Beta blocker: openclaw — createAgentSession({ tools: [] }) silently disables all tools in pi 0.68.1

Bug type

Regression (worked before, now fails)

Beta release blocker

Yes

Summary

After upgrading @mariozechner/pi-* packages from 0.67.68 to 0.68.1 (commit 0f4dc42767), all tool calls are silently disabled. The LLM API request body contains no tools field, so the model responds text-only and never invokes any tools (read, write, edit, exec, web_search, etc.).

Root cause

pi-coding-agent 0.68.1 changed createAgentSession({ tools }) from a built-in tool object list (AgentTool[]) to a tool-name allowlist (string[]). OpenClaw's splitSdkTools() always returns builtInTools: [] and passes all tools via customTools. In 0.68.1, the empty array [] is truthy, causing new Set([]) to be created as the allowlist — which rejects every tool name, including those in customTools.

// pi-coding-agent 0.68.1 — agent-session.js:128
this._allowedToolNames = config.allowedToolNames
  ? new Set(config.allowedToolNames)  // new Set([]) — truthy, empty set
  : undefined;

// line 1787
const isAllowedTool = (name) => !allowedToolNames || allowedToolNames.has(name);
// → !Set() || Set().has("read") → false || false → false
// ALL tools filtered out

Steps to reproduce

1. Install openclaw at version 2026.4.22 (includes pi packages 0.68.1)
2. Start any agent session via any channel (QQ Bot, CLI, gateway, etc.)
3. Send a message that would normally trigger tool use, e.g. "list files in current directory" or "read file X"
4. Observe: model responds with plain text describing what it would do, but never actually calls any tool

Expected behavior

Tools are injected into the API request. The model calls tools (read, exec, write, etc.) as needed.

Actual behavior

• The API request body has no tools field
• Model responds text-only, e.g. outputs <exec command="ls"> as literal text instead of calling the exec tool
• All tools silently disabled — no error logged, no warning

Evidence

Screenshot

Model outputs tool XML as text instead of executing:

Code trace

1. splitSdkTools() returns builtInTools: [] → passed to createAgentSession({ tools: [] })
2. SDK creates allowedToolNames = new Set([]) (truthy, empty)
3. isAllowedTool(name) returns false for all tools
4. context.tools is empty when reaching buildOpenAICompletionsParams()
5. if (context.tools) branch is skipped → no tools in request body

Rollback verification

Rolling back to pi-* 0.67.68 restores tool functionality.

Environment

Introduced by

Commit 0f4dc42767 — build: update pi packages to 0.68.1

Changed files:

• package.json: @mariozechner/pi-* 0.67.68 → 0.68.1
• src/agents/pi-tools.ts: codingTools → createCodingTools()
• src/agents/pi-embedded-runner/tool-split.ts: builtInTools: AnyAgentTool[] → string[]

Suggested fix

Option A — OpenClaw side (1 file, 3 lines)

--- a/src/agents/pi-embedded-runner/tool-split.ts
+++ b/src/agents/pi-embedded-runner/tool-split.ts
 export function splitSdkTools(options: { tools: AnyAgentTool[]; sandboxEnabled: boolean }): {
-  builtInTools: string[];
+  builtInTools: string[] | undefined;
   customTools: ReturnType<typeof toToolDefinitions>;
 } {
   const { tools } = options;
   return {
-    builtInTools: [],
+    builtInTools: undefined,
     customTools: toToolDefinitions(tools),
   };
 }

Option B — SDK side (pi-coding-agent)

// agent-session.js:128
this._allowedToolNames = config.allowedToolNames?.length
  ? new Set(config.allowedToolNames)
  : undefined;

Impact and severity

Critical — all tool calls silently disabled for all models, all providers, all channels. No error or warning emitted. Affects every user on version 2026.4.22.

Scope

NOT_ENOUGH_INFO on whether other consumers of createAgentSession are affected — only the embedded runner path (attempt.ts) was traced.