shellbrain 0.1.42__tar.gz → 0.1.44__tar.gz

{shellbrain-0.1.42 → shellbrain-0.1.44}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: shellbrain
-Version: 0.1.42
+Version: 0.1.44
 Summary: Repo-scoped Shellbrain CLI with explicit evidence-backed writes.
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown

shellbrain-0.1.44/app/infrastructure/host_apps/inner_agents/claude_cli.py

@@ -0,0 +1,477 @@
+"""Claude Code CLI-backed inner-agent provider adapter."""
+
+from __future__ import annotations
+
+import json
+import shutil
+import subprocess
+from time import perf_counter
+
+from app.core.ports.host_apps.inner_agents import (
+    BuildKnowledgeAgentRequest,
+    BuildKnowledgeAgentResult,
+    InnerAgentRunRequest,
+    InnerAgentRunResult,
+    TeachKnowledgeAgentRequest,
+    WikiSummaryAgentRequest,
+    WikiSummaryAgentResult,
+)
+from app.infrastructure.host_apps.inner_agents.codex_cli import (
+    _build_knowledge_result,
+    _duration_ms,
+    _estimate_tokens,
+    _inner_agent_env,
+    _result,
+    _truncate,
+    _usage_or_estimate,
+    _wiki_summary_result,
+)
+from app.infrastructure.host_apps.inner_agents.output_parser import (
+    InnerAgentOutputParseError,
+    parse_build_knowledge_output,
+    parse_inner_agent_response_output,
+    parse_wiki_summary_output,
+)
+from app.infrastructure.host_apps.inner_agents.prompt import (
+    render_build_context_prompt,
+    render_build_context_synthesis_prompt,
+    render_build_knowledge_prompt,
+    render_teach_knowledge_prompt,
+    render_wiki_summary_prompt,
+)
+
+
+class ClaudeCliInnerAgentRunner:
+    """Run Shellbrain inner-agent work through the local Claude Code CLI."""
+
+    def __init__(self, *, command: str) -> None:
+        self._command = command
+
+    def run(self, request: InnerAgentRunRequest) -> InnerAgentRunResult:
+        """Run one Claude Code build_context request."""
+
+        prompt = (
+            render_build_context_synthesis_prompt(request)
+            if request.synthesis_only
+            else render_build_context_prompt(request)
+        )
+        run = self._run_claude(
+            request,
+            prompt=prompt,
+            mode="build_context_synthesis" if request.synthesis_only else "build_context",
+            tool_profile="none" if request.synthesis_only else "shellbrain",
+        )
+        if run["status"] != "ok":
+            return _context_error_result(request, run)
+        final_message = str(run["result"])
+        try:
+            brief, read_trace = parse_inner_agent_response_output(final_message)
+        except InnerAgentOutputParseError as exc:
+            return _result(
+                request,
+                status="invalid_output",
+                fallback_used=True,
+                duration_ms=int(run["duration_ms"]),
+                **_usage_or_estimate(
+                    prompt=prompt, output=final_message, usage=run.get("usage")
+                ),
+                error_code="invalid_output",
+                error_message=str(exc),
+            )
+        return _result(
+            request,
+            status="ok",
+            brief=brief,
+            duration_ms=int(run["duration_ms"]),
+            **_usage_or_estimate(
+                prompt=prompt, output=final_message, usage=run.get("usage")
+            ),
+            read_trace=read_trace,
+        )
+
+    def run_build_knowledge(
+        self, request: BuildKnowledgeAgentRequest
+    ) -> BuildKnowledgeAgentResult:
+        """Run one Claude Code build_knowledge request."""
+
+        return self._run_knowledge(
+            request,
+            prompt=render_build_knowledge_prompt(request),
+            mode="build_knowledge",
+            tool_profile="knowledge",
+        )
+
+    def run_teach_knowledge(
+        self, request: TeachKnowledgeAgentRequest
+    ) -> BuildKnowledgeAgentResult:
+        """Run one Claude Code teach request."""
+
+        return self._run_knowledge(
+            request,
+            prompt=render_teach_knowledge_prompt(request),
+            mode="teach",
+            tool_profile="shellbrain",
+        )
+
+    def _run_knowledge(
+        self,
+        request: BuildKnowledgeAgentRequest | TeachKnowledgeAgentRequest,
+        *,
+        prompt: str,
+        mode: str,
+        tool_profile: str,
+    ) -> BuildKnowledgeAgentResult:
+        run = self._run_claude(
+            request,
+            prompt=prompt,
+            mode=mode,
+            tool_profile=tool_profile,
+            knowledge_build_run_id=request.run_id,
+        )
+        if run["status"] != "ok":
+            return _knowledge_error_result(request, run)
+        final_message = str(run["result"])
+        try:
+            parsed = parse_build_knowledge_output(final_message)
+        except InnerAgentOutputParseError as exc:
+            return _build_knowledge_result(
+                request,
+                status="invalid_output",
+                duration_ms=int(run["duration_ms"]),
+                **_usage_or_estimate(
+                    prompt=prompt, output=final_message, usage=run.get("usage")
+                ),
+                error_code="invalid_output",
+                error_message=str(exc),
+            )
+        return _build_knowledge_result(
+            request,
+            status=parsed["status"],
+            duration_ms=int(run["duration_ms"]),
