fix(agent): bound structured output generation

[BillLeoutsakosvl346]

Summary

• default Agent structured-output requests to 4,096 output tokens when Max Output Tokens is blank
• preserve explicit limits and leave unstructured requests on existing provider/model defaults
• fail explicitly token-limited structured responses with the non-retryable structured_output_token_limit error
• propagate Anthropic no-tool stream stop_reason into existing timing segments
• retain failed-generation content, timing, token, model, and cost diagnostics without exposing a successful downstream output

Incident / root cause

The affected Agent calls had no configured output cap, so Anthropic could continue generating until its provider/model limit. Existing transport deadlines govern connection/body transport rather than intentionally long, continuously progressing generations; there were no Sim retries, restarts, or timeout failures to stop these runs. Malformed structured output then fell through the compatibility fallback as a tiny effective answer.

Validation

• bun run test -- executor/handlers/agent/agent-handler.test.ts providers/anthropic/utils.test.ts executor/execution/block-executor.test.ts executor/execution/block-retry.test.ts — 158 passed
• bun run type-check — passed
• Biome check on the 10 changed files — passed
• frozen lockfile unchanged; no generated files or environment artifacts changed

Live in-app Browser validation was not run because this environment exposed no iab Browser target (only a Chrome extension target). The requester explicitly waived that gate before PR creation.

Intentionally deferred

This PR adds no universal wall-clock, time-to-first-token, or inter-token idle timeout. Legitimate reasoning and research calls can run for minutes; broader deadline policy should be designed separately.

TEAgent's proposed explicit 1,024-token workflow setting remains a separate workflow configuration change and is not included here.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	操作	Updated (UTC)
docs	Skipped		Aug 29, 2026 7:22pm

[greptile-apps]

Greptile Summary

This PR bounds structured Agent output by default, identifies token-limited structured generations as explicit failures, and preserves their provider diagnostics without exposing successful downstream output.

• Defaults structured responses to 4,096 output tokens while preserving explicit and unstructured behavior.
• Propagates provider finish reasons through streaming events and timing segments.
• Retains redacted content, model, token, timing, tool-call, and cost diagnostics for failed generations.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
apps/sim/executor/execution/block-executor.ts	Detects token-limited structured streams before downstream completion and safely retains scrubbed provider diagnostics on failure.
apps/sim/executor/handlers/agent/agent-handler.ts	Applies the structured-output token default and rejects token-limited non-streaming responses.
apps/sim/executor/handlers/shared/response-format.ts	Defines the shared default, finish-reason detection, and non-retryable structured-output limit error.
apps/sim/providers/stream-pump.ts	Carries terminal provider finish reasons through the stream-drain result.
apps/sim/providers/anthropic/utils.ts	Preserves Anthropic stop reasons when completing no-tool streams.

Sequence Diagram

sequenceDiagram
  participant A as Agent handler
  participant P as Model provider
  participant S as Stream pump
  participant E as Block executor
  participant W as Workflow state

  A->>P: Structured request with explicit/default token limit
  P-->>S: Text deltas and terminal finish reason
  S-->>E: Final text and finish reason
  E->>E: Attach finish reason to provider timing
  alt Token-limited structured output
    E->>E: Build and redact failure diagnostics
    E->>W: Record failed block output
  else Complete structured output
    E->>E: Parse structured response
    E->>W: Record successful block output
  end

Loading

_{Reviews (2): Last reviewed commit: "fix(agent): harden structured stream gua..." | Re-trigger Greptile}

[cubic-dev-ai]

All reported issues were addressed across 10 files

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>

[BillLeoutsakosvl346]

@cubic-dev-ai review this PR

[cubic-dev-ai]

@cubic-dev-ai review this PR

@BillLeoutsakosvl346 I have started the AI code review. It will take a few minutes to complete.

[BillLeoutsakosvl346]

@greptileai

[cubic-dev-ai]

No issues found across 22 files

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.

_{Re-trigger cubic}