor (Akka.Persistence) to support restart/offline gap recovery without full replay.
Drops stale out-of-order events using cursor monotonicity checks to prevent duplicate processing/replies.
Applies prompt-injection screening to backfilled history and fail-closes on detector failures (with user-visible warning).

Closes #575

Key Implementation Details

Removed IsBackfill from channel/session contracts (ChannelInput, SendUserMessage, and pipeline mapping).
Moved continuity logic to Slack channel layer; LlmSessionActor remains transport-agnostic.
SlackThreadBindingActor now:
- recovers persisted cursor,
- fetches thread history via conversations.replies,
- computes (cursor, currentEventTs) gap,
- merges gap + live message into one payload,
- advances cursor monotonically after successful enqueue,
- compacts old cursor events periodically.
Timeout model simplified in thread binding actor:
- inbound processing: 30s
- operation timeout: 10s

Tests and Validation

Added/updated Slack backfill integration coverage for:
- atomic merged turn assembly,
- restart catch-up behavior,
- high-risk backfill filtering,
- stale out-of-order event dropping.
Updated fetcher tests for root inclusion and current filtering behavior.
Removed obsolete IsBackfill propagation test.
Validation run:
- dotnet test src/Netclaw.Actors.Tests/Netclaw.Actors.Tests.csproj
- dotnet slopwatch analyze

Add OpenSpec change directory for backfilling Slack thread history when the bot is @-mentioned mid-thread. Channel-agnostic design targeting Slack first with Teams/Discord in mind.

…575) Proposal establishes the need for backfilling thread history when the bot is @-mentioned mid-thread, with channel-agnostic design targeting Slack first. Design documents key decisions: reuse existing multimodal inbound pipeline, IThreadHistoryFetcher interface, backfill-as-context injection, and no artificial caps (compaction handles overflow).

Specs: new thread-history-backfill capability plus delta specs for netclaw-input-adapters (IsBackfill flag), netclaw-slack-socket (conversations.replies fetch), and netclaw-session (context block injection). Tasks: 22 items across 6 groups covering abstraction layer, Slack implementation, adapter wiring, session injection, testing, and documentation sync.

…#575) Add thread history backfill so the bot sees prior messages (text + images) when @-mentioned in an existing Slack thread. Backfilled content flows through the existing multimodal inbound pipeline and is injected as a read-only context block before the first LLM turn. Key changes: - IsBackfill flag on ChannelInput/SendUserMessage for backfill detection - IThreadHistoryFetcher interface for channel-agnostic history fetch - SlackThreadHistoryFetcher using conversations.replies with pagination, bot filtering, image download, and content scanning - ThreadHistoryContextBuilder assembles backfill into delimited context - LlmSessionActor accumulates backfill messages and injects as context - SlackThreadBindingActor calls fetcher during first initialization

Two integration tests using full actor hierarchy: - Backfill text + images appear in LLM context before the mention message - Second message in same thread does not re-trigger backfill All 12 backfill-related tests passing, 29/29 tasks complete.

) Add timestamp formatting to ThreadHistoryContextBuilder so each backfilled message shows when it was sent, matching the spec requirement and design doc format: <user: alice, 2026-04-09 10:15 UTC>

) - Remove unused sessionsBasePath parameter from ThreadHistoryContextBuilder - Replace DateTimeOffset.UtcNow fallback with default in ParseSlackTs (violates TimeProvider rule and silently assigns wrong timestamp) - Replace _backfillComplete flag with generation == 1 check (eliminates redundant mutable state) - Give backfill its own 30s timeout instead of sharing pipeline init budget - Download images concurrently via Task.WhenAll instead of sequentially - Cap _pendingBackfill at 500 messages for defense in depth

…ion comments (#575) Consolidate duplicate file download logic from SlackThreadBindingActor and SlackThreadHistoryFetcher into a shared SlackFileDownloader helper. Both callers now delegate HTTP GET + Bearer token auth to the same method. Remove narration comment that restated the method name.

