bog-agents-cli 0.7.2__tar.gz → 0.7.4__tar.gz

{bog_agents_cli-0.7.2 → bog_agents_cli-0.7.4}/CHANGELOG.md

@@ -1,5 +1,27 @@
 # Changelog
 
+## [0.7.4](https://github.com/bogware/bog-agents/compare/bog-agents-cli==0.7.3...bog-agents-cli==0.7.4) (2026-04-30)
+
+
+### Bug Fixes
+
+* **bedrock:** add `auth_mode` toggle (`auto`/`sso`/`static`/`profile`/`iam`) plus auto-fallback from expired SSO to static credentials when `~/.aws/config` short-circuits the credential chain ([#54](https://github.com/bogware/bog-agents/issues/54)) ([fef8228](https://github.com/bogware/bog-agents/commit/fef82283e9fc07f5d286a26eea093e68d28cdb42))
+* **bedrock:** cache probe failures so a single expired SSO session no longer logs 20+ identical TokenRetrievalError tracebacks ([#53](https://github.com/bogware/bog-agents/issues/53)) ([fef8228](https://github.com/bogware/bog-agents/commit/fef82283e9fc07f5d286a26eea093e68d28cdb42))
+* **bedrock:** add `langchain-aws` + AWS credential probe to `--doctor`; pre-flight credential check in `-n` mode surfaces SSO-expired errors as a clean stderr line instead of a wrapped RemoteException ([fef8228](https://github.com/bogware/bog-agents/commit/fef82283e9fc07f5d286a26eea093e68d28cdb42))
+* **cli:** install atexit + signal handlers in `cli_main()` to emit terminal-restore sequences (disable mouse tracking, leave alternate screen, show cursor) so a Textual crash mid-launch no longer leaves SGR mouse-protocol garbage like `[<35;57;14M[` in the user's shell input line ([fef8228](https://github.com/bogware/bog-agents/commit/fef82283e9fc07f5d286a26eea093e68d28cdb42))
+
+
+### Catalog
+
+* refresh provider model lists against live docs (2026-04-30): Anthropic Opus 4.7 / Sonnet 4.6 / Haiku 4.5 + legacy 4.6/4.5/4.1; Bedrock `us.*` inference-profile IDs + base IDs for Anthropic + Amazon Nova (Premier/Pro/Lite/Micro) + Meta Llama 4 Maverick/Scout + 3.3 + Mistral Large 3 / Pixtral Large; Google Gemini 2.5 Pro/Flash/Flash-Lite + Gemini 3.1 preview family ([fef8228](https://github.com/bogware/bog-agents/commit/fef82283e9fc07f5d286a26eea093e68d28cdb42))
+
+## [0.7.3](https://github.com/bogware/bog-agents/compare/bog-agents-cli==0.7.2...bog-agents-cli==0.7.3) (2026-04-29)
+
+
+### Features
+
+* verify/call subcommands, shell reliability, cross-platform hardening, 12 CVEs closed ([#51](https://github.com/bogware/bog-agents/issues/51)) ([5f13fb4](https://github.com/bogware/bog-agents/commit/5f13fb4de5aa7cb50731b634796f0732a8a25f65))
+
 ## [0.7.2](https://github.com/bogware/bog-agents/compare/bog-agents-cli==0.7.1...bog-agents-cli==0.7.2) (2026-04-25)
 
 

