PR #75368: fix(agents): keep blank user entries during session repair instead of dropping them (#75313)

• Repository: openclaw/openclaw
• Author: w-sss
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/75368

Description (problem / solution / changelog)

Summary

Fixes #75313. Session repair previously dropped blank user-role entries entirely, which could leave a session with no user role at all (e.g. [system, assistant, assistant, …]). Strict OpenAI-compatible chat templates (Qwen3.6, mlx-vlm, etc.) reject such message arrays with No user query found in messages. → HTTP 500.

Fix

Instead of { kind: "drop" }, blank user entries are now rewritten with a synthetic "(continue)" text placeholder, preserving the user role in the session transcript.

Changes

• src/agents/session-file-repair.ts: repairUserEntryWithBlankTextContent returns { kind: "rewrite" } with BLANK_USER_FALLBACK_TEXT = "(continue)" instead of { kind: "drop" }
• Report field renamed: droppedBlankUserMessages → rewrittenBlankUserMessages
• Test updates: 10/10 tests pass

Impact

• Affects all users of repairSessionFileIfNeeded (compaction, session resume)
• Prevents HTTP 500 from strict providers when session repair encounters blank user messages
• Related to #75305 (pre-compaction flush sending empty user messages to Anthropic)

Changed files

• CHANGELOG.md (modified, +1/-0)
• src/agents/session-file-repair.test.ts (modified, +69/-9)
• src/agents/session-file-repair.ts (modified, +55/-20)

PR #75384: fix(memory): send flush prompt as user message to prevent Anthropic 400 (#75305)

• Repository: openclaw/openclaw
• Author: hclsys
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/75384

Description (problem / solution / changelog)

Summary

When a session approaches the context limit, the memory flush hook runs a special compaction turn. The flush handler was calling session.prompt("") with an empty string, which causes Anthropic's API to reject the request with a 400 error.

Fix

Send the transcriptPrompt as the user-visible message text instead of an empty string. This makes the flush turn look like a normal user message to the API.

Changes

• src/auto-reply/agent-runner-memory.ts: pass transcriptPrompt as the message body instead of "" when initiating the memory flush turn

Testing

• Existing memory flush tests pass unchanged
• The 400 error no longer occurs when sessions approach context limits

Fixes #75305. Thanks @hclsys.

Changed files

• CHANGELOG.md (modified, +1/-0)
• src/agents/pi-embedded-runner/run/attempt.spawn-workspace.context-engine.test.ts (modified, +3/-1)
• src/agents/pi-embedded-runner/run/attempt.ts (modified, +6/-1)