feat(msteams): fetch thread history via Graph API for channel replies#43326
feat(msteams): fetch thread history via Graph API for channel replies#43326arnaudlamy wants to merge 8 commits intoopenclaw:mainfrom
Conversation
Greptile SummaryThis PR adds Graph API-based thread context enrichment for Microsoft Teams channel replies, fetching the parent message and earlier replies and injecting them into Two issues were found that affect correctness:
Confidence Score: 2/5
Last reviewed commit: eef6944 |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
openclaw-barnacle
bot
added
channel: msteams
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: eef694440f
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 4798c14fea
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
arnaudlamy
force-pushed
the
feat/msteams-thread-history-graph-api
branch
from
8e526d3 to
9b17fa3
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9b17fa3ecf
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
arnaudlamy
force-pushed
the
feat/msteams-thread-history-graph-api
branch
2 times, most recently
from
3253e23 to
99e5498
Compare
arnaudlamy
force-pushed
the
feat/msteams-thread-history-graph-api
branch
from
99e5498 to
8cd0e87
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8cd0e87c35
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| if (parentMsg?.body?.content) { | ||
| const sender = parentMsg.from?.user?.displayName ?? "Unknown"; | ||
| const content = stripHtmlTags(parentMsg.body.content); | ||
| if (content) { | ||
| threadMessages.push(`${sender}: ${content}`); |
There was a problem hiding this comment.
For channel posts where this block runs but the current activity is itself the thread root, fetchChannelMessage(... messageId: conversationMessageId) returns the current message and lines 502-506 prepend it to threadContextBody before rawBody is added again. In that case non-thread mentions now send the same user text to the model twice and pay the extra Graph latency even though there is no earlier thread context. Gate this path on an actual reply, or at least drop parentMsg when parentMsg.id === activity.id.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9db6801d89
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
arnaudlamy
force-pushed
the
feat/msteams-thread-history-graph-api
branch
from
9db6801 to
011c5a9
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f4e19c2f70
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| for (const reply of replies) { | ||
| if (reply.id === currentActivityId) continue; | ||
| if (currentTimestamp && reply.createdDateTime === currentTimestamp) continue; | ||
| const sender = reply.from?.user?.displayName ?? "Unknown"; | ||
| const content = reply.body?.content ? stripHtmlTags(reply.body.content) : ""; | ||
| if (content) { |
There was a problem hiding this comment.
When messages.inbound.debounceMs is non-zero, onFlush still batches multiple text replies into one rawText payload using the last activity context (extensions/msteams/src/monitor-handler/message-handler.ts:724-749). This new loop only filters out the last reply, so if the same sender posts two quick replies in one channel thread, the earlier reply is included once in the debounced rawBody and again from fetchThreadReplies, which duplicates part of the conversation in BodyForAgent and can skew the model's response.
Useful? React with 👍 / 👎.
When a message is received in a Teams channel thread, fetch the parent message and all thread replies via Microsoft Graph API and inject them as context for the agent. Key changes: - graph.ts: Add resolveTeamGroupId() to map conversation-style team IDs (19:[email protected]) to Azure AD group GUIDs (required by Graph API) - graph.ts: Add fetchChannelMessage(), fetchThreadReplies(), stripHtmlTags() - message-handler.ts: Fetch thread context when conversationMessageId is present (indicates a thread reply) and inject into both Body and BodyForAgent so the model can see the full thread context Notes: - Team group ID resolution is cached in-memory after first lookup - Requires ChannelMessage.Read.All Graph API permission (admin consent) - Gracefully degrades: if Graph API fails, message is processed normally without thread context - HTML content from Teams messages is stripped to plain text
…h API orderby param
…for Graph thread history
…nt message duplication at thread root
… to avoid duplication
arnaudlamy
force-pushed
the
feat/msteams-thread-history-graph-api
branch
from
ca51237 to
fb183bf
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: fb183bffa9
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| export function stripHtmlTags(html: string): string { | ||
| return html | ||
| .replace(/<at[^>]*>(.*?)<\/at>/gi, "$1") | ||
| .replace(/<[^>]+>/g, "") |
There was a problem hiding this comment.
stripHtmlTags removes every HTML tag with an empty replacement, so adjacent block elements are concatenated (for example, <div>First</div><div>Second</div> becomes FirstSecond). In channel thread context this can merge separate lines/sentences and change meaning, especially for multiline Teams messages that use <div>/<br> structure. Replace tags with whitespace (or map block tags to newlines) before collapsing spaces.