{bog_agents_cli-0.7.2 → bog_agents_cli-0.7.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: bog-agents-cli
-Version: 0.7.2
+Version: 0.7.4
 Summary: A coding agent in your terminal. 50+ commands, any LLM provider, persistent memory, git workflow, code review, plan mode, remote sandboxes, and CI/CD automation. One install, no code required.
 Project-URL: Homepage, https://github.com/bogware/bog-agents
 Project-URL: Repository, https://github.com/bogware/bog-agents
@@ -24,25 +24,30 @@ Classifier: Topic :: Software Development :: Libraries :: Python Modules
 Classifier: Topic :: Terminals
 Requires-Python: <4.0,>=3.11
 Requires-Dist: aiosqlite<1.0.0,>=0.19.0
-Requires-Dist: bog-agents==0.7.2
+Requires-Dist: bog-agents==0.7.4
 Requires-Dist: croniter<4.0.0,>=3.0.0
+Requires-Dist: cryptography>=46.0.7
 Requires-Dist: httpx<1.0.0,>=0.28.1
 Requires-Dist: keyring<27.0.0,>=25.0.0
 Requires-Dist: langchain-mcp-adapters<1.0.0,>=0.2.0
-Requires-Dist: langchain-openai<2.0.0,>=1.1.8
+Requires-Dist: langchain-openai<2.0.0,>=1.1.14
 Requires-Dist: langchain<2.0.0,>=1.2.10
 Requires-Dist: langgraph-checkpoint-sqlite<4.0.0,>=3.0.0
 Requires-Dist: langgraph-cli[inmem]<1.0.0,>=0.4.15
 Requires-Dist: langgraph-sdk<1.0.0,>=0.3.11
 Requires-Dist: langgraph<2.0.0,>=1.1.2
-Requires-Dist: langsmith<1.0.0,>=0.7.7
+Requires-Dist: langsmith<1.0.0,>=0.7.31
 Requires-Dist: markdownify<2.0.0,>=0.13.0
-Requires-Dist: pillow<13.0.0,>=10.0.0
+Requires-Dist: pillow<13.0.0,>=12.2.0
 Requires-Dist: prompt-toolkit<4.0.0,>=3.0.52
+Requires-Dist: pyasn1>=0.6.3
+Requires-Dist: pygments>=2.20.0
+Requires-Dist: pyjwt>=2.12.0
 Requires-Dist: pyperclip<2.0.0,>=1.11.0
-Requires-Dist: python-dotenv<2.0.0,>=1.0.0
+Requires-Dist: python-dotenv<2.0.0,>=1.2.2
+Requires-Dist: python-multipart>=0.0.26
 Requires-Dist: pyyaml<7.0.0,>=6.0.0
-Requires-Dist: requests<3.0.0,>=2.0.0
+Requires-Dist: requests<3.0.0,>=2.33.0
 Requires-Dist: rich<15.0.0,>=14.0.0
 Requires-Dist: textual-autocomplete<5.0.0,>=3.0.0
 Requires-Dist: textual<9.0.0,>=8.0.0
@@ -197,9 +202,9 @@ the introductions in about thirty seconds.
 
 ```bash
 bog-agents -M claude-sonnet-4-6
-bog-agents -M gpt-4o
-bog-agents -M ollama:llama3              # local, free
-bog-agents -M bedrock_converse:anthropic.claude-sonnet-4-6
+bog-agents -M openai:gpt-5.4
+bog-agents -M ollama:gpt-oss:20b                          # local, free, tool-capable
+bog-agents -M bedrock_converse:us.anthropic.claude-sonnet-4-6
 ```
 
 Something feeling off? Ask it.
@@ -472,15 +477,15 @@ Use `provider:model` format. Any LangChain-compatible chat model works.
 | Provider | Extra | Example |
 |----------|--------------|---------|
 | Anthropic | `anthropic` | `anthropic:claude-sonnet-4-6` |
-| OpenAI | *(included)* | `openai:gpt-4o` |
-| AWS Bedrock | `bedrock` | `bedrock_converse:anthropic.claude-sonnet-4-6` |
+| OpenAI | *(included)* | `openai:gpt-5.4` |
+| AWS Bedrock | `bedrock` | `bedrock_converse:us.anthropic.claude-sonnet-4-6` |
 | Google AI | `google-genai` | `google_genai:gemini-2.5-pro` |
 | Vertex AI | `vertexai` | `google_vertexai:gemini-2.5-pro` |
-| Ollama | `ollama` | `ollama:llama3` |
+| Ollama | `ollama` | `ollama:gpt-oss:20b` |
 | Groq | `groq` | `groq:llama-3.3-70b` |
 | DeepSeek | `deepseek` | `deepseek:deepseek-chat` |
 | Fireworks | `fireworks` | `fireworks:llama-v3p3-70b` |
-| Mistral | `mistralai` | `mistralai:mistral-large` |
+| Mistral | `mistralai` | `mistralai:mistral-large-3-2411` |
 | NVIDIA | `nvidia` | `nvidia:nemotron-70b` |
 | OpenRouter | `openrouter` | `openrouter:meta-llama/llama-3` |
 | Perplexity | `perplexity` | `perplexity:sonar-pro` |
@@ -490,6 +495,48 @@ Use `provider:model` format. Any LangChain-compatible chat model works.
 | HuggingFace | `huggingface` | `huggingface:meta-llama/Llama-3` |
 | Azure OpenAI | *(via openai)* | `azure_openai:gpt-4o` |
 