+            **_usage_or_estimate(
+                prompt=prompt, output=final_message, usage=run.get("usage")
+            ),
+            write_count=int(parsed["write_count"]),
+            skipped_item_count=int(parsed["skipped_item_count"]),
+            run_summary=parsed["run_summary"],
+            read_trace=parsed["read_trace"],
+            code_trace=parsed["code_trace"],
+        )
+
+    def run_wiki_summary(
+        self, request: WikiSummaryAgentRequest
+    ) -> WikiSummaryAgentResult:
+        """Run one Claude Code wiki_summary request."""
+
+        prompt = render_wiki_summary_prompt(request)
+        run = self._run_claude(
+            request,
+            prompt=prompt,
+            mode="wiki_summary",
+            tool_profile="none",
+        )
+        if run["status"] != "ok":
+            return _wiki_error_result(request, run)
+        final_message = str(run["result"])
+        try:
+            body = parse_wiki_summary_output(final_message)
+        except InnerAgentOutputParseError as exc:
+            return _wiki_summary_result(
+                request,
+                status="invalid_output",
+                duration_ms=int(run["duration_ms"]),
+                **_usage_or_estimate(
+                    prompt=prompt, output=final_message, usage=run.get("usage")
+                ),
+                error_code="invalid_output",
+                error_message=str(exc),
+            )
+        return _wiki_summary_result(
+            request,
+            status="ok",
+            body=body,
+            duration_ms=int(run["duration_ms"]),
+            **_usage_or_estimate(
+                prompt=prompt, output=final_message, usage=run.get("usage")
+            ),
+        )
+
+    def _run_claude(
+        self,
+        request,
+        *,
+        prompt: str,
+        mode: str,
+        tool_profile: str,
+        knowledge_build_run_id: str | None = None,
+    ) -> dict[str, object]:
+        command_path = shutil.which(self._command)
+        if command_path is None:
+            return {
+                "status": "provider_unavailable",
+                "error_code": "command_not_found",
+                "error_message": f"Claude command not found: {self._command}",
+            }
+        started = perf_counter()
+        try:
+            completed = subprocess.run(
+                _claude_args(
+                    command_path=command_path,
+                    model=request.model,
+                    reasoning=request.reasoning,
+                    tool_profile=tool_profile,
+                ),
+                input=prompt,
+                text=True,
+                capture_output=True,
+                timeout=request.timeout_seconds,
+                check=False,
+                env=_inner_agent_env(
+                    mode=mode,
+                    knowledge_build_run_id=knowledge_build_run_id,
+                ),
+            )
+        except subprocess.TimeoutExpired:
+            return _claude_error(
+                status="timeout",
+                duration_ms=_duration_ms(started),
+                prompt=prompt,
+                error_code="timeout",
+                error_message="Claude CLI timed out",
+            )
+
+        duration_ms = _duration_ms(started)
+        if completed.returncode != 0:
+            return _claude_error(
+                status="error",
+                duration_ms=duration_ms,
+                prompt=prompt,
+                error_code="claude_nonzero_exit",
+                error_message=_truncate(completed.stderr or completed.stdout, 500),
+            )
+        envelope = _claude_envelope(completed.stdout, allowed_model=request.model)
+        if envelope.get("error_code"):
+            return _claude_error(
+                status=str(envelope["status"]),
+                duration_ms=duration_ms,
+                prompt=prompt,
+                error_code=str(envelope["error_code"]),
+                error_message=str(envelope["error_message"]),
+            )
+        return {
+            "status": "ok",
+            "result": envelope["result"],
+            "usage": envelope.get("usage"),
+            "duration_ms": duration_ms,
+        }
+
+
+def _claude_error(
+    *,
+    status: str,
+    duration_ms: int,
+    prompt: str,
+    error_code: str,
+    error_message: str,
+) -> dict[str, object]:
+    return {
+        "status": status,
+        "duration_ms": duration_ms,
+        "input_tokens": _estimate_tokens(prompt),
+        "capture_quality": "estimated",
+        "error_code": error_code,
+        "error_message": error_message,
+    }
+
+
+def _context_error_result(
+    request: InnerAgentRunRequest, run: dict[str, object]
+) -> InnerAgentRunResult:
+    return _result(
+        request,
+        status=str(run["status"]),
+        fallback_used=True,
+        duration_ms=int(run.get("duration_ms") or 0),
+        input_tokens=run.get("input_tokens"),
+        capture_quality=run.get("capture_quality"),
+        error_code=str(run["error_code"]),
+        error_message=str(run["error_message"]),
+    )
+
+
+def _knowledge_error_result(
+    request: BuildKnowledgeAgentRequest | TeachKnowledgeAgentRequest,
+    run: dict[str, object],
+) -> BuildKnowledgeAgentResult:
+    return _build_knowledge_result(
+        request,
+        status=str(run["status"]),
