gemcode 0.3.59__tar.gz → 0.3.65__tar.gz

{gemcode-0.3.59/src/gemcode.egg-info → gemcode-0.3.65}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gemcode
-Version: 0.3.59
+Version: 0.3.65
 Summary: Local-first coding agent on Google Gemini + ADK
 Author: GemCode Contributors
 License:                                  Apache License

{gemcode-0.3.59 → gemcode-0.3.65}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "gemcode"
-version = "0.3.59"
+version = "0.3.65"
 description = "Local-first coding agent on Google Gemini + ADK"
 readme = "README.md"
 requires-python = ">=3.11"

{gemcode-0.3.59 → gemcode-0.3.65}/src/gemcode/agent.py

@@ -29,6 +29,17 @@ from gemcode.tools import build_function_tools
 from gemcode.tool_prompt_manifest import build_tool_manifest
 
 
+def build_global_instruction() -> str:
+  """Global instruction applied to the entire agent tree (via ADK plugin)."""
+  return (
+    "You are GemCode, an expert software engineering agent powered by Google Gemini. "
+    "Think deeply about what the person actually wants before you do anything. "
+    "Use exactly as many tools as the task genuinely requires — no more. "
+    "Act fully and autonomously when action is needed. "
+    "Always use read-only tools before shell or write tools."
+  )
+
+
 def _chain_before_model_callbacks(*callbacks):
   cbs = [c for c in callbacks if c is not None]
   if not cbs:
@@ -197,7 +208,7 @@ def _build_runtime_facts(cfg: GemCodeConfig) -> str:
   if getattr(cfg, "enable_memory", False):
     mem_path = root / ".gemcode" / "memories.jsonl"
     mem_kind = "embedding-backed" if getattr(cfg, "enable_embeddings", False) else "keyword-backed"
-    caps.append(f"memory ON ({mem_kind}, stored at {mem_path}; ADK preload_memory injects relevant memories before each turn)")
+    caps.append(f"memory ON ({mem_kind}, stored at {mem_path}; ADK preload_memory auto-injects relevant memories before each turn; use load_memory(query) for explicit on-demand retrieval)")
   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):
@@ -515,11 +526,12 @@ You have native deep thinking capability — use it actively:
   - For **subfolders**: `bash("cargo build --release", cwd_subdir="backend")`
 
 - **Long-running servers / watchers** — use `bash` with `background=True`:
-  - `bash("npm run dev", background=True)` — start the dev server in background
+  - `bash("npm run dev", background=True)` — start the dev server in background → returns PID
   - `bash("python manage.py runserver", background=True)` — Django server
-  - `bash("tail -f logs/app.log", background=True)` — background log watcher
   - NEVER call `bash("npm run dev")` without `background=True` — it blocks forever and crashes the turn
-  - After starting a background process, confirm the port is ready with `bash("sleep 2 && curl -s http://localhost:3000 -o /dev/null && echo ready")`
+  - After starting: use `task_output(pid)` to read startup logs, then check if port is ready
+  - Use `list_tasks()` to see all running background processes
+  - Use `kill_task(pid)` to stop a background server when done
 
 - **`run_command`** — simple single-executable calls without shell features:
   - `run_command("npm", args=["install", "--legacy-peer-deps"])` — clean npm install
@@ -549,7 +561,21 @@ You have native deep thinking capability — use it actively:
 - **`move_file`** — rename or reorganize files/directories within the project.
 - **`delete_file`** — remove a single file.
 