+### AWS Bedrock: pick how you authenticate
+
+boto3's credential chain stops at the first config it sees. If `~/.aws/config`
+declares an SSO session that's expired but `~/.aws/credentials` has fresh static
+keys, the SSO leg short-circuits and the static keys never get a turn. The CLI
+handles this in `auto` mode (default) by retrying with a credentials-file-only
+session when the SSO probe fails.
+
+Force a specific path when you need to. Either set
+`BOG_AGENTS_BEDROCK_AUTH_MODE` in the env, or write to `~/.bog-agents/config.toml`:
+
+```toml
+[models.providers.bedrock]
+auth_mode = "static"            # auto | sso | static | profile | iam
+aws_profile = "dev"             # only when auth_mode = "profile"
+```
+
+`bog-agents --doctor` shows you which mode resolved and whether the credentials
+came back valid. New in 0.7.4.
+
+### Local Ollama: which model to use
+
+Ollama's chat API mimics OpenAI's tools-API JSON schema. Models trained
+against that exact schema engage tools cleanly; models trained against
+other formats (Mistral's `[TOOL_CALLS]{}`, Hermes' `<tool_call>{}</tool_call>`,
+Qwen's chat-template tool call) emit calls in the message text and Ollama's
+adapter doesn't translate them. The CLI ships a parser middleware that
+recovers most text-shaped tool calls automatically when you select an
+`ollama:` model, but recovery is best-effort.
+
+- **Recommended**: `ollama:gpt-oss:20b` — OpenAI tools-API native, works
+  end-to-end with no recovery needed. Fits in 16GB of VRAM.
+- **Recovers via parser**: `ollama:mistral-nemo:12b`, `ollama:hermes3:8b`,
+  some `ollama:qwen2.5-coder` runs.
+- **Doesn't work**: `ollama:deepseek-coder-v2:16b` (Ollama's manifest
+  doesn't expose the `tools` capability — see
+  [ollama/ollama#3303](https://github.com/ollama/ollama/issues) if you want
+  to nudge upstream), `ollama:starcoder2`, `ollama:codellama`.
+
+Run `bog-agents --doctor` to see whether your configured default Ollama
+model is on the known-good list.
+
 ---
 
 ## Recipes for CI and scripting
@@ -548,6 +595,9 @@ Commands:
   reset                         Reset an agent's prompt
   skills                        Manage skills (list/create/info/delete)
   threads                       Manage threads (list/delete)
+  daemon                        Manage the ambient daemon (start/stop/jobs/...)
+  verify                        Run typecheck + lint + tests; write verification_summary.md
+  call MESSAGE                  Talk to a running --serve instance (thin HTTP client)
 
 Core:
   -M, --model MODEL             Model to use

{bog_agents_cli-0.7.2 → bog_agents_cli-0.7.4}/README.md

@@ -44,9 +44,9 @@ the introductions in about thirty seconds.
 
 ```bash
 bog-agents -M claude-sonnet-4-6
-bog-agents -M gpt-4o
-bog-agents -M ollama:llama3              # local, free
-bog-agents -M bedrock_converse:anthropic.claude-sonnet-4-6
+bog-agents -M openai:gpt-5.4
+bog-agents -M ollama:gpt-oss:20b                          # local, free, tool-capable
+bog-agents -M bedrock_converse:us.anthropic.claude-sonnet-4-6
 ```
 