+        duration_ms=int(run.get("duration_ms") or 0),
+        input_tokens=run.get("input_tokens"),
+        capture_quality=run.get("capture_quality"),
+        error_code=str(run["error_code"]),
+        error_message=str(run["error_message"]),
+    )
+
+
+def _wiki_error_result(
+    request: WikiSummaryAgentRequest, run: dict[str, object]
+) -> WikiSummaryAgentResult:
+    return _wiki_summary_result(
+        request,
+        status=str(run["status"]),
+        duration_ms=int(run.get("duration_ms") or 0),
+        input_tokens=run.get("input_tokens"),
+        capture_quality=run.get("capture_quality"),
+        error_code=str(run["error_code"]),
+        error_message=str(run["error_message"]),
+    )
+
+
+def _claude_args(
+    *,
+    command_path: str,
+    model: str,
+    reasoning: str,
+    tool_profile: str,
+) -> list[str]:
+    """Return the Claude Code print-mode command for one inner-agent run."""
+
+    args = [
+        command_path,
+        "-p",
+        "--output-format",
+        "json",
+        "--no-session-persistence",
+        "--safe-mode",
+        "--model",
+        model,
+        "--settings",
+        json.dumps(
+            {
+                "model": model,
+                "availableModels": [model],
+                "enforceAvailableModels": True,
+            },
+            sort_keys=True,
+            separators=(",", ":"),
+        ),
+        "--effort",
+        _claude_effort(reasoning),
+        "--strict-mcp-config",
+        "--disallowedTools",
+        "mcp__*",
+    ]
+    args.extend(_claude_tool_args(tool_profile))
+    return args
+
+
+def _claude_tool_args(tool_profile: str) -> list[str]:
+    if tool_profile == "none":
+        return ["--tools", ""]
+    if tool_profile == "knowledge":
+        return [
+            "--tools",
+            "Bash,Read,Grep,Glob",
+            "--allowedTools",
+            "Bash(shellbrain *)",
+            "Read",
+            "Grep",
+            "Glob",
+        ]
+    return ["--tools", "Bash", "--allowedTools", "Bash(shellbrain *)"]
+
+
+def _claude_effort(reasoning: str) -> str:
+    if reasoning in {"low", "medium", "high", "xhigh"}:
+        return reasoning
+    return "low"
+
+
+def _claude_envelope(output: str, *, allowed_model: str) -> dict[str, object]:
+    try:
+        payload = json.loads(output)
+    except json.JSONDecodeError:
+        return {
+            "status": "invalid_output",
+            "error_code": "claude_invalid_json",
+            "error_message": "Claude output is not valid JSON",
+        }
+    if not isinstance(payload, dict):
+        return {
+            "status": "invalid_output",
+            "error_code": "claude_invalid_json",
+            "error_message": "Claude output must be a JSON object",
+        }
+    disallowed_model = _disallowed_reported_model(payload, allowed_model=allowed_model)
+    if disallowed_model is not None:
+        return {
+            "status": "error",
+            "error_code": "claude_disallowed_model",
+            "error_message": f"Claude reported disallowed model: {disallowed_model}",
+        }
+    result = payload.get("result")
+    if not isinstance(result, str):
+        return {
+            "status": "invalid_output",
+            "error_code": "claude_missing_result",
+            "error_message": "Claude output must include a string result",
+        }
+    return {"result": result, "usage": _claude_usage(payload)}
+
+
+def _disallowed_reported_model(
+    payload: dict[str, object], *, allowed_model: str
+) -> str | None:
+    for model in _reported_models(payload):
+        normalized = model.lower()
+        if allowed_model.lower() not in normalized:
+            return model
+    return None
+
+
+def _reported_models(payload: dict[str, object]) -> list[str]:
+    models: list[str] = []
+    for key in ("model", "active_model", "activeModel"):
+        value = payload.get(key)
+        if isinstance(value, str):
+            models.append(value)
+    model_usage = payload.get("modelUsage")
+    if isinstance(model_usage, dict):
+        for key, value in model_usage.items():
+            if isinstance(key, str) and "model" in key.lower() and isinstance(value, str):
+                models.append(value)
+    return models
+
+
+def _claude_usage(payload: dict[str, object]) -> dict[str, int] | None:
+    source = payload.get("usage")
+    if not isinstance(source, dict):
+        source = payload.get("modelUsage")
+    if not isinstance(source, dict):
+        return None
+    usage = {
+        "input_tokens": _first_int(source, "input_tokens", "inputTokens"),
+        "output_tokens": _first_int(source, "output_tokens", "outputTokens"),
+        "reasoning_output_tokens": _first_int(
+            source,
+            "reasoning_output_tokens",
+            "reasoningOutputTokens",
+            "thinking_tokens",
+        ),
+        "cached_input_tokens": _first_int(
+            source,
+            "cached_input_tokens",
+            "cachedInputTokens",
+            "cache_read_input_tokens",
+        ),
+    }
+    return usage if any(value is not None for value in usage.values()) else None
+
+
+def _first_int(value: dict[object, object], *names: str) -> int | None:
+    for name in names:
+        item = value.get(name)
+        if isinstance(item, int) and item >= 0:
+            return item
+    return None

