feat(workflows): inline sub-agent definitions on DAG nodes

CHANGELOG.md

@@ -9,6 +9,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
+- **Inline sub-agent definitions on DAG nodes (`agents:`).** Define Claude Agent SDK `AgentDefinition`s directly in workflow YAML, keyed by kebab-case agent ID. The main agent can spawn them in parallel via the `Task` tool — useful for map-reduce patterns where a cheap model (e.g. Haiku) briefs items and a stronger model reduces. Removes the need to author `.claude/agents/*.md` files for workflow-scoped helpers. Claude only; Codex and community providers that don't support inline agents emit a capability warning and ignore the field. Merges with the internal `dag-node-skills` wrapper set by `skills:` on the same node — user-defined agents win on ID collision (a warning is logged). (#1276)
 - **Pi community provider (`@mariozechner/pi-coding-agent`).** First community provider under the Phase 2 registry (`builtIn: false`). One adapter exposes ~20 LLM backends (Anthropic, OpenAI, Google, Groq, Mistral, Cerebras, xAI, OpenRouter, Hugging Face, and more) via a `<pi-provider-id>/<model-id>` model format. Reads credentials from `~/.pi/agent/auth.json` (populated by running `pi /login` for OAuth subscriptions like Claude Pro/Max, ChatGPT Plus, GitHub Copilot) AND from env vars (env vars take priority per-request). Per-node workflow options supported: `effort`/`thinking` → Pi `thinkingLevel`; `allowed_tools`/`denied_tools` → filter Pi's 7 built-in coding tools; `skills` → resolved against `.agents/skills`, `.claude/skills` (project + user-global); `systemPrompt`; codebase env vars; session resume via `sessionId` round-trip. Unsupported fields (MCP, hooks, structured output, cost limits, fallback model, sandbox) trigger an explicit dag-executor warning rather than silently dropping. Use in workflow YAML: `provider: pi` + `model: anthropic/claude-haiku-4-5`. (#1270)
 - **`registerCommunityProviders()` aggregator** in `@archon/providers`. Process entrypoints (CLI, server, config-loader) now call one function to register every bundled community provider. Adding a new community provider is a single-line edit to this aggregator rather than touching each entrypoint — makes the Phase 2 "community providers are a localized addition" promise real.
 - **`contributing/adding-a-community-provider.md` guide** — contributor-facing walkthrough of the Phase 2 registry pattern using Pi as the reference implementation.

