fix(sdk): custom agent loop parity for continuations, steering, and subtasks#3936
Merged
Conversation
🦋 Changeset detectedLatest commit: 04c89b9 The changes in this PR will be included in the next version bump. This PR includes changesets to release 25 packages
Not sure what this means? Click here to learn what changesets are. Click here if you're a maintainer who wants to add another changeset to this PR |
Contributor
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Repository UI Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (1)
🚧 Files skipped from review as they are similar to previous changes (1)
📜 Recent review details⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (34)
WalkthroughThis PR addresses three behavioral issues in custom agent loops and chat session handling. The primary fix introduces resume cursor seeding that scans session history before user code attaches listeners, preventing replay of already-answered messages on continuation. Tool subtask execution now threads parent session context so task-backed tools can stream progress to the root chat. Chat capture streaming receives explicit message ID generation to avoid text loss during mid-stream steering. The raw chat session iterator is reordered to seed cursors before stop-signal creation, with cleanup made safe against early termination. A changelog entry documents all three fixes. 🚥 Pre-merge checks | ✅ 3 | ❌ 2❌ Failed checks (1 warning, 1 inconclusive)
✅ Passed checks (3 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Warning There were issues while running some tools. Please review the errors and either fix the tool's configuration or disable the tool if it's a critical failure. 🔧 ESLint
ESLint install timed out. The project may have too many dependencies for the sandbox. Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
… steering, and subtasks
Three fixes for chat.customAgent raw loops and chat.createSession:
Continuation boots replayed already-answered user messages into the
first wait: the .in SSE tail attached (via createStopSignal or any
listener) before a resume cursor existed, so S2 replayed from seq 0.
The custom-agent wrapper and createChatSession's first next() now seed
both manager cursors from the latest turn-complete header before
anything attaches, the same boot logic chat.agent uses. Seeding only
setLastSeqNum after attach (the reverted earlier attempt) does not
work because dispatch is gated on the other cursor.
Steering a hand-rolled loop mid-stream wiped the in-flight assistant
text: pipeChatAndCapture called toUIMessageStream without
generateMessageId, so a prepareStep injection starting a new step
regenerated the assistant id and the frontend replaced the partial
message. It now stamps the server-generated id like chat.agent's pipe.
Task-backed tools (ai.toolExecute) failed from custom agent loops with
"session handle is not initialized" on the child run: the chatId only
threaded from the per-turn context that raw loops never set. It now
falls back to the session handle the customAgent wrapper binds at boot,
so child tasks can stream into the parent's chat with
chat.stream.writer({ target: "root" }).
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
The wire can omit the continuation flag on a run that still has prior turns. The cursor scan doubles as the prior-state probe (a fresh session has no turn-complete on .out and seeds nothing), so run it on every custom-loop boot instead of gating on continuation or attempt number, mirroring the snapshot-exists arm of chat.agent's boot check.
ericallam
force-pushed
the
fix/custom-agent-loop-fixes
branch
from
ce2b2d8 to
174ba12
Compare
@trigger.dev/build
trigger.dev
@trigger.dev/core
@trigger.dev/python
@trigger.dev/react-hooks
@trigger.dev/redis-worker
@trigger.dev/rsc
@trigger.dev/schema-to-json
@trigger.dev/sdk
commit: |
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
…itly Custom-loop subtask tool metadata read the chatId off the Session handle id. That value already equals the external chatId (the handle is opened on payload.chatId), but reading it from a dedicated locals slot set at run boot matches the documented ToolCallExecutionOptions contract directly and decouples from the handle id semantics.
