gemcode 0.3.28__tar.gz → 0.3.29__tar.gz

{gemcode-0.3.28/src/gemcode.egg-info → gemcode-0.3.29}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gemcode
-Version: 0.3.28
+Version: 0.3.29
 Summary: Local-first coding agent on Google Gemini + ADK
 Author: GemCode Contributors
 License:                                  Apache License

{gemcode-0.3.28 → gemcode-0.3.29}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "gemcode"
-version = "0.3.28"
+version = "0.3.29"
 description = "Local-first coding agent on Google Gemini + ADK"
 readme = "README.md"
 requires-python = ">=3.11"

{gemcode-0.3.28 → gemcode-0.3.29}/src/gemcode/agent.py

@@ -77,8 +77,12 @@ def _build_runtime_facts(cfg: GemCodeConfig) -> str:
     caps.append(f"memory ON ({mem_kind}, stored at {mem_path}; ADK preload_memory injects relevant memories before each turn)")
   if getattr(cfg, "enable_computer_use", False):
     caps.append("computer_use ON (tools: navigate, click_at, type_text_at, browser_screenshot, browser_find_element, etc.)")
+  if getattr(cfg, "enable_code_executor", False):
+    caps.append("code_executor ON — you can write Python code blocks and they will be executed safely via Gemini's built-in sandboxed executor; results appear as code_execution_result events. Use this for math, data processing, quick tests, and anything that would otherwise require a shell command.")
+  if getattr(cfg, "enable_artifacts", True):
+    caps.append("artifacts ON — use save_artifact(filename, bytes, mime_type) / load_artifact(filename) to store large/binary outputs (screenshots, PDFs, generated files) outside session history. Artifacts are keyed by filename; prefix 'user:' for cross-session persistence.")
   if not caps:
-    caps.append("none enabled (use /research on, /embeddings on, /memory on, /computer on to enable)")
+    caps.append("none enabled (use /research on, /embeddings on, /memory on, /computer on, /code on to enable)")
   caps_text = "\n".join(f"  - {c}" for c in caps)
 
   # ── Limits ───────────────────────────────────────────────────────────────
@@ -426,6 +430,41 @@ Use `run_subtask` when the work is better done in an isolated context:
 
 The sub-agent inherits your permission settings and returns its final text as `result`. Treat it as a trusted colleague returning a written summary.
 
+## ADK Special Tools (always available when ADK supports them)
+
+### `get_user_choice`
+Present the user with a structured multi-option prompt rather than open-ended questions.
+Use when you need the user to pick from 2–6 specific options (e.g. "Which framework would you like?", "Choose migration strategy: A, B, or C").
+This provides a better UX than asking them to type a free-form answer.
+
+### `load_artifacts`
+Load binary/large artifacts that were saved in a previous turn or by a sub-agent.
+Artifacts are keyed by filename (e.g. "report.pdf", "screenshot.png", "output.json").
+Use `user:filename` prefix for user-scoped artifacts that persist across sessions.
+After loading, the artifact bytes are available for further processing (display, analysis, transformation).
+
+### `exit_loop`
+Signal the surrounding LoopAgent to stop iterating and return the final result.
+Only meaningful when this agent is running inside an ADK LoopAgent pipeline.
+Call this when the task is complete and no further iterations are needed.
+
+## Artifacts — storing large outputs
+When `artifacts ON` (see Runtime facts above):
+- **Save** large generated content as artifacts instead of printing them inline:
+  - Screenshots from computer_use: save as "screenshot.png" artifact
+  - Generated reports/PDFs: save as "report.pdf" artifact
+  - Large JSON data: save as "data.json" artifact
+- **Reference** artifacts in instructions via `{artifact.filename?}` template syntax
+- Artifacts are keyed by filename; `user:` prefix = cross-session persistence
+
+## Code Executor (sandboxed Python)
+When `code_executor ON` (see Runtime facts above):
+- You can write Python code blocks in your response and the Gemini API executes them safely
+- The result appears as a `code_execution_result` event with stdout and the outcome
+- Best for: math calculations, data transformation, unit testing logic, quick experiments
+- The sandbox does NOT have internet access or filesystem access — use for pure computation
+- For file I/O or shell commands, use the standard tools (`bash`, `write_file`, etc.)
+
 ## Evaluator-optimizer loop
 For tasks where quality matters:
 1. Complete the task (execute tools, write code, run commands)
@@ -470,6 +509,17 @@ All file tools use paths **relative to the project root** (where GemCode was sta
   return base
 
