PR #72666: fix(gateway): strip trailing assistant messages before context-engine assembly returns

• Repository: openclaw/openclaw
• Author: Bojun-Vvibe
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/72666

Description (problem / solution / changelog)

Fixes #72556.

Problem

`installContextEngineLoopHook` (the per-iteration `afterTurn` + `assemble` wrapper for sessions where the context engine owns compaction) returns whatever the engine's `assemble` produced, falling back to the raw source messages on three paths: short-circuit (`hasNewMessages === false`), engine-failure (caught and swallowed), and assemble-returned-same-reference. None of those paths previously guarded against the transcript ending on an assistant turn.

When that view reaches the Anthropic SDK, the API hard-rejects with:

``` This model does not support assistant message prefill. The conversation must end with a user message. ```

The same defect breaks heartbeat injections, system-event bridge calls, and any session whose last persisted turn happens to be assistant. The reference-equality check (`assembled.messages !== sourceMessages`) the existing code uses to detect "no transform" has the additional sharp edge that plugins which short-circuit by returning the source array unchanged are silently treated as no-op — passing the raw, possibly-assistant-ending transcript straight through.

Fix

Add a small provider-agnostic helper `stripTrailingAssistantMessages` and apply it on all three return paths of the loop hook:

```ts export function stripTrailingAssistantMessages(messages: AgentMessage[]): AgentMessage[] { let lastNonAssistant = messages.length - 1; while (lastNonAssistant >= 0 && messages[lastNonAssistant]?.role === "assistant") { lastNonAssistant -= 1; } if (lastNonAssistant === messages.length - 1) { return messages; } return messages.slice(0, lastNonAssistant + 1); } ```

The helper preserves reference-equality on the no-trim fast path so downstream identity assumptions (`assembled.messages !== sourceMessages`) keep working unchanged. Stripping is provider-agnostic — Anthropic is the loudest case but other providers may also misbehave.

Scope is intentionally narrow: only the context-engine loop hook is touched. `installToolResultContextGuard` is left alone (its outputs always end on the inbound message which is whatever the agent loop just appended, not a separately-assembled view).

Tests

6 new unit tests for the helper:

• Fast path returns input reference when last is non-assistant
• Single trailing assistant dropped
• Multiple consecutive trailing assistants dropped (matches the #72556 "raw assistant prefill" reproducer)
• All-assistant input → empty array
• Interleaved (preserves middle assistants, drops only trailing run)
• Empty input returned unchanged

All 29 pre-existing tests in the file continue to pass (35/35 total). `pnpm tsgo:test` clean.

Diff

+89 / -3 across 1 production file + 1 test file + CHANGELOG (92 lines).

Changed files

• src/agents/pi-embedded-runner/tool-result-context-guard.test.ts (modified, +48/-0)
• src/agents/pi-embedded-runner/tool-result-context-guard.ts (modified, +40/-3)