CLAUDE.md

@@ -697,7 +697,7 @@ async function createSession(conversationId: string, codebaseId: string) {
 2. **Workflows** (YAML-based):
    - Stored in `.archon/workflows/` (searched recursively)
    - Multi-step AI execution chains, discovered at runtime
-   - **`nodes:` (DAG format)**: Nodes with explicit `depends_on` edges; independent nodes in the same topological layer run concurrently. Node types: `command:` (named command file), `prompt:` (inline prompt), `bash:` (shell script, stdout captured as `$nodeId.output`, no AI, receives managed per-project env vars in its subprocess environment when configured), `loop:` (iterative AI prompt until completion signal), `approval:` (human gate; pauses until user approves or rejects; `capture_response: true` stores the user's comment as `$<node-id>.output` for downstream nodes, default false), `script:` (inline TypeScript/Python or named script from `.archon/scripts/`, runs via `bun` or `uv`, stdout captured as `$nodeId.output`, no AI, receives managed per-project env vars in its subprocess environment when configured, supports `deps:` for dependency installation and `timeout:` in ms, requires `runtime: bun` or `runtime: uv`) . Supports `when:` conditions, `trigger_rule` join semantics, `$nodeId.output` substitution, `output_format` for structured JSON output (Claude and Codex), `allowed_tools`/`denied_tools` for per-node tool restrictions (Claude only), `hooks` for per-node SDK hook callbacks (Claude only), `mcp` for per-node MCP server config files (Claude only, env vars expanded at execution time), and `skills` for per-node skill preloading via AgentDefinition wrapping (Claude only), and `effort`/`thinking`/`maxBudgetUsd`/`systemPrompt`/`fallbackModel`/`betas`/`sandbox` for Claude SDK advanced options (Claude only, also settable at workflow level)
+   - **`nodes:` (DAG format)**: Nodes with explicit `depends_on` edges; independent nodes in the same topological layer run concurrently. Node types: `command:` (named command file), `prompt:` (inline prompt), `bash:` (shell script, stdout captured as `$nodeId.output`, no AI, receives managed per-project env vars in its subprocess environment when configured), `loop:` (iterative AI prompt until completion signal), `approval:` (human gate; pauses until user approves or rejects; `capture_response: true` stores the user's comment as `$<node-id>.output` for downstream nodes, default false), `script:` (inline TypeScript/Python or named script from `.archon/scripts/`, runs via `bun` or `uv`, stdout captured as `$nodeId.output`, no AI, receives managed per-project env vars in its subprocess environment when configured, supports `deps:` for dependency installation and `timeout:` in ms, requires `runtime: bun` or `runtime: uv`) . Supports `when:` conditions, `trigger_rule` join semantics, `$nodeId.output` substitution, `output_format` for structured JSON output (Claude and Codex), `allowed_tools`/`denied_tools` for per-node tool restrictions (Claude only), `hooks` for per-node SDK hook callbacks (Claude only), `mcp` for per-node MCP server config files (Claude only, env vars expanded at execution time), and `skills` for per-node skill preloading via AgentDefinition wrapping (Claude only), `agents` for inline sub-agent definitions invokable via the Task tool (Claude only), and `effort`/`thinking`/`maxBudgetUsd`/`systemPrompt`/`fallbackModel`/`betas`/`sandbox` for Claude SDK advanced options (Claude only, also settable at workflow level)
    - Provider inherited from `.archon/config.yaml` unless explicitly set; per-node `provider` and `model` overrides supported
    - Model and options can be set per workflow or inherited from config defaults
    - `interactive: true` at the workflow level forces foreground execution on web (required for approval-gate workflows in the web UI)

packages/docs-web/src/content/docs/getting-started/ai-assistants.md

@@ -308,6 +308,7 @@ nodes:
 | Tool restrictions | ✅ | `allowed_tools` / `denied_tools` (read, bash, edit, write, grep, find, ls) |
 | Thinking level | ✅ | `effort: low\|medium\|high\|max` (max → xhigh) |
 | Skills | ✅ | `skills: [name]` (searches `.agents/skills`, `.claude/skills`, user-global) |
+| Inline sub-agents | ❌ | `agents:` is Claude-only; ignored with a warning on Pi |
 | System prompt override | ✅ | `systemPrompt:` |
 | Codebase env vars (`envInjection`) | ✅ | `.archon/config.yaml` `env:` section |
 | MCP servers | ❌ | Pi rejects MCP by design |

packages/docs-web/src/content/docs/guides/authoring-workflows.md

@@ -196,6 +196,7 @@ nodes:
 | `hooks` | object | — | Per-node SDK hook callbacks. Claude only. See [Hooks](/guides/hooks/) |
 | `mcp` | string | — | Path to MCP server config JSON file. Claude only. See [MCP Servers](/guides/mcp-servers/) |
 | `skills` | string[] | — | Skills to preload. Claude only. See [Skills](/guides/skills/) |
+| `agents` | object | — | Inline sub-agent definitions keyed by kebab-case ID. Claude only. See [Inline sub-agents](#inline-sub-agents) |
 | `effort` | `'low'`\|`'medium'`\|`'high'`\|`'max'` | — | Reasoning depth. Claude only. Also settable at workflow level |
 | `thinking` | string \| object | — | Thinking mode: `'adaptive'`, `'disabled'`, or `{type:'enabled', budgetTokens:N}`. Claude only. Also settable at workflow level |
 | `maxBudgetUsd` | number | — | USD cost cap; node fails if exceeded. Claude only. Per-node only |
@@ -404,6 +405,43 @@ nodes:
 - `undefined` (field absent) and `[]` have different semantics — absent means use default tool set, `[]` means no tools
 - Claude only — Codex nodes/steps emit a warning and continue (Codex doesn't support per-call tool restrictions)
 
