gemcode 0.4.0__tar.gz → 0.4.1__tar.gz

{gemcode-0.4.0/src/gemcode.egg-info → gemcode-0.4.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gemcode
-Version: 0.4.0
+Version: 0.4.1
 Summary: Local-first coding agent on Google Gemini + ADK
 Author: GemCode Contributors
 License:                                  Apache License
@@ -195,14 +195,18 @@ Dynamic: license-file
 This document is the primary user-facing manual for GemCode. It explains the product at a high level and points to the subsystem-specific documentation pages that provide production-grade depth.
 
 ## What GemCode is
-GemCode is a local-first coding agent built around Google Gemini and the Google Agent Development Kit (ADK). It operates against a chosen project root and combines:
-- a configuration model
-- a runtime runner
-- a root language-model agent
-- a configurable tool inventory
-- project-local state under `.gemcode/`
+GemCode is a local-first, self-evolving coding agent built on Google Gemini and the Agent Development Kit (ADK). It operates against a chosen project root and provides:
 
-GemCode is designed for repository-native work rather than copy-paste chat workflows.
+- **Autonomous multi-agent orchestration** — agents delegate, verify, and fix each other's work
+- **Self-healing code** — changes are auto-verified; failures are auto-fixed
+- **Self-evolving tools** — the agent creates new reusable tools from repeated patterns
+- **Progressive learning** — gets smarter with every session (memory, skills, project map)
+- **Scheduled habits** — agents wake up on cron/interval to run tests, audits, checks
+- **Cross-machine agents** — expose/consume agents via Google A2A protocol
+- **58 built-in tools** — filesystem, shell, web, search, memory, orchestration, synthesis
+- **Full ADK integration** — native sub-agents, transfer_to_agent, output_key, workflow agents
+
+All state lives under `.gemcode/` in the project root. No external services required beyond a Gemini API key.
 
 ## Runtime modes
 
@@ -331,14 +335,20 @@ Reference:
 
 ### Super mode (fully autonomous)
 
-Use when you want GemCode to run without GemCode’s own confirmation prompts (mutations, shell, ADK confirmation handoffs, AFC stdin prompt, attachment gate, auto-trust on CLI, and non-interactive `get_user_choice` = first option).
+Use when you want GemCode to run with all powers unlocked and zero friction. In super mode:
+- All capabilities auto-enable (memory, web search, agents, habits, triggers)
+- Default org members auto-create (kaira + verifier)
+- Habits auto-generate based on project type
+- Auto-verification runs after risky changes
+- No confirmation prompts for any operation
 
 - **CLI:** `gemcode -C . --super "your task"`
 - **Env:** `GEMCODE_SUPER_MODE=1`
-- **REPL/TUI:** `/super` (use `/super off` to clear the flag only)
-- **Kaira:** `gemcode kaira -C . --super`
+- **REPL/TUI:** `/super` (use `/super off` to clear)
+
+In normal mode, GemCode asks on first run: "Enable autonomous mode? [Y/n]"
 
-Details and safety notes: [`../docs/tools-and-permissions.md`](../docs/tools-and-permissions.md#super-mode-fully-autonomous).
+Details: [`../docs/orchestration.md`](../docs/orchestration.md).
 
 ## Common commands
 
@@ -367,17 +377,22 @@ gemcode kaira -C .
 GemCode includes a built-in multi-agent orchestration system that works automatically — no separate daemon required.
 
 **Key features:**
-- **Agent Mesh** — in-process concurrent agent execution with full tool access
+- **Native ADK sub-agents** — org members are real ADK sub-agents with `transfer_to_agent` routing
+- **Agent Mesh** — async background execution with full GemCode sessions per agent
 - **Event Bus** — agents communicate via pub/sub (no Unix sockets needed)
-- **Self-Triggering Agents** — agents auto-activate on events (e.g., verifier reviews completed work)
+- **Self-Healing** — closed loop: change → verify → fix → verify → done
+- **Self-Triggers** — agents auto-activate on events (verification, failure recovery)
+- **Tool Synthesis** — agent creates new reusable tools from repeated patterns
 - **Delegation Learning** — remembers which agents succeed at which tasks
-- **A2A Bridge** — expose/consume agents across machines via Google's A2A protocol
+- **A2A Bridge** — expose/consume agents across machines via Google A2A protocol
 
 Quick example in the REPL:
 ```text
 > Analyze the auth module. Delegate security review to the verifier.
 ```
-The agent calls `org_delegate("verifier", ...)` → mesh runs a full-power verifier agent → result flows back automatically.
+The LLM calls `transfer_to_agent(agent_name='verifier')` → ADK routes natively → verifier runs → result saved to session state.
+
+For background work: `org_delegate("kaira", "run tests")` → mesh runs kaira as a full GemCode session → result flows back via fleet reports.
 