+### Memory (when memory is enabled)
+- **`preload_memory`** — automatically injects relevant past memories before each turn (runs in background).
+- **`load_memory`** — explicit on-demand memory search:
+  - `load_memory("authentication patterns used in this project")` — recall specific knowledge
+  - `load_memory("previous bugs fixed in the auth module")` — targeted retrieval
+  - Use when the preloaded context is missing something specific you know you've seen before.
+
 ### Research and documentation
+- **`web_search`** — search the web without any API key or research mode:
+  - `web_search("python asyncio tutorial 2025")` — general search
+  - `web_search("fastapi jwt authentication example")` — find code examples
+  - `web_search("react 19 breaking changes")` — check recent releases
+  - Returns titles, URLs, and snippets. Follow with `web_fetch(url)` to read full content.
+  - Use this for quick lookups; use `/research on` for deep multi-page research.
+
 - **`web_fetch`** — fetch docs, APIs, changelogs, READMEs from the web:
   - `web_fetch("https://docs.python.org/3/library/asyncio.html")` — official docs
   - `web_fetch("https://api.github.com/repos/owner/repo/releases/latest")` — API data
@@ -566,6 +592,23 @@ You have native deep thinking capability — use it actively:
 - **`todo_write`** — track work items. Use for any task with 3+ steps.
   - Create at task start, mark completed as you finish, merge updates.
 
+- **`todo_read`** — read the current session todo list.
+  - Call this to check progress, find task ids for a merge update, or verify what's pending.
+
+- **`notebook_read`** — read a Jupyter notebook (.ipynb) as structured cells.
+  - Always prefer this over `read_file` for `.ipynb` files — gives clean cell-by-cell output.
+  - `notebook_read("analysis.ipynb")` — shows all cells with source and outputs.
+
+- **`notebook_edit`** — edit a cell in a Jupyter notebook:
+  - `notebook_edit("nb.ipynb", cell_index=2, new_source="import pandas as pd")` — replace cell
+  - `notebook_edit("nb.ipynb", cell_index=0, new_source="# Title", cell_type="markdown", edit_mode="insert")` — insert
+  - `notebook_edit("nb.ipynb", cell_index=3, new_source="", edit_mode="delete")` — delete cell
+
+- **Background task management** — for processes started with `bash(..., background=True)`:
+  - `list_tasks()` — see all background tasks (PID, command, status: running/finished)
+  - `task_output(pid)` — read stdout/stderr captured from a background task
+  - `kill_task(pid)` — gracefully stop a background task (use `force=True` for SIGKILL)
+
 - **`run_subtask`** — spawn an isolated sub-agent with its own fresh context window.
   - The sub-agent has the same tools (bash, read_file, grep, etc.) but starts from scratch.
   - Use when a task would bloat your context too much: e.g. "read all 40 test files and find patterns"
@@ -587,14 +630,34 @@ One user message = many model↔tool rounds (up to 256 LLM calls by default). Th
 
 **Do not stop after step 2 or 3** — complete the full task.
 
-## Parallelism — batch independent work
+## Parallelism — batch independent work aggressively
 Issue independent tool calls **in the same turn** when outputs don't depend on each other.
-This is faster and costs fewer turns. Concrete examples:
-- Reading multiple files → send all `read_file` calls together
-- Grepping different patterns → one message, multiple `grep_content` calls
-- `list_directory` + `glob_files` → issue both at once
-- Exploring multiple subsystems → one `run_subtask` per subsystem in one turn
-- `git status` and `git log` → chain with `&&` or issue in parallel
+This is always faster. **Default to parallel; only serialize when you must.**
+
+Concrete patterns:
+
+**Parallel file exploration (always do this):**
+- Reading multiple files → emit all `read_file` calls in one turn, not one by one
+- Grepping different patterns → multiple `grep_content` in one response
+- `list_directory` + `glob_files` → both at once
+
+**Parallel sub-agent exploration (OpenClaude pattern):**
+When a task requires understanding several subsystems before acting:
+1. Spawn parallel `run_subtask` workers, one per subsystem
+2. Wait for all results to return in the same turn
+3. Synthesise findings and execute the change
+
+Example — understanding a codebase before a big refactor:
+```
+run_subtask("Analyse src/auth/ — how does authentication flow work? List all key files and patterns.")
+run_subtask("Analyse src/api/ — what endpoints exist? How are they protected?")
+run_subtask("Analyse tests/auth* — what is the test coverage for auth?")
+```
+All three run concurrently. Then synthesise and act.
+
+**Parallel git + build:**
+- `git status && git diff --stat` → one bash call
+- Running lint + type-check → `npm run lint && npm run typecheck` in one call
 
 Sequential only when step B genuinely needs step A's output.
 
