Fix fork-mode lifecycle test and default test model to gpt-5.6-luna

The fork test asked the model to call subagent with fork: true and no
agent, which the tool schema has never supported - it only passed when a
model improvised. It now spawns a real fork-mode agent (test-fork with
session-mode: fork).

The harness default model is now openai-codex/gpt-5.6-luna. The previous
active-session-model fallback broke under pi -ne, where extension-registered
providers do not resolve.

test/integration/agents/test-fork.md

@@ -0,0 +1,13 @@
+---
+name: test-fork
+description: Integration test agent — fork-mode child that inherits the parent session
+model: anthropic/claude-haiku-4-5
+tools: read, bash
+session-mode: fork
+auto-exit: true
+disable-model-invocation: true
+---
+
+You are a test agent. Complete the task given to you immediately. Be direct and concise.
+When asked to write content to a file, do it right away using the bash tool.
+Do not ask questions. Do not explain. Just execute the task.

test/integration/harness.ts

@@ -73,13 +73,14 @@ const EXTENSION_SOURCE = join(PROJECT_ROOT, "pi-extension", "subagents", "index.
 
 // ── Configuration ──
 
-/** Model used for integration tests. Override with PI_TEST_MODEL env var. */
-const activePiModel =
-  process.env.PI_PROVIDER && process.env.PI_MODEL
-    ? `${process.env.PI_PROVIDER}/${process.env.PI_MODEL}`
-    : undefined;
-export const TEST_MODEL =
-  process.env.PI_TEST_MODEL ?? activePiModel ?? "anthropic/claude-haiku-4-5";
+/**
+ * Model used for integration tests. Override with PI_TEST_MODEL env var.
+ *
+ * The default must be a provider that exists under `pi -ne`: startPi() launches
+ * with extensions disabled, so extension-registered providers (and therefore
+ * the parent session's active model) may not resolve in the child.
+ */
+export const TEST_MODEL = process.env.PI_TEST_MODEL ?? "openai-codex/gpt-5.6-luna";
 
 /** Per-test timeout in ms. Override with PI_TEST_TIMEOUT env var. */
 export const PI_TIMEOUT = Number(process.env.PI_TEST_TIMEOUT ?? "120000");

test/integration/subagent-lifecycle.test.ts

@@ -1,18 +1,18 @@
 /**
  * Integration tests for the full subagent lifecycle.
  *
- * These tests spawn REAL pi sessions with REAL LLM calls (haiku by default).
+ * These tests spawn REAL pi sessions with REAL LLM calls (gpt-5.6-luna by default).
  * Each test creates a terminal surface, runs pi with a task that uses the subagent
  * tool, and verifies the outcome via marker files and screen output.
  *
- * Costs: ~$0.01-0.05 per test run (haiku).
+ * Costs: ~$0.01-0.05 per test run.
  * Duration: ~30-90s per test.
  *
  * Run inside Herdr, inside an Orca terminal, or inside tmux:
  *   tmux new 'npm run test:integration'
  *
  * Configuration:
- *   PI_TEST_MODEL     — model for all pi sessions (default: anthropic/claude-haiku-4-5)
+ *   PI_TEST_MODEL     — model for all pi sessions (default: openai-codex/gpt-5.6-luna)
  *   PI_TEST_TIMEOUT   — per-test timeout in ms (default: 120000)
  */
 import { describe, it, before, after } from "node:test";
@@ -253,9 +253,9 @@ for (const backend of backends) {
       const task = [
         `Call the subagent tool with these EXACT parameters:`,
         `  name: "Fork-${id}"`,
-        `  fork: true`,
+        `  agent: "test-fork"`,
         `  task: "Run this bash command: echo 'FORK_OK_${id}' > '${markerFile}'"`,
-        `Do not set the agent parameter. Just set name, fork, and task.`,
+        `Do not do anything else. Just call the subagent tool once.`,
         `After you receive the result, say FORK_COMPLETE.`,
       ].join("\n");