 Docs:
 - [`../docs/orchestration.md`](../docs/orchestration.md)
@@ -428,6 +443,60 @@ Status note:
 | `/agent assign <member> <task>` | Delegate work to a member |
 | `/agent improve <member> <lessons>` | Improve a member's skill |
 
+## Intelligence features (automatic)
+
+These work without configuration. In super mode, everything is enabled silently. In normal mode, GemCode asks once on first run.
+
+| Feature | How it works |
+|---|---|
+| **Codebase awareness** | Builds a persistent understanding of the project from every tool call — file structure, imports/exports, recent changes, learned facts. The agent starts each turn already knowing the project. |
+| **Self-healing** | After file changes, auto-runs tests/lint. If they fail, auto-fixes (up to 2 attempts). Closed loop: change → verify → fix → verify → done. |
+| **Tool synthesis** | When the agent repeats multi-step operations, it can create a reusable tool (bash/python script) stored in `.gemcode/synthesized_tools/`. |
+| **Self-improving skills** | When a delegation succeeds, the member's skill file gets a "Learned pattern" appended. Future invocations benefit from past successes. |
+| **Proactive memory** | After exploring 5+ files or running 3+ commands, key discoveries are auto-saved to curated memory. Future sessions start with this knowledge. |
+| **Impact analysis** | When a file changes, GemCode knows which other files are affected (via import tracking + learned correlations). Self-healing runs only relevant tests. |
+| **Auto-verification** | After 3+ file writes, the verifier agent auto-checks for syntax errors, broken imports, and logic bugs. |
+| **Delegation suggestions** | `suggest_delegate(task)` recommends the best agent based on historical success patterns. |
+| **Capability auto-enable** | If a project consistently uses web search or memory, those capabilities auto-enable in future sessions. |
+
+## Tool Synthesis (self-evolving)
+
+The agent can create new reusable tools when it detects repeated patterns:
+
+```text
+# Create a tool
+synthesize_tool("run-tests", "Run pytest with coverage", "pytest --cov=src -q")
+synthesize_tool("deploy-staging", "Deploy to staging", "git push origin main && ssh staging 'cd app && git pull'")
+
+# Use it later
+run_synthesized_tool("run-tests")
+run_synthesized_tool("deploy-staging")
+
+# List all synthesized tools
+list_synthesized_tools()
+```
+
+Tools persist in `.gemcode/synthesized_tools/` across sessions.
+
+## Agent Habits (scheduled tasks)
+
+Agents can run recurring tasks on a schedule — no daemon needed, runs inside the main GemCode process.
+
+```text
+# From the agent (tools):
+habits_add("test-watch", "kaira", "Run pytest -q and report", every_minutes=30)
+habits_add("nightly-audit", "verifier", "Full security review", daily_at="02:00")
+habits_add("hourly-status", "self", "Summarize recent changes", cron="0 * * * *")
+
+# Management:
+habits_list()
+habits_pause("test-watch")
+habits_resume("test-watch")
+habits_remove("test-watch")
+```
+
+In super mode, GemCode auto-creates habits based on project type (test-watch for Python, lint-watch for Node).
+
 Detailed behavior:
 - [`../docs/cli-and-repl.md`](../docs/cli-and-repl.md)
 
@@ -435,15 +504,23 @@ Detailed behavior:
 
