PR #74476: fix(providers): guard anthropic-messages reader against empty-content/token-mismatch turns

• Repository: openclaw/openclaw
• Author: hclsys
• State: closed | merged: False
• Link: https://github.com/openclaw/openclaw/pull/74476

Description (problem / solution / changelog)

Problem

When the anthropic-messages SSE stream reports output_tokens > 0 in message_delta but no content_block_start events were processed (stream format mismatch, or a proxy like meridian using a subtly different SSE format), the stream reader silently produced:

content: []
usage.output: 0
stopReason: "stop"

This empty-content turn triggered resolveEmptyResponseRetryInstruction → fallback cascade:

1. Retry with meridian/anthropic (same reader, same empty result)
2. Retry to codex (400s on the broken-payload JSONL state)
3. ~$44 of cache_write thrash overnight (confirmed in issue #74410)

Fix

Add a guard between the stream loop and finalizeTransportStream:

if (output.usage.output > 0 && output.content.length === 0) {
  throw new Error(
    `anthropic-messages: ${output.usage.output} output token(s) reported but no content blocks parsed — stream format mismatch`,
  );
}

When triggered, failTransportStream sets stopReason: "error". The "error" stop reason short-circuits isEmptyResponseAssistantTurn (returns false at the stopReason === "error" guard), breaking the cascade.

Tests

Two new tests in src/agents/anthropic-transport-stream.test.ts:

1. [thinking, text] shaped response (#74410 regression): verifies both the thinking block and text block are preserved in result.content with correct text/signature values and usage.output = 9.

2. Guard test: verifies that when message_delta reports output_tokens: 9 but no content_block_start events were sent, the result has stopReason: "error" with a descriptive mismatch message.

Acceptance criteria

• pnpm test src/agents/anthropic-transport-stream.test.ts — 44/44 pass
• pnpm test src/agents/pi-embedded-runner/run.incomplete-turn.test.ts — 166/166 pass
• npx oxlint --threads=1 src/agents/anthropic-transport-stream.ts src/agents/anthropic-transport-stream.test.ts — 0 warnings, 0 errors

Closes #74410

Changed files

• CHANGELOG.md (modified, +156/-1)
• extensions/telegram/src/action-runtime.ts (modified, +16/-5)
• extensions/telegram/src/send.test.ts (modified, +63/-0)
• extensions/telegram/src/send.ts (modified, +24/-2)
• src/agents/anthropic-transport-stream.test.ts (modified, +71/-0)
• src/agents/anthropic-transport-stream.ts (modified, +5/-0)

PR #74557: fix(agents): preserve seeded Anthropic text blocks

• Repository: openclaw/openclaw
• Author: vyctorbrzezowski
• State: closed | merged: True
• Link: https://github.com/openclaw/openclaw/pull/74557

Description (problem / solution / changelog)

Summary

• Problem: Anthropic-compatible anthropic-messages streams can seed text/thinking content on content_block_start; the transport only applied later deltas, so a [thinking, text] response could persist empty blocks.
• Why it matters: Meridian/Anthropic responses with visible text could be treated as empty assistant turns, triggering empty-response fallback behavior despite upstream output.
• What changed: Seeded text and thinking content is now copied into the mutable assistant output and emitted as matching stream deltas; signature_delta replaces the seeded thinking signature instead of concatenating it.
• What did NOT change (scope boundary): No fallback policy, replay policy, provider selection, or non-Anthropic transport behavior changed.

AI-assisted: yes.

Change Type (select all)

• Bug fix
• Feature
• Refactor required for the fix
• Docs
• Security hardening
• Chore/infra

Scope (select all touched areas)

• Gateway / orchestration
• Skills / tool execution
• Auth / tokens
• Memory / storage
• Integrations
• API / contracts
• UI / DX
• CI/CD / infra

Linked Issue/PR

• Closes #74410
• This PR fixes a bug or regression

Root Cause (if applicable)

• Root cause: The Anthropic stream parser initialized text and thinking blocks with empty strings on content_block_start and only appended later content_block_delta payloads. Some compatible streams include non-empty content in the start block itself.
• Missing detection / guardrail: There was no unit coverage for a thinking block followed by a text block where the visible text is seeded on content_block_start.
• Contributing context (if known): Meridian and Anthropic-compatible Messages streams can produce [thinking, text] content where the start event carries the initial block payload.

Regression Test Plan (if applicable)

• Coverage level that should have caught this:
  • Unit test
  • Seam / integration test
  • End-to-end test
  • Existing coverage already sufficient
• Target test or file: src/agents/anthropic-transport-stream.test.ts
• Scenario the test should lock in: A Meridian-style Anthropic Messages SSE response starts a thinking block and then a text block with seeded NO_REPLY; the final assistant content and stream events preserve the text.
• Why this is the smallest reliable guardrail: The bug is in the transport parser, so a mocked SSE unit test exercises the exact event shape without live Anthropic/Meridian credentials.
• Existing test that already covers this (if any): N/A
• If no new test is added, why not: N/A

User-visible / Behavior Changes

Anthropic-compatible streamed replies that include visible text in content_block_start are no longer treated as empty assistant turns.

Diagram (if applicable)

Before:
[thinking start with content] -> [text start with content] -> persisted empty blocks

After:
[thinking start with content] -> [text start with content] -> persisted thinking + text

Security Impact (required)

• New permissions/capabilities? (Yes/No) No
• Secrets/tokens handling changed? (Yes/No) No
• New/changed network calls? (Yes/No) No
• Command/tool execution surface changed? (Yes/No) No
• Data access scope changed? (Yes/No) No
• If any Yes, explain risk + mitigation: N/A

Repro + Verification

Environment

• OS: macOS
• Runtime/container: local Node/pnpm checkout
• Model/provider: mocked meridian model with api: "anthropic-messages"
• Integration/channel (if any): N/A
• Relevant config (redacted): N/A

Steps

1. Stream Anthropic Messages SSE events with content_block_start index 0 containing { type: "thinking", thinking: "checking", signature: "sig_1" }.
2. Emit a signature_delta for index 0.
3. Stream content_block_start index 1 containing { type: "text", text: "NO_REPLY" }.
4. Complete the message with non-zero output_tokens.

Expected

• The assistant message contains both the thinking block and { type: "text", text: "NO_REPLY" }.
• The stream emits text_delta and text_end for NO_REPLY.
• The thinking signature is the final signature_delta value.

Actual

• Before this patch, the seeded start-block content was ignored and the assistant content could persist as empty blocks.

Evidence

• Failing test/log before + passing after
• Trace/log snippets
• Screenshot/recording
• Perf numbers (if relevant)

Human Verification (required)

• Verified scenarios: targeted Anthropic transport stream unit test; changed-gate for core/coreTests; local Codex review.
• Edge cases checked: seeded text after thinking; seeded thinking with later signature_delta replacement.
• What you did not verify: Live Anthropic or Meridian API traffic.

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.

Compatibility / Migration

• Backward compatible? (Yes/No) Yes
• Config/env changes? (Yes/No) No
• Migration needed? (Yes/No) No
• If yes, exact upgrade steps: N/A

Risks and Mitigations

• Risk: Some providers may emit both seeded start content and later deltas for the same text/thinking block.
  • Mitigation: The transport appends later deltas after the seeded content, matching Anthropic SDK snapshot behavior for text/thinking accumulation; signature deltas replace the seeded signature.

Changed files

• CHANGELOG.md (modified, +1/-0)
• src/agents/anthropic-transport-stream.test.ts (modified, +76/-0)
• src/agents/anthropic-transport-stream.ts (modified, +33/-6)