+### Inline sub-agents
+
+Define Claude sub-agents directly in the workflow YAML, without authoring `.claude/agents/*.md` files. The main agent can spawn them in parallel via the `Task` tool — useful for map-reduce patterns where a cheap model (e.g. Haiku) briefs items and a stronger model reduces.
+
+```yaml
+nodes:
+  - id: triage
+    prompt: |
+      Fetch open issues via `gh issue list ...`. For each issue, spawn the
+      brief-gen sub-agent in parallel (one message, multiple Task tool calls)
+      to produce a 2-3 sentence brief. Then cluster briefs for duplicates.
+    model: sonnet
+    allowed_tools: [Bash, Read, Write, Task]
+    agents:
+      brief-gen:
+        description: Summarises a single GitHub issue in 2-3 sentences
+        prompt: |
+          You are concise. Read the issue provided in the caller's prompt.
+          Return JSON { summary, primarySymptom, affectedArea }.
+        model: haiku
+        tools: [Bash, Read]
+```
+
+Keys:
+
+- Agent IDs must be **kebab-case** (`^[a-z0-9]+(-[a-z0-9]+)*$`)
+- Each definition requires `description` and `prompt`; `model`, `tools`, `disallowedTools`, `skills`, and `maxTurns` are optional
+- Map is merged with any SDK-level agents and with the internal `dag-node-skills` wrapper created by `skills:` — user-defined agents win on ID collision (a warning is logged when this happens)
+- Claude only. Codex and community providers that don't support inline agents emit a warning and ignore the field
+
+**When to use `agents:` vs `.claude/agents/*.md` files:**
+
+- **`agents:` (inline)** — use when the sub-agent is specific to ONE workflow's needs. Keeps the workflow self-contained in a single YAML file; travels cleanly in PRs and forks.
+- **`.claude/agents/*.md` (on-disk)** — use when the sub-agent is shared across multiple workflows OR the whole project (for example, a `triage-agent` used by several maintenance workflows). On-disk agents live outside workflow YAMLs and are picked up automatically by the Claude Agent SDK.
+
+Both sources coexist — inline agents and on-disk agents are both available to `Task(subagent_type=...)` at runtime.
+
 ---
 
 ## Retry Configuration
@@ -1126,10 +1164,11 @@ Before deploying a workflow:
 10. **`hooks`** — attach SDK hook callbacks to Claude nodes for tool control and context injection
 11. **`mcp:`** — attach per-node MCP servers via JSON config (Claude only)
 12. **`skills:`** — preload skills into Claude nodes for domain expertise
-13. **`effort` / `thinking`** — control reasoning depth and thinking mode per node or workflow (Claude only)
-14. **`maxBudgetUsd`** — set a USD cost cap per node; fails with error if exceeded (Claude only)
-15. **`systemPrompt`** — override the default system prompt per node (Claude only)
-16. **`sandbox`** — OS-level filesystem/network restrictions per node or workflow (Claude only)
-17. **Loop nodes** — use `loop:` within a DAG node for iterative execution until completion signal
-18. **Defaults as templates** — browse `.archon/workflows/defaults/` for real examples to copy and modify
-19. **Test thoroughly** — each command, the artifact flow, and edge cases
+13. **`agents:`** — inline Claude sub-agent definitions invokable via the `Task` tool
+14. **`effort` / `thinking`** — control reasoning depth and thinking mode per node or workflow (Claude only)
+15. **`maxBudgetUsd`** — set a USD cost cap per node; fails with error if exceeded (Claude only)
+16. **`systemPrompt`** — override the default system prompt per node (Claude only)
+17. **`sandbox`** — OS-level filesystem/network restrictions per node or workflow (Claude only)
+18. **Loop nodes** — use `loop:` within a DAG node for iterative execution until completion signal
+19. **Defaults as templates** — browse `.archon/workflows/defaults/` for real examples to copy and modify
+20. **Test thoroughly** — each command, the artifact flow, and edge cases

packages/docs-web/src/content/docs/guides/skills.md

@@ -235,6 +235,7 @@ To use skills, ensure the node uses Claude (the default provider, or set
 