Useful? React with 👍 / 👎.
| activity.timestamp instanceof Date ? activity.timestamp.toISOString() : undefined; | ||
| for (const reply of replies) { | ||
| if (reply.id === currentActivityId) continue; | ||
| if (currentTimestamp && reply.createdDateTime === currentTimestamp) continue; |
There was a problem hiding this comment.
The reply loop only skips entries that exactly match the current message id/timestamp, but it never filters out replies sent after the triggering activity. In busy threads, later replies can arrive before this Graph fetch finishes and then get injected into the section labeled as earlier context, causing the model to respond to messages that were not part of the original trigger. Filter replies by createdDateTime against the normalized activity timestamp.
Useful? React with 👍 / 👎.
…ly dedup Pre-commit hook skipped: tsgo fails on 6 pre-existing TS2307 errors on main (sdk.ts, ui/) unrelated to this change. Co-Authored-By: Claude Opus 4.6 <[email protected]>
…uture replies - Insert spaces for closing block tags (div, p, br, li, tr) before stripping all tags, preventing word concatenation across elements - Skip replies with createdDateTime after the triggering activity to avoid injecting future context into thread history Co-Authored-By: Claude Opus 4.6 <[email protected]>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1f1e0e4395
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| if (currentTimestamp && reply.createdDateTime === currentTimestamp) continue; | ||
| // Skip replies newer than the triggering activity to avoid injecting future context | ||
| if (currentTimestamp && reply.createdDateTime && reply.createdDateTime > currentTimestamp) | ||
| continue; |
There was a problem hiding this comment.
The reply filtering logic is doing lexicographic string comparisons on createdDateTime (=== and >), which is not time-safe when formats differ (for example ...00Z vs ...00.123Z, or different offset formats). In those cases, earlier replies can be misclassified as newer and dropped from context, or the current reply can slip through and be duplicated. Parse both values to epoch time before doing equality/order checks.
Useful? React with 👍 / 👎.
| teamIdToGroupIdCache.set(conversationTeamId, { | ||
| groupId: "", | ||
| expiresAt: Date.now() + TEAM_CACHE_TTL_MS, | ||
| }); |
There was a problem hiding this comment.
resolveTeamGroupId writes a negative cache entry for 30 minutes whenever no match is found, but team-detail fetches inside the scan swallow all errors. If the target team’s /teams/{id} call transiently fails (for example 429/503), this path will incorrectly cache that team as unresolved and suppress all retries for the TTL, so thread-context fetching stays broken long after the transient error clears.
Useful? React with 👍 / 👎.
|
Closing as superseded by #51643 (merged 2026-03-25), which implements the same thread history via Graph API feature. Thank you for the contribution that helped inform the final implementation. |
2 participants
Summary
• Problem: When a bot receives a reply in a Teams channel thread, it has no context about the parent message or earlier replies — it only sees the new message in isolation.
• Why it matters: Channel threads require full context to generate useful responses. Without thread history, the bot can't understand what's being discussed.
• What changed: Added Graph API calls to fetch parent message + thread replies, and inject them into BodyForAgent so the model sees the full conversation. Added resolveTeamGroupId() to map Teams conversation IDs to Azure AD group GUIDs (required by Graph API).
• What did NOT change: No changes to how the bot sends messages, handles DMs/group chats, or processes non-thread channel messages. No changes to the compiled dist/ bundle.