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
The docstring showed a plain task(), but createSession resolves its stop signal and message channels through the session handle that only the chat.customAgent (or chat.agent) wrapper binds at run boot. A plain task() would throw "session handle is not initialized" on the first turn. Aligns the example with the reference project and the docs.
nicktrn
approved these changes
Jun 14, 2026
ericallam
added a commit
that referenced
this pull request
Jun 15, 2026
…locks (#3954) ## Summary The script that generates the changeset release PR description was silently dropping some changelog entries and stripping code examples. In [#3932](#3932), entry [#3937](#3937) was missing entirely from the Improvements list and [#3952](#3952 code block was gone, even though both were present in the raw changeset output. ## Root cause `parsePrBody` parsed the raw changeset body line by line: - The dependency-bump filter matched any entry whose text *began* with a backticked package name, so a real changelog entry like `` `@trigger.dev/sdk` now bundles... `` got thrown out along with the genuine version-bump lines. - Only the first line of each bullet was kept, so fenced code blocks, sub-bullets, and continuation paragraphs were discarded. ## Fix Group each top-level bullet with its indented continuation (code blocks, sub-bullets, paragraphs), dedent it, and re-emit it intact. The dependency filter is now anchored so it only matches lines that are *entirely* a package bump, leaving real entries that merely start with a package name. Verified by replaying #3932's raw body through the script: #3937 returns to the list, #3952's code block is preserved, and #3936's sub-bullets nest correctly under their parent.
ericallam
pushed a commit
that referenced
this pull request
Jun 17, 2026
## Summary 7 improvements. ## Improvements - `@trigger.dev/sdk` now bundles the Trigger.dev agent skills and a curated snapshot of the docs those skills reference. The skills that `trigger skills` installs into your coding agent read this content from node_modules, so the guidance your AI assistant follows is pinned to the SDK version installed in your project and stays current across upgrades instead of going stale until the next reinstall. ([#3937](#3937)) - Running a CLI command like `dev`, `deploy`, `preview`, or `update` before initializing a project no longer crashes with a raw `Cannot find matching package.json` stack trace. The CLI now detects the missing project and points you to `npx trigger.dev@latest init` instead. ([#3929](#3929)) - The agent skills installed by `trigger skills` are now namespaced with a `trigger-` prefix (e.g. `trigger-authoring-tasks`, `trigger-getting-started`) so they don't collide with unrelated skills in your coding agent's skills directory. Adds a `trigger-cost-savings` skill for auditing and reducing compute spend (right-sizing machines, `maxDuration`, batching, debounce), and `@trigger.dev/sdk` now bundles the full Trigger.dev documentation so your agent can read the complete, version-pinned reference directly from node_modules. ([#3970](#3970)) - The run span API response now includes `cachedCost` and `cacheCreationCost` on the `ai` object, alongside the existing `inputCost` / `outputCost` / `totalCost`. `inputCost` reflects only the non-cached input, so these fields let you reconstruct the full cost breakdown for prompt-cached calls. ([#3958](#3958)) - `chat.headStart` now works with the `chat.customAgent` and `chat.createSession` backends, not only `chat.agent`. The warm step-1 response hands over to your loop the same way it does for a managed agent. ([#3963](#3963)) In a `chat.customAgent` loop, consume the handover on turn 0: ```ts const conversation = new chat.MessageAccumulator(); const { isFinal, skipped } = await conversation.consumeHandover({ payload }); if (skipped) return; // warm handler aborted, so exit without a turn if (isFinal) { await chat.writeTurnComplete(); // step 1 is the response, no streamText } else { const result = streamText({ model, messages: conversation.modelMessages, tools }); // Pass