 Something feeling off? Ask it.
@@ -319,15 +319,15 @@ Use `provider:model` format. Any LangChain-compatible chat model works.
 | Provider | Extra | Example |
 |----------|--------------|---------|
 | Anthropic | `anthropic` | `anthropic:claude-sonnet-4-6` |
-| OpenAI | *(included)* | `openai:gpt-4o` |
-| AWS Bedrock | `bedrock` | `bedrock_converse:anthropic.claude-sonnet-4-6` |
+| OpenAI | *(included)* | `openai:gpt-5.4` |
+| AWS Bedrock | `bedrock` | `bedrock_converse:us.anthropic.claude-sonnet-4-6` |
 | Google AI | `google-genai` | `google_genai:gemini-2.5-pro` |
 | Vertex AI | `vertexai` | `google_vertexai:gemini-2.5-pro` |
-| Ollama | `ollama` | `ollama:llama3` |
+| Ollama | `ollama` | `ollama:gpt-oss:20b` |
 | Groq | `groq` | `groq:llama-3.3-70b` |
 | DeepSeek | `deepseek` | `deepseek:deepseek-chat` |
 | Fireworks | `fireworks` | `fireworks:llama-v3p3-70b` |
-| Mistral | `mistralai` | `mistralai:mistral-large` |
+| Mistral | `mistralai` | `mistralai:mistral-large-3-2411` |
 | NVIDIA | `nvidia` | `nvidia:nemotron-70b` |
 | OpenRouter | `openrouter` | `openrouter:meta-llama/llama-3` |
 | Perplexity | `perplexity` | `perplexity:sonar-pro` |
@@ -337,6 +337,48 @@ Use `provider:model` format. Any LangChain-compatible chat model works.
 | HuggingFace | `huggingface` | `huggingface:meta-llama/Llama-3` |
 | Azure OpenAI | *(via openai)* | `azure_openai:gpt-4o` |
 
+### AWS Bedrock: pick how you authenticate
+
+boto3's credential chain stops at the first config it sees. If `~/.aws/config`
+declares an SSO session that's expired but `~/.aws/credentials` has fresh static
+keys, the SSO leg short-circuits and the static keys never get a turn. The CLI
+handles this in `auto` mode (default) by retrying with a credentials-file-only
+session when the SSO probe fails.
+
+Force a specific path when you need to. Either set
+`BOG_AGENTS_BEDROCK_AUTH_MODE` in the env, or write to `~/.bog-agents/config.toml`:
+
+```toml
+[models.providers.bedrock]
+auth_mode = "static"            # auto | sso | static | profile | iam
+aws_profile = "dev"             # only when auth_mode = "profile"
+```
+
+`bog-agents --doctor` shows you which mode resolved and whether the credentials
+came back valid. New in 0.7.4.
+
+### Local Ollama: which model to use
+
+Ollama's chat API mimics OpenAI's tools-API JSON schema. Models trained
+against that exact schema engage tools cleanly; models trained against
+other formats (Mistral's `[TOOL_CALLS]{}`, Hermes' `<tool_call>{}</tool_call>`,
+Qwen's chat-template tool call) emit calls in the message text and Ollama's
+adapter doesn't translate them. The CLI ships a parser middleware that
+recovers most text-shaped tool calls automatically when you select an
+`ollama:` model, but recovery is best-effort.
+
+- **Recommended**: `ollama:gpt-oss:20b` — OpenAI tools-API native, works
+  end-to-end with no recovery needed. Fits in 16GB of VRAM.
+- **Recovers via parser**: `ollama:mistral-nemo:12b`, `ollama:hermes3:8b`,
+  some `ollama:qwen2.5-coder` runs.
+- **Doesn't work**: `ollama:deepseek-coder-v2:16b` (Ollama's manifest
+  doesn't expose the `tools` capability — see
+  [ollama/ollama#3303](https://github.com/ollama/ollama/issues) if you want
+  to nudge upstream), `ollama:starcoder2`, `ollama:codellama`.
+
+Run `bog-agents --doctor` to see whether your configured default Ollama
+model is on the known-good list.
+
 ---
 