@@ -642,10 +705,36 @@ When `code_executor ON` (see Runtime facts above):
 - 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.)
 
+## Verification contract (mandatory for non-trivial tasks)
+
+After completing any implementation that touches **3 or more files**, introduces a new feature, or fixes a bug, you **MUST** run a verification pass before calling the task done.
+
+**How to verify:**
+
+Option A — Run tests/build (preferred when tests exist):
+```
+bash("npm run build 2>&1 | tail -30")
+bash("pytest tests/ -x -q --tb=short 2>&1 | head -80")
+```
+
+Option B — Spawn a verification sub-agent (for complex multi-file changes):
+```
+run_subtask(
+  task="You are a strict code reviewer. Verify the following changes are correct, complete, and consistent. Check: (1) syntax errors, (2) logic bugs, (3) broken imports, (4) missing edge cases, (5) consistency across all modified files. Report PASS or FAIL with specific findings.",
+  context="Files changed: [list them]. Change summary: [what you did]."
+)
+```
+
+**Rules:**
+- If verification finds issues → fix them → verify again. Never stop at a failed verification.
+- Only report "done" after a clean verification pass.
+- For destructive changes (delete, refactor) always run both Option A and Option B.
+- For simple single-file edits, a quick `bash("python3 -c 'import <module>'")` or syntax check is sufficient.
+
 ## Evaluator-optimizer loop
 For tasks where quality matters:
 1. Complete the task (execute tools, write code, run commands)
-2. Spawn a verification `run_subtask` or use `bash` to run tests/lint
+2. Verify — run tests, build, or spawn a verification sub-agent (see Verification contract above)
 3. If verification fails, read the error, fix, re-verify
 4. Report done only when verified
 
@@ -704,7 +793,9 @@ Use `gh pr create` via `bash`. When asked to create a PR:
 - Do NOT retry failing commands in a sleep loop — diagnose the root cause first
 
 ## Communication
-- One short line before the first tool call in a turn (e.g. "Reading the auth module and checking the test suite...").
+- **ACT FIRST, narrate after.** Do NOT write out a multi-step numbered plan as prose and then stop. Execute immediately — use tools right away.
+- One short line before the first tool call is fine (e.g. "Reading the auth module..."). That's it. No verbose announcements.
+- If you want to plan, use the **`think` tool privately** — never dump a plan into your text response before acting. The user cannot run plan text; they need results.
 - Summarize tool results in plain language — the user doesn't see raw tool internals.
 - After completing a task: clear summary of what changed, where, and why.
 - If the user pastes UI copy / noise / error output, extract the real intent and act on source files.
@@ -851,22 +942,10 @@ def build_root_agent(
       tool_config=tool_cfg,
     )
 
-  # 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. "
-    "Think deeply about what the person actually wants before you do anything. "
-    "Use exactly as many tools as the task genuinely requires — no more. "
-    "Act fully and autonomously when action is needed. "
-    "Always use read-only tools before shell or 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,
@@ -876,6 +955,17 @@ def build_root_agent(
   if code_executor is not None:
     agent_kwargs["code_executor"] = code_executor
 
+  # Optional: ADK PlanReActPlanner — injects a structured "plan then act" pass
+  # into every turn at the framework level (not just via prompting).
+  # Enable with: GEMCODE_PLANREACT=1
+  import os as _os
+  if _os.environ.get("GEMCODE_PLANREACT", "").lower() in ("1", "true", "yes", "on"):
+    try:
+      from google.adk.planners import PlanReActPlanner
+      agent_kwargs["planner"] = PlanReActPlanner()
+    except Exception:
+      pass
+
   return LlmAgent(**agent_kwargs)
 
 