originalMessages so the handed-over tool round merges into the // step-1 assistant instead of starting a new message. const response = await chat.pipeAndCapture(result, { originalMessages: conversation.uiMessages, }); if (response) await conversation.addResponse(response); } ``` With `chat.createSession`, the iterator surfaces it as `turn.handover`; call `turn.complete()` with no argument on a final handover. The lower-level `chat.waitForHandover()` and `accumulator.applyHandover()` are also exported for hand-rolled loops. - Cache your chat agent's system prompt with Anthropic prompt caching. `chat.toStreamTextOptions()` now emits the system prompt as a cacheable message when you opt in, so a large, stable system block is billed at cache-read rates on every turn instead of full price. ([#3952](#3952)) ```ts // at the streamText call site (Anthropic sugar) streamText({ ...chat.toStreamTextOptions({ cacheControl: { type: "ephemeral" } }), messages, }); // provider-agnostic equivalent chat.toStreamTextOptions({ systemProviderOptions: { anthropic: { cacheControl: { type: "ephemeral" } } }, }); // or where the prompt is defined chat.prompt.set(SYSTEM_PROMPT, { providerOptions: { anthropic: { cacheControl: { type: "ephemeral" } } }, }); ``` Without an option, `system` stays a plain string. Pairs with a `prepareMessages` cache breakpoint to cache the conversation prefix across turns too. - Three fixes for custom agent loops (`chat.customAgent`, `chat.createSession`, and hand-rolled `MessageAccumulator` loops): ([#3936](#3936)) - Continuation runs no longer replay already-answered user messages into the first turn. The `.in` resume cursor is now seeded before any listener attaches (the same boot logic `chat.agent` uses), so a chat that continues after a cancel, crash, or upgrade only sees genuinely new messages. - Steering a hand-rolled loop mid-stream no longer wipes the in-flight assistant response. `chat.pipeAndCapture` now stamps a server-generated message id on the stream, so a `prepareStep` injection keeps the partial text instead of replacing the message. - Task-backed tools (`ai.toolExecute`) now work from custom agent loops: the parent's session is threaded to the child run, so child tasks can stream progress into the chat with `chat.stream.writer({ target: "root" })` instead of failing with "session handle is not initialized". <details> <summary>Raw changeset output</summary>⚠️ ⚠️ ⚠️ ⚠️ ⚠️ ⚠️ `main` is currently in **pre mode** so this branch has prereleases rather than normal releases. If you want to exit prereleases, run `changeset pre exit` on `main`.⚠️ ⚠️ ⚠️ ⚠️ ⚠️ ⚠️ # Releases ## @trigger.dev/build@4.5.0-rc.7 ### Patch Changes - Updated dependencies: - `@trigger.dev/core@4.5.0-rc.7` ## trigger.dev@4.5.0-rc.7 ### Patch Changes - `@trigger.dev/sdk` now bundles the Trigger.dev agent skills and a curated snapshot of the docs those skills reference. The skills that `trigger skills` installs into your coding agent read this content from node_modules, so the guidance your AI assistant follows is pinned to the SDK version installed in your project and stays current across upgrades instead of going stale until the next reinstall. ([#3937](#3937)) - Running a CLI command like `dev`, `deploy`, `preview`, or `update` before initializing a project no longer crashes with a raw `Cannot find matching package.json` stack trace. The CLI now detects the missing project and points you to `npx trigger.dev@latest init` instead. ([#3929](#3929)) - The agent skills installed by `trigger skills` are now namespaced with a `trigger-` prefix (e.g. `trigger-authoring-tasks`, `trigger-getting-started`) so they don't collide with unrelated skills in your coding agent's skills directory. Adds a `trigger-cost-savings` skill for auditing and reducing compute spend (right-sizing machines, `maxDuration`, batching, debounce), and `@trigger.dev/sdk` now bundles the full Trigger.dev documentation so your agent can read the complete, version-pinned reference directly from node_modules. ([#3970](#3970)) - Updated dependencies: - `@trigger.dev/core@4.5.0-rc.7` - `@trigger.dev/build@4.5.0-rc.7` - `@trigger.dev/schema-to-json@4.5.0-rc.7` ## @trigger.dev/core@4.5.0-rc.7 ### Patch Changes - The run span API response now includes `cachedCost` and `cacheCreationCost` on the `ai` object, alongside the existing `inputCost` / `outputCost` / `totalCost`. `inputCost` reflects only the non-cached input, so these fields let you reconstruct the full cost breakdown for prompt-cached calls. ([#3958](#3958)) ## @trigger.dev/python@4.5.0-rc.7 ### Patch Changes - Updated dependencies: - `@trigger.dev/sdk@4.5.0-rc.7` - `@trigger.dev/core@4.5.0-rc.7` - `@trigger.dev/build@4.5.0-rc.7` ## @trigger.dev/react-hooks@4.5.0-rc.7 ### Patch Changes - Updated dependencies: - `@trigger.dev/core@4.5.0-rc.7` ## @trigger.dev/redis-worker@4.5.0-rc.7 ### Patch Changes - Updated dependencies: - `@trigger.dev/core@4.5.0-rc.7` ## @trigger.dev/rsc@4.5.0-rc.7 ### Patch Changes - Updated dependencies: - `@trigger.dev/core@4.5.0-rc.7` ## @trigger.dev/schema-to-json@4.5.0-rc.7 ### Patch Changes - Updated dependencies: - `@trigger.dev/core@4.5.0-rc.7` ## @trigger.dev/sdk@4.5.0-rc.7 ### Patch Changes - `@trigger.dev/sdk` now bundles the Trigger.dev agent skills and a curated snapshot of the docs those skills reference. The skills that `trigger skills` installs into your coding agent read this content from node_modules, so the guidance your AI assistant follows is pinned to the SDK version installed in your project and stays current across upgrades instead of going stale until the next reinstall. ([#3937](#3937)) - `chat.headStart` now works with the `chat.customAgent` and `chat.createSession` backends, not only `chat.agent`. The warm step-1 response hands over to your loop the same way it does for a managed agent. ([#3963](#3963)) In a `chat.customAgent` loop, consume the handover on turn 0: ```ts const conversation = new chat.MessageAccumulator(); const { isFinal, skipped } = await conversation.consumeHandover({ payload }); if (skipped) return; // warm handler aborted, so exit without a turn if (isFinal) { await chat.writeTurnComplete(); // step 1 is the response, no streamText } else { const result = streamText({ model, messages: conversation.modelMessages, tools }); // Pass originalMessages so the handed-over tool round merges into the // step-1 assistant instead of starting a new message. const response = await chat.pipeAndCapture(result, { originalMessages: conversation.uiMessages, }); if (response) await conversation.addResponse(response); } ``` With `chat.createSession`, the iterator surfaces it as `turn.handover`; call `turn.complete()` with no argument on a final handover. The lower-level `chat.waitForHandover()` and `accumulator.applyHandover()` are also exported for hand-rolled loops. - Add `triggerConfig` support to `chat.headStart()` and `chat.openSession()`, so the auto-triggered handover-prepare run inherits tags, queue, machine, and other session trigger options the same way `chat.createStartSessionAction()` does. The `chat:{chatId}` tag is prepended automatically. ([#3963](#3963)) ```ts export const POST = chat.headStart({ agentId: "my-agent", triggerConfig: { tags: ["org:acme"], queue: "chat" }, run: async ({ chat }) => streamText({ ...chat.toStreamTextOptions(), model }), }); ``` Because the session is created once on the first head-start turn and is idempotent on the chat id, this is the only place to set those options for a head-start chat's lifetime. `chat.createStartSessionAction()` now also forwards `maxDuration`, `region`, and `lockToVersion` so both session entry points stay consistent. - Cache your chat agent's system prompt with Anthropic prompt caching. `chat.toStreamTextOptions()` now emits the system prompt as a cacheable message when you opt in, so a large, stable system block is billed at cache-read rates on every turn instead of full price. ([#3952](#3952)) ```ts // at the streamText call site (Anthropic sugar) streamText({ ...chat.toStreamTextOptions({ cacheControl: { type: "ephemeral" } }), messages, }); // provider-agnostic equivalent chat.toStreamTextOptions({ systemProviderOptions: { anthropic: { cacheControl: { type: "ephemeral" } } }, }); // or where the prompt is defined