 | Capability | What it adds |
 |---|---|
-| Agent Mesh | In-process multi-agent orchestration (automatic, no daemon needed) |
-| A2A Bridge | Cross-machine agent communication via Google A2A protocol |
-| Self-Triggers | Agents auto-activate on events (verification, failure recovery) |
-| Delegation Learning | Remembers successful patterns, suggests optimal routing |
-| Deep research | Research-focused tool routing and optional dedicated model path |
-| Embeddings | Semantic search and optional embedding-backed memory |
-| Memory | Retrieval-oriented persistent memory |
-| Browser/computer use | Playwright-backed browser automation and inspection |
-| Live audio | Gemini Live microphone sessions (experimental) |
+| **Codebase Awareness** | Persistent understanding of the project — structure graph, change journal, insight cache. Compounds over time, zero extra cost |
+| **Agent Mesh** | In-process multi-agent orchestration — each agent is a full GemCode session with own workspace, memory, and persistent history |
+| **Self-Healing** | Closed loop: change → verify → fix → verify → done. Code repairs itself automatically |
+| **Tool Synthesis** | Agent creates new reusable tools at runtime from repeated patterns |
+| **Agent Habits** | Scheduled recurring tasks (cron/interval/daily) — agents wake up and do work autonomously |
+| **Self-Triggers** | Agents auto-activate on events (verification after changes, failure recovery) |
+| **Self-Improving Skills** | Skills evolve — successful patterns are appended automatically |
+| **Delegation Learning** | Remembers which agents succeed at which tasks, suggests optimal routing |
+| **Proactive Memory** | Auto-saves important discoveries to curated memory without being asked |
+| **A2A Bridge** | Cross-machine agent communication via Google A2A protocol |
+| **Event Bus** | In-memory pub/sub for agent-to-agent communication |
+| **Deep research** | Research-focused tool routing and optional dedicated model path |
+| **Embeddings** | Semantic search and optional embedding-backed memory |
+| **Memory** | Retrieval-oriented persistent memory across sessions |
+| **Browser/computer use** | Playwright-backed browser automation and inspection |
+| **Checkpoints** | File mutations are reversible — undo any agent edit |
+| **Live audio** | Gemini Live microphone sessions (experimental) |
 
 Detailed behavior:
 - [`../docs/capabilities.md`](../docs/capabilities.md)

{gemcode-0.4.0 → gemcode-0.4.1}/README.md

@@ -3,14 +3,18 @@
 This document is the primary user-facing manual for GemCode. It explains the product at a high level and points to the subsystem-specific documentation pages that provide production-grade depth.
 
 ## What GemCode is
-GemCode is a local-first coding agent built around Google Gemini and the Google Agent Development Kit (ADK). It operates against a chosen project root and combines:
-- a configuration model
-- a runtime runner
-- a root language-model agent
-- a configurable tool inventory
-- project-local state under `.gemcode/`
+GemCode is a local-first, self-evolving coding agent built on Google Gemini and the Agent Development Kit (ADK). It operates against a chosen project root and provides:
 
-GemCode is designed for repository-native work rather than copy-paste chat workflows.
+- **Autonomous multi-agent orchestration** — agents delegate, verify, and fix each other's work
+- **Self-healing code** — changes are auto-verified; failures are auto-fixed
+- **Self-evolving tools** — the agent creates new reusable tools from repeated patterns
+- **Progressive learning** — gets smarter with every session (memory, skills, project map)
+- **Scheduled habits** — agents wake up on cron/interval to run tests, audits, checks
+- **Cross-machine agents** — expose/consume agents via Google A2A protocol
+- **58 built-in tools** — filesystem, shell, web, search, memory, orchestration, synthesis
+- **Full ADK integration** — native sub-agents, transfer_to_agent, output_key, workflow agents
+
+All state lives under `.gemcode/` in the project root. No external services required beyond a Gemini API key.
 
 ## Runtime modes
 
@@ -139,14 +143,20 @@ Reference:
 
 ### Super mode (fully autonomous)
 
-Use when you want GemCode to run without GemCode’s own confirmation prompts (mutations, shell, ADK confirmation handoffs, AFC stdin prompt, attachment gate, auto-trust on CLI, and non-interactive `get_user_choice` = first option).
+Use when you want GemCode to run with all powers unlocked and zero friction. In super mode:
+- All capabilities auto-enable (memory, web search, agents, habits, triggers)
+- Default org members auto-create (kaira + verifier)
+- Habits auto-generate based on project type
+- Auto-verification runs after risky changes
+- No confirmation prompts for any operation
 
 - **CLI:** `gemcode -C . --super "your task"`
 - **Env:** `GEMCODE_SUPER_MODE=1`
-- **REPL/TUI:** `/super` (use `/super off` to clear the flag only)
-- **Kaira:** `gemcode kaira -C . --super`
+- **REPL/TUI:** `/super` (use `/super off` to clear)
+
+In normal mode, GemCode asks on first run: "Enable autonomous mode? [Y/n]"
 
-Details and safety notes: [`../docs/tools-and-permissions.md`](../docs/tools-and-permissions.md#super-mode-fully-autonomous).
+Details: [`../docs/orchestration.md`](../docs/orchestration.md).
 
