SPEC-3 — cross-harness peers (CC + Pi backends, profile routing, session_key resume)

agents/general-purpose-cc.md

@@ -0,0 +1,11 @@
+---
+name: general-purpose-cc
+description: A focused general-purpose CC subagent. Use for any task needing Claude Code as the worker.
+backend: claude
+todoSync: true
+memoryHydrate: true
+vision: true
+---
+You are a focused subagent delegate running under Claude Code. Complete the assigned task
+thoroughly, work autonomously to completion, and return a concise result summary.
+Do not call the `todo` tool — the fleet engine manages todo tracking for you.

docs/SPEC-3-smoke-checklist.md

@@ -0,0 +1,26 @@
+# SPEC-3 smoke checklist (real-pi, term-driven)
+
+Rows that need no `claude` call are run via `term` inside a real pi session; rows 2-4 are the script.
+
+## How to run
+- Script (rows 2-4): `node --import tsx scripts/spec-3-smoke.mts` (skips CC rows if claude absent)
+- Term rows (1/5/6/7): spawn pi in `~/local-dev/getpipher/armory-fleet`, drive via `term`
+
+## Rows
+| # | Action | Expected |
+|---|---|---|
+| 1 | extension loads with `claude` absent | `/fleet` Backends view shows `claude: ✗ (not installed)`; `pi: ✓` |
+| 2 | `subagent(general-purpose, "reply OK")` (pi) | run completes; armory chip `t✓ m✓ v✓` |
+| 3 | `subagent(general-purpose-cc, "reply OK")` (claude, if available) | run completes via `claude -p`; `backendSessionId` set; chip `t✓ m✓ v~` |
+| 4 | re-spawn `general-purpose-cc` same `sessionKey` | `--resume <id>` passed; CC replays history |
+| 5 | `backend: invalid` profile in `.pi/agents/` | load warning surfaced; profile excluded from registry |
+| 6 | `claude` schema drift (point FLEET_CLAUDE_BIN at a fake) | Backends view shows `schema ✗`; spawn fails fast with actionable error |
+| 7 | Backends view `r:Refresh` + `i:Info` | refresh notifies "restart pi to re-detect"; info shows flag matrix + hook mechanism notes |
+
+## How to inspect the CC invocation
+- The `i:Info` pane on the `claude` backend row shows the flag-support matrix probed at init.
+- Set `DEBUG=fleet:cc` (or equivalent) to log the composed `claude -p` args + the NDJSON events received.
+
+## Pass bar
+- Rows 1, 5, 6, 7 pass (term-driven, no CC call).
+- Rows 2-4 pass when `claude` is installed; skipped (exit 0) otherwise.

scripts/spec-3-smoke.mts

@@ -0,0 +1,59 @@
+// scripts/spec-3-smoke.mts — SPEC-3 full-run smoke (rows 2-4).
+// Exercises the REAL CC backend (spawn a real `claude -p`) when claude is installed; skips cleanly otherwise.
+// Run: node --import tsx scripts/spec-3-smoke.mts
+import { spawnSubagent } from "../src/engine/spawnSubagent.ts";
+import { RunRegistry } from "../src/engine/run-registry.ts";
+import { createSingleSlotLock } from "../src/engine/concurrency-lock.ts";
+import { ArmoryTodoAdapter } from "../src/todo-sync/adapter.ts";
+import { ArmoryMemoryAdapter } from "../src/memory-hydrate/adapter.ts";
+import { BackendRegistry, PI_HOOK_PARITY, CLAUDE_HOOK_PARITY } from "../src/backend/port.ts";
+import { detectClaude } from "../src/backend/claude-detector.ts";
+import { createClaudeChildFactory } from "../src/backend/claude-factory.ts";
+import { ResumeStore } from "../src/backend/resume-store.ts";
+import { ModelRuntime } from "@earendil-works/pi-coding-agent";
+import { createChildSessionFactory } from "../src/index.ts";
+import type { AgentDef } from "../src/registry/frontmatter.ts";
+
+let pass = 0, fail = 0;
+function check(name: string, cond: boolean, detail = ""): void {
+  if (cond) { console.log(`  ✔ ${name}`); pass++; }
+  else { console.log(`  ✖ ${name} ${detail}`); fail++; }
+}
+
+const resumeStore = new ResumeStore();
+const claudeInfo = await detectClaude();
+if (!claudeInfo?.schemaOk) {
+  console.log("⏭  claude not available (not installed or schema drift) — skipping CC rows. Pi row 2 still runs.");
+}
+
+const runtime = await ModelRuntime.create();
+const reg = new BackendRegistry();
+reg.register({ id: "pi", factory: createChildSessionFactory(runtime, new ArmoryMemoryAdapter(), resumeStore), available: () => true, versionInfo: () => null, hookParity: PI_HOOK_PARITY });
+if (claudeInfo) reg.register({ id: "claude", factory: createClaudeChildFactory(claudeInfo, resumeStore), available: () => claudeInfo.schemaOk, versionInfo: () => claudeInfo, hookParity: CLAUDE_HOOK_PARITY });
+
+const piAgent: AgentDef = { name: "general-purpose", description: "d", rolePrompt: "Reply minimally.", todoSync: true, memoryHydrate: true, vision: true, backend: "pi", sessionKey: "general-purpose", source: "builtin", filePath: "/x" };
+const ccAgent: AgentDef = { name: "general-purpose-cc", description: "d", rolePrompt: "Reply minimally.", todoSync: true, memoryHydrate: true, vision: true, backend: "claude", sessionKey: "general-purpose-cc", source: "builtin", filePath: "/x" };
+const registry = new Map<string, AgentDef>([["general-purpose", piAgent], ["general-purpose-cc", ccAgent]]);
+
+// Row 2: pi backend (real Ollama Cloud session.prompt())
+{
+  console.log("Row 2: pi backend spawn");
+  const res = await spawnSubagent({ agent: "general-purpose", task: "Reply with exactly: OK", registry, todoSync: new ArmoryTodoAdapter(), runRegistry: new RunRegistry(), lock: createSingleSlotLock(), backendRegistry: reg, parentModel: { provider: "Ollama", id: "glm-5.2:cloud" }, parentCwd: process.cwd() });
+  check("pi run completes", res.status === "completed", res.error ?? "");
+  check("pi run produced finalText", res.finalText.length > 0);
+}
+
+// Rows 3-4: CC backend + resume (only if claude available)
+if (claudeInfo?.schemaOk) {
+  console.log("Row 3: claude backend spawn");
+  const res = await spawnSubagent({ agent: "general-purpose-cc", task: "Reply with exactly: OK", registry, todoSync: new ArmoryTodoAdapter(), runRegistry: new RunRegistry(), lock: createSingleSlotLock(), backendRegistry: reg, parentModel: { provider: "x", id: "y" }, parentCwd: process.cwd() });
+  check("cc run completes", res.status === "completed", res.error ?? "");
+  console.log("Row 4: claude resume (re-spawn same sessionKey)");
+  const res2 = await spawnSubagent({ agent: "general-purpose-cc", task: "What did I just say?", registry, todoSync: new ArmoryTodoAdapter(), runRegistry: new RunRegistry(), lock: createSingleSlotLock(), backendRegistry: reg, parentModel: { provider: "x", id: "y" }, parentCwd: process.cwd() });
+  check("cc resume run completes", res2.status === "completed", res2.error ?? "");
+} else {
+  console.log("Rows 3-4: skipped (claude unavailable)");
+}
+
+console.log(`\n${pass} pass / ${fail} fail`);
+process.exit(fail === 0 ? 0 : 1);