chat.prompt.set(SYSTEM_PROMPT, { providerOptions: { anthropic: { cacheControl: { type: "ephemeral" } } }, }); ``` Without an option, `system` stays a plain string. Pairs with a `prepareMessages` cache breakpoint to cache the conversation prefix across turns too. - Three fixes for custom agent loops (`chat.customAgent`, `chat.createSession`, and hand-rolled `MessageAccumulator` loops): ([#3936](#3936)) - Continuation runs no longer replay already-answered user messages into the first turn. The `.in` resume cursor is now seeded before any listener attaches (the same boot logic `chat.agent` uses), so a chat that continues after a cancel, crash, or upgrade only sees genuinely new messages. - Steering a hand-rolled loop mid-stream no longer wipes the in-flight assistant response. `chat.pipeAndCapture` now stamps a server-generated message id on the stream, so a `prepareStep` injection keeps the partial text instead of replacing the message. - Task-backed tools (`ai.toolExecute`) now work from custom agent loops: the parent's session is threaded to the child run, so child tasks can stream progress into the chat with `chat.stream.writer({ target: "root" })` instead of failing with "session handle is not initialized". - The agent skills installed by `trigger skills` are now namespaced with a `trigger-` prefix (e.g. `trigger-authoring-tasks`, `trigger-getting-started`) so they don't collide with unrelated skills in your coding agent's skills directory. Adds a `trigger-cost-savings` skill for auditing and reducing compute spend (right-sizing machines, `maxDuration`, batching, debounce), and `@trigger.dev/sdk` now bundles the full Trigger.dev documentation so your agent can read the complete, version-pinned reference directly from node_modules. ([#3970](#3970)) - Updated dependencies: - `@trigger.dev/core@4.5.0-rc.7` ## @trigger.dev/plugins@4.5.0-rc.7 ### Patch Changes - Updated dependencies: - `@trigger.dev/core@4.5.0-rc.7` </details> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Three fixes that bring custom agent loops (
chat.customAgenthand-rolled loops andchat.createSession) up to the behaviorchat.agentusers already get, and that the docs already promise:.inresume cursor is now seeded before any listener attaches, using the same boot logic aschat.agent.chat.pipeAndCapture(also backingturn.complete()) streamed without a server-generated message id, so aprepareStepinjection regenerated the assistant id mid-stream and the frontend replaced the partial message, discarding everything streamed before the injection.ai.toolExecutefailed with "chat.agent session handle is not initialized" because the parent's chatId only threaded from the per-turn context that hand-rolled loops never set. It now falls back to the session handle thechat.customAgentwrapper binds at run boot, so children can stream progress into the chat withchat.stream.writer({ target: "root" })(the documented sub-agent pattern).Root cause on the replay fix
Attaching any
.inlistener (chat.createStopSignal,chat.messages.on, the first wait) opens the SSE tail withLast-Event-IDtaken from the seq cursor at attach time. Custom loops attached before any cursor existed, so S2 replayed from seq 0. The fix resolves the cursor from the latest turn-complete header and seeds both manager cursors (setLastSeqNumdrives the SSE resume point,setLastDispatchedSeqNumgates waiter dispatch) before attach;chat.createSessionnow creates its stop signal lazily on the first iteration, after the seed. Seeding only the first cursor after attach does not work, which is why the earlier attempt at this was reverted.All three were reproduced red-green against the references ai-chat project: the replay repro showed the continuation wait consuming a stale message in 403ms with the real message arriving via steering injection; post-fix the wait consumes the real message directly with no injection. Steering now preserves the full in-flight response, and the deepResearch sub-agent streams its progress parts into a raw-loop parent. Existing behavior verified unchanged: full SDK unit suite,
chat.agentsteering, and stop-then-continue onchat.createSession.