Incompatible with OpenClaw 2026.4.5: model call never fires after before_prompt_build hook

[fkc0p]

Bug Report

Environment

• OpenViking plugin: 2026.3.25 (installed via ClawHub)
• OpenClaw: 2026.4.5
• Platform: Linux x86_64, Kali

Symptom

After upgrading OpenClaw from 4.2 → 4.5, the OpenViking context-engine plugin causes the agent pipeline to silently stop. Messages are received, OpenViking auto-recall runs and injects memories, but the model call never happens. The typing indicator times out after 2 minutes with no response.

Slash commands (INTERACTION_CREATE) still work because they bypass the context-engine pipeline.

Works on: OpenClaw 2026.4.2, 2026.4.3
Broken on: OpenClaw 2026.4.5

Root Cause Analysis

OpenViking registers both a before_prompt_build hook (for auto-recall memory injection) and a context-engine (via api.registerContextEngine()).

In OpenClaw 4.5, the execution order changed:

1. contextEngine.assemble() runs first — assembles and locks the message array
2. before_prompt_build hook runs second — returns { prependContext: "..." }

The hook's prependContext return conflicts with the already-assembled message array. The runtime detects this inconsistency and silently aborts — the model API call is never made.

Relevant OpenClaw 4.5 Changes

From the changelog:

• assemble() now receives a new prompt field (PR #50848) — legacy engines get a retry without it, but the retry path may conflict with the hook
• before_prompt_build hook result merging logic changed (PR #58375) — higher-priority systemPrompt preservation may affect how prependContext is applied

OpenViking Server Status

OpenViking server itself works fine (HTTP 200, rerank completes, memories returned). The issue is purely in the plugin-side hook/context-engine interaction with OpenClaw 4.5's runtime.

Suggested Fix

When OpenViking is registered as a context-engine in OpenClaw 4.5+:

1. Move auto-recall memory injection logic from the before_prompt_build hook into assemble()
2. Skip the before_prompt_build hook when context-engine mode is active
3. This avoids the dual hook + context-engine conflict

Workaround

Setting contextEngine: "legacy" and disabling OpenViking (enabled: false) restores normal operation. Agents respond normally without OpenViking's semantic recall.

Logs

[plugins] openviking: auto-recall failed: Error: openviking: auto-recall search timeout
[plugins] openviking: injecting 3 memories (~54 tokens, budget=1000)
[gateway] openviking: session message POST ... 
// No [agent] embedded run or model call logs after this
// 2 minutes later:
typing TTL reached (2m); stopping typing indicator

[hubofvalley]

up