Import AITURK IDE 1.0.0-beta.1 from Hermes 63279301; preserve MIT license

agent/transports/__init__.py

@@ -0,0 +1,68 @@
+"""Transport layer types and registry for provider response normalization.
+
+Usage:
+    from agent.transports import get_transport
+    transport = get_transport("anthropic_messages")
+    result = transport.normalize_response(raw_response)
+"""
+
+from agent.transports.types import (
+    NormalizedResponse,
+    ToolCall,
+    Usage,
+    build_tool_call,
+    map_finish_reason,
+)  # noqa: F401
+
+_REGISTRY: dict = {}
+_discovered: bool = False
+
+
+def register_transport(api_mode: str, transport_cls: type) -> None:
+    """Register a transport class for an api_mode string."""
+    _REGISTRY[api_mode] = transport_cls
+
+
+def get_transport(api_mode: str):
+    """Get a transport instance for the given api_mode.
+
+    Returns None if no transport is registered for this api_mode.
+    This allows gradual migration — call sites can check for None
+    and fall back to the legacy code path.
+    """
+    global _discovered
+    if not _discovered:
+        _discover_transports()
+    cls = _REGISTRY.get(api_mode)
+    if cls is None:
+        # The registry can be partially populated when a specific transport
+        # module was imported directly (for example chat_completions before
+        # codex).  Discover on misses, not only when the registry is empty, so
+        # test/order-dependent imports do not make valid api_modes unavailable.
+        _discover_transports()
+        cls = _REGISTRY.get(api_mode)
+    if cls is None:
+        return None
+    return cls()
+
+
+def _discover_transports() -> None:
+    """Import all transport modules to trigger auto-registration."""
+    global _discovered
+    _discovered = True
+    try:
+        import agent.transports.anthropic  # noqa: F401
+    except ImportError:
+        pass
+    try:
+        import agent.transports.codex  # noqa: F401
+    except ImportError:
+        pass
+    try:
+        import agent.transports.chat_completions  # noqa: F401
+    except ImportError:
+        pass
+    try:
+        import agent.transports.bedrock  # noqa: F401
+    except ImportError:
+        pass

agent/transports/anthropic.py

@@ -0,0 +1,261 @@
+"""Anthropic Messages API transport.
+
+Delegates to the existing adapter functions in agent/anthropic_adapter.py.
+This transport owns format conversion and normalization — NOT client lifecycle.
+"""
+
+from typing import Any, Dict, List, Optional
+
+from agent.transports.base import ProviderTransport
+from agent.transports.types import NormalizedResponse
+
+
+class AnthropicTransport(ProviderTransport):
+    """Transport for api_mode='anthropic_messages'.
+
+    Wraps the existing functions in anthropic_adapter.py behind the
+    ProviderTransport ABC.  Each method delegates — no logic is duplicated.
+    """
+
+    @property
+    def api_mode(self) -> str:
+        return "anthropic_messages"
+
+    def convert_messages(self, messages: List[Dict[str, Any]], **kwargs) -> Any:
+        """Convert OpenAI messages to Anthropic (system, messages) tuple.
+
+        kwargs:
+            base_url: Optional[str] — affects thinking signature handling.
+        """
+        from agent.anthropic_adapter import convert_messages_to_anthropic
+
+        base_url = kwargs.get("base_url")
+        return convert_messages_to_anthropic(messages, base_url=base_url)
+
+    def convert_tools(self, tools: List[Dict[str, Any]]) -> Any:
+        """Convert OpenAI tool schemas to Anthropic input_schema format."""
+        from agent.anthropic_adapter import convert_tools_to_anthropic
+
+        return convert_tools_to_anthropic(tools)
+
+    def build_kwargs(
+        self,
+        model: str,
+        messages: List[Dict[str, Any]],
+        tools: Optional[List[Dict[str, Any]]] = None,
+        **params,
+    ) -> Dict[str, Any]:
+        """Build Anthropic messages.create() kwargs.
+
+        Calls convert_messages and convert_tools internally.
+
+        params (all optional):
+            max_tokens: int
+            reasoning_config: dict | None
+            tool_choice: str | None
+            is_oauth: bool
+            preserve_dots: bool
+            context_length: int | None
+            base_url: str | None
+            fast_mode: bool
+            drop_context_1m_beta: bool
+        """
+        from agent.anthropic_adapter import build_anthropic_kwargs
+
+        return build_anthropic_kwargs(
+            model=model,
+            messages=messages,
+            tools=tools,
+            max_tokens=params.get("max_tokens", 16384),
+            reasoning_config=params.get("reasoning_config"),
+            tool_choice=params.get("tool_choice"),
+            is_oauth=params.get("is_oauth", False),
+            preserve_dots=params.get("preserve_dots", False),
+            context_length=params.get("context_length"),
+            base_url=params.get("base_url"),
+            fast_mode=params.get("fast_mode", False),
+            drop_context_1m_beta=params.get("drop_context_1m_beta", False),
+        )
+
+    def normalize_response(self, response: Any, **kwargs) -> NormalizedResponse:
+        """Normalize Anthropic response to NormalizedResponse.
+
+        Parses content blocks (text, thinking, tool_use), maps stop_reason
+        to OpenAI finish_reason, and collects reasoning_details in provider_data.
+        """
+        import json
+        from agent.anthropic_adapter import (
+            _OAUTH_TOOL_NAME_REVERSE_ALIASES,
+            _sanitize_replay_block,
+            _to_plain_data,
+        )
+        from agent.transports.types import ToolCall
+
+        strip_tool_prefix = kwargs.get("strip_tool_prefix", False)
+        _MCP_PREFIX = "mcp__"
+
+        text_parts = []
+        reasoning_parts = []
+        reasoning_details = []
+        tool_calls = []
+        # Verbatim, order-preserving copy of every content block in the turn.
+        # Anthropic signs each thinking block against the turn content that
+        # PRECEDES it at its position; when a turn interleaves thinking and
+        # tool_use (adaptive/interleaved thinking, Claude 4.6+), the parallel
+        # reasoning_details + tool_calls lists below lose that cross-type
+        # ordering. Replaying the latest assistant message in the wrong order
+        # invalidates the signatures -> HTTP 400 "thinking ... blocks in the
+        # latest assistant message cannot be modified". Preserve the exact
+        # block sequence here so the adapter can replay it unchanged. See
+        # tests/agent/test_anthropic_thinking_block_order.py.
+        ordered_blocks = []
+
+        for block in response.content:
+            block_dict = _to_plain_data(block)
+            clean_block = None
+            if isinstance(block_dict, dict):
+                # Sanitize at capture so output-only SDK fields (parsed_output,
+                # caller, citations=None, …) never persist to state.db and leak
+                # back as request input on replay → HTTP 400 "Extra inputs are
+                # not permitted". Defence-in-depth with the replay-side sanitize.
+                clean_block = _sanitize_replay_block(block_dict)
+                if clean_block is not None:
+                    ordered_blocks.append(clean_block)
+            if block.type == "text":
+                text_parts.append(block.text)
+            elif block.type in ("thinking", "redacted_thinking"):
+                if block.type == "thinking":
+                    reasoning_parts.append(block.thinking)
+                # Use the sanitized block (clean_block) for reasoning_details too,
+                # since _extract_preserved_thinking_blocks replays these on the
+                # non-ordered path. Falls back to raw only if sanitize dropped it.
+                if isinstance(clean_block, dict):
+                    reasoning_details.append(clean_block)
+                elif isinstance(block_dict, dict):
+                    reasoning_details.append(block_dict)
+            elif block.type == "tool_use":
+                name = block.name
+                if strip_tool_prefix and name.startswith(_MCP_PREFIX):
+                    # On the OAuth wire every tool carries a double-underscore
+                    # ``mcp__`` prefix (added in build_anthropic_kwargs to avoid
+                    # Anthropic's single-underscore third-party classifier).
+                    # Reverse it back to the name the registry/dispatcher knows.
+                    # Two original forms map onto the same ``mcp__`` wire name:
+                    #   ``mcp__read_file``       <- bare native tool ``read_file``
+                    #   ``mcp__linear_get_issue`` <- MCP server tool
+                    #                                ``mcp_linear_get_issue``
+                    # Resolve by registry lookup, preferring whichever original
+                    # is actually registered; never rewrite a name the LLM used
+                    # that already resolves natively. GH-25255.
+                    from tools.registry import registry as _tool_registry
+                    if not _tool_registry.get_entry(name):
+                        bare = name[len(_MCP_PREFIX):]            # read_file
+                        single = "mcp_" + bare                    # mcp_read_file / mcp_linear_get_issue
+                        if _tool_registry.get_entry(single):
+                            name = single
+                        elif _tool_registry.get_entry(bare):
+                            name = bare
+                        elif bare in _OAUTH_TOOL_NAME_REVERSE_ALIASES:
+                            # OAuth wire alias (e.g. chat_history_lookup ->
+                            # session_search, #65365). Checked LAST so a real
+                            # tool actually registered under the wire name
+                            # still wins — same GH-25255 precedence.
+                            name = _OAUTH_TOOL_NAME_REVERSE_ALIASES[bare]
+                tool_calls.append(
+                    ToolCall(
+                        id=block.id,
+                        name=name,
+                        arguments=json.dumps(block.input),
+                    )
+                )
+
+        finish_reason = self._STOP_REASON_MAP.get(response.stop_reason, "stop")
+
+        provider_data = {}
+        if reasoning_details:
+            provider_data["reasoning_details"] = reasoning_details
+        # Only worth carrying the ordered-blocks channel when the turn
+        # actually interleaves signed thinking with tool_use — that's the
+        # only shape the parallel lists reconstruct incorrectly. A turn that
+        # is purely text, or thinking-then-tools with a single leading
+        # thinking block, replays correctly without it.
+        _has_signed_thinking = any(
+            isinstance(b, dict)
+            and b.get("type") in ("thinking", "redacted_thinking")
+            and (b.get("signature") or b.get("data"))
+            for b in ordered_blocks
+        )
+        _has_tool_use = any(
+            isinstance(b, dict) and b.get("type") == "tool_use"
+            for b in ordered_blocks
+        )
+        if _has_signed_thinking and _has_tool_use:
+            provider_data["anthropic_content_blocks"] = ordered_blocks
+
+        return NormalizedResponse(
+            content="\n".join(text_parts) if text_parts else None,
+            tool_calls=tool_calls or None,
+            finish_reason=finish_reason,
+            reasoning="\n\n".join(reasoning_parts) if reasoning_parts else None,
+            usage=None,
+            provider_data=provider_data or None,
+        )
+
+    def validate_response(self, response: Any) -> bool:
+        """Check Anthropic response structure is valid.
+
+        An empty content list is legitimate for terminal stop reasons that
+        carry no text payload:
+
+        - ``end_turn`` — the model's canonical "nothing more to add" after a
+          tool turn that already delivered the user-facing text.
+        - ``refusal`` — the model declined to respond (Claude 4.5+). The
+          Messages API returns an empty ``content`` list with this stop
+          reason. Treating it as invalid sends a deterministic refusal into
+          the invalid-response retry loop, which reproduces the refusal on
+          every attempt and surfaces a misleading "rate limited / invalid
+          response" error instead of the refusal. ``normalize_response`` maps
+          ``refusal`` → ``content_filter`` so the agent loop's refusal handler
+          can surface it.
+
+        Treating either as invalid falsely retries a completed response.
+        """
+        if response is None:
+            return False
+        content_blocks = getattr(response, "content", None)
+        if not isinstance(content_blocks, list):
+            return False
+        if not content_blocks:
+            return getattr(response, "stop_reason", None) in {"end_turn", "refusal"}
+        return True
+
+    def extract_cache_stats(self, response: Any) -> Optional[Dict[str, int]]:
+        """Extract Anthropic cache_read and cache_creation token counts."""
+        usage = getattr(response, "usage", None)
+        if usage is None:
+            return None
+        cached = getattr(usage, "cache_read_input_tokens", 0) or 0
+        written = getattr(usage, "cache_creation_input_tokens", 0) or 0
+        if cached or written:
+            return {"cached_tokens": cached, "creation_tokens": written}
+        return None
+
+    # Promote the adapter's canonical mapping to module level so it's shared
+    _STOP_REASON_MAP = {
+        "end_turn": "stop",
+        "tool_use": "tool_calls",
+        "max_tokens": "length",
+        "stop_sequence": "stop",
+        "refusal": "content_filter",
+        "model_context_window_exceeded": "length",
+    }
+
+    def map_finish_reason(self, raw_reason: str) -> str:
+        """Map Anthropic stop_reason to OpenAI finish_reason."""
+        return self._STOP_REASON_MAP.get(raw_reason, "stop")
+
+
+# Auto-register on import
+from agent.transports import register_transport  # noqa: E402
+
+register_transport("anthropic_messages", AnthropicTransport)

