fix: handle Claude Agent SDK 0.3.x system messages to stop runtime-warning flood#2872
Conversation
- no-op thinking_tokens to stop flooding work log - map permission_denied to runtime warning with tool + reason - map mirror_error to runtime.error
- new ToolDeniedPayload schema and ProviderRuntimeToolDeniedEvent in contracts - permission_denied system subtype emits tool.denied instead of runtime warning - ingestion maps tool.denied to a distinct error-tone work-log entry
|
Important Review skippedAuto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Repository UI Review profile: CHILL Plan: Pro Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
✨ Finishing Touches🧪 Generate unit tests (beta)
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 |
github-actions
Bot
added
vouch:unvouched
ApprovabilityVerdict: Needs human review Unable to check for correctness in 423af15. You can customize Macroscope's approvability policy. Learn more. |
|
@juliusmarminge -- kind ping to review this PR since the issue has been gaining some traction 🙂 |
3 participants
What Changed
Teach the Claude adapter about the
systemmessage subtypes added in the@anthropic-ai/claude-agent-sdk0.3.xbump, so they stop falling through to thegeneric "Unhandled subtype" runtime warning:
thinking_tokens→ no-op. This is the one the SDK streams continuously while amodel is in its extended-thinking phase, and is responsible for the nonstop
warnings on every prompt.
permission_denied→ emits a new dedicatedtool.deniedprovider-runtime event.mirror_error→ mapped toruntime.error.Adds the
tool.deniedruntime event (contract + ingestion) so a denied tool showsup as a distinct, error-toned work-log entry (tool name + reason) instead of a
generic "Runtime warning".
Why
The
0.3.xSDK introduced newsystemmessage subtypes.ClaudeAdapter'shandleSystemMessageswitch only knows the older set, so anything new hits thedefaultbranch and emits aruntime.warning.thinking_tokensis emittedrepeatedly throughout a model's thinking, so every prompt floods the work log with
"Runtime warning" entries — exactly what's reported in #2870.
Handling the three new subtypes explicitly removes the noise while keeping the
defaultbranch as a catch-all so genuinely-unknown future messages are stillsurfaced.
permission_deniedandmirror_errorare real signals, so they're routedto meaningful events rather than silently dropped.
Fixes #2870
UI Changes
No visual/component changes. Behavioral only: the work log no longer fills with
repeated "Runtime warning" entries on every prompt (the "before" is the screenshot
in #2870), and a denied tool now renders as a distinct error-tone entry.
Checklist
Note
Handle Claude Agent SDK 0.3.x system messages to stop runtime-warning flood
thinking_tokens,permission_denied, andmirror_errorsubtypes to the system message handler in ClaudeAdapter.ts:thinking_tokensis silently ignored,permission_deniedemits a newtool.deniedruntime event, andmirror_erroremits a runtime error instead of an unhandled warning.tool.deniedevent type in providerRuntime.ts with a payload carryingtoolName,toolUseId,reason, andagentId.tool.deniedevents to an error activity in ProviderRuntimeIngestion.ts, surfacing the tool name and denial reason in the activity summary.Macroscope summarized 29784f6.
Note
Low Risk
Focused adapter and contract extension with no auth or persistence changes; main risk is new event type consumers if any strict schema lists omit tool.denied.
Overview
Stops the work-log runtime warning flood from Claude Agent SDK 0.3.x by handling new
systemmessage subtypes inhandleSystemMessageinstead of falling through to the generic unhandled-subtype warning.thinking_tokensis ignored (no runtime event).permission_deniedbecomes a newtool.deniedprovider-runtime event (tool name, optional use id, reason, agent id).mirror_erroris emitted asruntime.error.Contracts add
tool.deniedwith payload schema and union membership. Ingestion maps it to an error-toned thread activity (Tool denied: …) with optional detail, separate from generic runtime warnings.Reviewed by Cursor Bugbot for commit 29784f6. Bugbot is set up for automated code reviews on this repo. Configure here.