 
+def _build_code_executor(cfg: GemCodeConfig):
+  """Return an ADK BuiltInCodeExecutor when enable_code_executor=True, else None."""
+  if not getattr(cfg, "enable_code_executor", False):
+    return None
+  try:
+    from google.adk.code_executors import BuiltInCodeExecutor
+    return BuiltInCodeExecutor()
+  except Exception:
+    return None
+
+
 def build_root_agent(
   cfg: GemCodeConfig,
   extra_tools: list | None = None,
@@ -488,12 +538,19 @@ def build_root_agent(
   if _tools is not None:
     tools = list(_tools)
   else:
-  tools = build_function_tools(cfg)
+    tools = build_function_tools(cfg)
   if getattr(cfg, "enable_memory", False):
     # ADK preload_memory injects retrieved memories into the next llm_request.
     from google.adk.tools import preload_memory
-
     tools = [preload_memory, *tools]
+
+  # ADK built-in interactive + artifact tools — always available when ADK supports them.
+  try:
+    from google.adk.tools import get_user_choice, load_artifacts, exit_loop
+    tools = [*tools, get_user_choice, load_artifacts, exit_loop]
+  except Exception:
+    pass
+
   if extra_tools:
     tools = [*tools, *extra_tools]
 
@@ -548,15 +605,31 @@ def build_root_agent(
       tool_config=tool_cfg,
     )
 
-  return LlmAgent(
+  # global_instruction applies to the entire agent tree (including sub-agents
+  # spawned via run_subtask or multi-agent delegation).  Keep it short — it's
+  # prepended to every agent's effective instruction.
+  global_instr = (
+    "You are GemCode, an expert software engineering agent powered by Google Gemini. "
+    "Act, don't advise. Complete tasks fully and autonomously. "
+    "Think before destructive actions. Use read-only tools before shell/write tools."
+  )
+
+  agent_kwargs: dict = dict(
       model=cfg.model,
       name="gemcode",
       instruction=build_instruction(cfg),
+      global_instruction=global_instr,
       tools=tools,
       generate_content_config=gen_cfg,
       **cb_kwargs,
   )
 
+  code_executor = _build_code_executor(cfg)
+  if code_executor is not None:
+    agent_kwargs["code_executor"] = code_executor
+
+  return LlmAgent(**agent_kwargs)
+
 
 def create_runner(cfg: GemCodeConfig, extra_tools: list | None = None):
   """Backward-compatible: prefer `gemcode.session_runtime.create_runner`."""

{gemcode-0.3.28 → gemcode-0.3.29}/src/gemcode/config.py

@@ -251,6 +251,22 @@ class GemCodeConfig:
     default_factory=lambda: _truthy_env("GEMCODE_SHOW_FULL_THINKING", default=False)
   )
 
+  # Enable ADK BuiltInCodeExecutor for safe sandboxed Python execution via
+  # Gemini's code execution API. When on, the agent can write and run Python
+  # snippets inline (math, data processing, quick tests) without requiring
+  # bash/shell permissions. Requires a Gemini model that supports code execution
+  # (gemini-2.5-flash, gemini-2.5-pro, gemini-3.x).
+  enable_code_executor: bool = field(
+    default_factory=lambda: _truthy_env("GEMCODE_ENABLE_CODE_EXECUTOR", default=False)
+  )
+
+  # Enable ADK artifact service for storing large/binary outputs (screenshots,
+  # generated files, reports) outside of session history.
+  # When on, the agent can save_artifact / load_artifact keyed by filename.
+  enable_artifacts: bool = field(
+    default_factory=lambda: _truthy_env("GEMCODE_ENABLE_ARTIFACTS", default=True)
+  )
+
   def __post_init__(self) -> None:
     self.project_root = self.project_root.resolve()
     # Default agentic depth when env omits GEMCODE_MAX_LLM_CALLS (was: None → SDK default).