{gemcode-0.3.59 → gemcode-0.3.65}/src/gemcode/callbacks.py

@@ -626,29 +626,44 @@ def make_on_tool_error_callback(cfg: GemCodeConfig):
 
 
 def make_on_model_error_callback(cfg: GemCodeConfig):
-  """Structured model errors to the user + audit trail."""
+  """Structured model errors to the user + audit trail.
+
+  For transient errors (HTTP 503, 429, server-overloaded) we return None so the
+  exception propagates to invoke.py, which retries with exponential backoff.
+  For permanent errors we absorb and return a user-friendly LlmResponse.
+  """
 
   async def on_model_error(*, callback_context, llm_request, error: Exception):
-    try:
-      st = callback_context.state
-      if st is not None and not st.get(TERMINAL_REASON_KEY):
-        st[TERMINAL_REASON_KEY] = "model_error"
-    except Exception:
-      pass
+    from gemcode.model_errors import is_transient_error
+
     append_audit(
         cfg.project_root,
         {
             "phase": "model_exception",
             "error": f"{type(error).__name__}: {error}",
+            "transient": is_transient_error(error),
         },
     )
+
+    # Transient errors (503, 429, server-overloaded): let the exception propagate
+    # so invoke.py can retry with backoff. Do NOT set terminal state here — the
+    # turn is not over yet.
+    if is_transient_error(error):
+      return None
+
+    # Permanent errors: mark session terminal and return a user-friendly message.
+    try:
+      st = callback_context.state
+      if st is not None and not st.get(TERMINAL_REASON_KEY):
+        st[TERMINAL_REASON_KEY] = "model_error"
+    except Exception:
+      pass
+
     if _truthy_env("GEMCODE_VERBOSE_MODEL_ERRORS", default=False):
       import traceback
-
       traceback.print_exception(type(error), error, error.__traceback__, file=sys.stderr)
 
     user_text = format_model_error_for_user(error)
-    # Scrollback/TUI already prints "GemCode:" before assistant text — avoid "GemCode: GemCode:".
     from google.adk.models.llm_response import LlmResponse
     from google.genai import types
 

{gemcode-0.3.59 → gemcode-0.3.65}/src/gemcode/config.py

@@ -245,10 +245,10 @@ class GemCodeConfig:
   )
 
   # Controls how the TUI renders model thinking: True = full Rich Markdown,
-  # False = collapsed one-line excerpt (default, like OpenClaude).
+  # False = collapsed one-line excerpt.
   # Toggled at runtime via /thinking verbose|brief.
   show_full_thinking: bool = field(
-    default_factory=lambda: _truthy_env("GEMCODE_SHOW_FULL_THINKING", default=False)
+    default_factory=lambda: _truthy_env("GEMCODE_SHOW_FULL_THINKING", default=True)
   )
 
   # Enable ADK BuiltInCodeExecutor for safe sandboxed Python execution via

{gemcode-0.3.59 → gemcode-0.3.65}/src/gemcode/invoke.py

@@ -6,6 +6,7 @@ CLI and tests call `run_turn` with a Runner already bound to app + session servi
 
 from __future__ import annotations
 
+import asyncio
 import os
 import sys
 from typing import Any
@@ -16,6 +17,11 @@ from google.adk.runners import Runner
 from google.genai import types
 
 
+# Delays (seconds) between successive transient-error retries: 2s, 5s, 12s.
+# Three retries = up to ~19 seconds of total wait before giving up.
+_TRANSIENT_RETRY_DELAYS = [2.0, 5.0, 12.0]
+
+
 _HITL_PROMPT_LOCK = Lock()
 
 
@@ -160,11 +166,43 @@ async def run_turn(
       # Runner handoff loop: if tools request confirmations, we pause here to
       # ask HITL, then send back function responses so ADK can re-execute the
       # tools.
+      #
+      # Transient API errors (HTTP 503, 429) are retried here with exponential
+      # backoff. on_model_error returns None for these, so the exception
+      # propagates from runner.run_async and we catch it below.
       do_reset = True
+      transient_attempts = 0
       while True:
-        events = await _await_runner_events(
-          next_message=current_message, do_reset=do_reset
-        )
+        try:
+          events = await _await_runner_events(
+            next_message=current_message, do_reset=do_reset
+          )
+        except Exception as _exc:
+          from gemcode.model_errors import is_transient_error
+          if is_transient_error(_exc) and transient_attempts < len(_TRANSIENT_RETRY_DELAYS):
+            delay = _TRANSIENT_RETRY_DELAYS[transient_attempts]
+            transient_attempts += 1
+            _tui_active = os.environ.get("GEMCODE_TUI_ACTIVE", "0").lower() in ("1", "true", "yes", "on")
+            _msg = (
+              f"\n[gemcode] Transient API error ({type(_exc).__name__}). "
+              f"Retrying in {delay:.0f}s (attempt {transient_attempts}/{len(_TRANSIENT_RETRY_DELAYS)})...\n"
+            )
+            print(_msg, file=sys.stderr)
+            # Surface retry notice in TUI if available.
+            if _tui_active:
+              try:
+                from gemcode.tui import scrollback as _sb
+                _sb._transient_retry_notice = _msg  # type: ignore[attr-defined]
+              except Exception:
+                pass
+            await asyncio.sleep(delay)
+            # Retry the same message from scratch (session history is intact in SQLite).
+            continue
+          # Non-transient or out of retries: re-raise so the TUI surfaces it.
+          raise
+
+        # Reset transient counter after a successful model call.
+        transient_attempts = 0
         collected.extend(events)
 
