feat(gateway): add optional role param to chat.inject RPC by donut-wenzhang · Pull Request #60706 · openclaw/openclaw

donut-wenzhang · 2026-04-04T05:56:48Z

Summary

Add optional role parameter to chat.inject RPC, allowing callers to inject messages with role=user (in addition to the existing role=assistant default).

Motivation: After d0-cli hot-update, we inject a message telling the bot to re-run d0 list. With the previous hardcoded role=assistant, the bot ignored it (doesn't treat its own past messages as instructions). With role=user, the bot executes it as a user command.

Changes:

logs-chat.ts: Add optional role to ChatInjectParamsSchema ("assistant" | "user")
chat.ts: Pass role through handler → appendAssistantTranscriptMessage
chat-transcript-inject.ts: Use role param (default "assistant"); omit assistant-specific fields (stopReason, usage, api, provider, model) for user-role messages

Backward compatible — existing callers without role param get the same behavior.

Test plan

Existing chat.inject calls without role still produce role=assistant messages
chat.inject with role=user produces user-role messages in session JSONL
Bot treats injected role=user message as instruction and executes it

Allow overriding the Telegram Bot API base URL via `channels.telegram.apiRoot`, enabling custom Bot API servers and bot routing proxies. Co-Authored-By: Claude Opus 4.6 <[email protected]>

Co-Authored-By: Claude Opus 4.6 <[email protected]>

…ons (#1) Allow hook API callers to pass sessionTarget="reuse" so that multiple hook invocations with the same sessionKey share the same agent session, enabling multi-turn conversation context retention. Previously sessionTarget was hardcoded to "isolated" for all hook calls, meaning every invocation created a brand new session regardless of sessionKey. With "reuse", the existing session freshness evaluation is used to determine whether to continue an existing session. Changes: - Add HookSessionTarget type and sessionTarget field to hook payload - Pass sessionTarget through to CronJob creation (default: "isolated") - Add "reuse" to CronSessionTarget union type across all validators - Add "reuse" to CronSessionTargetSchema for gateway cron API - Add assertSupportedJobSpec validation for "reuse" + "agentTurn" - Allow channel delivery for "reuse" sessions (same as "isolated") Co-authored-by: Claude Opus 4.6 <[email protected]>

When skills/ changes on main, triggers openclaw-updated repository_dispatch to DonutLabs-ai/donut-backend, which builds a skills tarball and hot-updates running D0 Gateway containers without restart. Co-authored-by: Regison <[email protected]> Co-authored-by: Claude Opus 4.6 <[email protected]>

…spaceFiles Allow operators to extend the built-in workspace file whitelist (bootstrap + memory files) by setting agents.defaults.extraWorkspaceFiles in openclaw.json. Co-Authored-By: Claude Opus 4.6 <[email protected]>

When senderId or senderUsername is available in the agent run context, inject X-TG-User-Id and X-TG-Username headers into outgoing LLM requests. This allows downstream proxies to attribute token usage to individual end users (e.g. Telegram senders). Co-authored-by: Claude Opus 4.6 <[email protected]>

Add two new streamFn wrappers that inject metadata into every outbound LLM request so the proxy can apply quota and usage tracking: - trigger_source: indicates what initiated the agent run ("user", "cron", "heartbeat"), derived from the existing PluginHookAgentContext.trigger field that is already set by cron jobs and user message handlers. - turn_id: a per-turn UUID generated before activeSession.prompt(), shared across all LLM calls within a single agent turn so the proxy can correlate multi-call agent loops back to one user message. Both values are injected via HTTP headers (X-Trigger-Source, X-Turn-Id) and request body fields (trigger_source, turn_id) for maximum proxy compatibility regardless of transport. Co-authored-by: Regison <[email protected]>

@shakkernerd

Merged via squash. Prepared head SHA: e4183b3 Co-authored-by: Sid-Qin <[email protected]> Co-authored-by: shakkernerd <[email protected]> Reviewed-by: @shakkernerd