agent/transports/base.py

@@ -0,0 +1,89 @@
+"""Abstract base for provider transports.
+
+A transport owns the data path for one api_mode:
+  convert_messages → convert_tools → build_kwargs → normalize_response
+
+It does NOT own: client construction, streaming, credential refresh,
+prompt caching, interrupt handling, or retry logic.  Those stay on AIAgent.
+"""
+
+from abc import ABC, abstractmethod
+from typing import Any, Dict, List, Optional
+
+from agent.transports.types import NormalizedResponse
+
+
+class ProviderTransport(ABC):
+    """Base class for provider-specific format conversion and normalization."""
+
+    @property
+    @abstractmethod
+    def api_mode(self) -> str:
+        """The api_mode string this transport handles (e.g. 'anthropic_messages')."""
+        ...
+
+    @abstractmethod
+    def convert_messages(self, messages: List[Dict[str, Any]], **kwargs) -> Any:
+        """Convert OpenAI-format messages to provider-native format.
+
+        Returns provider-specific structure (e.g. (system, messages) for Anthropic,
+        or the messages list unchanged for chat_completions).
+        """
+        ...
+
+    @abstractmethod
+    def convert_tools(self, tools: List[Dict[str, Any]]) -> Any:
+        """Convert OpenAI-format tool definitions to provider-native format.
+
+        Returns provider-specific tool list (e.g. Anthropic input_schema format).
+        """
+        ...
+
+    @abstractmethod
+    def build_kwargs(
+        self,
+        model: str,
+        messages: List[Dict[str, Any]],
+        tools: Optional[List[Dict[str, Any]]] = None,
+        **params,
+    ) -> Dict[str, Any]:
+        """Build the complete API call kwargs dict.
+
+        This is the primary entry point — it typically calls convert_messages()
+        and convert_tools() internally, then adds model-specific config.
+
+        Returns a dict ready to be passed to the provider's SDK client.
+        """
+        ...
+
+    @abstractmethod
+    def normalize_response(self, response: Any, **kwargs) -> NormalizedResponse:
+        """Normalize a raw provider response to the shared NormalizedResponse type.
+
+        This is the only method that returns a transport-layer type.
+        """
+        ...
+
+    def validate_response(self, response: Any) -> bool:
+        """Optional: check if the raw response is structurally valid.
+
+        Returns True if valid, False if the response should be treated as invalid.
+        Default implementation always returns True.
+        """
+        return True
+
+    def extract_cache_stats(self, response: Any) -> Optional[Dict[str, int]]:
+        """Optional: extract provider-specific cache hit/creation stats.
+
+        Returns dict with 'cached_tokens' and 'creation_tokens', or None.
+        Default returns None.
+        """
+        return None
+
+    def map_finish_reason(self, raw_reason: str) -> str:
+        """Optional: map provider-specific stop reason to OpenAI equivalent.
+
+        Default returns the raw reason unchanged.  Override for providers
+        with different stop reason vocabularies.
+        """
+        return raw_reason