Aaronontheweb

Needs work

Aaronontheweb · 2026-04-10T13:21:07Z

+    /// When true, this message is backfilled thread history — not a live message.
+    /// Backfilled messages are injected as read-only context before the first LLM turn.
+    /// </summary>
+    public bool IsBackfill { get; init; }


Doesn't seem necessary any more - when we do a channel sync prior to the first turn, we're just supposed to lump the user messages together into a batch (like how we normally do with user messages that accumulate between LLM turns)

Aaronontheweb · 2026-04-10T13:21:19Z

            MediaReferences = mediaRefs,
-            Source = MessageSourceFactory.Create(input, options, turnId)
+            Source = MessageSourceFactory.Create(input, options, turnId),
+            IsBackfill = input.IsBackfill


Probably also not necessary

Aaronontheweb · 2026-04-10T13:22:35Z

    private static readonly TimeSpan FileDownloadTimeout = TimeSpan.FromSeconds(10);
    private static readonly TimeSpan ContentScanTimeout = TimeSpan.FromSeconds(5);
    private static readonly TimeSpan PipelineInitTimeout = TimeSpan.FromSeconds(15);
+    private static readonly TimeSpan BackfillTimeout = TimeSpan.FromSeconds(30);


why do we need a separate timeout for this?

Aaronontheweb · 2026-04-10T13:27:40Z

+        }
+        catch (Exception ex)
+        {
+            _log.Warning(ex, "Prompt injection detector failed for backfill message; allowing message through");


shouldn't this be fail-closed instead of fail-open? we just need to make sure the user gets visual feedback that there was a problem.

Aaronontheweb · 2026-04-10T13:30:15Z

+        _log.Debug("Slack thread cursor did not advance stream={StreamKey} ts={Ts}", _sessionId.Value, ts);
+    }
+
+    private static List<AIContent> MergeGapWithLiveContents(IReadOnlyList<ChannelInput> gap, IReadOnlyList<AIContent> liveContents)


this all gets batched with the live event that triggered this session to be re-activated right?

Aaronontheweb · 2026-04-10T13:31:15Z

+        return merged;
+    }
+
+    private static string? TryExtractTsFromEventId(string eventId)


what information gets exposed by the strongly typed SlackNet APIs? is there a way to get message datetime from that? or are we operating based on monotonic event ids instead?

Aaronontheweb · 2026-04-10T13:33:56Z

+
+namespace Netclaw.Channels.Slack;
+
+public interface ISlackThreadCursorStore


imho, this entire concept should not exist. this should be Akka.Persistence inside the SlackThreadBindingActor. each time we advance the cursor we save that asynchronously using PersistAsync - and we can check LastSeqNr % 10 and delete all prior messages (LastSeqNr - 1) in order to keep the Akka.Persistence state small. All we'd be persisting is the cursor position.

Aaronontheweb · 2026-04-10T13:35:55Z

    private static readonly TimeSpan FileDownloadTimeout = TimeSpan.FromSeconds(10);
    private static readonly TimeSpan ContentScanTimeout = TimeSpan.FromSeconds(5);
    private static readonly TimeSpan PipelineInitTimeout = TimeSpan.FromSeconds(15);
+    private static readonly TimeSpan BackfillTimeout = TimeSpan.FromSeconds(30);


Also, just generally - there are way too many different timeout types - that's ridiculous. Consolidate to having two:

Inbound processing timeout (covers backfill too) - 30s

Operation timeout - 10s, covers everything else.

…nd archive Rewrite proposal, design, delta specs, and tasks to match the cursor-based hydration implementation (persistent binding actor, merge-into-trigger semantics, injection gate, stale-event drop). Sync deltas to main specs and move the change to archive.

