Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

Active-memory sub-agent's LLM request includes tools=[] instead of [memory_search, memory_get], preventing the model from calling memory tools.

Steps to reproduce

1. Upgrade from 2026.4.14 to 2026.4.15
2. Enable active-memory plugin with any OpenAI-compatible model
3. Send a message that should trigger memory recall (e.g. "What is my favorite programming language?")
4. Check sub-agent transcript in ~/.openclaw/plugins/active-memory/transcripts/

Observed: transcript shows 6 lines only (no tool_use, no tool_result). Model's thinking says "tool not listed". Strict-validation providers like Qwen return 400: "[] is too short - 'tools'".

In 2026.4.14, the same query produces 8+ line transcripts with memory_search tool_use and tool_result entries.

Expected behavior

In 2026.4.14, the active-memory sub-agent transcript contains tool_use (memory_search) and tool_result entries (8+ lines), and memory recall succeeds.

Actual behavior

In 2026.4.15, the sub-agent transcript contains only 6 lines with no tool_use or tool_result. The model's thinking output says "tool not listed", confirming tools=[] was sent.
Providers with strict validation (e.g. Qwen via OpenCode Go) reject with: 400 [] is too short - 'tools'. The plugin still reports status=ok by misinterpreting the thinking text or error message as the summary (summaryChars=56), masking the failure.

OpenClaw version

2026.4.15 (041266a) — regression from 2026.4.14 (323493f)

Operating system

macOS 14.7 (Sonoma)

Install method

npm global

Model

openrouter/openai/gpt-oss-120b:free, opencode-go/qwen3.5-plus

Provider / routing chain

openclaw -> openrouter (gpt-oss-120b:free), openclaw -> opencode-go (qwen3.5-plus)

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Transcript comparison (same query: "What is my favorite programming language?"):

**2026.4.14 (working)** — 8 lines, tool call present:

L5 thinking: "Need to search memory. Use memory_search."
L5 tool_use: memory_search input={"query":"favorite programming language"}
L6 tool_result: [{path: "...", score: 0.69, snippet: "..."}]
L7 text: "User's favorite programming language is Python."


**2026.4.15 (broken)** — 6 lines, no tool call:

L5 thinking: "Need to search memory. Use memory_search tool? Not listed. Return NONE."
L5 text: "NONE"


**2026.4.15 with Qwen (strict validation)** — explicit 400 error:

L5 content: [], stopReason: "error"
L5 errorMessage: "400 [] is too short - 'tools'"


Note: plugin reports status=ok summaryChars=56 even on failure, misinterpreting thinking
text as summary.

Impact and severity

No response

Additional information

No response