 ## Related
 
+- [Inline sub-agents](/guides/authoring-workflows/#inline-sub-agents) — `agents:` field for workflow-scoped sub-agents (composes with `skills:` on the same node; user-defined agents win on ID collision with the internal `dag-node-skills` wrapper)
 - [Per-Node MCP Servers](/guides/mcp-servers/) — `mcp:` field for external tool access
 - [Hooks](/guides/hooks/) — `hooks:` field for tool permission control
 - [skills.sh](https://skills.sh) — marketplace for discovering skills

packages/providers/src/claude/capabilities.ts

@@ -5,6 +5,7 @@ export const CLAUDE_CAPABILITIES: ProviderCapabilities = {
   mcp: true,
   hooks: true,
   skills: true,
+  agents: true,
   toolRestrictions: true,
   structuredOutput: true,
   envInjection: true,

packages/providers/src/claude/provider.test.ts

@@ -97,6 +97,7 @@ describe('ClaudeProvider', () => {
         mcp: true,
         hooks: true,
         skills: true,
+        agents: true,
         toolRestrictions: true,
         structuredOutput: true,
         envInjection: true,
@@ -1165,4 +1166,128 @@ describe('sendQuery decomposition behaviors', () => {
       'claude.result_is_error'
     );
   });
+
+  describe('inline agents (nodeConfig.agents)', () => {
+    test('passes inline agents map through to SDK options.agents', async () => {
+      mockQuery.mockImplementation(async function* () {
+        yield { type: 'result', session_id: 'sid' };
+      });
+
+      const agents = {
+        'brief-gen': {
+          description: 'Summarises issues',
+          prompt: 'Be concise.',
+          model: 'haiku',
+          tools: ['Bash', 'Read'],
+        },
+      };
+
+      for await (const _ of client.sendQuery('test', '/workspace', undefined, {
+        nodeConfig: { agents },
+      })) {
+        // consume
+      }
+
+      expect(mockQuery).toHaveBeenCalledTimes(1);
+      const callArgs = mockQuery.mock.calls[0][0] as { options: Record<string, unknown> };
+      expect(callArgs.options.agents).toMatchObject(agents);
+    });
+
+    test('does not set options.agent when only inline agents are present', async () => {
+      mockQuery.mockImplementation(async function* () {
+        yield { type: 'result', session_id: 'sid' };
+      });
+
+      for await (const _ of client.sendQuery('test', '/workspace', undefined, {
+        nodeConfig: {
+          agents: {
+            'sub-a': { description: 'd', prompt: 'p' },
+          },
+        },
+      })) {
+        // consume
+      }
+
+      const callArgs = mockQuery.mock.calls[0][0] as { options: Record<string, unknown> };
+      // agent (singular) is set by skills wrapper; inline-only must leave it unset
+      expect(callArgs.options.agent).toBeUndefined();
+    });
+
+    test('merges inline agents with skills wrapper; user wins on ID collision', async () => {
+      mockQuery.mockImplementation(async function* () {
+        yield { type: 'result', session_id: 'sid' };
+      });
+
+      for await (const _ of client.sendQuery('test', '/workspace', undefined, {
+        nodeConfig: {
+          skills: ['my-skill'],
+          agents: {
+            // Intentionally collides with the internal 'dag-node-skills' wrapper ID
+            'dag-node-skills': {
+              description: 'user override',
+              prompt: 'user-defined prompt',
+            },
+            'extra-sub': { description: 'd', prompt: 'p' },
+          },
+        },
+      })) {
+        // consume
+      }
+
+      const callArgs = mockQuery.mock.calls[0][0] as { options: Record<string, unknown> };
+      const outAgents = callArgs.options.agents as Record<
+        string,
+        { description: string; prompt: string }
+      >;
+      // Both entries present
+      expect(Object.keys(outAgents).sort()).toEqual(['dag-node-skills', 'extra-sub']);
+      // User's definition wins the collision
+      expect(outAgents['dag-node-skills'].description).toBe('user override');
+      expect(outAgents['dag-node-skills'].prompt).toBe('user-defined prompt');
+    });
+
+    test('logs a warning when user-defined dag-node-skills overrides the skills wrapper', async () => {
+      mockQuery.mockImplementation(async function* () {
+        yield { type: 'result', session_id: 'sid' };
+      });
+
+      for await (const _ of client.sendQuery('test', '/workspace', undefined, {
+        nodeConfig: {
+          skills: ['my-skill'],
+          agents: {
+            'dag-node-skills': { description: 'user override', prompt: 'p' },
+          },
+        },
+      })) {
+        // consume
+      }
+
+      expect(mockLogger.warn).toHaveBeenCalledWith(
+        expect.objectContaining({ nodeSkills: ['my-skill'] }),
+        'claude.inline_agents_override_skills_wrapper'
+      );
+    });
+
+    test('does NOT warn when inline agents do not collide with the skills wrapper', async () => {
+      mockQuery.mockImplementation(async function* () {
+        yield { type: 'result', session_id: 'sid' };
+      });
+
+      for await (const _ of client.sendQuery('test', '/workspace', undefined, {
+        nodeConfig: {
+          skills: ['my-skill'],
+          agents: {
+            'brief-gen': { description: 'd', prompt: 'p' },
+          },
+        },
+      })) {
+        // consume
+      }
+
+      const warnCalls = mockLogger.warn.mock.calls.filter(
+        (args: unknown[]) => args[1] === 'claude.inline_agents_override_skills_wrapper'
+      );
+      expect(warnCalls).toHaveLength(0);
+    });
+  });
 });