agent/transports/bedrock.py

@@ -0,0 +1,161 @@
+"""AWS Bedrock Converse API transport.
+
+Delegates to the existing adapter functions in agent/bedrock_adapter.py.
+Bedrock uses its own boto3 client (not the OpenAI SDK), so the transport
+owns format conversion and normalization, while client construction and
+boto3 calls stay on AIAgent.
+"""
+
+from typing import Any, Dict, List, Optional
+
+from agent.transports.base import ProviderTransport
+from agent.transports.types import NormalizedResponse, ToolCall, Usage
+
+
+class BedrockTransport(ProviderTransport):
+    """Transport for api_mode='bedrock_converse'."""
+
+    @property
+    def api_mode(self) -> str:
+        return "bedrock_converse"
+
+    def convert_messages(self, messages: List[Dict[str, Any]], **kwargs) -> Any:
+        """Convert OpenAI messages to Bedrock Converse format."""
+        from agent.bedrock_adapter import convert_messages_to_converse
+        return convert_messages_to_converse(messages)
+
+    def convert_tools(self, tools: List[Dict[str, Any]]) -> Any:
+        """Convert OpenAI tool schemas to Bedrock Converse toolConfig."""
+        from agent.bedrock_adapter import convert_tools_to_converse
+        return convert_tools_to_converse(tools)
+
+    def build_kwargs(
+        self,
+        model: str,
+        messages: List[Dict[str, Any]],
+        tools: Optional[List[Dict[str, Any]]] = None,
+        **params,
+    ) -> Dict[str, Any]:
+        """Build Bedrock converse() kwargs.
+
+        Calls convert_messages and convert_tools internally.
+
+        params:
+            max_tokens: int — output token limit (default 4096)
+            temperature: float | None
+            guardrail_config: dict | None — Bedrock guardrails
+            region: str — AWS region (default 'us-east-1')
+        """
+        from agent.bedrock_adapter import build_converse_kwargs
+
+        region = params.get("region", "us-east-1")
+        guardrail = params.get("guardrail_config")
+
+        kwargs = build_converse_kwargs(
+            model=model,
+            messages=messages,
+            tools=tools,
+            max_tokens=params.get("max_tokens", 4096),
+            temperature=params.get("temperature"),
+            guardrail_config=guardrail,
+        )
+        # Sentinel keys for dispatch — agent pops these before the boto3 call
+        kwargs["__bedrock_converse__"] = True
+        kwargs["__bedrock_region__"] = region
+        return kwargs
+
+    def normalize_response(self, response: Any, **kwargs) -> NormalizedResponse:
+        """Normalize Bedrock response to NormalizedResponse.
+
+        Handles two shapes:
+        1. Raw boto3 dict (from direct converse() calls)
+        2. Already-normalized SimpleNamespace with .choices (from dispatch site)
+        """
+        from agent.bedrock_adapter import normalize_converse_response
+
+        # Normalize to OpenAI-compatible SimpleNamespace
+        if hasattr(response, "choices") and response.choices:
+            # Already normalized at dispatch site
+            ns = response
+        else:
+            # Raw boto3 dict
+            ns = normalize_converse_response(response)
+
+        choice = ns.choices[0]
+        msg = choice.message
+        finish_reason = choice.finish_reason or "stop"
+
+        tool_calls = None
+        if msg.tool_calls:
+            tool_calls = [
+                ToolCall(
+                    id=tc.id,
+                    name=tc.function.name,
+                    arguments=tc.function.arguments,
+                )
+                for tc in msg.tool_calls
+            ]
+
+        usage = None
+        if hasattr(ns, "usage") and ns.usage:
+            u = ns.usage
+            usage = Usage(
+                prompt_tokens=getattr(u, "prompt_tokens", 0) or 0,
+                completion_tokens=getattr(u, "completion_tokens", 0) or 0,
+                total_tokens=getattr(u, "total_tokens", 0) or 0,
+            )
+
+        reasoning = getattr(msg, "reasoning", None) or getattr(msg, "reasoning_content", None)
+
+        provider_data = {}
+        if getattr(msg, "reasoning_details", None):
+            provider_data["reasoning_details"] = msg.reasoning_details
+        if getattr(msg, "bedrock_content_blocks", None):
+            provider_data["bedrock_content_blocks"] = msg.bedrock_content_blocks
+
+        return NormalizedResponse(
+            content=msg.content,
+            tool_calls=tool_calls,
+            finish_reason=finish_reason,
+            reasoning=reasoning,
+            usage=usage,
+            provider_data=provider_data or None,
+        )
+
+    def validate_response(self, response: Any) -> bool:
+        """Check Bedrock response structure.
+
+        After normalize_converse_response, the response has OpenAI-compatible
+        .choices — same check as chat_completions.
+        """
+        if response is None:
+            return False
+        # Raw Bedrock dict response — check for 'output' key
+        if isinstance(response, dict):
+            return "output" in response
+        # Already-normalized SimpleNamespace
+        if hasattr(response, "choices"):
+            return bool(response.choices)
+        return False
+
+    def map_finish_reason(self, raw_reason: str) -> str:
+        """Map Bedrock stop reason to OpenAI finish_reason.
+
+        The adapter already does this mapping inside normalize_converse_response,
+        so this is only used for direct access to raw responses.
+        """
+        _MAP = {
+            "end_turn": "stop",
+            "tool_use": "tool_calls",
+            "max_tokens": "length",
+            "stop_sequence": "stop",
+            "guardrail_intervened": "content_filter",
+            "content_filtered": "content_filter",
+        }
+        return _MAP.get(raw_reason, "stop")
+
+
+# Auto-register on import
+from agent.transports import register_transport  # noqa: E402
+
+register_transport("bedrock_converse", BedrockTransport)

