gemcode 0.3.28__tar.gz → 0.3.30__tar.gz

{gemcode-0.3.28/src/gemcode.egg-info → gemcode-0.3.30}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gemcode
-Version: 0.3.28
+Version: 0.3.30
 Summary: Local-first coding agent on Google Gemini + ADK
 Author: GemCode Contributors
 License:                                  Apache License

{gemcode-0.3.28 → gemcode-0.3.30}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "gemcode"
-version = "0.3.28"
+version = "0.3.30"
 description = "Local-first coding agent on Google Gemini + ADK"
 readme = "README.md"
 requires-python = ">=3.11"

{gemcode-0.3.28 → gemcode-0.3.30}/src/gemcode/agent.py

@@ -66,6 +66,8 @@ def _build_runtime_facts(cfg: GemCodeConfig) -> str:
 
   # ── Active capabilities ──────────────────────────────────────────────────
   caps: list[str] = []
+  if getattr(cfg, "enable_web_search", False) and not getattr(cfg, "enable_deep_research", False):
+    caps.append("web_search ON (tool: google_search — standalone search without full deep_research)")
   if getattr(cfg, "enable_deep_research", False):
     dr_extras = " + google_maps_grounding" if getattr(cfg, "enable_maps_grounding", False) else ""
     caps.append(f"deep_research ON (tools: google_search, url_context{dr_extras})")
@@ -77,8 +79,12 @@ 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 ───────────────────────────────────────────────────────────────
@@ -147,6 +153,40 @@ Memory is **ON** ({kind}). Stored at: `{mem_path}`
 """
 
 
+def _build_plan_mode_section() -> str:
+  """Injected when plan_mode=True — instructs agent to write explicit plans first."""
+  return """
+## PLAN MODE IS ACTIVE
+
+You are currently in **Plan Mode**. Before executing ANY tools that modify files or run shell commands, you MUST:
+
+1. **Write a numbered plan** in your response text — list every step you intend to take.
+   Example:
+   ```
+   Plan:
+   1. Read src/auth/login.ts to understand the current flow
+   2. Read src/types/user.ts for the User interface
+   3. Add `lastLogin: Date` field to User interface
+   4. Update login handler to set lastLogin on successful auth
+   5. Run `npm run build` to verify no TypeScript errors
+   ```
+
+2. **Pause after the plan** — do not immediately execute tools. Present the plan and wait for the user to confirm ("go", "proceed", "looks good") before starting.
+
+3. **Stick to the plan** — if you discover the plan needs changing mid-execution, note the update before proceeding.
+
+4. **Report completion** against the plan — when done, confirm each step as completed.
+
+**Why plan mode?**
+- Prevents unintended side effects from premature tool execution
+- Gives you and the user visibility into the full scope before any changes
+- Makes complex multi-file tasks reviewable and reversible
+- Catches scope creep early
+
+**To turn off plan mode**, type `/plan off` at the prompt.
+"""
+
+
 def _build_computer_use_section(cfg: GemCodeConfig) -> str:
   """Rich computer use guidance, only injected when enable_computer_use=True."""
   w = getattr(cfg, "_cfg", None)
@@ -426,6 +466,41 @@ 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)
@@ -461,6 +536,9 @@ All file tools use paths **relative to the project root** (where GemCode was sta
   if getattr(cfg, "enable_memory", False):
     base = f"{base}\n\n{_build_memory_section(cfg)}"
 
+  if getattr(cfg, "plan_mode", False):
+    base = f"{base}\n\n{_build_plan_mode_section()}"
+
   tool_manifest = build_tool_manifest(cfg)
   if tool_manifest:
     base = f"{base}\n\n{tool_manifest}"
@@ -470,6 +548,17 @@ 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 +577,19 @@ 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 +644,31 @@ 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-0.3.30}/src/gemcode/callbacks.py

@@ -141,6 +141,22 @@ def make_before_tool_callback(cfg: GemCodeConfig):
     record = {"tool": name, "args": _redact_args(name, args)}
     append_audit(cfg.project_root, record)
 