* feat(telegram): add apiRoot config passthrough to grammY Allow overriding the Telegram Bot API base URL via `channels.telegram.apiRoot`, enabling custom Bot API servers and bot routing proxies. Co-Authored-By: Claude Opus 4.6 <[email protected]> * docs(telegram): document apiRoot config option Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(hooks): add sessionTarget="reuse" for multi-turn hook conversations (#1) Allow hook API callers to pass sessionTarget="reuse" so that multiple hook invocations with the same sessionKey share the same agent session, enabling multi-turn conversation context retention. Previously sessionTarget was hardcoded to "isolated" for all hook calls, meaning every invocation created a brand new session regardless of sessionKey. With "reuse", the existing session freshness evaluation is used to determine whether to continue an existing session. Changes: - Add HookSessionTarget type and sessionTarget field to hook payload - Pass sessionTarget through to CronJob creation (default: "isolated") - Add "reuse" to CronSessionTarget union type across all validators - Add "reuse" to CronSessionTargetSchema for gateway cron API - Add assertSupportedJobSpec validation for "reuse" + "agentTurn" - Allow channel delivery for "reuse" sessions (same as "isolated") Co-authored-by: Claude Opus 4.6 <[email protected]> * feat(skills): scan sub-skills from index.md for hierarchical display Add scanSubSkills() to skills-status that reads subdirectory index.md files under each skill's baseDir. Router skills (SKILL.md) now include a children array in the skills.status RPC response, enabling the HIL Dashboard to display skills in a categorized hierarchy. Co-Authored-By: Claude Opus 4.6 <[email protected]> --------- Co-authored-by: arucil <[email protected]> Co-authored-by: Claude Opus 4.6 <[email protected]>

* feat(agents): inject feature_context and skill_names into LLM proxy requests Add onPayload wrapper that automatically classifies each LLM API call: - chat_response: user message triggered the call - tool_result: processing tool execution results - system: fallback for unclassifiable calls Also injects active skill_names array for per-skill token attribution. Both fields are read by the donut-backend LLM proxy for analytics. Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * refactor(agents): remove client-side feature_context, keep skill_names only feature_context is now inferred server-side by the donut-backend LLM proxy from the messages array. Remove the client-side inference logic and rename the wrapper to createSkillNamesWrapper since that is its only remaining responsibility. Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> --------- Co-authored-by: Regison <[email protected]> Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]>

* feat: add triggerSource to GetReplyOptions Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * feat: use triggerSource from opts in agent runner trigger assignment Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * feat: read x-trigger-source header and attach to webhook update body Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * feat: extract _triggerSource from grammy update in message context Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * feat: thread triggerSource from context into replyOptions Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> --------- Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]>

buildSyntheticContext() creates a new object with only {message, me, getFile}, discarding ctx.update. So _triggerSource on the update root was always lost. Fix by attaching it to body.value.message and reading from primaryCtx.message in the context builder.

…jection (#10) The `createSkillNamesWrapper` was injecting `skill_names` into the LLM request body via `onPayload`. When the gateway sends requests directly to Anthropic (bypassing the LLM proxy), Anthropic rejects the unknown body field with "Extra inputs are not permitted". Switch to injecting via `X-Skill-Names` HTTP header instead, which is safely ignored by strict APIs. The donut-backend LLM proxy will be updated to read from this header (with body field as legacy fallback). Co-authored-by: Claude Opus 4.6 <[email protected]>

…_id wrappers (#11) Same issue as skill_names — strict APIs (Anthropic) reject unknown body fields. Both wrappers already inject via HTTP headers (X-Trigger-Source, X-Turn-Id), so the redundant onPayload body injection is removed. Co-authored-by: Claude Opus 4.6 <[email protected]>

…sion (TES-642) (#12) Cherry-pick upstream 7f2708a — fixes the core issue where messagingToolSentTargets with missing provider/to params caused shouldSuppressMessagingToolReplies to return false, leaking all assistant text blocks to Telegram. Key changes: - resolveTargetProviderForComparison: falls back to current provider when target provider is empty or "message" - targetsMatchForSuppression: Telegram-aware chatId+threadId matching - mergeDeliveryContext: prevents cross-channel route field inheritance - chat.ts: only inherit delivery routes for channel-scoped sessions Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]>