agent/transports/codex_app_server.py

@@ -0,0 +1,418 @@
+"""Codex app-server JSON-RPC client.
+
+Speaks the protocol documented in codex-rs/app-server/README.md (codex 0.125+).
+Transport is newline-delimited JSON-RPC 2.0 over stdio: spawn `codex app-server`,
+do an `initialize` handshake, then drive `thread/start` + `turn/start` and
+consume streaming `item/*` notifications until `turn/completed`.
+
+This module is the wire-level speaker only. Higher-level concerns (event
+projection into Hermes' display, approval bridging, transcript projection into
+AIAgent.messages, plugin migration) live in sibling modules.
+
+Status: optional opt-in runtime gated behind `model.openai_runtime ==
+"codex_app_server"`. Hermes' default tool dispatch is unchanged when this
+runtime is not selected.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import queue
+import subprocess
+import threading
+import time
+from dataclasses import dataclass, field
+from typing import Any, Optional
+
+from tools.environments.local import hermes_subprocess_env
+
+# Default minimum codex version we test against. The PR sets this from the
+# `codex --version` parsed at install time; bumping is a one-line change here.
+MIN_CODEX_VERSION = (0, 125, 0)
+
+
+@dataclass
+class CodexAppServerError(RuntimeError):
+    """Raised on JSON-RPC errors from the app-server."""
+
+    code: int
+    message: str
+    data: Optional[Any] = None
+
+    def __str__(self) -> str:  # pragma: no cover - trivial
+        return f"codex app-server error {self.code}: {self.message}"
+
+
+@dataclass
+class _Pending:
+    queue: queue.Queue
+    method: str
+    sent_at: float = field(default_factory=time.time)
+
+
+class CodexAppServerClient:
+    """Minimal JSON-RPC 2.0 client for `codex app-server` over stdio.
+
+    Threading model:
+      - Spawning thread (caller) drives request/response pairs synchronously.
+      - One reader thread parses stdout, dispatches replies to the right
+        pending future, and routes notifications + server-initiated requests
+        to bounded queues that the caller drains on their own cadence.
+      - One reader thread captures stderr for diagnostics; codex emits
+        tracing logs there at RUST_LOG-controlled levels.
+
+    Intentionally NOT async. AIAgent.run_conversation() is synchronous and
+    runs on the main thread; layering asyncio just to drive a stdio child
+    creates surprising interrupt semantics. We use blocking queues with
+    timeouts and rely on `turn/interrupt` for cancellation.
+    """
+
+    def __init__(
+        self,
+        codex_bin: str = "codex",
+        codex_home: Optional[str] = None,
+        extra_args: Optional[list[str]] = None,
+        env: Optional[dict[str, str]] = None,
+    ) -> None:
+        self._codex_bin = codex_bin
+        # codex app-server is a model-driving CLI executor: it runs a
+        # model-chosen agentic loop that executes shell commands, so it
+        # legitimately needs LLM provider credentials (inherit_credentials=True)
+        # to authenticate against the model endpoint. But the previous
+        # `os.environ.copy()` also handed it every Tier-1 Hermes secret — gateway
+        # bot tokens, GitHub auth, Modal/Daytona infra tokens, the dashboard
+        # session token, AUXILIARY_* side-LLM keys, GATEWAY_RELAY_* auth — none
+        # of which a coding subprocess has any use for. Route through the
+        # centralized helper so Tier-1 + dynamic-internal secrets are always
+        # stripped while provider creds still flow, matching copilot_acp_client
+        # (#29157 sibling spawn-site gap).
+        spawn_env = hermes_subprocess_env(inherit_credentials=True)
+        if env:
+            spawn_env.update(env)
+        if codex_home:
+            spawn_env["CODEX_HOME"] = codex_home
+
+        app_server_args = list(extra_args or [])
+        # Kanban workers must be able to write their handoff/status back to
+        # the board DB, which lives outside the per-task workspace. Keep the
+        # Codex sandbox on, but add the Kanban root as the only extra writable
+        # root. Without this, codex-runtime workers finish their actual work
+        # but crash/block when kanban_complete/kanban_block writes SQLite.
+        if spawn_env.get("HERMES_KANBAN_TASK"):
+            kanban_db = spawn_env.get("HERMES_KANBAN_DB")
+            kanban_root = (
+                os.path.dirname(kanban_db)
+                if kanban_db
+                else spawn_env.get(
+                    "HERMES_KANBAN_ROOT",
+                    os.path.join(
+                        spawn_env.get("HERMES_HOME", os.path.expanduser("~/.hermes")),
+                        "kanban",
+                    ),
+                )
+            )
+            app_server_args.extend(
+                [
+                    "-c",
+                    'sandbox_mode="workspace-write"',
+                    "-c",
+                    f'sandbox_workspace_write.writable_roots=["{kanban_root}"]',
+                    "-c",
+                    "sandbox_workspace_write.network_access=false",
+                ]
+            )
+
+        cmd = [codex_bin, "app-server"] + app_server_args
+        # Codex emits tracing to stderr; default WARN keeps it quiet for users.
+        spawn_env.setdefault("RUST_LOG", "warn")
+
+        # Hide the console the codex child would otherwise flash on Windows
+        # (#56747). Hide-only — stdio pipes stay intact for the app-server wire.
+        from hermes_cli._subprocess_compat import windows_hide_flags
+
+        self._proc = subprocess.Popen(
+            cmd,
+            stdin=subprocess.PIPE,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
+            bufsize=0,
+            env=spawn_env,
+            creationflags=windows_hide_flags(),
+        )
+        self._next_id = 1
+        self._pending: dict[int, _Pending] = {}
+        self._pending_lock = threading.Lock()
+        self._notifications: queue.Queue = queue.Queue()
+        self._server_requests: queue.Queue = queue.Queue()
+        self._stderr_lines: list[str] = []
+        self._stderr_lock = threading.Lock()
+        self._closed = False
+        self._initialized = False
+
+        self._reader = threading.Thread(target=self._read_stdout, daemon=True)
+        self._reader.start()
+        self._stderr_reader = threading.Thread(target=self._read_stderr, daemon=True)
+        self._stderr_reader.start()
+
+    # ---------- lifecycle ----------
+
+    def initialize(
+        self,
+        client_name: str = "hermes",
+        client_title: str = "Hermes Agent",
+        client_version: str = "0.1",
+        capabilities: Optional[dict] = None,
+        timeout: float = 10.0,
+    ) -> dict:
+        """Send `initialize` + `initialized` handshake. Returns the server's
+        InitializeResponse (userAgent, codexHome, platformFamily, platformOs)."""
+        if self._initialized:
+            raise RuntimeError("already initialized")
+        params = {
+            "clientInfo": {
+                "name": client_name,
+                "title": client_title,
+                "version": client_version,
+            },
+            "capabilities": capabilities or {},
+        }
+        result = self.request("initialize", params, timeout=timeout)
+        self.notify("initialized")
+        self._initialized = True
+        return result
+
+    def close(self, timeout: float = 3.0) -> None:
+        """Close stdin and wait for the subprocess to exit, escalating to kill."""
+        if self._closed:
+            return
+        self._closed = True
+        try:
+            if self._proc.stdin and not self._proc.stdin.closed:
+                self._proc.stdin.close()
+        except Exception:
+            pass
+        try:
+            self._proc.terminate()
+            self._proc.wait(timeout=timeout)
+        except subprocess.TimeoutExpired:
+            try:
+                self._proc.kill()
+                self._proc.wait(timeout=1.0)
+            except Exception:
+                pass
+
+    def __enter__(self) -> "CodexAppServerClient":
+        return self
+
+    def __exit__(self, *exc: Any) -> None:
+        self.close()
+
+    # ---------- send/receive ----------
+
+    def request(
+        self,
+        method: str,
+        params: Optional[dict] = None,
+        timeout: float = 30.0,
+    ) -> dict:
+        """Send a JSON-RPC request and block on the response. Returns `result`,
+        raises CodexAppServerError on `error`."""
+        rid = self._take_id()
+        q: queue.Queue = queue.Queue(maxsize=1)
+        with self._pending_lock:
+            self._pending[rid] = _Pending(queue=q, method=method)
+        self._send({"id": rid, "method": method, "params": params or {}})
+        try:
+            msg = q.get(timeout=timeout)
+        except queue.Empty:
+            with self._pending_lock:
+                self._pending.pop(rid, None)
+            raise TimeoutError(
+                f"codex app-server method {method!r} timed out after {timeout}s"
+            )
+        if "error" in msg:
+            err = msg["error"]
+            raise CodexAppServerError(
+                code=err.get("code", -1),
+                message=err.get("message", ""),
+                data=err.get("data"),
+            )
+        return msg.get("result", {})
+
+    def notify(self, method: str, params: Optional[dict] = None) -> None:
+        """Send a JSON-RPC notification (no id, no response expected)."""
+        self._send({"method": method, "params": params or {}})
+
+    def respond(self, request_id: Any, result: dict) -> None:
+        """Reply to a server-initiated request (e.g. approval prompts)."""
+        self._send({"id": request_id, "result": result})
+
+    def respond_error(
+        self, request_id: Any, code: int, message: str, data: Optional[Any] = None
+    ) -> None:
+        """Reply to a server-initiated request with an error."""
+        err: dict[str, Any] = {"code": code, "message": message}
+        if data is not None:
+            err["data"] = data
+        self._send({"id": request_id, "error": err})
+
+    def take_notification(self, timeout: float = 0.0) -> Optional[dict]:
+        """Pop the next streaming notification, or return None on timeout.
+
+        timeout=0.0 means non-blocking. Use small positive timeouts inside the
+        AIAgent turn loop to interleave reads with interrupt checks."""
+        try:
+            if timeout <= 0:
+                return self._notifications.get_nowait()
+            return self._notifications.get(timeout=timeout)
+        except queue.Empty:
+            return None
+
+    def take_server_request(self, timeout: float = 0.0) -> Optional[dict]:
+        """Pop the next server-initiated request (e.g. exec/applyPatch approval)."""
+        try:
+            if timeout <= 0:
+                return self._server_requests.get_nowait()
+            return self._server_requests.get(timeout=timeout)
+        except queue.Empty:
+            return None
+
+    # ---------- diagnostics ----------
+
+    def stderr_tail(self, n: int = 20) -> list[str]:
+        """Return last n lines of codex's stderr (for error reports)."""
+        with self._stderr_lock:
+            return list(self._stderr_lines[-n:])
+
+    def is_alive(self) -> bool:
+        return self._proc.poll() is None
+
+    # ---------- internals ----------
+
+    def _take_id(self) -> int:
+        # JSON-RPC ids only need to be unique per-connection. A simple
+        # monotonically increasing int is the common choice and matches what
+        # codex's own clients use.
+        rid = self._next_id
+        self._next_id += 1
+        return rid
+
+    def _send(self, obj: dict) -> None:
+        if self._closed:
+            raise RuntimeError("codex app-server client is closed")
+        if self._proc.stdin is None:
+            raise RuntimeError("codex app-server stdin not available")
+        try:
+            self._proc.stdin.write((json.dumps(obj) + "\n").encode("utf-8"))
+            self._proc.stdin.flush()
+        except (BrokenPipeError, ValueError) as exc:
+            raise RuntimeError(
+                f"codex app-server stdin closed unexpectedly: {exc}"
+            ) from exc
+
+    def _read_stdout(self) -> None:
+        if self._proc.stdout is None:
+            return
+        try:
+            for line in iter(self._proc.stdout.readline, b""):
+                if not line:
+                    break
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    msg = json.loads(line)
+                except json.JSONDecodeError:
+                    # Non-JSON output is unexpected on stdout; tracing belongs
+                    # on stderr. Surface it via stderr buffer for diagnostics.
+                    with self._stderr_lock:
+                        self._stderr_lines.append(
+                            f"<non-json on stdout> {line[:200]!r}"
+                        )
+                    continue
+                self._dispatch(msg)
+        except Exception as exc:
+            with self._stderr_lock:
+                self._stderr_lines.append(f"<stdout reader error> {exc}")
+
+    def _dispatch(self, msg: dict) -> None:
+        # Reply (has id + result/error, no method)
+        if "id" in msg and ("result" in msg or "error" in msg):
+            with self._pending_lock:
+                pending = self._pending.pop(msg["id"], None)
+            if pending is not None:
+                try:
+                    pending.queue.put_nowait(msg)
+                except queue.Full:  # pragma: no cover - defensive
+                    pass
+            return
+        # Server-initiated request (has id + method)
+        if "id" in msg and "method" in msg:
+            self._server_requests.put(msg)
+            return
+        # Notification (no id)
+        if "method" in msg:
+            self._notifications.put(msg)
+
+    def _read_stderr(self) -> None:
+        if self._proc.stderr is None:
+            return
+        try:
+            for line in iter(self._proc.stderr.readline, b""):
+                if not line:
+                    break
+                with self._stderr_lock:
+                    self._stderr_lines.append(
+                        line.decode("utf-8", "replace").rstrip()
+                    )
+                    # Bound memory: keep last 500 lines.
+                    if len(self._stderr_lines) > 500:
+                        self._stderr_lines = self._stderr_lines[-500:]
+        except Exception:  # pragma: no cover
+            pass
+
+
+def parse_codex_version(output: str) -> Optional[tuple[int, int, int]]:
+    """Parse `codex --version` output. Returns (major, minor, patch) or None."""
+    # Output format: "codex-cli 0.130.0" possibly followed by metadata.
+    import re
+
+    match = re.search(r"(\d+)\.(\d+)\.(\d+)", output or "")
+    if not match:
+        return None
+    return (int(match.group(1)), int(match.group(2)), int(match.group(3)))
+
+
+def check_codex_binary(
+    codex_bin: str = "codex", min_version: tuple[int, int, int] = MIN_CODEX_VERSION
+) -> tuple[bool, str]:
+    """Verify codex CLI is installed and meets minimum version.
+
+    Returns (ok, message). Used by setup wizard and runtime startup."""
+    try:
+        proc = subprocess.run(
+            [codex_bin, "--version"],
+            capture_output=True,
+            text=True, encoding='utf-8', errors='replace',
+            timeout=10,
+            stdin=subprocess.DEVNULL,
+        )
+    except FileNotFoundError:
+        return False, (
+            f"codex CLI not found at {codex_bin!r}. Install with: "
+            f"npm i -g @openai/codex"
+        )
+    except subprocess.TimeoutExpired:
+        return False, "codex --version timed out"
+    if proc.returncode != 0:
+        return False, f"codex --version exited {proc.returncode}: {proc.stderr.strip()}"
+    version = parse_codex_version(proc.stdout)
+    if version is None:
+        return False, f"could not parse codex version from: {proc.stdout!r}"
+    if version < min_version:
+        return False, (
+            f"codex {'.'.join(map(str, version))} is older than required "
+            f"{'.'.join(map(str, min_version))}. Run: npm i -g @openai/codex"
+        )
+    return True, ".".join(map(str, version))