 ## Recipes for CI and scripting
@@ -395,6 +437,9 @@ Commands:
   reset                         Reset an agent's prompt
   skills                        Manage skills (list/create/info/delete)
   threads                       Manage threads (list/delete)
+  daemon                        Manage the ambient daemon (start/stop/jobs/...)
+  verify                        Run typecheck + lint + tests; write verification_summary.md
+  call MESSAGE                  Talk to a running --serve instance (thin HTTP client)
 
 Core:
   -M, --model MODEL             Model to use

bog_agents_cli-0.7.4/bog_agents_cli/_proc.py

@@ -0,0 +1,96 @@
+"""Cross-platform process inspection helpers.
+
+Centralises the Windows-vs-POSIX awkwardness around `os.kill(pid, 0)`.
+
+On POSIX, sending signal 0 to a non-existent PID raises
+`ProcessLookupError` and the standard try/except idiom works.
+
+On Windows, `os.kill(pid, 0)` for a dead PID can raise a generic
+`OSError [WinError 87] "The parameter is incorrect"` *and* CPython
+sometimes propagates this as a `SystemError: returned a result with
+an exception set` — a known C-level quirk we hit during pass-2
+validation. The robust answer on Windows is to ask `tasklist` directly.
+
+`signal.SIGKILL` doesn't exist on Windows either, so the kill helper
+falls back to TerminateProcess via `os.kill(pid, signal.SIGTERM)` (which
+on Windows is mapped to TerminateProcess in CPython) and skips the
+SIGKILL escalation gracefully.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+import signal
+import subprocess  # noqa: S404
+import sys
+from typing import Final
+
+logger = logging.getLogger(__name__)
+
+_TASKLIST_TIMEOUT_SECS: Final[int] = 3
+
+
+def is_running(pid: int) -> bool:
+    """Return True when *pid* names a live process.
+
+    Args:
+        pid: Process ID to probe.
+
+    Returns:
+        True if the process exists, False otherwise. Catches
+        `ProcessLookupError`, `PermissionError`, generic `OSError`
+        (Windows WinError 87 family), and `SystemError` (CPython
+        Windows quirk) — never raises.
+    """
+    if pid <= 0:
+        return False
+
+    if sys.platform == "win32":
+        try:
+            result = subprocess.run(  # noqa: S603
+                ["tasklist", "/FI", f"PID eq {pid}", "/NH", "/FO", "CSV"],
+                capture_output=True,
+                text=True,
+                timeout=_TASKLIST_TIMEOUT_SECS,
+                check=False,
+            )
+        except (OSError, subprocess.SubprocessError):
+            return False
+        return f'"{pid}"' in (result.stdout or "")
+
+    try:
+        os.kill(pid, 0)
+    except (ProcessLookupError, PermissionError, OSError):
+        return False
+    return True
+
+
+def terminate(pid: int, *, force: bool = False) -> bool:
+    """Best-effort termination of *pid*.
+
+    Args:
+        pid: Process ID to terminate.
+        force: When True, escalate to SIGKILL on POSIX after SIGTERM.
+            Ignored on Windows (no SIGKILL — `os.kill(pid, SIGTERM)`
+            already maps to TerminateProcess and is hard-stop).
+
+    Returns:
+        True if the kill call ran without raising; the caller should
+        re-check `is_running(pid)` to confirm exit. False when the
+        kill itself failed (PID gone, permission denied, etc.).
+    """
+    if not is_running(pid):
+        return True
+    try:
+        os.kill(pid, signal.SIGTERM)
+    except (ProcessLookupError, PermissionError, OSError) as exc:
+        logger.debug("SIGTERM to pid %d failed: %s", pid, exc)
+        return False
+
+    if force and hasattr(signal, "SIGKILL"):
+        try:
+            os.kill(pid, signal.SIGKILL)
+        except (ProcessLookupError, PermissionError, OSError) as exc:
+            logger.debug("SIGKILL to pid %d failed: %s", pid, exc)
+    return True

bog_agents_cli-0.7.4/bog_agents_cli/_version.py

@@ -0,0 +1,3 @@
+"""Version information for `bog-agents-cli`."""
+
+__version__ = "0.7.4"  # x-release-please-version

{bog_agents_cli-0.7.2 → bog_agents_cli-0.7.4}/bog_agents_cli/agent.py

@@ -62,7 +62,15 @@ DEFAULT_AGENT_NAME = "agent"
 REQUIRE_COMPACT_TOOL_APPROVAL: bool = True
 """When `True`, `compact_conversation` requires HITL approval like other gated tools."""
 
