gemcode 0.3.47__tar.gz → 0.3.49__tar.gz

{gemcode-0.3.47/src/gemcode.egg-info → gemcode-0.3.49}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gemcode
-Version: 0.3.47
+Version: 0.3.49
 Summary: Local-first coding agent on Google Gemini + ADK
 Author: GemCode Contributors
 License:                                  Apache License

{gemcode-0.3.47 → gemcode-0.3.49}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "gemcode"
-version = "0.3.47"
+version = "0.3.49"
 description = "Local-first coding agent on Google Gemini + ADK"
 readme = "README.md"
 requires-python = ">=3.11"

{gemcode-0.3.47 → gemcode-0.3.49}/src/gemcode/agent.py

@@ -398,33 +398,19 @@ You run locally via the GemCode CLI. You are the same agent the user launched
 
 ## Core identity and approach
 
-### Intent-aware workflow
+Before you respond to anything, **think through what the person is actually trying to achieve**. Not what category their message fits into — what outcome they want, what they already know, and what the most useful response looks like.
 
-Before you see this message, a **gemini-2.5-flash-lite** intent classifier already determined the user's intent
-and stored it in session state as `_gemcode_intent`. Read it and adapt your behaviour:
+That thinking should drive everything: how much you use tools, how deep you go, how long you reply, what tone you take. A one-line social message deserves a one-line reply. A vague half-formed request might need a clarifying question before acting. A complex multi-file task needs systematic exploration first. A debugging session mid-error needs a hypothesis, not a search. There is no fixed list of intent types — the space of what people ask is open-ended, and your judgment should be too.
 
-| `_gemcode_intent` | Meaning | What to do |
-|---|---|---|
-| `GREETING` | Social / chitchat | *(You will never see this — handled upstream before reaching you.)* |
-| `CONCEPT` | General knowledge question | Answer from your training knowledge. No tools unless the answer truly requires this repo's files. |
-| `PROJECT_QUESTION` | Question about this codebase | Use 1–2 focused read-only tools (`list_directory`, `read_file`, `grep_content`) to find the answer, then reply concisely. |
-| `ENGINEERING_TASK` | Code modification / fix / build | Full workflow: **Orient → Plan → Execute → Verify**. |
-| `ANALYSIS` | Systematic audit or summarisation | Thorough tool sweep: list → read → grep across the affected area, then synthesise findings. |
+The one hard rule: **only reach for tools when they genuinely serve the response**. If the answer is in your knowledge, give it. If project context is needed, use the minimal set of read-only tools to get it. If you need to execute something, do it. But never open a tool call just to appear busy.
 
-If `_gemcode_intent` is not set (first turn or classifier unavailable), infer the intent yourself
-from the message before acting — the categories above still apply.
+When you do need to act on the codebase:
+1. **Understand first** — explore with `list_directory`, `glob_files`, `grep_content`, `read_file` before touching anything. These are instant and need no permission.
+2. **Plan for anything complex** — use `todo_write` to structure multi-step work before starting.
+3. **Execute completely** — don't stop at the first success. Finish the whole task.
+4. **Verify before you call it done** — check your own work.
 
-**Under no circumstances call `list_directory`, `read_project_notes`, or any tool in
-response to a greeting or a simple general question that requires no project context.**
-
-### Engineering task workflow
-When the intent is `ENGINEERING_TASK`:
-1. **Orient** — use `list_directory`, `glob_files`, `grep_content`, `read_file` to understand structure. These tools need **no permission** and are instant.
-2. **Plan** — for complex tasks, call `todo_write` upfront to map out the work.
-3. **Execute** — make the changes, run the checks, iterate.
-4. **Verify** — confirm the result is correct before reporting done.
-
-Never stop mid-task just because the first tool call succeeded. Keep going until the full task is complete or you hit a genuine blocker.
+The depth of each step should match the complexity of what was asked. Don't run a four-step engineering workflow for a one-sentence question.
 
 ## CRITICAL: Read-only tools first — never bash for exploration
 `bash` and `run_command` require permission confirmation by default. Always start with the **zero-permission** read-only tools:
@@ -447,12 +433,10 @@ You have native deep thinking capability — use it actively:
 - **For trade-off decisions** (which library, which pattern, which approach): reason through the pros/cons given this specific codebase.
 
 ## Interpreting requests
-- **Let `_gemcode_intent` guide you.** The pre-classifier already did the hard work. Trust it and act accordingly (see the table above).
-- **Engineering tasks** ("fix", "add", "refactor", "analyse", "debug"): infer from the repo — search, read, then act. Do not give abstract advice when concrete files exist.
-- If the user refers to symbols or behaviors, **find them** with `glob_files`/`grep_content`/`list_directory` — never ask them to paste paths you can discover yourself.
+- If the user refers to symbols, files, or behaviors they expect you to know — **find them** with `glob_files`/`grep_content`/`list_directory`. Never ask them to paste paths you can discover yourself.
 - **Never propose edits to files you haven't read.** Read first, then edit.
 - When something fails, diagnose (re-read the error, check assumptions) before switching strategy. Do not repeat the same failed call.
-- For analysis tasks ("analyse X", "explain X", "what does X do"): immediately start reading files with `list_directory` + `read_file` + `grep_content`. Produce concrete findings, not hypotheses.
+- When asked to analyse or explain something: read the actual files, produce concrete findings, not hypotheses.
 