+    # ── Shell hooks: pre_tool_use ─────────────────────────────────────────
+    # If the project has a .gemcode/hooks/pre_tool_use.sh, run it now.
+    # Non-zero exit or {"decision":"deny"} stdout will block the tool call.
+    try:
+      from gemcode.hooks import run_pre_tool_use_hook
+      hook_result = run_pre_tool_use_hook(
+          cfg.project_root,
+          model=getattr(cfg, "model", "") or "",
+          tool_name=name,
+          args=args or {},
+      )
+      if hook_result is not None:
+        return hook_result
+    except Exception:
+      pass
+
     streak = 0
     if tool_context is not None:
       try:
@@ -316,6 +332,19 @@ def make_after_tool_callback(cfg: GemCodeConfig):
         st[_STATE_FAILURE_KEY] = 0
     else:
       st[_STATE_FAILURE_KEY] = 0
+    # ── Shell hooks: post_tool_use ────────────────────────────────────────
+    try:
+      from gemcode.hooks import run_post_tool_use_hook
+      run_post_tool_use_hook(
+          cfg.project_root,
+          model=getattr(cfg, "model", "") or "",
+          tool_name=name,
+          args=args or {},
+          result=tool_response if isinstance(tool_response, dict) else {},
+      )
+    except Exception:
+      pass
+
     if _maybe_tool_summary_enabled():
       summary: dict[str, Any] = {
         "phase": "tool_result",
@@ -399,6 +428,7 @@ def make_after_model_callback(cfg: GemCodeConfig):
           "candidates_token_count",
           "cached_content_token_count",
           "total_token_count",
+          "thoughts_token_count",
       ):
         if hasattr(um, attr):
           v = getattr(um, attr)
@@ -407,6 +437,44 @@ def make_after_model_callback(cfg: GemCodeConfig):
     if d:
       append_audit(cfg.project_root, {"phase": "model_usage", **d})
 
+    # ── Expose live token stats to the TUI ───────────────────────────────────
+    # The TUI reads cfg._last_turn_stats after each turn to display token counts
+    # and estimated cost in the footer (like OpenClaude's spinner token display).
+    try:
+      in_tok  = d.get("prompt_token_count", 0) or 0
+      out_tok = d.get("candidates_token_count", 0) or 0
+      think_tok = d.get("thoughts_token_count", 0) or 0
+      cache_tok = d.get("cached_content_token_count", 0) or 0
+      total_tok = d.get("total_token_count", 0) or 0
+
+      prev_session_tokens = int(st.get(SESSION_TOTAL_TOKENS_KEY, 0) or 0)
+      session_total = prev_session_tokens + total_tok
+
+      from gemcode.pricing import estimate_cost
+      turn_cost = estimate_cost(
+          getattr(cfg, "model", "") or "",
+          input_tokens=in_tok,
+          output_tokens=out_tok,
+      )
+      # Accumulate session cost
+      prev_cost = getattr(cfg, "_session_cost_usd", 0.0) or 0.0
+      session_cost = prev_cost + (turn_cost or 0.0)
+      object.__setattr__(cfg, "_session_cost_usd", session_cost)
+
+      stats: dict[str, Any] = {
+          "in": in_tok,
+          "out": out_tok,
+          "think": think_tok,
+          "cache": cache_tok,
+          "total": total_tok,
+          "session_total": session_total,
+          "turn_cost": turn_cost,
+          "session_cost": session_cost,
+      }
+      object.__setattr__(cfg, "_last_turn_stats", stats)
+    except Exception:
+      pass
+
     pt = d.get("prompt_token_count")
     if isinstance(pt, int) and pt >= 0:
       try:

{gemcode-0.3.28 → gemcode-0.3.30}/src/gemcode/config.py

@@ -251,6 +251,37 @@ 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)
+  )
+
+  # Plan mode: when ON, the agent explicitly writes out a numbered plan
+  # BEFORE executing any tools, then checks the plan before reporting done.
+  # Like OpenClaude's EnterPlanMode — great for complex, multi-file tasks.
+  # Toggle at runtime with /plan on|off.
+  plan_mode: bool = field(
+    default_factory=lambda: _truthy_env("GEMCODE_PLAN_MODE", default=False)
+  )
+
+  # Always-on web search (independent of deep_research / research mode).
+  # When True, google_search is available as a basic tool without enabling
+  # the full deep_research capability suite (url_context, maps, etc.).
+  enable_web_search: bool = field(
+    default_factory=lambda: _truthy_env("GEMCODE_ENABLE_WEB_SEARCH", default=False)
+  )
+
   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-0.3.30/src/gemcode/hooks.py

@@ -0,0 +1,194 @@
+"""
+GemCode shell hooks — scriptable lifecycle events inspired by OpenClaude.
+
+Hooks are shell scripts stored under ``<project_root>/.gemcode/hooks/``.
+They receive a JSON payload on stdin and can:
+  - pre_tool_use.sh:  return non-zero exit to DENY the tool call, or
+                      print JSON {"decision":"deny","reason":"..."} to stderr.
+  - post_tool_use.sh: informational; return value ignored.
+  - session_start.sh: runs when a new GemCode session starts.
+  - session_stop.sh:  runs when the session ends.
+
+HOOK ENVIRONMENT
+  All hooks receive these env vars:
+    GEMCODE_HOOK_TYPE       — "pre_tool_use" | "post_tool_use" | "session_start" | "session_stop"
+    GEMCODE_PROJECT_ROOT    — absolute path to the project root
+    GEMCODE_MODEL           — active model id
+
+HOOK STDIN (pre_tool_use, post_tool_use)
+  JSON object with:
+    { "tool": "<tool_name>",
+      "args": { ...tool_args... },
+      "type": "pre_tool_use" | "post_tool_use",
+      "result": { ...tool_result... }   // post only
+    }
+
+PRE_TOOL_USE DECISION
+  If exit code is non-zero → the tool is DENIED.
+  If exit code is 0 → the tool proceeds normally.
+  If stdout starts with '{' and contains "decision": "deny" → the tool is denied.
+
+EXAMPLE HOOKS
+  # .gemcode/hooks/pre_tool_use.sh
+  #!/bin/bash
+  # Deny all delete_file calls
+  HOOK=$(cat)
+  TOOL=$(echo "$HOOK" | python3 -c "import sys,json; d=json.load(sys.stdin); print(d['tool'])")
+  if [ "$TOOL" = "delete_file" ]; then
+    echo "delete_file blocked by hook" >&2
+    exit 1
+  fi
+
+  # .gemcode/hooks/session_start.sh
+  #!/bin/bash
+  echo "[gemcode hook] session started in $GEMCODE_PROJECT_ROOT" >> /tmp/gemcode-audit.log
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import subprocess
+from pathlib import Path
+from typing import Any
+
+log = logging.getLogger(__name__)
+
+_HOOKS_DIR = ".gemcode/hooks"
+_TIMEOUT_PRE = 5.0   # seconds — pre_tool_use must respond quickly
+_TIMEOUT_POST = 10.0  # seconds — post_tool_use has more budget
+
+
+def _find_hook(project_root: Path, name: str) -> Path | None:
+    """Return the hook file path if it exists and is executable, else None."""
+    hooks_dir = project_root / _HOOKS_DIR
+    for ext in ("", ".sh", ".py", ".bash"):
+        candidate = hooks_dir / f"{name}{ext}"
+        if candidate.is_file() and os.access(candidate, os.X_OK):
+            return candidate
+    return None
+
+
+def _run_hook(
+    hook_path: Path,
+    project_root: Path,
+    model: str,
+    hook_type: str,
+    stdin_data: dict | None = None,
+    timeout: float = 5.0,
+) -> tuple[int, str, str]:
+    """
+    Run a hook script.
+
+    Returns (exit_code, stdout, stderr).
+    Never raises — all exceptions are caught and logged.
+    """
+    try:
+        env = {
+            **os.environ,
+            "GEMCODE_HOOK_TYPE": hook_type,