         confirmation_fcs = _get_confirmation_requests(events)

{gemcode-0.3.59 → gemcode-0.3.65}/src/gemcode/model_errors.py

@@ -5,6 +5,38 @@ from __future__ import annotations
 import re
 
 
+def is_transient_error(error: Exception) -> bool:
+  """Return True for HTTP 503 / 429 and similar transient API errors that are safe to retry.
+
+  Transient means: the request was fine, the server was temporarily unavailable or
+  rate-limited. Retrying the same request (with backoff) will likely succeed.
+  """
+  try:
+    from google.genai import errors as genai_errors
+    if isinstance(error, genai_errors.APIError):
+      code = int(getattr(error, "code", None) or 0) or None
+      if code in (429, 503):
+        return True
+      # Some 500-range server errors are also transient (502 Bad Gateway, etc.)
+      if code is not None and 500 <= code < 600 and code not in (400, 401, 403, 404):
+        return True
+  except Exception:
+    pass
+
+  # gRPC / google-api-core equivalents
+  et = type(error).__name__
+  if "ResourceExhausted" in et or "ServiceUnavailable" in et or "DeadlineExceeded" in et:
+    return True
+
+  msg = str(error)
+  # Match the specific phrases Gemini uses in 503 responses
+  if "503" in msg and any(p in msg for p in ("high demand", "service unavailable", "overloaded")):
+    return True
+  if "429" in msg and any(p in msg for p in ("rate limit", "quota", "resource exhausted")):
+    return True
+  return False
+
+
 def _sanitize_api_text(s: str) -> str:
   """Strip likely API key material from strings shown to the user."""
   if not s:

{gemcode-0.3.59 → gemcode-0.3.65}/src/gemcode/plugins/terminal_hooks_plugin.py

@@ -141,6 +141,19 @@ class GemCodeTerminalHooksPlugin(BasePlugin):
         },
       )
 
+      # Surface suggestion to the TUI by storing it on cfg.
+      # The TUI reads cfg._last_prompt_suggestion after each turn and displays it.
+      try:
+        object.__setattr__(self.cfg, "_last_prompt_suggestion", suggestion)
+      except Exception:
+        pass
+    else:
+      # Clear any stale suggestion from the previous turn.
+      try:
+        object.__setattr__(self.cfg, "_last_prompt_suggestion", None)
+      except Exception:
+        pass
+
     if getattr(self.cfg, "enable_memory", False):
       try:
         await callback_context.add_session_to_memory()

{gemcode-0.3.59 → gemcode-0.3.65}/src/gemcode/repl_commands.py

@@ -253,7 +253,7 @@ def slash_help_lines() -> list[str]:
       "  Thinking:",
       "  /thinking             Show current thinking config",
       "  /thinking verbose     Show full thinking text each turn",
-      "  /thinking brief       Show collapsed one-line excerpt (default)",
+      "  /thinking brief       Show collapsed one-line excerpt",
       "  /thinking off         Disable model thinking",
       "  /thinking on          Re-enable thinking (auto budget/level)",
       "  /thinking budget <N>  Set thinking token budget (Gemini 2.5, 0=off, -1=dynamic)",