agent/transports/codex_event_projector.py

@@ -0,0 +1,314 @@
+"""Projects codex app-server events into Hermes' messages list.
+
+The translator that lets Hermes' memory/skill review keep working under the
+Codex runtime: it converts Codex `item/*` notifications into the standard
+OpenAI-shaped `{role, content, tool_calls, tool_call_id}` entries that
+`agent/curator.py` already knows how to read.
+
+Codex emits items with a discriminator field `type`:
+  - userMessage         → {role: "user", content}
+  - agentMessage        → {role: "assistant", content}
+  - reasoning           → stashed in the assistant's "reasoning" field
+  - commandExecution    → assistant tool_call(name="exec") + tool result
+  - fileChange          → assistant tool_call(name="apply_patch") + tool result
+  - mcpToolCall         → assistant tool_call(name=f"mcp.{server}.{tool}") + tool result
+  - dynamicToolCall     → assistant tool_call(name=tool) + tool result
+  - plan/hookPrompt/collabAgentToolCall → recorded as opaque assistant notes
+
+Each item maps to AT MOST one assistant entry + one tool entry, preserving
+Hermes' message-alternation invariants (system → user → assistant → user/tool
+→ assistant → ...). Multiple Codex tool calls within one Codex turn produce
+multiple consecutive (assistant, tool) pairs, which is the same shape Hermes
+already produces for parallel tool calls.
+
+Counters tracked alongside projection:
+  - tool_iterations: ticks once per completed tool-shaped item. Used by
+    AIAgent._iters_since_skill (skill nudge gate, default threshold 10).
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+from dataclasses import dataclass, field
+from typing import Any, Optional
+
+
+def _deterministic_call_id(item_type: str, item_id: str) -> str:
+    """Stable id for tool_call message correlation.
+
+    Uses the codex item id directly when present (already a uuid); falls back
+    to a content hash so replay produces the same id across sessions and
+    prefix caches stay valid. See AGENTS.md Pitfall #16 (deterministic IDs in
+    tool call history)."""
+    if item_id:
+        return f"codex_{item_type}_{item_id}"
+    digest = hashlib.sha256(f"{item_type}".encode()).hexdigest()[:16]
+    return f"codex_{item_type}_{digest}"
+
+
+def _format_tool_args(d: dict) -> str:
+    """Format a dict as JSON the way Hermes' existing tool_calls path does."""
+    return json.dumps(d, ensure_ascii=False, sort_keys=True)
+
+
+@dataclass
+class ProjectionResult:
+    """Output of projecting one Codex item.
+
+    `messages` is a list because some Codex items produce two messages
+    (assistant tool_call + tool result). Empty list = item ignored (e.g. a
+    streaming `outputDelta` that doesn't materialize into messages until the
+    `item/completed` event)."""
+
+    messages: list[dict] = field(default_factory=list)
+    is_tool_iteration: bool = False
+    final_text: Optional[str] = None  # Set when an agentMessage completes
+
+
+class CodexEventProjector:
+    """Stateful projector consuming Codex notifications in arrival order.
+
+    Owns the in-progress reasoning content (codex emits reasoning as separate
+    items but Hermes stashes it on the next assistant message)."""
+
+    def __init__(self) -> None:
+        self._pending_reasoning: list[str] = []
+
+    def project(self, notification: dict) -> ProjectionResult:
+        """Project a single notification. Idempotent for non-completion events;
+        only `item/completed` and `turn/completed` materialize messages."""
+        method = notification.get("method", "")
+        params = notification.get("params", {}) or {}
+
+        # We only materialize messages on `item/completed`. Streaming deltas
+        # (`item/<type>/outputDelta`, `item/<type>/delta`) are display-only and
+        # don't enter the messages list — same way Hermes already only writes
+        # the assistant message after the streaming completion event.
+        if method != "item/completed":
+            return ProjectionResult()
+
+        item = params.get("item") or {}
+        item_type = item.get("type") or ""
+        item_id = item.get("id") or ""
+
+        if item_type == "agentMessage":
+            return self._project_agent_message(item)
+        if item_type == "reasoning":
+            self._pending_reasoning.extend(item.get("summary") or [])
+            self._pending_reasoning.extend(item.get("content") or [])
+            return ProjectionResult()
+        if item_type == "commandExecution":
+            return self._project_command(item, item_id)
+        if item_type == "fileChange":
+            return self._project_file_change(item, item_id)
+        if item_type == "mcpToolCall":
+            return self._project_mcp_tool_call(item, item_id)
+        if item_type == "dynamicToolCall":
+            return self._project_dynamic_tool_call(item, item_id)
+        if item_type == "userMessage":
+            return self._project_user_message(item)
+
+        # Unknown / rare items (plan, hookPrompt, collabAgentToolCall, etc.)
+        # — record as opaque assistant note so memory review can still see
+        # *something* happened, but don't fabricate tool_call structure.
+        return self._project_opaque(item, item_type)
+
+    # ---------- per-type projections ----------
+
+    def _project_agent_message(self, item: dict) -> ProjectionResult:
+        text = item.get("text") or ""
+        msg: dict[str, Any] = {"role": "assistant", "content": text}
+        if self._pending_reasoning:
+            msg["reasoning"] = "\n".join(self._pending_reasoning)
+            self._pending_reasoning = []
+        return ProjectionResult(messages=[msg], final_text=text)
+
+    def _project_user_message(self, item: dict) -> ProjectionResult:
+        # codex's userMessage content is a list of UserInput variants. For
+        # projection purposes we flatten any text fragments and ignore
+        # non-text parts (images, etc.) — Hermes' messages store text only.
+        text_parts: list[str] = []
+        for fragment in item.get("content") or []:
+            if isinstance(fragment, dict):
+                if fragment.get("type") == "text":
+                    text_parts.append(fragment.get("text") or "")
+                elif "text" in fragment:
+                    text_parts.append(str(fragment["text"]))
+        return ProjectionResult(
+            messages=[{"role": "user", "content": "\n".join(text_parts)}]
+        )
+
+    def _project_command(self, item: dict, item_id: str) -> ProjectionResult:
+        call_id = _deterministic_call_id("exec", item_id)
+        args = {
+            "command": item.get("command") or "",
+            "cwd": item.get("cwd") or "",
+        }
+        assistant_msg = {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [
+                {
+                    "id": call_id,
+                    "type": "function",
+                    "function": {
+                        "name": "exec_command",
+                        "arguments": _format_tool_args(args),
+                    },
+                }
+            ],
+        }
+        if self._pending_reasoning:
+            assistant_msg["reasoning"] = "\n".join(self._pending_reasoning)
+            self._pending_reasoning = []
+        output = item.get("aggregatedOutput") or ""
+        exit_code = item.get("exitCode")
+        if exit_code is not None and exit_code != 0:
+            output = f"[exit {exit_code}]\n{output}"
+        tool_msg = {
+            "role": "tool",
+            "tool_call_id": call_id,
+            "content": output,
+        }
+        return ProjectionResult(
+            messages=[assistant_msg, tool_msg], is_tool_iteration=True
+        )
+
+    def _project_file_change(self, item: dict, item_id: str) -> ProjectionResult:
+        call_id = _deterministic_call_id("apply_patch", item_id)
+        # Reduce the codex changes array to a digest the agent loop will
+        # find readable. We record per-file change kinds (Add/Update/Delete)
+        # without inlining full file contents — those can be huge.
+        changes_summary = []
+        for change in item.get("changes") or []:
+            kind = (change.get("kind") or {}).get("type") or "update"
+            path = change.get("path") or ""
+            changes_summary.append({"kind": kind, "path": path})
+        args = {"changes": changes_summary}
+        assistant_msg = {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [
+                {
+                    "id": call_id,
+                    "type": "function",
+                    "function": {
+                        "name": "apply_patch",
+                        "arguments": _format_tool_args(args),
+                    },
+                }
+            ],
+        }
+        if self._pending_reasoning:
+            assistant_msg["reasoning"] = "\n".join(self._pending_reasoning)
+            self._pending_reasoning = []
+        status = item.get("status") or "unknown"
+        n = len(changes_summary)
+        tool_msg = {
+            "role": "tool",
+            "tool_call_id": call_id,
+            "content": f"apply_patch status={status}, {n} change(s)",
+        }
+        return ProjectionResult(
+            messages=[assistant_msg, tool_msg], is_tool_iteration=True
+        )
+
+    def _project_mcp_tool_call(self, item: dict, item_id: str) -> ProjectionResult:
+        server = item.get("server") or "mcp"
+        tool = item.get("tool") or "unknown"
+        # Mirror the native MCP tool-name convention (mcp__server__tool) so the
+        # deterministic call_id input stays consistent with registration names.
+        call_id = _deterministic_call_id(f"mcp__{server}__{tool}", item_id)
+        args = item.get("arguments") or {}
+        if not isinstance(args, dict):
+            args = {"arguments": args}
+        assistant_msg = {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [
+                {
+                    "id": call_id,
+                    "type": "function",
+                    "function": {
+                        "name": f"mcp.{server}.{tool}",
+                        "arguments": _format_tool_args(args),
+                    },
+                }
+            ],
+        }
+        if self._pending_reasoning:
+            assistant_msg["reasoning"] = "\n".join(self._pending_reasoning)
+            self._pending_reasoning = []
+        result = item.get("result")
+        error = item.get("error")
+        if error:
+            content = f"[error] {json.dumps(error, ensure_ascii=False)[:1000]}"
+        elif result is not None:
+            content = json.dumps(result, ensure_ascii=False)[:4000]
+        else:
+            content = ""
+        tool_msg = {
+            "role": "tool",
+            "tool_call_id": call_id,
+            "content": content,
+        }
+        return ProjectionResult(
+            messages=[assistant_msg, tool_msg], is_tool_iteration=True
+        )
+
+    def _project_dynamic_tool_call(
+        self, item: dict, item_id: str
+    ) -> ProjectionResult:
+        tool = item.get("tool") or "unknown"
+        call_id = _deterministic_call_id(f"dyn_{tool}", item_id)
+        args = item.get("arguments") or {}
+        if not isinstance(args, dict):
+            args = {"arguments": args}
+        assistant_msg = {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [
+                {
+                    "id": call_id,
+                    "type": "function",
+                    "function": {
+                        "name": tool,
+                        "arguments": _format_tool_args(args),
+                    },
+                }
+            ],
+        }
+        if self._pending_reasoning:
+            assistant_msg["reasoning"] = "\n".join(self._pending_reasoning)
+            self._pending_reasoning = []
+        content_items = item.get("contentItems") or []
+        if isinstance(content_items, list) and content_items:
+            content = json.dumps(content_items, ensure_ascii=False)[:4000]
+        else:
+            success = item.get("success")
+            content = f"success={success}"
+        tool_msg = {
+            "role": "tool",
+            "tool_call_id": call_id,
+            "content": content,
+        }
+        return ProjectionResult(
+            messages=[assistant_msg, tool_msg], is_tool_iteration=True
+        )
+
+    def _project_opaque(self, item: dict, item_type: str) -> ProjectionResult:
+        # Record the existence of the item without inventing tool_calls.
+        # Memory review will see this and may or may not save anything.
+        try:
+            payload = json.dumps(item, ensure_ascii=False)[:1500]
+        except (TypeError, ValueError):
+            payload = repr(item)[:1500]
+        return ProjectionResult(
+            messages=[
+                {
+                    "role": "assistant",
+                    "content": f"[codex {item_type}] {payload}",
+                }
+            ]
+        )