gemcode-0.3.29/src/gemcode/refine.py

@@ -0,0 +1,204 @@
+"""
+Iterative refinement pipelines using ADK LoopAgent + SequentialAgent.
+
+This module exposes factory functions for building multi-agent orchestration
+patterns that GemCode can invoke for complex, multi-step tasks:
+
+  - build_refine_loop(): LoopAgent that loops write→test→fix until tests pass
+    or max_iterations is reached (escalates via exit_loop tool).
+
+  - build_sequential_pipeline(): SequentialAgent that runs N specialist agents
+    in order, passing results through session state via output_key.
+
+  - build_parallel_research(): ParallelAgent that fans out N research sub-agents
+    then collects their outputs.
+
+These are "heavy" patterns; they are NOT wired by default — callers instantiate
+them on demand (e.g. via /refine slash command or run_subtask).
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Any
+
+log = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Guard: check what's available in the installed ADK version
+# ---------------------------------------------------------------------------
+try:
+    from google.adk.agents import LoopAgent, SequentialAgent, ParallelAgent, LlmAgent  # type: ignore
+    _ADK_WORKFLOW_AGENTS_OK = True
+except ImportError:
+    _ADK_WORKFLOW_AGENTS_OK = False
+    log.warning("google.adk workflow agents (LoopAgent/SequentialAgent/ParallelAgent) "
+                "not available — refine pipelines disabled.")
+
+
+def _require_workflow() -> None:
+    if not _ADK_WORKFLOW_AGENTS_OK:
+        raise RuntimeError(
+            "ADK workflow agents (LoopAgent / SequentialAgent / ParallelAgent) are not "
+            "available in the installed google-adk version. "
+            "Try: pip install -U google-adk"
+        )
+
+
+# ---------------------------------------------------------------------------
+# Write → Test → Fix Loop
+# ---------------------------------------------------------------------------
+
+def build_refine_loop(
+    *,
+    model: str,
+    task_description: str,
+    test_command: str,
+    max_iterations: int = 8,
+    extra_tools: list | None = None,
+) -> Any:
+    """
+    Build a LoopAgent that iteratively writes code, runs tests, and fixes
+    failures until all tests pass or ``max_iterations`` is exhausted.
+
+    Architecture:
+      LoopAgent(max_iterations=N)
+        ├─ writer_agent   — writes / edits code based on current task + test output
+        ├─ tester_agent   — runs the test command via bash, stores result in state
+        └─ checker_agent  — inspects test result; calls exit_loop if tests pass
+
+    State keys used:
+      "task"          — original task description (seeded by caller)
+      "test_command"  — shell command to run tests
+      "test_output"   — stdout/stderr from last test run
+      "test_passed"   — "yes" if tests passed
+      "iteration"     — current loop count (auto-incremented by writer)
+
+    Returns the LoopAgent instance, ready to run via a Runner.
+
+    Example usage:
+      loop = build_refine_loop(model="gemini-2.5-flash",
+                               task_description="add unit tests for parser.py",
+                               test_command="pytest tests/test_parser.py -x")
+      runner = Runner(agent=loop, app_name="refine", session_service=...)
+      async for event in runner.run_async(...):
+          ...
+    """
+    _require_workflow()
+    from google.adk.tools import exit_loop  # type: ignore
+
+    from gemcode.tools import build_function_tools  # type: ignore
+
+    # Shared tools that the sub-agents may use (no run_subtask to avoid recursion)
+    tools = extra_tools or []
+
+    writer = LlmAgent(
+        model=model,
+        name="writer_agent",
+        description="Writes or edits source code to fulfil the task, using test failure output as feedback.",
+        instruction=(
+            "You are a coding agent. Your goal: {task}.\n\n"
+            "Test command: {test_command}\n"
+            "Last test output:\n{test_output?}\n\n"
+            "Write or fix code to make the tests pass. "
+            "After making your edits, update state key 'iteration' by incrementing it."
+        ),
+        tools=tools,
+        output_key="writer_notes",
+    )
+
+    tester = LlmAgent(
+        model=model,
+        name="tester_agent",
+        description="Runs the test suite and stores the output in session state.",
+        instruction=(
+            "Run the test command stored in state: {test_command}.\n"
+            "Use the bash tool to execute it. Store ALL stdout+stderr verbatim in "
+            "state key 'test_output'. "
+            "If the command exits with code 0 (success), set state key 'test_passed' to 'yes'. "
+            "If it fails, set 'test_passed' to 'no'."
+        ),
+        tools=tools,
+        output_key="test_output",
+    )
+
+    checker = LlmAgent(
+        model=model,
+        name="checker_agent",
+        description=(
+            "Checks whether tests passed. If yes, calls exit_loop to terminate the "
+            "refinement cycle. If no, does nothing (loop continues)."
+        ),
+        instruction=(
+            "Check the value of state key 'test_passed'.\n"
+            "If 'test_passed' is 'yes': call the exit_loop tool to stop the loop — "
+            "the task is complete.\n"
+            "If 'test_passed' is 'no': output a brief summary of the failures so the "
+            "writer knows what to fix next. Do NOT call exit_loop."
+        ),
+        tools=[exit_loop, *tools],
+    )
+
+    return LoopAgent(
+        name="refine_loop",
+        description=(
+            f"Iterative write→test→fix loop for: {task_description[:120]}"
+        ),
+        sub_agents=[writer, tester, checker],
+        max_iterations=max_iterations,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Sequential Pipeline
+# ---------------------------------------------------------------------------
+
+def build_sequential_pipeline(
+    agents: list[Any],
+    *,
+    name: str = "sequential_pipeline",
+    description: str = "Runs specialist agents sequentially, passing results via state.",
+) -> Any:
+    """
+    Wrap a list of LlmAgents in a SequentialAgent.
+
+    Each agent should set ``output_key`` so subsequent agents can read its output
+    via {key} template substitution in their instructions.
+
+    Example:
+      pipeline = build_sequential_pipeline([
+          planner_agent,   # output_key="plan"
+          coder_agent,     # reads {plan}, output_key="code"
+          reviewer_agent,  # reads {code}, output_key="review"
+      ])
+    """
+    _require_workflow()
+    return SequentialAgent(name=name, description=description, sub_agents=agents)
+
+
+# ---------------------------------------------------------------------------
+# Parallel Fan-Out
+# ---------------------------------------------------------------------------
+
+def build_parallel_research(
+    agents: list[Any],
+    *,
+    name: str = "parallel_research",
+    description: str = "Runs research agents in parallel, collecting results via state.",
+) -> Any:
+    """
+    Wrap a list of LlmAgents in a ParallelAgent.
+
+    Each agent MUST use a distinct ``output_key`` to avoid state key collisions.
+
+    Example:
+      research = build_parallel_research([
+          web_researcher,   # output_key="web_results"
+          code_searcher,    # output_key="code_results"
+          doc_reader,       # output_key="doc_results"
+      ])
+      # Wrap in a SequentialAgent to collect after:
+      pipeline = build_sequential_pipeline([research, synthesis_agent])
+    """
+    _require_workflow()
+    return ParallelAgent(name=name, description=description, sub_agents=agents)