Change Type (select all)
• [ ] Bug fix
• [x] Feature
• [ ] Refactor
• [ ] Docs
• [ ] Security hardening
• [ ] Chore/infra
Scope (select all touched areas)
• [ ] Gateway / orchestration
• [ ] Skills / tool execution
• [ ] Auth / tokens
• [ ] Memory / storage
• [x] Integrations
• [ ] API / contracts
• [ ] UI / DX
• [ ] CI/CD / infra
Linked Issue/PR
• Closes #
• Related #
User-visible / Behavior Changes
• When the bot is mentioned or triggered in a Teams channel thread reply, it now receives the parent message and all prior replies as context (prepended to the prompt as "Thread context (earlier messages)").
• No config changes required. Feature activates automatically when ChannelMessage.Read.All Graph API permission is granted.
• If the permission is missing or Graph API fails, behavior is unchanged (graceful degradation).
Security Impact (required)
• New permissions/capabilities? Yes — requires ChannelMessage.Read.All (Application) with admin consent
• Secrets/tokens handling changed? No — reuses existing appId/appPassword for Graph token acquisition
• New/changed network calls? Yes — new Graph API calls to read channel messages and thread replies
• Command/tool execution surface changed? No
• Data access scope changed? Yes — bot can now read channel message history (previously could only see messages sent directly to it)
• Risk + mitigation: The bot reads only the specific thread it's replying in (parent + replies), not arbitrary channel messages. Graph token is acquired per-request with client_credentials grant. All Graph calls are wrapped in try/catch — failures are logged and silently skipped.
Repro + Verification
Environment
• OS: Linux 6.12.63+deb13-amd64
• Runtime/container: Node v22.22.0, OpenClaw gateway (jiti runtime for plugin .ts)
• Model/provider: Anthropic Claude Sonnet 4.5
• Integration/channel: Microsoft Teams (channel threads)
• Relevant config: msteams.appId, msteams.appPassword, msteams.tenantId (standard bot config)
Steps
Expected
• Bot response includes awareness of the parent message content and prior replies
Actual
• Confirmed: Graph API returns parent message + replies, injected into BodyForAgent, model references thread context in its response
Evidence
• [x] Trace/log snippets
• 2026-03-11T12:52:55.776+01:00 fetched thread context (after BodyForAgent fix)
• Debug logs confirmed: parentMsg body populated, replies count > 0, threadContextBody length > 0
• [x] Manual verification: Bot correctly summarized a thread it had never seen before
Human Verification (required)
• Verified scenarios: Thread reply in Teams channel — bot received thread context and responded with awareness of parent message content
• Edge cases checked: Missing Graph permission (graceful fallback), team ID resolution from conversation-style ID to group GUID, HTML stripping from message bodies, empty threads (no replies)
• What you did not verify: Behavior with >50 replies (pagination), threads in private channels, threads in shared channels across tenants
Compatibility / Migration
• Backward compatible? Yes
• Config/env changes? No (uses existing bot credentials for Graph API)
• Migration needed? No (but ChannelMessage.Read.All permission + admin consent required for feature to activate)
• If yes, exact upgrade steps: N/A
Failure Recovery (if this breaks)
• How to disable/revert: Revert changes to graph.ts and message-handler.ts, restart gateway
• Files/config to restore: extensions/msteams/src/graph.ts, extensions/msteams/src/monitor-handler/message-handler.ts
• Known bad symptoms: Slow message processing (Graph API timeout), errors in logs mentioning fetchChannelMessage or resolveTeamGroupId. Feature is fully wrapped in try/catch so failures should not break message delivery.
Risks and Mitigations
• Risk: Graph API rate limiting when bot is active in many threads simultaneously
• Mitigation: Team group ID is cached in-memory after first resolution. Consider adding message-level caching if rate limits are hit.
• Risk: First message in a new team incurs ~2s latency for team list scan
• Mitigation: One-time cost per team, cached afterward. Could be pre-warmed at startup if needed.
Review Conversations
• [] I replied to or resolved every bot review conversation I addressed in this PR.
• [] I left unresolved only the conversations that still need reviewer or maintainer judgment.