agent/transports/hermes_tools_mcp_server.py

@@ -0,0 +1,289 @@
+"""Hermes-tools-as-MCP server for the codex_app_server runtime.
+
+When the user runs `openai/*` turns through the codex app-server, codex
+owns the loop and builds its own tool list. By default, that means
+Hermes' richer tool surface — web search, browser automation,
+delegate_task subagents, vision analysis, persistent memory, skills,
+cross-session search, image generation, TTS — is unreachable.
+
+This module exposes a curated subset of those Hermes tools to the
+spawned codex subprocess via stdio MCP. Codex registers it as a normal
+MCP server (per `~/.codex/config.toml [mcp_servers.hermes-tools]`) and
+the user gets full Hermes capability inside a Codex turn.
+
+Scope (what we expose):
+  - web_search, web_extract              — Firecrawl, no codex equivalent
+  - browser_navigate / _click / _type /  — Camofox/Browserbase automation
+    _snapshot / _scroll / _back / _press /
+    _get_images / _console / _vision
+  - vision_analyze                       — image inspection by vision model
+  - image_generate                       — image generation
+  - skill_view, skills_list              — Hermes' skill library
+  - text_to_speech                       — TTS
+  - kanban_* (complete/block/comment/    — kanban worker + orchestrator
+    heartbeat/show/list/create/            handoff (stateless: read env var,
+    unblock/link)                          write ~/.hermes/kanban.db)
+
+What we DO NOT expose:
+  - terminal / shell                     — codex's own shell tool
+  - read_file / write_file / patch       — codex's apply_patch + shell
+  - search_files / process               — codex's shell
+  - clarify                              — codex's own UX
+  - delegate_task / memory /             — `_AGENT_LOOP_TOOLS` in Hermes
+    session_search / todo                  (model_tools.py). They require
+                                           the running AIAgent context to
+                                           dispatch (mid-loop state), so a
+                                           stateless MCP callback can't
+                                           drive them. See the inline
+                                           comment on EXPOSED_TOOLS below.
+
+Run with: python -m agent.transports.hermes_tools_mcp_server
+Spawned by: CodexAppServerSession.ensure_started() when the runtime is
+            active and config opts in.
+"""
+
+from __future__ import annotations
+
+import inspect
+import json
+import logging
+import os
+import sys
+from typing import Any, Optional
+
+logger = logging.getLogger(__name__)
+
+# JSON Schema type -> Python type mapping for signature generation
+_JSON_TO_PY = {
+    "string": str,
+    "integer": int,
+    "number": float,
+    "boolean": bool,
+    "array": list,
+    "object": dict,
+}
+
+
+def _signature_from_schema(schema: dict | None) -> tuple[inspect.Signature, dict[str, type]]:
+    """Build a Python function signature and annotations from a JSON schema.
+
+    Args:
+        schema: JSON Schema dict with "properties" and "required" keys.
+
+    Returns:
+        (signature, annotations_dict) where signature has KEYWORD_ONLY params
+        and annotations maps param names to Python types.
+    """
+    props = (schema or {}).get("properties") or {}
+    required = set((schema or {}).get("required") or [])
+    params, annots = [], {}
+
+    for pname, pspec in props.items():
+        if pname.startswith("_"):
+            continue
+        py = _JSON_TO_PY.get((pspec or {}).get("type"), Any)
+        ann, default = (
+            (py, inspect.Parameter.empty)
+            if pname in required
+            else (Optional[py], None)
+        )
+        annots[pname] = ann
+        params.append(
+            inspect.Parameter(
+                pname, inspect.Parameter.KEYWORD_ONLY, annotation=ann, default=default
+            )
+        )
+
+    return inspect.Signature(params, return_annotation=str), annots
+
+
+# Tools we expose. Each name MUST match a registered Hermes tool that
+# `model_tools.handle_function_call()` can dispatch.
+#
+# What we deliberately DO NOT expose:
+#   - terminal / shell / read_file / write_file / patch / search_files /
+#     process — codex's built-ins cover these and approval routes through
+#     codex's own UI.
+#   - delegate_task / memory / session_search / todo — these are
+#     `_AGENT_LOOP_TOOLS` in Hermes (model_tools.py:493). They require
+#     the running AIAgent context to dispatch (mid-loop state), so a
+#     stateless MCP callback can't drive them. Hermes' default runtime
+#     keeps these working; the codex_app_server runtime cannot.
+EXPOSED_TOOLS: tuple[str, ...] = (
+    "web_search",
+    "web_extract",
+    "browser_navigate",
+    "browser_click",
+    "browser_type",
+    "browser_press",
+    "browser_snapshot",
+    "browser_scroll",
+    "browser_back",
+    "browser_get_images",
+    "browser_console",
+    "browser_vision",
+    "vision_analyze",
+    "image_generate",
+    "skill_view",
+    "skills_list",
+    "text_to_speech",
+    # Kanban worker handoff tools — gated on HERMES_KANBAN_TASK env var
+    # (set by the kanban dispatcher when spawning a worker). Without these
+    # in the callback, a worker spawned with openai_runtime=codex_app_server
+    # could do the work but couldn't report completion back to the kernel,
+    # making it hang until timeout. Stateless dispatch — they just read
+    # the env var and write to ~/.hermes/kanban.db.
+    "kanban_complete",
+    "kanban_block",
+    "kanban_request_review",
+    "kanban_request_changes",
+    "kanban_comment",
+    "kanban_heartbeat",
+    "kanban_show",
+    "kanban_list",
+    # NOTE: kanban_create / kanban_unblock / kanban_link are orchestrator-
+    # only — the kanban tool gates them on HERMES_KANBAN_TASK being unset.
+    # They're exposed here for orchestrator agents running on the codex
+    # runtime that need to dispatch new tasks.
+    "kanban_create",
+    "kanban_unblock",
+    "kanban_link",
+)
+
+
+def _build_server() -> Any:
+    """Create the MCP server with Hermes tools attached. Lazy imports
+    so the module can be imported without the mcp package installed
+    (we degrade to a clear error only when actually run)."""
+    try:
+        # mcp 2.0 removed `mcp.server.fastmcp`; `mcp.server.MCPServer` is the
+        # same decorator/add_tool surface under the new name.
+        from mcp.server import MCPServer
+    except ImportError as exc:  # pragma: no cover - install hint
+        raise ImportError(
+            f"hermes-tools MCP server requires the 'mcp' package: {exc}"
+        ) from exc
+
+    # Discover Hermes tools so dispatch works.
+    from model_tools import (
+        get_tool_definitions,
+        handle_function_call,
+    )
+
+    mcp = MCPServer(
+        "hermes-tools",
+        instructions=(
+            "Hermes Agent's tool surface, exposed for use inside a Codex "
+            "session. Use these for capabilities Codex's built-in toolset "
+            "doesn't cover: web search/extract, browser automation, "
+            "subagent delegation, vision, image generation, persistent "
+            "memory, skills, and cross-session search."
+        ),
+    )
+
+    # Pull authoritative Hermes tool schemas for the ones we expose, so
+    # MCP clients see the same parameter docs Hermes gives the model.
+    all_defs = {
+        td["function"]["name"]: td["function"]
+        for td in (get_tool_definitions(quiet_mode=True) or [])
+        if isinstance(td, dict) and td.get("type") == "function"
+    }
+
+    exposed_count = 0
+
+    for name in EXPOSED_TOOLS:
+        spec = all_defs.get(name)
+        if spec is None:
+            logger.debug(
+                "skipping %s — not registered in this Hermes process", name
+            )
+            continue
+
+        description = spec.get("description") or f"Hermes {name} tool"
+        params_schema = spec.get("parameters") or {"type": "object", "properties": {}}
+
+        # The SDK wants a Python callable and derives the input schema from
+        # its signature — there is no inputSchema parameter on either the
+        # decorator or add_tool(). So build a closure that takes the arguments
+        # dict, dispatches via handle_function_call, returns the result
+        # string, and carries a __signature__ synthesized from the Hermes
+        # JSON Schema (see _signature_from_schema) for the SDK to read.
+        def _make_handler(tool_name: str, schema: dict | None):
+            sig, annots = _signature_from_schema(schema)
+
+            def _dispatch(**kwargs: Any) -> str:
+                try:
+                    # Filter out None values before dispatch so unset optionals
+                    # aren't forwarded to the handler.
+                    args = {k: v for k, v in kwargs.items() if v is not None}
+                    return handle_function_call(tool_name, args or {})
+                except Exception as exc:
+                    logger.exception("tool %s raised", tool_name)
+                    return json.dumps({"error": str(exc), "tool": tool_name})
+
+            _dispatch.__name__ = tool_name
+            _dispatch.__doc__ = description
+            _dispatch.__signature__ = sig
+            _dispatch.__annotations__ = {**annots, "return": str}
+            return _dispatch
+
+        try:
+            mcp.add_tool(
+                _make_handler(name, params_schema),
+                name=name,
+                description=description,
+            )
+        except TypeError:
+            # Older mcp SDK signature — fall back to decorator-style. The
+            # synthesized __signature__ on the handler still drives schema
+            # generation there.
+            handler = _make_handler(name, params_schema)
+            handler = mcp.tool(name=name, description=description)(handler)
+
+        exposed_count += 1
+
+    logger.info(
+        "hermes-tools MCP server registered %d/%d tools",
+        exposed_count,
+        len(EXPOSED_TOOLS),
+    )
+    return mcp
+
+
+def main(argv: Optional[list[str]] = None) -> int:
+    """Entry point for `python -m agent.transports.hermes_tools_mcp_server`."""
+    argv = argv or sys.argv[1:]
+    verbose = "--verbose" in argv or "-v" in argv
+
+    log_level = logging.INFO if verbose else logging.WARNING
+    logging.basicConfig(
+        level=log_level,
+        stream=sys.stderr,  # MCP uses stdio for protocol — logs MUST go to stderr
+        format="%(asctime)s [%(levelname)s] %(name)s: %(message)s",
+    )
+
+    # Quiet mode: keep Hermes' own banners off stdout (which is the MCP wire).
+    os.environ.setdefault("HERMES_QUIET", "1")
+    os.environ.setdefault("HERMES_REDACT_SECRETS", "true")
+
+    try:
+        server = _build_server()
+    except ImportError as exc:
+        sys.stderr.write(f"hermes-tools MCP server cannot start: {exc}\n")
+        return 2
+
+    # MCPServer.run() defaults to stdio transport, which is what codex
+    # spawns us on.
+    try:
+        server.run()
+    except KeyboardInterrupt:
+        return 0
+    except Exception as exc:
+        logger.exception("hermes-tools MCP server crashed")
+        sys.stderr.write(f"hermes-tools MCP server error: {exc}\n")
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