-_RESERVED_AGENT_HOME_DIRS = frozenset({"logs", "plugins", "skills"})
+_RESERVED_AGENT_HOME_DIRS = frozenset(
+    {
+        "daemon",  # bog-agents-daemon state (token, runs/, daemon.pid)
+        "logs",
+        "pipelines",  # CLI pipeline definitions, not an agent
+        "plugins",
+        "skills",
+    }
+)
 """Directories under `~/.bog-agents` reserved for global CLI state, not agents."""
 
 
@@ -732,9 +740,16 @@ def create_cli_agent(
         else settings.get_project_agents_dir()
     )
 
+    # Bundled-agents seeding: if the project is Python/Node/Rust/Go and
+    # the user hasn't authored their own subagents, this pulls in
+    # code-reviewer, test-author, and language-specific specialists from
+    # the package's bundled_agents/ tree. User and project subagents
+    # override on name conflict.
+    project_root_for_bundled = effective_cwd if effective_cwd is not None else None
     for subagent_meta in list_subagents(
         user_agents_dir=user_agents_dir,
         project_agents_dir=project_agents_dir,
+        project_root=project_root_for_bundled,
     ):
         subagent: SubAgent = {
             "name": subagent_meta["name"],
@@ -749,10 +764,24 @@ def create_cli_agent(
     agent_middleware = []
     agent_middleware.append(ConfigurableModelMiddleware())
 
-    # Add ask_user middleware (must be early so its tool is available)
-    from bog_agents_cli.ask_user import AskUserMiddleware
+    # Auto-enable tool-call parser for Ollama models. Many local models emit
+    # tool calls as text (Mistral [TOOL_CALLS], Hermes <tool_call>, fenced
+    # JSON) instead of using OpenAI's structured tool_calls field; the parser
+    # recovers them so the agent loop can proceed. No-op for cloud providers.
+    if (settings.model_provider or "").lower() == "ollama":
+        from bog_agents.middleware import ToolCallParserMiddleware
+
+        agent_middleware.append(ToolCallParserMiddleware())
+
+    # Add ask_user middleware (must be early so its tool is available).
+    # Skip in non-interactive mode: there is no user to answer, and a stray
+    # `ask_user` call mid-run produces a malformed HITL interrupt that the
+    # CLI rejects, which derails the agent without recourse. Headless agents
+    # should make a best-effort decision and proceed instead.
+    if interactive:
+        from bog_agents_cli.ask_user import AskUserMiddleware
 
-    agent_middleware.append(AskUserMiddleware())
+        agent_middleware.append(AskUserMiddleware())
 
     # Add memory middleware
     if enable_memory:

{bog_agents_cli-0.7.2 → bog_agents_cli-0.7.4}/bog_agents_cli/app.py

@@ -10059,6 +10059,18 @@ class BogAgentsApp(App):
                 model_id = (self._base_model_spec or "").lower()
                 if model_id.startswith("ollama:"):
                     is_tool_capability_error = True
+            # Bedrock-specific: when the langgraph remote stream wraps a
+            # botocore TokenRetrievalError it ends up here as a generic
+            # 'internal error occurred' message. Detect by exception
+            # *name* (RemoteException carries the original class name in
+            # str(e) on most paths) AND by checking whether the active
+            # model is bedrock-flavoured. We surface a Bedrock-specific
+            # action hint below.
+            is_bedrock_model = (
+                (self._base_model_spec or "")
+                .lower()
+                .startswith(("bedrock:", "bedrock_converse:"))
+            )
             is_auth_error = any(
                 keyword in err_name.lower() or keyword in err_str
                 for keyword in (
@@ -10069,7 +10081,15 @@ class BogAgentsApp(App):
                     "accessdenied",
                     "expired",
                     "sso",
+                    "tokenretrievalerror",
+                    "nocredentialserror",
                 )
+            ) or (
+                # Generic 'internal error' from langgraph + bedrock model =
+                # almost always an SSO/credential issue at the AWS layer.
+                err_name == "RemoteException"
+                and "internal error" in err_str
+                and is_bedrock_model
             )
             if is_tool_capability_error:
                 await self._mount_message(