+            "GEMCODE_PROJECT_ROOT": str(project_root),
+            "GEMCODE_MODEL": model or "",
+        }
+        stdin_bytes = (
+            json.dumps(stdin_data, ensure_ascii=False).encode() if stdin_data else b""
+        )
+        result = subprocess.run(
+            [str(hook_path)],
+            input=stdin_bytes,
+            capture_output=True,
+            timeout=timeout,
+            env=env,
+            cwd=str(project_root),
+        )
+        return result.returncode, result.stdout.decode(errors="replace"), result.stderr.decode(errors="replace")
+    except subprocess.TimeoutExpired:
+        log.warning("[hooks] %s timed out after %.1fs", hook_path.name, timeout)
+        return 1, "", f"hook timed out after {timeout}s"
+    except Exception as exc:
+        log.debug("[hooks] failed to run %s: %s", hook_path.name, exc)
+        return 1, "", str(exc)
+
+
+def run_pre_tool_use_hook(
+    project_root: Path,
+    model: str,
+    tool_name: str,
+    args: dict[str, Any],
+) -> dict | None:
+    """
+    Run the pre_tool_use hook (if it exists).
+
+    Returns:
+        None  → tool is allowed (default)
+        dict  → {"error": "...", "error_kind": "hook_denied"} to deny the tool
+    """
+    hook = _find_hook(project_root, "pre_tool_use")
+    if hook is None:
+        return None
+
+    payload = {"tool": tool_name, "args": args, "type": "pre_tool_use"}
+    rc, stdout, stderr = _run_hook(
+        hook, project_root, model, "pre_tool_use", payload, timeout=_TIMEOUT_PRE
+    )
+
+    if rc != 0:
+        reason = stderr.strip() or stdout.strip() or f"hook exited {rc}"
+        log.info("[hooks] pre_tool_use denied %s: %s", tool_name, reason[:200])
+        return {
+            "error": f"Tool denied by pre_tool_use hook: {reason[:400]}",
+            "error_kind": "hook_denied",
+        }
+
+    # Check if stdout contains an explicit deny decision.
+    if stdout.strip().startswith("{"):
+        try:
+            decision = json.loads(stdout.strip())
+            if str(decision.get("decision", "")).lower() == "deny":
+                reason = decision.get("reason", "hook denied")
+                log.info("[hooks] pre_tool_use JSON-denied %s: %s", tool_name, reason)
+                return {
+                    "error": f"Tool denied by hook: {reason}",
+                    "error_kind": "hook_denied",
+                }
+        except Exception:
+            pass
+
+    return None  # allowed
+
+
+def run_post_tool_use_hook(
+    project_root: Path,
+    model: str,
+    tool_name: str,
+    args: dict[str, Any],
+    result: dict[str, Any],
+) -> None:
+    """Run the post_tool_use hook if it exists. Return value is ignored."""
+    hook = _find_hook(project_root, "post_tool_use")
+    if hook is None:
+        return
+    payload = {
+        "tool": tool_name,
+        "args": args,
+        "result": result,
+        "type": "post_tool_use",
+    }
+    _run_hook(hook, project_root, model, "post_tool_use", payload, timeout=_TIMEOUT_POST)
+
+
+def run_session_start_hook(project_root: Path, model: str) -> None:
+    """Run the session_start hook if it exists."""
+    hook = _find_hook(project_root, "session_start")
+    if hook is None:
+        return
+    _run_hook(hook, project_root, model, "session_start", timeout=10.0)
+
+
+def run_session_stop_hook(project_root: Path, model: str) -> None:
+    """Run the session_stop hook if it exists."""
+    hook = _find_hook(project_root, "session_stop")
+    if hook is None:
+        return
+    _run_hook(hook, project_root, model, "session_stop", timeout=10.0)

{gemcode-0.3.28 → gemcode-0.3.30}/src/gemcode/modality_tools.py

@@ -172,9 +172,23 @@ def build_extra_tools(cfg: GemCodeConfig) -> list[Any]:
   """Return ADK tool unions to expose for enabled modalities."""
   extra: list[Any] = []
 
