specsmith 0.3.6.dev176__tar.gz → 0.3.6.dev178__tar.gz

{specsmith-0.3.6.dev176/src/specsmith.egg-info → specsmith-0.3.6.dev178}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: specsmith
-Version: 0.3.6.dev176
+Version: 0.3.6.dev178
 Summary: Applied Epistemic Engineering toolkit — AEE agent sessions, execution profiles, FPGA/HDL governance, tool installer, 50+ CLI commands.
 Author: BitConcepts
 License-Expression: MIT

{specsmith-0.3.6.dev176 → specsmith-0.3.6.dev178}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "specsmith"
-version = "0.3.6.dev176"
+version = "0.3.6.dev178"
 description = "Applied Epistemic Engineering toolkit — AEE agent sessions, execution profiles, FPGA/HDL governance, tool installer, 50+ CLI commands."
 readme = "README.md"
 license = "MIT"

{specsmith-0.3.6.dev176 → specsmith-0.3.6.dev178}/src/specsmith/agent/runner.py

@@ -153,6 +153,36 @@ H13: All proposals must state their epistemic boundaries. Hidden assumptions are
         agents_md or f"Spec version: {spec_version}. AGENTS.md not found — run specsmith audit."
     )
 
+    # Snapshot of VCS state at session start so the agent knows what’s pending
+    vcs_section = ""
+    try:
+        import subprocess as _sp
+
+        _st = _sp.run(
+            ["git", "status", "--short"],
+            cwd=project_dir,
+            capture_output=True,
+            text=True,
+            timeout=5,
+        )
+        _log = _sp.run(
+            ["git", "log", "--oneline", "-5"],
+            cwd=project_dir,
+            capture_output=True,
+            text=True,
+            timeout=5,
+        )
+        if _st.returncode == 0:
+            status_text = _st.stdout.strip() or "(clean — no uncommitted changes)"
+            log_text = _log.stdout.strip() if _log.returncode == 0 else "(no commits yet)"
+            vcs_section = (
+                "\n## VCS State at Session Start\n"
+                f"```\n{status_text}\n```\n"
+                f"Recent commits:\n```\n{log_text}\n```\n"
+            )
+    except Exception:  # noqa: BLE001
+        pass
+
     # Load tool-specific rules for this project type
     tool_rules_section = ""
     if project_type:
@@ -166,6 +196,7 @@ H13: All proposals must state their epistemic boundaries. Hidden assumptions are
             pass
 
     prompt = f"""SYSTEM LANGUAGE DIRECTIVE — ABSOLUTE HARD RULE — HIGHEST PRIORITY:
+{vcs_section}
 You MUST respond in English ONLY. This overrides all other instructions.
 Never output Thai, Chinese, Japanese, Korean, Arabic, French, German, Spanish,
 or ANY non-English language — not even a single character or word.
@@ -257,8 +288,12 @@ class AgentRunner:
     QUICK_COMMANDS = {
         "start": (
             "[RESPOND IN ENGLISH ONLY] "
-            "Run session start protocol: sync, load AGENTS.md, read last LEDGER.md entries. "
-            "Translate any non-English context internally if needed, but respond only in English."
+            "Run the session start protocol in this order:\n"
+            "1. Run: git status (report staged, modified, and untracked files)\n"
+            "2. Run: git log --oneline -5 (report the 5 most recent commits)\n"
+            "3. Read AGENTS.md (confirm your role and governance rules)\n"
+            "4. Read the last 10 lines of LEDGER.md (confirm current project state)\n"
+            "Summarize findings in 3-4 plain sentences, then propose the most logical next action."
         ),
         "resume": (
             "[RESPOND IN ENGLISH ONLY] Resume from last LEDGER.md entry"
@@ -480,6 +515,11 @@ class AgentRunner:
         # Add user message
         self._state.messages.append(Message(role=Role.USER, content=user_input))
 