) * fix: suppress block reply for text+toolCall mixed messages (TES-642) When an assistant message contains both text and tool call blocks, the text is intermediate reasoning (e.g. "Let me read the skill...") that should NOT be forwarded to the user. Real user-facing replies go through the messaging tool explicitly. Add hasToolCall check in handleMessageEnd to skip block reply emission when the message content includes toolCall or tool_use blocks. Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * fix: track message tool sends without explicit target for suppression (TES-642) When the message tool is called without to/target (e.g. Telegram DMs where there's only one chat), extractMessagingToolSend returned undefined — so no target was tracked in messagingToolSentTargets. This caused shouldSuppressMessagingToolReplies to return false (empty targets array), allowing ALL assistant text blocks to leak as final payloads to Telegram. Fix: - extractMessagingToolSend: return { to: "" } instead of undefined when to/target is missing - shouldSuppressMessagingToolReplies: treat empty to as "matches origin target" since it means the message was implicitly sent to the originating conversation Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * fix(telegram): delete draft preview on tool start to prevent reasoning leak (TES-642) When the agent streams intermediate reasoning text ("Let me fetch data..."), the Telegram draft preview shows it as a live-updating message. Previously, these previews were only archived on the next assistant message start and deleted in the finally block — meaning users saw all intermediate thinking during the agent loop. Now when tool execution starts, the onToolStart callback immediately deletes the draft preview message. This is correct because: - Text before a tool call is always intermediate reasoning - The actual user-facing response comes via the message() tool - The draft preview for reasoning text should never persist Also made onToolStart always available (not gated on statusReactionController) so the cleanup works even without reaction support. Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> --------- Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]>

…nchain [codex] Gateway: inject donut runtime analytics IDs

…eded (#17) * feat(heartbeat): add beforeRun gate to skip LLM calls when no work needed Add a `beforeRun` config option that executes a user-supplied script before each heartbeat LLM call. If the script exits non-zero, the heartbeat is skipped with zero token cost. Bypassed for cron, exec-event, and wake-triggered heartbeats to ensure time-sensitive events always run. Made-with: Cursor * fix(heartbeat): use async execFile and improve beforeRun error handling - Replace execFileSync with async execFile to avoid blocking the event loop - Distinguish timeout, ENOENT, and intentional rejection in error logging - Strengthen workspace dir test assertion to verify actual path - Add test for non-existent script path (ENOENT) Made-with: Cursor

* feat(cron): add beforeRun gate to skip LLM for cron jobs Add a `beforeRun` field to CronJobBase that accepts a script path. The script runs before the LLM session — exit 0 proceeds, non-zero skips the entire LLM call (zero token cost). This allows cron jobs to run data scripts deterministically and only invoke the LLM when there is meaningful data to format and deliver. Changes: - types-shared.ts: add optional `beforeRun` field - normalize.ts: trim/validate beforeRun during normalization - service/jobs.ts: handle beforeRun in createJob and applyJobPatch - service/timer.ts: execute beforeRun gate before runIsolatedAgentJob with 120s timeout, tilde expansion, and structured logging Made-with: Cursor * test(cron): add unit tests for beforeRun gate 10 test cases covering: - Gate script exit 0 → LLM proceeds - Gate script exit 1/2 → LLM skipped - Missing script → LLM skipped (logged as rejection) - No beforeRun configured → normal execution - Job id passed as $1 argument - CRON_JOB_ID env var set - Logging: pass/reject/not-found Made-with: Cursor * fix(cron): use err.code instead of err.status for exit code logging execFile error uses `code` (number for exit codes, string for system errors like ENOENT), not `status`. The previous code always logged undefined for the exit code in gate rejection messages. Made-with: Cursor * fix(cron): normalize beforeRun field during store deserialization Add beforeRun normalization in store.ts ensureLoaded loop, consistent with how description and sessionKey are handled. Prevents malformed beforeRun values (whitespace, wrong type) from persisting in the store file. Made-with: Cursor

…tlabs [codex] Report TG tool lifecycle to backend

The cron system fully supports beforeRun (storage, execution, update patching) but CronAddParamsSchema was missing the field. Because additionalProperties is false, passing beforeRun to cron.add was rejected with "unexpected property 'beforeRun'". This forced a workaround of add + update to set the gate, which is error-prone for LLM agents creating dynamic cron jobs (e.g. price alerts). Adding beforeRun to CronCommonOptionalFields makes it available in both add and patch schemas. Co-authored-by: Claude Opus 4.6 <[email protected]>

…nchain feat(observability): emit D0 runtime run traces (PRODUCT-2628)

[codex] add tg tool detail analytics

When the LLM outputs NO_REPLY (a silent control token), the streaming partial "NO" enters the draft buffer. On finalization, stop() flushed this text to Telegram before clear() deleted it, causing a visible flash. Skip stop() when shouldClear=true (no final content delivered) and call clear() directly, which abandons pending content without flushing. Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]>