{shellbrain-0.1.42 → shellbrain-0.1.44}/app/infrastructure/host_apps/inner_agents/codex_cli.py

@@ -635,7 +635,7 @@ def _inner_agent_env(
 
 
 def _inherit_parent_caller_identity(env: dict[str, str]) -> None:
-    """Preserve the outer host identity across a nested Codex inner-agent run."""
+    """Preserve the outer host identity across a nested inner-agent run."""
 
     if env.get("CODEX_THREAD_ID"):
         env["SHELLBRAIN_PARENT_HOST_APP"] = "codex"

{shellbrain-0.1.42 → shellbrain-0.1.44}/app/settings/internal-agents/defaults.yaml

@@ -1,7 +1,7 @@
 # Default model/runtime settings for Shellbrain inner agents.
 build_context:
   strategy: deterministic_synthesis
-  provider: codex
+  provider: auto
   model: gpt-5.4-mini
   reasoning: medium
   timeout_seconds: 90
@@ -9,7 +9,7 @@ build_context:
   max_brief_tokens: 1800
 
 build_knowledge:
-  provider: codex
+  provider: auto
   model: gpt-5.4-mini
   reasoning: xhigh
   timeout_seconds: 600
@@ -20,7 +20,7 @@ build_knowledge:
   running_run_stale_seconds: 3600
 
 teach:
-  provider: codex
+  provider: auto
   model: gpt-5.4-mini
   reasoning: medium
   timeout_seconds: 600
@@ -29,7 +29,7 @@ teach:
   max_write_commands: 12
 
 wiki_summary:
-  provider: codex
+  provider: auto
   model: gpt-5.4-mini
   reasoning: medium
   timeout_seconds: 120
@@ -43,3 +43,7 @@ wiki_summary:
 providers:
   codex:
     command: codex
+    model: gpt-5.4-mini
+  claude:
+    command: claude
+    model: sonnet

{shellbrain-0.1.42 → shellbrain-0.1.44}/app/startup/internal_agent_config.py

@@ -21,6 +21,7 @@ class InnerAgentProviderConfig(_StrictModel):
     """Provider-specific runtime settings selected by startup."""
 
     command: str = Field(min_length=1)
+    model: str = Field(min_length=1)
 
 
 class InternalAgentsConfig(_StrictModel):
@@ -50,6 +51,8 @@ class InternalAgentsConfig(_StrictModel):
 
         for agent_name in ("build_context", "build_knowledge", "teach", "wiki_summary"):
             agent = getattr(self, agent_name)
+            if agent.provider == "auto":
+                continue
             if agent.provider not in self.providers:
                 raise ValueError(
                     f"internal_agents.{agent_name}.provider is not configured"