{gemcode-0.3.59 → gemcode-0.3.65}/src/gemcode/repl_slash.py

@@ -1001,7 +1001,7 @@ async def process_repl_slash(
         out("  /thinking level <minimal|low|medium|high>")
       out("Display commands (all models):")
       out("  /thinking verbose  — show full thinking text each turn")
-      out("  /thinking brief    — show collapsed one-line excerpt (default)")
+      out("  /thinking brief    — show collapsed one-line excerpt")
       out()
       return ReplSlashResult(skip_model_turn=True)
 
@@ -1016,7 +1016,7 @@ async def process_repl_slash(
 
     if sub in ("brief", "short", "collapsed"):
       cfg.show_full_thinking = False
-      out("thinking display: brief — collapsed one-line excerpt (default)")
+      out("thinking display: brief — collapsed one-line excerpt")
       out()
       return ReplSlashResult(skip_model_turn=True)
 

{gemcode-0.3.59 → gemcode-0.3.65}/src/gemcode/session_runtime.py

@@ -22,7 +22,7 @@ warnings.filterwarnings("ignore", category=UserWarning, message=".*EXPERIMENTAL.
 from google.adk.runners import Runner
 from google.adk.sessions.sqlite_session_service import SqliteSessionService
 
-from gemcode.agent import build_root_agent
+from gemcode.agent import build_global_instruction, build_root_agent
 from gemcode.config import GemCodeConfig
 from gemcode.modality_tools import build_extra_tools as build_modality_extra_tools
 from gemcode.memory.embedding_memory_service import EmbeddingFileMemoryService
@@ -31,6 +31,52 @@ from gemcode.plugins.terminal_hooks_plugin import GemCodeTerminalHooksPlugin
 from gemcode.plugins.tool_recovery_plugin import GemCodeReflectAndRetryToolPlugin
 
 
+# ---------------------------------------------------------------------------
+# ADK App-level feature helpers
+# ---------------------------------------------------------------------------
+
+def _build_context_cache_config():
+  """Return ContextCacheConfig if context caching is enabled, else None.
+
+  Context caching lets Gemini reuse the compiled representation of a stable
+  prefix (system prompt + tools) across multiple turns, cutting ~75% of input
+  token costs on long sessions.
+
+  Opt-out: set ``GEMCODE_CONTEXT_CACHE=0`` in the environment.
+  """
+  if os.environ.get("GEMCODE_CONTEXT_CACHE", "1").lower() in ("0", "false", "no", "off"):
+    return None
+  try:
+    from google.adk.agents.context_cache_config import ContextCacheConfig
+    return ContextCacheConfig(
+      cache_intervals=10,   # refresh the cache every 10 invocations
+      ttl_seconds=1800,     # cache lives 30 minutes
+      min_tokens=1024,      # skip caching tiny sessions (< ~1 K tokens)
+    )
+  except Exception:
+    return None
+
+
+def _build_app(agent, plugins, cfg: GemCodeConfig):
+  """Wrap the root agent in an ADK App for modern plugin + context-cache support.
+
+  Using ``App`` instead of passing ``agent`` + ``plugins`` directly to ``Runner``
+  is the recommended ADK pattern as of ADK 1.x (``plugins=`` on ``Runner`` is
+  officially deprecated).
+  """
+  try:
+    from google.adk.apps.app import App
+    return App(
+      name="gemcode",
+      root_agent=agent,
+      plugins=plugins,
+      context_cache_config=_build_context_cache_config(),
+    )
+  except Exception:
+    # Fall back silently — Runner still accepts the legacy kwargs.
+    return None
+
+
 def session_db_path(cfg: GemCodeConfig) -> Path:
   return cfg.project_root / ".gemcode" / "sessions.sqlite"
 