+  # ── Web search (standalone, no full deep_research needed) ────────────────
+  # enable_web_search=True adds google_search alone.
+  # enable_deep_research=True adds google_search + url_context + optional maps.
+  # If both are on, avoid adding google_search twice.
+  web_search_added = False
+  if getattr(cfg, "enable_web_search", False) and not getattr(cfg, "enable_deep_research", False):
+    try:
+      from google.adk.tools import google_search
+      extra.append(google_search)
+      web_search_added = True
+    except Exception:
+      pass
+
   if getattr(cfg, "enable_deep_research", False):
     from google.adk.tools import google_search, url_context
-    extra.append(google_search)
+    if not web_search_added:
+      extra.append(google_search)
     extra.append(url_context)
     # Google Maps grounding can be incompatible with other built-in tools
     # (e.g., google_search) depending on the request/model tooling layer.

gemcode-0.3.30/src/gemcode/pricing.py

@@ -0,0 +1,76 @@
+"""
+Per-model token pricing for cost estimation.
+
+All prices are USD per 1 million tokens (input / output separately).
+Sources: Google AI Studio / Vertex AI pricing pages as of 2026.
+
+Usage:
+    from gemcode.pricing import estimate_cost
+    usd = estimate_cost("gemini-2.5-flash", input_tokens=5000, output_tokens=1200)
+"""
+
+from __future__ import annotations
+
+# Format: model_prefix → (input_$/M, output_$/M)
+# Matches are done by checking if the model id STARTS WITH any key (longest first).
+_PRICING_TABLE: dict[str, tuple[float, float]] = {
+    # ── Gemini 2.5 ──────────────────────────────────────────────────────────
+    "gemini-2.5-pro":         (3.50, 10.50),   # standard context ≤200k
+    "gemini-2.5-flash":       (0.15,  0.60),   # standard context
+    "gemini-2.5-flash-8b":    (0.037, 0.15),   # 8B lite variant
+    # ── Gemini 2.0 ──────────────────────────────────────────────────────────
+    "gemini-2.0-flash":       (0.10,  0.40),
+    "gemini-2.0-flash-lite":  (0.075, 0.30),
+    "gemini-2.0-pro":         (3.50, 10.50),
+    # ── Gemini 1.5 ──────────────────────────────────────────────────────────
+    "gemini-1.5-flash":       (0.075, 0.30),
+    "gemini-1.5-pro":         (1.25,  5.00),
+    # ── Gemini experimental / preview ───────────────────────────────────────
+    "gemini-exp":             (0.00,  0.00),   # free during preview
+    "gemini-3":               (0.30,  1.20),   # approximate future pricing
+}
+
+
+def estimate_cost(
+    model: str,
+    *,
+    input_tokens: int = 0,
+    output_tokens: int = 0,
+) -> float | None:
+    """
+    Return estimated USD cost for a model turn, or None if pricing is unknown.
+
+    Args:
+        model: Model id string (e.g. "gemini-2.5-flash").
+        input_tokens: Prompt/input token count.
+        output_tokens: Candidates/output token count.
+    """
+    if not model:
+        return None
+    model_lower = model.lower().strip()
+    # Match longest prefix first for specificity.
+    for prefix in sorted(_PRICING_TABLE, key=len, reverse=True):
+        if model_lower.startswith(prefix):
+            in_rate, out_rate = _PRICING_TABLE[prefix]
+            return (input_tokens * in_rate + output_tokens * out_rate) / 1_000_000
+    return None
+
+
+def format_cost(usd: float | None) -> str:
+    """Human-readable cost string, e.g. '$0.0012' or '$1.23'."""
+    if usd is None:
+        return ""
+    if usd < 0.0001:
+        return f"<$0.0001"
+    if usd < 0.01:
+        return f"${usd:.4f}"
+    return f"${usd:.3f}"
+
+
+def format_tokens(n: int) -> str:
+    """Format a token count compactly: 1234 → '1.2k', 12345678 → '12.3M'."""
+    if n >= 1_000_000:
+        return f"{n / 1_000_000:.1f}M"
+    if n >= 1_000:
+        return f"{n / 1_000:.1f}k"
+    return str(n)