bog_agents_cli-0.7.4/bog_agents_cli/bundled_agents/go/code-reviewer/AGENTS.md

@@ -0,0 +1,43 @@
+---
+name: code-reviewer
+description: |
+  Critical Go code review. Reads the most recent diff, reports bugs
+  by severity with concrete fixes. Read-only.
+model: anthropic:claude-sonnet-4-6
+---
+
+You are a senior Go reviewer.
+
+## Inputs (read-only)
+- `git diff HEAD~1 HEAD`
+- `go.mod`
+- `go vet ./...`
+- `golangci-lint run` if configured
+
+## What to look for
+1. **Goroutine leaks**: spawned goroutines without context-cancel
+   path, missing `defer wg.Done()`.
+2. **Error handling**: `if err != nil { return nil, err }` with the
+   wrong sentinel comparison; missing `errors.Is`/`errors.As`.
+3. **Race conditions**: shared map without `sync.Mutex`,
+   `time.After` in a tight loop.
+4. **Resource leaks**: `defer rsp.Body.Close()` missing,
+   `defer file.Close()` missing.
+5. **Test gaps**: new exported function without `func TestXxx`.
+
+## Output
+
+```markdown
+## Code review summary
+- Files reviewed: N
+- Bugs found: critical=N, major=N, minor=N
+
+## Findings
+
+### [critical|major|minor] <one-line title>
+**File:** path/to/file.go:LINE
+**Symptom:** what's wrong
+**Fix:** before/after snippet
+```
+
+End with `--- review complete ---`.

bog_agents_cli-0.7.4/bog_agents_cli/bundled_agents/node/code-reviewer/AGENTS.md

@@ -0,0 +1,49 @@
+---
+name: code-reviewer
+description: |
+  Critical code review for a Node.js / TypeScript / JavaScript project.
+  Reads the most recent diff, reports bugs by severity with concrete
+  fixes. Read-only — never edits code.
+model: anthropic:claude-sonnet-4-6
+---
+
+You are a senior Node.js / TypeScript reviewer. Find real bugs, not
+style nits.
+
+## Inputs (read-only)
+- `git diff HEAD~1 HEAD`
+- `package.json` for scripts + dependency versions
+- `tsconfig.json` for type-check strictness flags
+
+## What to look for
+1. **Async correctness**: missing `await`, unhandled promise
+   rejections, `forEach` over async functions, race conditions on
+   shared state.
+2. **TypeScript holes**: `any` leaks, `as` casts without runtime
+   guards, exhaustiveness gaps in `switch` over union types.
+3. **Null/undefined**: optional chaining where strict access is
+   required, missing nullish-coalescing, eager `.toString()` on
+   maybe-undefined.
+4. **React / Ink**: missing `key` props in lists, stale closures in
+   `useEffect`, side effects in render, Ink components rendered
+   without mount-time keys.
+5. **Memory + resource leaks**: unsubscribed event listeners,
+   uncleared timers, open file handles.
+6. **Test gaps**: new public function without coverage.
+
+## Output
+
+```markdown
+## Code review summary
+- Files reviewed: N
+- Bugs found: critical=N, major=N, minor=N
+
+## Findings
+
+### [critical|major|minor] <one-line title>
+**File:** path/to/file.ts:LINE
+**Symptom:** what's wrong
+**Fix:** before/after snippet
+```
+
+End with `--- review complete ---`.

bog_agents_cli-0.7.4/bog_agents_cli/bundled_agents/node/react-ink-artist/AGENTS.md