Block reply messages (sent via deliverReplies) were not tracked after delivery, so onToolStart could not delete them. This caused intermediate reasoning text ("Let me load the skill...") to leak as visible TG messages. - Add onMessageSent callback to deliverReplies/deliverTextReply to track sent message IDs - In onToolStart, delete all pending block reply messages before processing the tool (mirrors existing preview deletion logic) Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]>

When apiRoot is configured (e.g. for TG proxy mode), file downloads still hardcoded https://api.telegram.org, causing 404s with virtual tokens. Now downloadAndSaveTelegramFile respects apiRoot, and SSRF policy dynamically includes the proxy hostname. Co-authored-by: Claude Opus 4.6 <[email protected]>

Telegram webhook channels suffered two compounding bugs: 1. The health monitor's stale-socket heuristic falsely triggered on webhook-mode channels every 30 minutes because webhook channels receive messages via HTTP POST, not socket events. 2. The webhook shutdown() was fire-and-forget — server.close() and bot.stop() were not awaited, so a health-monitor restart could call startChannel before the old server released its port, causing EADDRINUSE or a broken message pipeline (HTTP 200 but zero dispatch). Changes: - channel-health-policy: add mode to snapshot, skip stale-socket for webhook-mode channels - webhook.ts: make shutdown await server.close() and bot.stop(), expose shutdownComplete() promise - monitor.ts: await shutdownComplete() after abort in webhook path Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Allow callers to specify role ("assistant" | "user") for injected messages. Defaults to "assistant" for backward compatibility. When role=user, assistant-specific fields (stopReason, usage, api, provider, model) are omitted from the message body. This enables the backend to inject user-role messages that the bot treats as instructions to execute (e.g. refreshing tool knowledge after d0-cli hot-update). Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

greptile-apps · 2026-04-04T05:58:56Z

Greptile Summary

Adds an optional role parameter ("assistant" | "user", defaulting to "assistant") to the chat.inject RPC, threading it through the handler into appendInjectedAssistantMessageToTranscript, which correctly omits assistant-specific metadata fields (stopReason, usage, api, provider, model) for user-role messages. The change is backward compatible and consistent with existing RPC schema patterns.

Confidence Score: 5/5

Safe to merge — focused, backward-compatible addition with correct conditional logic for role-specific metadata.

The only finding is a P2 function naming issue; the logic, schema, and metadata-gating are all correct.

No files require special attention.

Comments Outside Diff (1)

src/gateway/server-methods/chat-transcript-inject.ts, line 18 (link)

Misleading function name after role expansion

appendInjectedAssistantMessageToTranscript now handles both assistant and user roles, so "Assistant" in the name is no longer accurate. Consider renaming to something like appendInjectedMessageToTranscript to avoid confusion for future callers who read the name and assume it only writes assistant-role entries.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/gateway/server-methods/chat-transcript-inject.ts
Line: 18

Comment:
**Misleading function name after role expansion**

`appendInjectedAssistantMessageToTranscript` now handles both `assistant` and `user` roles, so "Assistant" in the name is no longer accurate. Consider renaming to something like `appendInjectedMessageToTranscript` to avoid confusion for future callers who read the name and assume it only writes assistant-role entries.

How can I resolve this? If you propose a fix, please make it concise.

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}

Prompt To Fix All With AI

This is a comment left during a code review.
Path: src/gateway/server-methods/chat-transcript-inject.ts
Line: 18

Comment:
**Misleading function name after role expansion**

`appendInjectedAssistantMessageToTranscript` now handles both `assistant` and `user` roles, so "Assistant" in the name is no longer accurate. Consider renaming to something like `appendInjectedMessageToTranscript` to avoid confusion for future callers who read the name and assume it only writes assistant-role entries.

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "feat(gateway): add optional role param t..." | Re-trigger Greptile}

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5998c3aa19

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".

chatgpt-codex-connector · 2026-04-04T06:03:54Z

+      const isChannelScopedSession = sessionPeerShapeCandidates.some((part) =>
+        CHANNEL_SCOPED_SESSION_SHAPES.has(part),


Handle legacy channel keys when inheriting delivery route

This new gate only treats sessions as channel-scoped when one of the next two tokens is direct|dm|group|channel, so legacy keys like agent:main:telegram:12345 (still present in existing stores/tests) will fail canInheritDeliverableRoute even when lastChannel/lastTo are valid. In chat.send, that makes OriginatingChannel fall back to webchat, so follow-up sends from those sessions stop routing back to the original external conversation.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-04-04T06:03:54Z