agent/transports/types.py

@@ -0,0 +1,180 @@
+"""Shared types for normalized provider responses.
+
+These dataclasses define the canonical shape that all provider adapters
+normalize responses to.  The shared surface is intentionally minimal —
+only fields that every downstream consumer reads are top-level.
+Protocol-specific state goes in ``provider_data`` dicts (response-level
+and per-tool-call) so that protocol-aware code paths can access it
+without polluting the shared type.
+"""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass, field
+from typing import Any
+
+
+@dataclass
+class ToolCall:
+    """A normalized tool call from any provider.
+
+    ``id`` is the protocol's canonical identifier — what gets used in
+    ``tool_call_id`` / ``tool_use_id`` when constructing tool result
+    messages.  May be ``None`` when the provider omits it; the agent
+    fills it via ``_deterministic_call_id()`` before storing in history.
+
+    ``provider_data`` carries per-tool-call protocol metadata that only
+    protocol-aware code reads:
+
+    * Codex: ``{"call_id": "call_XXX", "response_item_id": "fc_XXX"}``
+    * Gemini: ``{"extra_content": {"google": {"thought_signature": "..."}}}``
+    * Others: ``None``
+    """
+
+    id: str | None
+    name: str
+    arguments: str  # JSON string
+    provider_data: dict[str, Any] | None = field(default=None, repr=False)
+
+    # ── Backward compatibility ──────────────────────────────────
+    # The agent loop reads tc.function.name / tc.function.arguments
+    # throughout run_agent.py (45+ sites).  These properties let
+    # NormalizedResponse pass through without the _nr_to_assistant_message
+    # shim, while keeping ToolCall's canonical fields flat.
+    @property
+    def type(self) -> str:
+        return "function"
+
+    @property
+    def function(self) -> ToolCall:
+        """Return self so tc.function.name / tc.function.arguments work."""
+        return self
+
+    @property
+    def call_id(self) -> str | None:
+        """Codex call_id from provider_data, accessed via getattr by _build_assistant_message."""
+        return (self.provider_data or {}).get("call_id")
+
+    @property
+    def response_item_id(self) -> str | None:
+        """Codex response_item_id from provider_data."""
+        return (self.provider_data or {}).get("response_item_id")
+
+    @property
+    def extra_content(self) -> dict[str, Any] | None:
+        """Gemini extra_content (thought_signature) from provider_data.
+
+        Gemini 3 thinking models attach ``extra_content`` with a
+        ``thought_signature`` to each tool call.  This signature must be
+        replayed on subsequent API calls — without it the API rejects the
+        request with HTTP 400.  The chat_completions transport stores this
+        in ``provider_data["extra_content"]``; this property exposes it so
+        ``_build_assistant_message`` can ``getattr(tc, "extra_content")``
+        uniformly.
+        """
+        return (self.provider_data or {}).get("extra_content")
+
+
+@dataclass
+class Usage:
+    """Token usage from an API response."""
+
+    prompt_tokens: int = 0
+    completion_tokens: int = 0
+    total_tokens: int = 0
+    cached_tokens: int = 0
+
+
+@dataclass
+class NormalizedResponse:
+    """Normalized API response from any provider.
+
+    Shared fields are truly cross-provider — every caller can rely on
+    them without branching on api_mode.  Protocol-specific state goes in
+    ``provider_data`` so that only protocol-aware code paths read it.
+
+    Response-level ``provider_data`` examples:
+
+    * Anthropic: ``{"reasoning_details": [...]}``
+    * Codex: ``{"codex_reasoning_items": [...], "codex_message_items": [...]}``
+    * Others: ``None``
+    """
+
+    content: str | None
+    tool_calls: list[ToolCall] | None
+    finish_reason: str  # "stop", "tool_calls", "length", "content_filter"
+    reasoning: str | None = None
+    usage: Usage | None = None
+    provider_data: dict[str, Any] | None = field(default=None, repr=False)
+
+    # ── Backward compatibility ──────────────────────────────────
+    # The shim _nr_to_assistant_message() mapped these from provider_data.
+    # These properties let NormalizedResponse pass through directly.
+    @property
+    def reasoning_content(self) -> str | None:
+        pd = self.provider_data or {}
+        return pd.get("reasoning_content")
+
+    @property
+    def reasoning_details(self):
+        pd = self.provider_data or {}
+        return pd.get("reasoning_details")
+
+    @property
+    def anthropic_content_blocks(self):
+        """Verbatim, order-preserving Anthropic content blocks for a turn.
+
+        Present only when an Anthropic turn interleaves signed thinking with
+        tool_use — the one shape the parallel reasoning_details + tool_calls
+        lists reconstruct in the wrong order, invalidating thinking-block
+        signatures on replay. See agent/transports/anthropic.py.
+        """
+        pd = self.provider_data or {}
+        return pd.get("anthropic_content_blocks")
+
+    @property
+    def bedrock_content_blocks(self):
+        """Verbatim, order-preserving Bedrock Converse content blocks."""
+        pd = self.provider_data or {}
+        return pd.get("bedrock_content_blocks")
+
+    @property
+    def codex_reasoning_items(self):
+        pd = self.provider_data or {}
+        return pd.get("codex_reasoning_items")
+
+    @property
+    def codex_message_items(self):
+        pd = self.provider_data or {}
+        return pd.get("codex_message_items")
+
+
+# ---------------------------------------------------------------------------
+# Factory helpers
+# ---------------------------------------------------------------------------
+
+
+def build_tool_call(
+    id: str | None,
+    name: str,
+    arguments: Any,
+    **provider_fields: Any,
+) -> ToolCall:
+    """Build a ``ToolCall``, auto-serialising *arguments* if it's a dict.
+
+    Any extra keyword arguments are collected into ``provider_data``.
+    """
+    args_str = json.dumps(arguments) if isinstance(arguments, dict) else str(arguments)
+    pd = dict(provider_fields) if provider_fields else None
+    return ToolCall(id=id, name=name, arguments=args_str, provider_data=pd)
+
+
+def map_finish_reason(reason: str | None, mapping: dict[str, str]) -> str:
+    """Translate a provider-specific stop reason to the normalised set.
+
+    Falls back to ``"stop"`` for unknown or ``None`` reasons.
+    """
+    if reason is None:
+        return "stop"
+    return mapping.get(reason, "stop")