 ## Common commands
 
@@ -175,17 +185,22 @@ gemcode kaira -C .
 GemCode includes a built-in multi-agent orchestration system that works automatically — no separate daemon required.
 
 **Key features:**
-- **Agent Mesh** — in-process concurrent agent execution with full tool access
+- **Native ADK sub-agents** — org members are real ADK sub-agents with `transfer_to_agent` routing
+- **Agent Mesh** — async background execution with full GemCode sessions per agent
 - **Event Bus** — agents communicate via pub/sub (no Unix sockets needed)
-- **Self-Triggering Agents** — agents auto-activate on events (e.g., verifier reviews completed work)
+- **Self-Healing** — closed loop: change → verify → fix → verify → done
+- **Self-Triggers** — agents auto-activate on events (verification, failure recovery)
+- **Tool Synthesis** — agent creates new reusable tools from repeated patterns
 - **Delegation Learning** — remembers which agents succeed at which tasks
-- **A2A Bridge** — expose/consume agents across machines via Google's A2A protocol
+- **A2A Bridge** — expose/consume agents across machines via Google A2A protocol
 
 Quick example in the REPL:
 ```text
 > Analyze the auth module. Delegate security review to the verifier.
 ```
-The agent calls `org_delegate("verifier", ...)` → mesh runs a full-power verifier agent → result flows back automatically.
+The LLM calls `transfer_to_agent(agent_name='verifier')` → ADK routes natively → verifier runs → result saved to session state.
+
+For background work: `org_delegate("kaira", "run tests")` → mesh runs kaira as a full GemCode session → result flows back via fleet reports.
 
 Docs:
 - [`../docs/orchestration.md`](../docs/orchestration.md)
@@ -236,6 +251,60 @@ Status note:
 | `/agent assign <member> <task>` | Delegate work to a member |
 | `/agent improve <member> <lessons>` | Improve a member's skill |
 
+## Intelligence features (automatic)
+
+These work without configuration. In super mode, everything is enabled silently. In normal mode, GemCode asks once on first run.
+
+| Feature | How it works |
+|---|---|
+| **Codebase awareness** | Builds a persistent understanding of the project from every tool call — file structure, imports/exports, recent changes, learned facts. The agent starts each turn already knowing the project. |
+| **Self-healing** | After file changes, auto-runs tests/lint. If they fail, auto-fixes (up to 2 attempts). Closed loop: change → verify → fix → verify → done. |
+| **Tool synthesis** | When the agent repeats multi-step operations, it can create a reusable tool (bash/python script) stored in `.gemcode/synthesized_tools/`. |
+| **Self-improving skills** | When a delegation succeeds, the member's skill file gets a "Learned pattern" appended. Future invocations benefit from past successes. |
+| **Proactive memory** | After exploring 5+ files or running 3+ commands, key discoveries are auto-saved to curated memory. Future sessions start with this knowledge. |
+| **Impact analysis** | When a file changes, GemCode knows which other files are affected (via import tracking + learned correlations). Self-healing runs only relevant tests. |
+| **Auto-verification** | After 3+ file writes, the verifier agent auto-checks for syntax errors, broken imports, and logic bugs. |
+| **Delegation suggestions** | `suggest_delegate(task)` recommends the best agent based on historical success patterns. |
+| **Capability auto-enable** | If a project consistently uses web search or memory, those capabilities auto-enable in future sessions. |
+
+## Tool Synthesis (self-evolving)
+
+The agent can create new reusable tools when it detects repeated patterns:
+
+```text
+# Create a tool
+synthesize_tool("run-tests", "Run pytest with coverage", "pytest --cov=src -q")
+synthesize_tool("deploy-staging", "Deploy to staging", "git push origin main && ssh staging 'cd app && git pull'")
+
+# Use it later
+run_synthesized_tool("run-tests")
+run_synthesized_tool("deploy-staging")
+
+# List all synthesized tools
+list_synthesized_tools()
+```
+
+Tools persist in `.gemcode/synthesized_tools/` across sessions.
+
+## Agent Habits (scheduled tasks)
+
+Agents can run recurring tasks on a schedule — no daemon needed, runs inside the main GemCode process.
+
+```text
+# From the agent (tools):
+habits_add("test-watch", "kaira", "Run pytest -q and report", every_minutes=30)
+habits_add("nightly-audit", "verifier", "Full security review", daily_at="02:00")
+habits_add("hourly-status", "self", "Summarize recent changes", cron="0 * * * *")
+
+# Management:
+habits_list()
+habits_pause("test-watch")
+habits_resume("test-watch")
+habits_remove("test-watch")
+```
+
+In super mode, GemCode auto-creates habits based on project type (test-watch for Python, lint-watch for Node).
+
 Detailed behavior:
 - [`../docs/cli-and-repl.md`](../docs/cli-and-repl.md)
 