@@ -242,15 +288,23 @@ def _make_safe_computer_toolset(computer):
 
 
 def _build_artifact_service(cfg: GemCodeConfig):
-  """
-  Return an ADK ArtifactService for this session, or None if disabled.
+  """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.
+  Uses ``FileArtifactService`` backed by ``.gemcode/artifacts/`` so that
+  artifacts (screenshots, generated reports, diffs, etc.) survive session
+  restarts.  Falls back to ``InMemoryArtifactService`` if the file-based
+  service is unavailable (older ADK).
   """
   if not getattr(cfg, "enable_artifacts", True):
     return None
+  try:
+    from google.adk.artifacts import FileArtifactService
+    artifacts_dir = cfg.project_root / ".gemcode" / "artifacts"
+    artifacts_dir.mkdir(parents=True, exist_ok=True)
+    return FileArtifactService(root_dir=artifacts_dir)
+  except Exception:
+    pass
+  # Fallback for older ADK versions that don't have FileArtifactService.
   try:
     from google.adk.artifacts import InMemoryArtifactService
     return InMemoryArtifactService()
@@ -344,11 +398,32 @@ def create_runner(cfg: GemCodeConfig, extra_tools: list | None = None) -> Runner
   db.parent.mkdir(parents=True, exist_ok=True)
   session_service = SqliteSessionService(str(db))
 
-  plugins = [GemCodeTerminalHooksPlugin(cfg)]
-  # Place recovery plugin before terminal hooks so it can influence tool results
-  # during the invocation.
-  if True:
-    plugins.insert(0, GemCodeReflectAndRetryToolPlugin(cfg))
+  # ── Plugins ──────────────────────────────────────────────────────────────
+  # Recovery plugin first so it can intercept tool errors before terminal hooks.
+  plugins = [GemCodeReflectAndRetryToolPlugin(cfg), GemCodeTerminalHooksPlugin(cfg)]
+
+  # Global instruction is now applied via ADK's GlobalInstructionPlugin (the
+  # modern replacement for the deprecated LlmAgent.global_instruction field).
+  try:
+    from google.adk.plugins.global_instruction_plugin import GlobalInstructionPlugin
+    plugins.insert(0, GlobalInstructionPlugin(build_global_instruction()))
+  except Exception:
+    pass
+
+  # Optional: rich YAML debug log (every LLM request/response + tool calls).
+  # Enable with: GEMCODE_DEBUG_LOG=1
+  if os.environ.get("GEMCODE_DEBUG_LOG", "").lower() in ("1", "true", "yes", "on"):
+    try:
+      from google.adk.plugins.debug_logging_plugin import DebugLoggingPlugin
+      debug_log_path = cfg.project_root / ".gemcode" / "debug.yaml"
+      plugins.append(DebugLoggingPlugin(
+        output_path=str(debug_log_path),
+        include_session_state=True,
+      ))
+    except Exception:
+      pass
+
+  # ── Memory service ────────────────────────────────────────────────────────
   memory_service = None
   if getattr(cfg, "enable_memory", False):
     mem_path = cfg.project_root / ".gemcode" / "memories.jsonl"
@@ -361,14 +436,31 @@ def create_runner(cfg: GemCodeConfig, extra_tools: list | None = None) -> Runner
 
   artifact_service = _build_artifact_service(cfg)
 
-  runner_kwargs: dict = dict(
-      app_name="gemcode",
-      agent=agent,
-      session_service=session_service,
-      plugins=plugins,
-      memory_service=memory_service,
-      auto_create_session=True,
-  )
+  # ── Runner via ADK App (modern pattern) ──────────────────────────────────
+  # App is the recommended top-level container as of ADK 1.x.  It owns the
+  # plugin list and context-cache config so Runner stays clean.
+  # ``plugins=`` on Runner is officially deprecated; using App avoids the
+  # DeprecationWarning and enables context caching + future App-level features.
+  app = _build_app(agent, plugins, cfg)
+
+  if app is not None:
+    runner_kwargs: dict = dict(
+        app=app,
+        session_service=session_service,
+        memory_service=memory_service,
+        auto_create_session=True,
+    )
+  else:
+    # Legacy fallback if App is unavailable (very old ADK installs).
+    runner_kwargs = dict(
+        app_name="gemcode",
+        agent=agent,
+        session_service=session_service,
+        plugins=plugins,
+        memory_service=memory_service,
+        auto_create_session=True,
+    )
+
   if artifact_service is not None:
     runner_kwargs["artifact_service"] = artifact_service