 ## Tool selection guide
 
@@ -769,8 +753,10 @@ def build_root_agent(
   # 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."
+    "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(

{gemcode-0.3.47 → gemcode-0.3.49}/src/gemcode/cli.py

@@ -154,17 +154,6 @@ async def _run_prompt(
   # MCP and OpenAPI toolsets are now loaded inside create_runner() directly.
   runner = create_runner(cfg, extra_tools=None)
   try:
-    # LLM intent pre-classifier: greetings bypass the main agent entirely.
-    try:
-      from gemcode.intent_classifier import (
-          classify_intent, generate_greeting_reply, INTENT_GREETING
-      )
-      _intent = await classify_intent(prompt)
-      if _intent == INTENT_GREETING:
-        return await generate_greeting_reply(prompt)
-    except Exception:
-      pass
-
     collected = await run_turn(
         runner,
         user_id="local",

{gemcode-0.3.47 → gemcode-0.3.49}/src/gemcode/config.py

@@ -68,7 +68,7 @@ class GemCodeConfig:
   """Runtime options (CLI + env)."""
 
   project_root: Path
-  model: str = field(default_factory=lambda: os.environ.get("GEMCODE_MODEL", "gemini-2.5-flash"))
+  model: str = field(default_factory=lambda: os.environ.get("GEMCODE_MODEL", "gemini-3.1-pro-preview"))
   # Model mode: fast|balanced|quality|auto
   model_mode: str = field(
     default_factory=lambda: os.environ.get("GEMCODE_MODEL_MODE", "fast")

{gemcode-0.3.47 → gemcode-0.3.49}/src/gemcode/pricing.py

@@ -14,6 +14,10 @@ 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 3 ────────────────────────────────────────────────────────────
+    "gemini-3.1-pro":         (0.30,  1.20),   # preview pricing (approximate)
+    "gemini-3.0-pro":         (0.30,  1.20),
+    "gemini-3":               (0.30,  1.20),
     # ── Gemini 2.5 ──────────────────────────────────────────────────────────
     "gemini-2.5-pro":         (3.50, 10.50),   # standard context ≤200k
     "gemini-2.5-flash-lite":  (0.10,  0.40),   # lowest-cost 2.5 (must precede flash)
@@ -28,7 +32,6 @@ _PRICING_TABLE: dict[str, tuple[float, float]] = {
     "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
 }
 
 

{gemcode-0.3.47 → gemcode-0.3.49}/src/gemcode/tui/scrollback.py

@@ -582,61 +582,6 @@ async def run_gemcode_scrollback_tui(
     except Exception:
       pass
 
-    # ── LLM intent pre-classifier ────────────────────────────────────────────
-    # gemini-2.5-flash-lite classifies the message (same lane as Thinking)
-    try:
-      from gemcode.intent_classifier import (
-          classify_intent_with_source,
-          generate_greeting_reply,
-          show_intentifying_line,
-          INTENT_GREETING,
-          INTENT_DESCRIPTIONS,
-      )
-      # Show as a transient spinner status (do not print a permanent line).
-      _start_anim("Intentifying\u2026")
-      try:
-        _intent, _intent_src = await classify_intent_with_source(prompt)
-      finally:
-        _stop_anim()
-      # If disabled, show nothing; otherwise the spinner already conveyed it.
-      _ = show_intentifying_line(_intent_src)
-
-      if _intent == INTENT_GREETING:
-        _start_anim("Replying\u2026")
-        try:
-          _reply = await generate_greeting_reply(prompt)
-        finally:
-          _stop_anim()
-        print(f"  \u23bf  {ansi.bold}GemCode{ansi.reset}:")
-        console.print(_RichPadding(_RichMarkdown(_reply), (0, 0, 0, 4)))
-        print("")
-        if os.environ.get("GEMCODE_TUI_TURN_RULE", "1").lower() in (
-            "1", "true", "yes", "on"
-        ):
-          print(f"{ansi.dim}{_hr(ch='─')}{ansi.reset}")
-        print("")
-        continue
-
-      # Non-greeting: store classified intent in session state so the main
-      # agent can read it and adapt its tool-use strategy accordingly.
-      try:
-        ss = runner.session_service
-        app = getattr(runner, "app_name", None) or getattr(cfg, "app_name", "gemcode")
-        sess = await ss.get_session(
-            app_name=app, user_id="local", session_id=current_session_id
-        )
-        if sess is not None:
-          _desc = INTENT_DESCRIPTIONS.get(_intent, _intent)
-          sess.state["_gemcode_intent"] = _intent
-          sess.state["_gemcode_intent_desc"] = _desc
-          await ss.update_session(sess)
-      except Exception:
-        pass  # State injection is best-effort
-
-    except Exception:
-      pass  # Classifier unavailable — fall through to main agent unchanged
-    # ─────────────────────────────────────────────────────────────────────────
-
     # Snapshot pre-turn capability state so we can detect routing-triggered changes.
     _pre_dr  = cfg.enable_deep_research
     _pre_emb = cfg.enable_embeddings

{gemcode-0.3.47 → gemcode-0.3.49/src/gemcode.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gemcode
-Version: 0.3.47
+Version: 0.3.49
 Summary: Local-first coding agent on Google Gemini + ADK
 Author: GemCode Contributors
 License:                                  Apache License