@@ -0,0 +1,68 @@
+import { FREEBUFF_PROJECTS_SUBAGENT_ID } from '@codebuff/common/constants/free-agents'
+
+import { publisher } from './constants'
+
+import type { SecretAgentDefinition } from './types/secret-agent-definition'
+
+/**
+ * The subagent a Freebuff Projects account's threads may spawn (the
+ * Projects allowlist only; the completions route refuses the id for anyone
+ * else). Unlike fast mode's worker it is a full agent: it has the root's
+ * coding tools, terminal included, and may spawn more of itself, so a thread
+ * can fan out as wide and as deep as the work calls for.
+ *
+ * It runs the root's model with `inheritParentSystemPrompt` and
+ * `includeMessageHistory`, like the fast-mode worker (agents/base3-fast.ts):
+ * the request is the root's prefix plus the assignment, a prompt-cache hit on
+ * a prefix-cached provider, and the session gate accepts it because the model
+ * is the session's own.
+ *
+ * There is no count cap. What bounds a runaway tree is unchanged: every agent
+ * has its step budget, and the server refuses a run more than
+ * MAX_ANCESTOR_RUN_IDS (32) levels deep (web/src/app/api/v1/agent-runs/_post.ts).
+ */
+export function createProjectsSubagent(options: {
+  model: SecretAgentDefinition['model']
+  /** The tools the subagent may run: the root's own, minus any that address
+   *  the human. `spawn_agents` is added here. */
+  toolNames: readonly string[]
+  /** Other agents it may spawn besides itself (they must be in the run's
+   *  agent definitions). */
+  extraSpawnableAgents?: readonly string[]
+}): SecretAgentDefinition {
+  return {
+    id: FREEBUFF_PROJECTS_SUBAGENT_ID,
+    publisher,
+    model: options.model,
+    displayName: 'Buffy Subagent',
+    spawnerPrompt:
+      'A full coding agent that takes one piece of the current task and works it to the end in parallel with any siblings: it reads, edits, runs commands and tests, and can spawn subagents of its own. It sees this whole conversation and inherits your system prompt, so brief it rather than re-explaining. Spawn as many as the work splits into. Its last message is its report.',
+    inputSchema: {
+      prompt: {
+        type: 'string',
+        description:
+          'The assignment: what this subagent owns (files, area, question), what it must not touch because a sibling owns it, and what to report back. Brief: it already has the context above.',
+      },
+    },
+    outputMode: 'last_message',
+    includeMessageHistory: true,
+    inheritParentSystemPrompt: true,
+    windowedFileReads: true,
+    toolNames: Array.from(
+      new Set([...options.toolNames, 'spawn_agents']),
+    ) as SecretAgentDefinition['toolNames'],
+    spawnableAgents: Array.from(
+      new Set([
+        FREEBUFF_PROJECTS_SUBAGENT_ID,
+        ...(options.extraSpawnableAgents ?? []),
+      ]),
+    ),
+    instructionsPrompt: `You are a SUBAGENT spawned by the agent whose conversation appears above, possibly one of several running in parallel. The plan to spawn you has already happened: you are that spawn.
+
+- Do your assignment and only that. Stay inside what it says you own; if the work needs a change a sibling owns, leave it and say so in your report.
+- Do not re-explore what the conversation above already established.
+- If your piece itself splits into independent parts, you may spawn subagents of your own for them, in one spawn_agents call.
+- Verify what you changed when the assignment calls for it. Never commit, push or open a PR unless the assignment says to.
+- Your last message is your report: what you did, every file you changed, and anything left undone or worth checking, in a few sentences.`,
+  }
+}