+        # In json_events mode we defer the llm_chunk emission so we can run the
+        # language check BEFORE the UI sees the text.  Without this, a Thai/Chinese
+        # response would be displayed even when the correction turn later succeeds.
+        _defer_emit = not silent and self._json_events
+
         final_response = ""
         for _iteration in range(self._max_iterations):
             messages_with_system = [
@@ -487,7 +527,9 @@ class AgentRunner:
             ] + self._state.messages
 
             try:
-                response = self._call_provider(messages_with_system, silent=silent)
+                response = self._call_provider(
+                    messages_with_system, silent=silent, defer_emit=_defer_emit
+                )
             except Exception as e:  # noqa: BLE001
                 error_msg = f"[Provider error] {e}"
                 if not silent:
@@ -515,9 +557,12 @@ class AgentRunner:
                 final_response = response.content
 
             if not response.has_tool_calls:
-                # Non-English correction: if response appears to be in another language,
-                # issue a single correction turn rather than showing the wrong-language response.
-                if response.content and self._has_non_english(response.content) and _iteration == 0:
+                # Non-English correction: if the final response is in a non-English
+                # language, silently issue a correction turn instead of surfacing the
+                # bad text.  _defer_emit ensures the UI never sees the Thai/Chinese
+                # content — we only emit llm_chunk once we have an English reply.
+                # The iteration limit acts as a natural loop-break (max_iterations).
+                if response.content and self._has_non_english(response.content):
                     correction = (
                         "[LANG:EN] CRITICAL: Your last response was in a non-English language. "
                         "You MUST respond in English ONLY. Please re-answer in English."
@@ -526,12 +571,22 @@ class AgentRunner:
                         Message(role=Role.ASSISTANT, content=response.content)
                     )
                     self._state.messages.append(Message(role=Role.USER, content=correction))
-                    # Continue the loop to get an English response
+                    # Continue the loop to get an English response (don't emit bad content)
                     continue
+                # Language OK — emit the deferred llm_chunk now
+                if _defer_emit and response.content:
+                    self._emit_event(type="llm_chunk", text=response.content)
                 # Final response — add to history
                 self._state.messages.append(Message(role=Role.ASSISTANT, content=response.content))
                 break
 
+            # Has tool calls — emit any deferred partial text (planning/thinking text
+            # that precedes the tool invocation), but only if it is in English.
+            # Some Qwen/DeepSeek variants emit Thai/Chinese thinking text before tool
+            # calls; silently drop that rather than surfacing it to the UI.
+            if _defer_emit and response.content and not self._has_non_english(response.content):
+                self._emit_event(type="llm_chunk", text=response.content)
+
             # Process tool calls
             self._hard_stop = False  # reset before each batch
             tool_results = self._execute_tool_calls(response.tool_calls, silent=silent)
@@ -571,12 +626,20 @@ class AgentRunner:
 
         return final_response
 
-    def _call_provider(self, messages: list[Message], silent: bool = False) -> CompletionResponse:
+    def _call_provider(
+        self,
+        messages: list[Message],
+        silent: bool = False,
+        defer_emit: bool = False,
+    ) -> CompletionResponse:
         """Call the LLM provider with optimization engine pre/post hooks.
 
         Streaming is disabled when tools are registered (streaming drops tool_call blocks).
         When an OptimizationEngine is active, pre_call() may return a cache hit
         or transform messages/model/tools before the actual provider call.
+
+        ``defer_emit=True`` suppresses the ``llm_chunk`` event so the caller can
+        validate the response (e.g. language check) before surfacing it to the UI.
         """
         provider: Any = self._provider
         tools = self._tools
@@ -613,7 +676,7 @@ class AgentRunner:
             response = CompletionResponse(content=accumulated, model=str(provider.model))
         else:
             response = cast(CompletionResponse, provider.complete(messages, tools=tools))
-            if not silent and response.content:
+            if not silent and not defer_emit and response.content:
                 if self._json_events:
                     self._emit_event(type="llm_chunk", text=response.content)
                 else:

{specsmith-0.3.6.dev176 → specsmith-0.3.6.dev178/src/specsmith.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: specsmith
-Version: 0.3.6.dev176
+Version: 0.3.6.dev178
 Summary: Applied Epistemic Engineering toolkit — AEE agent sessions, execution profiles, FPGA/HDL governance, tool installer, 50+ CLI commands.
 Author: BitConcepts
 License-Expression: MIT