{gemcode-0.3.28 → gemcode-0.3.29}/src/gemcode/repl_commands.py

@@ -216,4 +216,5 @@ def slash_help_lines() -> list[str]:
       "  /permissions          Show permission / HITL settings",
       "  /hooks                Show post-turn hook configuration",
       "  /kairos               How to launch the background parallel job scheduler",
+      "  /code [on|off]        Toggle sandboxed Python executor (ADK BuiltInCodeExecutor)",
   ]

{gemcode-0.3.28 → gemcode-0.3.29}/src/gemcode/repl_slash.py

@@ -357,6 +357,49 @@ async def process_repl_slash(
     out()
     return ReplSlashResult(skip_model_turn=True)
 
+  # ── /code ─────────────────────────────────────────────────────────────────
+  if name == "code":
+    args_s = (sc.args or "").strip().lower()
+    if args_s in ("on", "enable", "1", "true"):
+      cfg.enable_code_executor = True
+      out("Code executor: ON  (sandboxed Python via Gemini built-in executor)")
+      out("The agent can now write Python code blocks and execute them safely,")
+      out("without requiring bash/shell permission. Best for: math, data, quick tests.")
+      out()
+      out("Note: the sandbox has no internet/filesystem access — use bash for I/O.")
+      return ReplSlashResult(skip_model_turn=True, force_rebuild_runner=True)
+    elif args_s in ("off", "disable", "0", "false"):
+      cfg.enable_code_executor = False
+      out("Code executor: OFF")
+      return ReplSlashResult(skip_model_turn=True, force_rebuild_runner=True)
+    else:
+      # Status
+      status = "ON" if getattr(cfg, "enable_code_executor", False) else "OFF"
+      out(f"Code executor: {status}")
+      out()
+      out("ADK BuiltInCodeExecutor — safe sandboxed Python execution via Gemini API.")
+      out()
+      out("What it does:")
+      out("  When ON, the agent can write Python code blocks directly in its response")
+      out("  and the Gemini API executes them in a sandboxed environment. The output")
+      out("  (stdout, result) is sent back so the agent can use it for further reasoning.")
+      out()
+      out("Best for:")
+      out("  - Math and numerical calculations")
+      out("  - Data processing and transformations")
+      out("  - Quick tests of logic or algorithms")
+      out("  - Anything that doesn't need filesystem or network access")
+      out()
+      out("Not for:")
+      out("  - Shell commands (use bash)")
+      out("  - Reading/writing files (use read_file / write_file)")
+      out("  - Internet requests (use web_fetch)")
+      out()
+      out("Toggle: /code on   /code off")
+      out()
+      out("Supported models: gemini-2.5-flash, gemini-2.5-pro, gemini-3.x and newer.")
+      return ReplSlashResult(skip_model_turn=True)
+
   # ── /computer ────────────────────────────────────────────────────────────
   if name in ("computer", "browser"):
     args_s = (sc.args or "").strip().lower()