@@ -0,0 +1,31 @@
+---
+name: react-ink-artist
+description: |
+  Designs ASCII sprites and Ink layouts for terminal UIs. Delegate for
+  new screens, banners, sprites, terminal-color polish.
+model: anthropic:claude-sonnet-4-6
+---
+
+You are a TUI specialist. Output renders inside a typical 80x24
+terminal — assume that constraint.
+
+## Rules
+- Sprites: ≤20 rows × 60 cols. Multi-line string constants ending
+  with `\n`-newlines. ASCII only — no emoji unless the project's
+  existing sprites use them.
+- Use Ink's `<Box>` and `<Text>` for layout. Use the `color` prop for
+  accent. Never inject ANSI escape sequences directly.
+- Match the visual idiom of existing screens. Read at least one
+  existing screen file before designing a new one.
+
+## Workflow
+1. Read existing `src/render/sprites.ts` or equivalent.
+2. Add new exports — don't modify shipped sprites unless asked.
+3. If wiring into a screen, edit only the layout. Leave game state
+   alone.
+4. Run `npx tsc --noEmit` before declaring done.
+
+## Output
+- Sprite name(s) and a 5-10 line preview in code-block
+- Files touched
+- tsc result (run, don't claim)

bog_agents_cli-0.7.4/bog_agents_cli/bundled_agents/node/test-author/AGENTS.md

@@ -0,0 +1,30 @@
+---
+name: test-author
+description: |
+  Writes deterministic vitest / jest unit tests for Node.js projects.
+  Picks one uncovered code path per invocation, adds a focused test,
+  and verifies it passes.
+model: anthropic:claude-sonnet-4-6
+---
+
+You are a Node.js QA engineer.
+
+## Toolchain
+- Detect: vitest first (fast, ESM-native), then jest. The runner is
+  whichever is in `package.json` `devDependencies`.
+- Place tests next to source as `<name>.test.ts` OR under `__tests__/`
+  matching the existing convention — mirror what's already there.
+- Prefer real implementations over mocks. Use `vi.useFakeTimers()` /
+  `jest.useFakeTimers()` only when wall time matters.
+
+## Workflow
+1. Read `git diff HEAD~1 HEAD` and the existing `*.test.ts` files.
+   Pick ONE uncovered branch.
+2. Write a deterministic test (`vi.spyOn` over module mocks, fixed
+   seed if RNG is involved).
+3. Run `npx vitest run path/to/test.test.ts` (or `jest`) via the
+   execute tool. Iterate the TEST until it passes — don't change
+   the source.
+4. Report: file path, count added, pass/fail counts.
+
+End with `--- test added ---`.

bog_agents_cli-0.7.4/bog_agents_cli/bundled_agents/python/code-reviewer/AGENTS.md

@@ -0,0 +1,47 @@
+---
+name: code-reviewer
+description: |
+  Critical code review for a Python codebase. Reads the most recent diff
+  (or specific files), reports bugs by severity with concrete fixes.
+  Read-only — never edits code.
+model: anthropic:claude-sonnet-4-6
+---
+
+You are a senior Python code reviewer. Your job: find real bugs, not
+style nits.
+
+## Inputs (read-only)
+- `git diff HEAD~1 HEAD` — most recent commit's changes
+- `pyproject.toml` / `setup.cfg` for type-check + lint config
+- `tests/` for what's covered
+
+## What to look for
+1. **Mutability + reference bugs**: shared default args, mutating
+   inputs, dict/list aliasing.
+2. **Async pitfalls**: blocking I/O inside async functions, missing
+   `await`, races on shared state.
+3. **Type-system holes**: `Any` leaks, `# type: ignore` without a
+   reason, missing `Protocol` boundary checks.
+4. **Error handling**: broad `except Exception`, swallowed
+   `KeyboardInterrupt`, silent failures in loops.
+5. **Resource leaks**: missing `with` for files/sockets/subprocess,
+   unawaited tasks, leaking fd's on Windows.
+6. **Test gaps**: new public function without a test, branch coverage
+   on the unhappy path missing.
+
+## Output
+
+```markdown
+## Code review summary
+- Files reviewed: N
+- Bugs found: critical=N, major=N, minor=N
+
+## Findings
+
+### [critical|major|minor] <one-line title>
+**File:** path/to/file.py:LINE
+**Symptom:** what's wrong
+**Fix:** before/after snippet
+```
+
+End with `--- review complete ---`.