@@ -243,15 +312,23 @@ Detailed behavior:
 
 | Capability | What it adds |
 |---|---|
-| Agent Mesh | In-process multi-agent orchestration (automatic, no daemon needed) |
-| A2A Bridge | Cross-machine agent communication via Google A2A protocol |
-| Self-Triggers | Agents auto-activate on events (verification, failure recovery) |
-| Delegation Learning | Remembers successful patterns, suggests optimal routing |
-| Deep research | Research-focused tool routing and optional dedicated model path |
-| Embeddings | Semantic search and optional embedding-backed memory |
-| Memory | Retrieval-oriented persistent memory |
-| Browser/computer use | Playwright-backed browser automation and inspection |
-| Live audio | Gemini Live microphone sessions (experimental) |
+| **Codebase Awareness** | Persistent understanding of the project — structure graph, change journal, insight cache. Compounds over time, zero extra cost |
+| **Agent Mesh** | In-process multi-agent orchestration — each agent is a full GemCode session with own workspace, memory, and persistent history |
+| **Self-Healing** | Closed loop: change → verify → fix → verify → done. Code repairs itself automatically |
+| **Tool Synthesis** | Agent creates new reusable tools at runtime from repeated patterns |
+| **Agent Habits** | Scheduled recurring tasks (cron/interval/daily) — agents wake up and do work autonomously |
+| **Self-Triggers** | Agents auto-activate on events (verification after changes, failure recovery) |
+| **Self-Improving Skills** | Skills evolve — successful patterns are appended automatically |
+| **Delegation Learning** | Remembers which agents succeed at which tasks, suggests optimal routing |
+| **Proactive Memory** | Auto-saves important discoveries to curated memory without being asked |
+| **A2A Bridge** | Cross-machine agent communication via Google A2A protocol |
+| **Event Bus** | In-memory pub/sub for agent-to-agent communication |
+| **Deep research** | Research-focused tool routing and optional dedicated model path |
+| **Embeddings** | Semantic search and optional embedding-backed memory |
+| **Memory** | Retrieval-oriented persistent memory across sessions |
+| **Browser/computer use** | Playwright-backed browser automation and inspection |
+| **Checkpoints** | File mutations are reversible — undo any agent edit |
+| **Live audio** | Gemini Live microphone sessions (experimental) |
 
 Detailed behavior:
 - [`../docs/capabilities.md`](../docs/capabilities.md)

{gemcode-0.4.0 → gemcode-0.4.1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "gemcode"
-version = "0.4.0"
+version = "0.4.1"
 description = "Local-first coding agent on Google Gemini + ADK"
 readme = "README.md"
 requires-python = ">=3.11"