- Live prompt-injection detector path now fails closed: detector exceptions drop the message and post a visible warning instead of silently allowing it through (CLAUDE.md no-silent-fallback). - IThreadHistoryFetcher is now a required dependency on SlackChannel and SlackGatewayDependencies. Tests opt in with an explicit EmptyThreadHistoryFetcher fake. - Unify the live and backfill detector invocations behind one ClassifyAsync helper returning Allow/Block/DetectorUnavailable. - Move Slack ts parsing onto SlackEventTs (IComparable, TryToDecimal, ToDateTimeOffset) and SlackEventId (TryGetEventTs), replacing four copies of decimal/double parsing in the actor and fetcher. - Parse the inbound event ts once per HandleInboundAsync and pass through stale-check, gap computation, and cursor advance. - Hydrate once per runtime by setting the fetch-attempted flag up front so a downstream throw can't re-fetch the whole thread. - Parallelize gap classification with Task.WhenAll — the detector call was the dominant first-turn latency on long gaps. - Gate DeleteMessages on !IsRecovering so journal truncation only runs on genuine persists, not replay. - Drop gap image bytes from the merged ChannelInput — the `[image attachments: N]` summary line already represents them and copying the raw payloads duplicated multi-MB arrays per turn. - Revert whitespace-only edits on ChannelInput and Commands.

Aaronontheweb added 11 commits April 9, 2026 17:39

feat(openspec): scaffold thread-history-backfill change (#575)

0819813

Add OpenSpec change directory for backfilling Slack thread history when the bot is @-mentioned mid-thread. Channel-agnostic design targeting Slack first with Teams/Discord in mind.

fix(slack): include UTC timestamps in thread history context block (#575

52408db

) Add timestamp formatting to ThreadHistoryContextBuilder so each backfilled message shows when it was sent, matching the spec requirement and design doc format: <user: alice, 2026-04-09 10:15 UTC>

Merge branch 'dev' into claude-wt-slack-threads

749de3b

fix(slack): add cursor-based atomic thread catch-up

6bd3a9c

Merge branch 'dev' into claude-wt-slack-threads

43054bb

Aaronontheweb commented Apr 10, 2026

View reviewed changes

Aaronontheweb added 4 commits April 10, 2026 14:07

fix(slack): persist thread cursor in binding actor

3a824e0

fix(slack): drop stale out-of-order thread events

b38de89

Aaronontheweb marked this pull request as ready for review April 10, 2026 18:41

Aaronontheweb merged commit 29f65e7 into dev Apr 10, 2026
3 checks passed

Aaronontheweb deleted the claude-wt-slack-threads branch April 10, 2026 18:41

Aaronontheweb mentioned this pull request Apr 10, 2026

fix(slack): restore gap image DataContent in thread hydration #583

Merged

3 tasks

Aaronontheweb mentioned this pull request May 14, 2026

fix(channels): propagate audience on Slack history-fetched ChannelInput #993

Merged

3 tasks

Navigation Menu

Uh oh!

feat(slack): backfill thread history on mid-thread mention - #576

feat(slack): backfill thread history on mid-thread mention#576
Aaronontheweb merged 15 commits into
devfrom
claude-wt-slack-threads

Aaronontheweb commented Apr 9, 2026 •

edited

Loading

Uh oh!

Aaronontheweb left a comment

Uh oh!

Aaronontheweb Apr 10, 2026

Uh oh!

Aaronontheweb Apr 10, 2026

Uh oh!

Aaronontheweb Apr 10, 2026

Uh oh!

Aaronontheweb Apr 10, 2026

Uh oh!

Aaronontheweb Apr 10, 2026

Uh oh!

Aaronontheweb Apr 10, 2026

Uh oh!

Aaronontheweb Apr 10, 2026

Uh oh!

Aaronontheweb Apr 10, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant


		namespace Netclaw.Channels.Slack;

		public interface ISlackThreadCursorStore

Uh oh!

Conversation

Aaronontheweb commented Apr 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Key Implementation Details

Tests and Validation

Uh oh!

Aaronontheweb left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Aaronontheweb commented Apr 9, 2026 •

edited

Loading