{gemcode-0.3.28 → gemcode-0.3.29}/src/gemcode/session_runtime.py

@@ -30,6 +30,23 @@ def session_db_path(cfg: GemCodeConfig) -> Path:
   return cfg.project_root / ".gemcode" / "sessions.sqlite"
 
 
+def _build_artifact_service(cfg: GemCodeConfig):
+  """
+  Return an ADK ArtifactService for this session, or None if disabled.
+
+  Uses InMemoryArtifactService so artifacts are available within the session
+  without requiring GCS credentials. The agent can save screenshots, generated
+  files, large reports, etc. as artifacts to avoid bloating session history.
+  """
+  if not getattr(cfg, "enable_artifacts", True):
+    return None
+  try:
+    from google.adk.artifacts import InMemoryArtifactService
+    return InMemoryArtifactService()
+  except Exception:
+    return None
+
+
 def create_runner(cfg: GemCodeConfig, extra_tools: list | None = None) -> Runner:
   """Construct Runner + SQLite session service + root LlmAgent."""
   modality_tools = build_modality_extra_tools(cfg)
@@ -86,7 +103,9 @@ def create_runner(cfg: GemCodeConfig, extra_tools: list | None = None) -> Runner
     else:
       memory_service = FileMemoryService(mem_path)
 
-  return Runner(
+  artifact_service = _build_artifact_service(cfg)
+
+  runner_kwargs: dict = dict(
       app_name="gemcode",
       agent=agent,
       session_service=session_service,
@@ -94,3 +113,7 @@ def create_runner(cfg: GemCodeConfig, extra_tools: list | None = None) -> Runner
       memory_service=memory_service,
       auto_create_session=True,
   )
+  if artifact_service is not None:
+    runner_kwargs["artifact_service"] = artifact_service
+
+  return Runner(**runner_kwargs)

{gemcode-0.3.28 → gemcode-0.3.29}/src/gemcode/tool_prompt_manifest.py

@@ -126,6 +126,9 @@ def build_tool_manifest(cfg: GemCodeConfig) -> str | None:
   - Returns the sub-agent's final text as `result`.
   - Use for: context-heavy exploration (reading 50+ files), parallel investigation of independent subsystems, verification passes after you finish work.
   - Always give the sub-agent enough context to work independently; end the task with "Summarise your findings clearly."
+- **`get_user_choice(question, choices)`** — present the user with a structured list of options (2–6 items). Returns the user's selected option. Use instead of open-ended questions when the set of valid responses is known and bounded (e.g. framework choice, migration strategy, naming convention).
+- **`load_artifacts(filenames)`** — load one or more named artifacts (binary/large files) saved in this or a previous session. Returns the content of each. Use `user:filename` prefix for cross-session artifacts.
+- **`exit_loop()`** — signal the surrounding LoopAgent to stop iterating and emit the final result. Only relevant when running inside a LoopAgent pipeline (e.g. the `/refine` write→test→fix loop).
 
 ### Read-only tools (no permission needed): {_fmt_list(read_only)}
 Use proactively. Never use bash/run_command just to list or read files — these are instant and require zero approval.

{gemcode-0.3.28 → gemcode-0.3.29}/src/gemcode/tui/input_handler.py

@@ -56,6 +56,7 @@ SLASH_COMMANDS: list[tuple[str, str]] = [
     ("budget",      "Set per-turn token budget  ·  /budget <N>  ·  /budget off"),
     ("limits",      "Show/set execution limits (max_llm_calls, context, etc.)"),
     ("kairos",      "Background parallel job scheduler — how to run gemcode kairos"),
+    ("code",        "Toggle sandboxed Python code executor (ADK BuiltInCodeExecutor)"),
     ("tools",       "List all tools and their permission categories"),
     ("config",      "Show full active configuration (all fields)"),
     ("permissions", "Show current permission mode (default / strict / yes)"),

{gemcode-0.3.28 → gemcode-0.3.29/src/gemcode.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gemcode
-Version: 0.3.28
+Version: 0.3.29
 Summary: Local-first coding agent on Google Gemini + ADK
 Author: GemCode Contributors
 License:                                  Apache License

{gemcode-0.3.28 → gemcode-0.3.29}/src/gemcode.egg-info/SOURCES.txt

@@ -29,6 +29,7 @@ src/gemcode/model_routing.py
 src/gemcode/paths.py
 src/gemcode/permissions.py
 src/gemcode/prompt_suggestions.py
+src/gemcode/refine.py
 src/gemcode/repl_commands.py
 src/gemcode/repl_slash.py
 src/gemcode/session_runtime.py