{gemcode-0.4.0 → gemcode-0.4.1}/src/gemcode/agent.py

@@ -1287,37 +1287,90 @@ def build_root_agent(
   sub_agents = []
   if getattr(cfg, "enable_adk_agent_transfer", True) and _tools is None:
     try:
-      # Explorer: read-only, fast, low-risk. Keep instruction short.
-      explorer_tools = build_function_tools(cfg, include_subtask=False)
-      explorer_tools = [t for t in explorer_tools if getattr(t, "__name__", "") not in ("write_file", "search_replace", "delete_file", "move_file", "bash", "run_command")]
-      explorer = LlmAgent(
-        name="explorer",
-        model=getattr(cfg, "model_alt", None) or cfg.model,
-        instruction=(
-          "You are Explorer. Your job is to quickly map the codebase and answer: "
-          "what files/symbols matter and where to look next. Use read-only tools only. "
-          "Return concise findings with file paths and symbol names."
-        ),
-        tools=explorer_tools,
-        generate_content_config=gen_cfg,
-        **cb_kwargs,
-      )
-      # Verifier: focuses on checking, tests, and consistency.
-      verifier_tools = build_function_tools(cfg, include_subtask=False)
-      verifier_tools = [t for t in verifier_tools if getattr(t, "__name__", "") not in ("write_file", "search_replace", "delete_file", "move_file")]
-      verifier = LlmAgent(
-        name="verifier",
-        model=getattr(cfg, "model_alt", None) or cfg.model,
-        instruction=(
-          "You are Verifier. Your job is to verify changes: run checks/tests when needed, "
-          "spot inconsistencies, and report PASS/FAIL with concrete evidence. "
-          "Prefer minimal commands and short outputs."
-        ),
-        tools=verifier_tools,
-        generate_content_config=gen_cfg,
-        **cb_kwargs,
-      )
-      sub_agents = [explorer, verifier]
+      from gemcode.org import list_members, resolve_fleet_root
+
+      fleet_root = resolve_fleet_root(cfg.project_root)
+      members = list_members(fleet_root)
+
+      for m in members:
+        # Build a specialized sub-agent for each org member
+        member_tools = build_function_tools(cfg, include_subtask=False)
+
+        # Restrict tools based on member kind
+        if m.kind == "subagent" and m.name in ("verifier", "reviewer"):
+          # Verifiers don't need write tools
+          member_tools = [t for t in member_tools if getattr(t, "__name__", "") not in (
+            "write_file", "search_replace", "delete_file", "move_file",
+          )]
+
+        # Build instruction from member's skill if available
+        member_instruction = (
+          f"You are {m.name} ({m.title}).\n"
+          f"Role: {m.description or 'General assistant'}\n"
+          f"Reports to: {m.reports_to or 'manager'}\n\n"
+          "Complete assigned tasks concisely. Return structured JSON when possible.\n"
+        )
+
+        # Load skill content if available
+        if m.skill_name:
+          try:
+            skill = load_skill(fleet_root, m.skill_name)
+            if skill is not None:
+              member_instruction += "\n" + expand_skill_text(skill, arguments="", session_id="")
+          except Exception:
+            pass
+
+        member_agent = LlmAgent(
+          name=m.name,
+          model=getattr(cfg, "model_alt", None) or cfg.model,
+          description=m.description or f"{m.name} — {m.title}",
+          instruction=member_instruction,
+          tools=member_tools,
+          generate_content_config=gen_cfg,
+          output_key=f"agent_{m.name}_result",
+          **cb_kwargs,
+        )
+        sub_agents.append(member_agent)
+
+      # If no org members exist, create default explorer + verifier
+      if not sub_agents:
+        explorer_tools = build_function_tools(cfg, include_subtask=False)
+        explorer_tools = [t for t in explorer_tools if getattr(t, "__name__", "") not in (
+          "write_file", "search_replace", "delete_file", "move_file", "bash", "run_command",
+        )]
+        explorer = LlmAgent(
+          name="explorer",
+          model=getattr(cfg, "model_alt", None) or cfg.model,
+          description="Quickly maps the codebase and finds relevant files/symbols.",
+          instruction=(
+            "You are Explorer. Your job is to quickly map the codebase and answer: "
+            "what files/symbols matter and where to look next. Use read-only tools only. "
+            "Return concise findings with file paths and symbol names."
+          ),
+          tools=explorer_tools,
+          generate_content_config=gen_cfg,
+          output_key="explorer_result",
+          **cb_kwargs,
+        )
+        verifier_tools = build_function_tools(cfg, include_subtask=False)
+        verifier_tools = [t for t in verifier_tools if getattr(t, "__name__", "") not in (
+          "write_file", "search_replace", "delete_file", "move_file",
+        )]
+        verifier = LlmAgent(
+          name="verifier",
+          model=getattr(cfg, "model_alt", None) or cfg.model,
+          description="Verifies changes: runs checks/tests, spots inconsistencies, reports PASS/FAIL.",
+          instruction=(
+            "You are Verifier. Your job is to verify changes: run checks/tests when needed, "
+            "spot inconsistencies, and report PASS/FAIL with concrete evidence. "
+            "Prefer minimal commands and short outputs."
+          ),
+          tools=verifier_tools,
+          generate_content_config=gen_cfg,
+          output_key="verifier_result",
+          **cb_kwargs,
+        )
+        sub_agents = [explorer, verifier]
     except Exception:
       sub_agents = []
 
@@ -1327,8 +1380,8 @@ def build_root_agent(
       instruction=build_instruction(cfg),
       tools=tools,
       generate_content_config=gen_cfg,
-      # ADK expects a list; passing None can fail validation on some versions.
       sub_agents=sub_agents,
+      output_key="gemcode_last_output",
       **cb_kwargs,
   )