packages/providers/src/claude/provider.ts

@@ -450,6 +450,32 @@ async function applyNodeConfig(
     getLog().info({ skills, agentId }, 'claude.skills_agent_created');
   }
 
+  // agents → inline AgentDefinition pass-through.
+  // Runs AFTER skills: so user-defined agents win on ID collision with
+  // the internal 'dag-node-skills' wrapper.
+  // options.agent is intentionally left alone — inline agents are sub-agents
+  // invokable via the Task tool, not the primary agent for the query.
+  if (nodeConfig.agents) {
+    // Warn loudly when a user-defined agent overrides the internal
+    // 'dag-node-skills' wrapper set by the skills: block above. The
+    // merge is by design (user wins) but silent capability removal
+    // is the exact failure mode we want to avoid.
+    if (
+      Object.hasOwn(nodeConfig.agents, 'dag-node-skills') &&
+      options.agents?.['dag-node-skills'] !== undefined
+    ) {
+      getLog().warn(
+        { nodeSkills: nodeConfig.skills ?? [] },
+        'claude.inline_agents_override_skills_wrapper'
+      );
+    }
+    options.agents = {
+      ...(options.agents ?? {}),
+      ...(nodeConfig.agents as NonNullable<Options['agents']>),
+    };
+    getLog().info({ agentIds: Object.keys(nodeConfig.agents) }, 'claude.inline_agents_registered');
+  }
+
   // effort
   if (nodeConfig.effort !== undefined) {
     options.effort = nodeConfig.effort as Options['effort'];

packages/providers/src/codex/capabilities.ts

@@ -5,6 +5,7 @@ export const CODEX_CAPABILITIES: ProviderCapabilities = {
   mcp: false,
   hooks: false,
   skills: false,
+  agents: false,
   toolRestrictions: false,
   structuredOutput: true,
   envInjection: true,

packages/providers/src/codex/provider.test.ts

@@ -75,6 +75,7 @@ describe('CodexProvider', () => {
         mcp: false,
         hooks: false,
         skills: false,
+        agents: false,
         toolRestrictions: false,
         structuredOutput: true,
         envInjection: true,

packages/providers/src/community/pi/capabilities.ts

@@ -14,6 +14,7 @@ export const PI_CAPABILITIES: ProviderCapabilities = {
   mcp: false,
   hooks: false,
   skills: true,
+  agents: false,
   toolRestrictions: true,
   structuredOutput: false,
   envInjection: true,

packages/providers/src/registry.test.ts

@@ -24,6 +24,7 @@ function makeMockProvider(id: string): IAgentProvider {
       mcp: false,
       hooks: false,
       skills: false,
+      agents: false,
       toolRestrictions: false,
       structuredOutput: false,
       envInjection: false,