+    if (!toRaw) {
+      return { tool: toolName, provider, accountId, to: "" };
+    }
    const to = normalizeTargetForProvider(provider, toRaw);
    return to ? { tool: toolName, provider, accountId, to } : undefined;


Preserve threadId when tracking message tool deliveries

The message tool path now supports thread-aware suppression, but this return branch drops threadId entirely, so a send like { action: "send", provider: "telegram", to: "-100123", threadId: "77" } is recorded without topic context. Downstream suppression logic then cannot match topic-targeted sends correctly, which causes duplicate final replies in threaded Telegram conversations.

Useful? React with 👍 / 👎.

arucil and others added 30 commits March 12, 2026 09:32

feat(telegram): add apiRoot config passthrough to grammY

e991d3e

Allow overriding the Telegram Bot API base URL via `channels.telegram.apiRoot`, enabling custom Bot API servers and bot routing proxies. Co-Authored-By: Claude Opus 4.6 <[email protected]>

docs(telegram): document apiRoot config option

cf383a3

Co-Authored-By: Claude Opus 4.6 <[email protected]>

fix(model): propagate custom provider headers to model objects (#27490)

9f37271

Merged via squash. Prepared head SHA: e4183b3 Co-authored-by: Sid-Qin <[email protected]> Co-authored-by: shakkernerd <[email protected]> Reviewed-by: @shakkernerd

fix: resolve TS2345 type error in hasToolCall check (#14)

3f44ec1

Gateway: inject donut runtime analytics IDs

19d3279

Gateway: trim donut analytics payload

542fc92

Merge pull request #16 from DonutLabs-ai/codex/openclaw-donutlabs-mai…

6cd579d

…nchain [codex] Gateway: inject donut runtime analytics IDs

feat(d0): report tg tool lifecycle

8c898c7

Merge pull request #19 from DonutLabs-ai/codex/tg-tool-lifecycle-donu…

ea2ecf8

…tlabs [codex] Report TG tool lifecycle to backend

feat(observability): emit D0 runtime run traces

ac09fbd

Merge pull request #21 from DonutLabs-ai/codex/openclaw-donutlabs-mai…

d391614

…nchain feat(observability): emit D0 runtime run traces (PRODUCT-2628)

add tg tool detail analytics

3075906

Merge pull request #22 from DonutLabs-ai/codex/tg-tool-usage-analytics

3754a21

[codex] add tg tool detail analytics

donut-wenzhang and others added 4 commits March 27, 2026 17:49

donut-wenzhang requested a review from a team as a code owner April 4, 2026 05:56

openclaw-barnacle Bot added docs Improvements or additions to documentation channel: telegram Channel integration: telegram app: web-ui App: web-ui gateway Gateway runtime agents Agent runtime and tooling size: XL labels Apr 4, 2026

donut-wenzhang closed this Apr 4, 2026

chatgpt-codex-connector Bot reviewed Apr 4, 2026

View reviewed changes

donut-wenzhang deleted the feat/chat-inject-role-param branch April 8, 2026 15:00

clawsweeper Bot mentioned this pull request Jun 22, 2026

gateway: support user-role chat.inject with idempotent replays #81513

Draft

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

feat(gateway): add optional role param to chat.inject RPC#60706

feat(gateway): add optional role param to chat.inject RPC#60706
donut-wenzhang wants to merge 34 commits into
openclaw:mainfrom
DonutLabs-ai:feat/chat-inject-role-param

donut-wenzhang commented Apr 4, 2026

Uh oh!

greptile-apps Bot commented Apr 4, 2026 •

edited

Loading

Comments Outside Diff (1)

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

chatgpt-codex-connector Bot Apr 4, 2026

Uh oh!

chatgpt-codex-connector Bot Apr 4, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

		const isChannelScopedSession = sessionPeerShapeCandidates.some((part) =>
		CHANNEL_SCOPED_SESSION_SHAPES.has(part),

Uh oh!

Uh oh!

Conversation

donut-wenzhang commented Apr 4, 2026

Summary

Test plan

Uh oh!

greptile-apps Bot commented Apr 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Greptile Summary

Confidence Score: 5/5

Comments Outside Diff (1)

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Apr 4, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Apr 4, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

greptile-apps Bot commented Apr 4, 2026 •

edited

Loading