Add agentic harness integration: types, HarnessEnvironment, OpenClaw adapter (RFC 005)

[Darktex]

Summary

Implements the full agentic harness integration from RFC 005 (#387). This adds support for wrapping external harnesses (OpenClaw, Claude Code, Gemini CLI, etc.) that own their own ReAct control loop.

What's included

Foundation types (src/openenv/core/harnesses/types.py):

• HarnessConfig — Pydantic model for harness process configuration
• HarnessTransport — Enum: stdio, streamable HTTP, MCP
• HarnessEvent / HarnessEventType — Standard event schema for turn trajectories (8 event types)
• HarnessResponse — Complete turn response with events and done signal
• HarnessAction — Action type for sending messages to harnesses

Adapter ABC (src/openenv/core/harnesses/adapter.py):

• HarnessAdapter — Abstract base for harness-specific lifecycle (start, stop, inject_tools, send_message, send_message_streaming)

HarnessEnvironment (src/openenv/core/harnesses/environment.py):

• reset() starts fresh harness process and conversation
• step(HarnessAction) sends one conversational turn, harness does ReAct loop
• Multi-turn: harness maintains context across step() calls
• Trajectory accumulated across turns via .trajectory property
• Rubric integration for reward computation

OpenClaw adapter (src/openenv/core/harnesses/adapters/openclaw.py):

• Process lifecycle via asyncio subprocess
• MCP tool injection via .openclaw/openclaw.json config
• JSON-line protocol over stdin/stdout
• Environment variable inheritance (merges with parent env)

Utilities (src/openenv/core/harnesses/tools.py):

• resolve_tool_conflicts() — Tool name collision handling via env_ prefixing

Test plan

• 92 unit tests covering all types, environment lifecycle, multi-turn conversations, trajectory, OpenClaw adapter
• Lint clean (usort + ruff format + ruff check)
• No regressions in existing tests

Closes #385

[greptile-apps]

Greptile Summary

This PR implements PR 2 from RFC 005's implementation plan: the foundation type system for agentic harness integration. It introduces Pydantic models (HarnessConfig, HarnessEvent, HarnessResponse, HarnessAction), enums (HarnessTransport, HarnessEventType), an abstract base class (HarnessAdapter), and a tool conflict resolution utility (resolve_tool_conflicts). These types define the contract that concrete harness adapters (OpenClaw, Claude Code, etc.) will implement in subsequent PRs.

The changes are additive and self-contained — no existing code is modified. The types align well with the project's existing Pydantic-based type system and follow the same ConfigDict(extra="forbid") pattern used throughout env_server/types.py.

• All new types live in src/openenv/core/harnesses/, cleanly separated from existing modules
• HarnessAction correctly extends the base Action class with a type discriminator field, consistent with ListToolsAction and CallToolAction in mcp_types.py
• HarnessAdapter ABC defines the lifecycle interface (start, stop, inject_tools, send_message, send_message_streaming, is_alive) that concrete adapters must implement
• resolve_tool_conflicts() handles env_ prefixing for name collisions, though the RFC's "error on ambiguity" behavior (raising errors for same-name-different-schema cases) is not yet implemented — likely deferred to the core implementation PR
• inject_tools on HarnessAdapter uses a bare List without type parameter, losing type information for implementers
• Timeout fields on HarnessConfig lack gt=0 validation, unlike existing timeout fields in the codebase

Alignment Review Report

Automated Checks

• Lint: SKIPPED — uv not available in review environment
• Debug code: CLEAN — no debugger statements found in new files

Tier 1: Fixes Required

• None identified

Tier 2: Alignment Discussion

None identified — the types are purely additive, respect the dual API boundary (harness messages go through step(), not exposed via MCP), and follow the RFC 005 design faithfully.

Confidence Score: 4/5

• This PR is safe to merge — it adds new types with no modifications to existing code and no alignment violations.
• Score of 4 reflects that this is a clean, additive foundation PR with thorough test coverage (43 tests). The two style suggestions (untyped List, missing timeout validation) are minor and don't affect correctness. No architectural invariants are violated. Deducted 1 point for the type safety gap in the abstract interface that downstream implementers will depend on.
• src/openenv/core/harnesses/adapter.py — inject_tools bare List type should be parameterized before concrete adapters depend on it

Important Files Changed

Filename	Overview
src/openenv/core/harnesses/types.py	Defines core Pydantic models and enums (HarnessTransport, HarnessEventType, HarnessEvent, HarnessResponse, HarnessConfig, HarnessAction). Well-structured and follows existing codebase patterns. Minor: timeout fields lack gt=0 validation.
src/openenv/core/harnesses/adapter.py	ABC for harness adapters with well-documented abstract methods. inject_tools uses bare List type without parameterization, losing type safety for implementers.
src/openenv/core/harnesses/tools.py	Tool conflict resolution utility. Clean implementation using model_copy. Correctly uses List[Tool]. RFC mentions "error on ambiguity" for same-name-different-schema cases, not yet implemented (may be future work).
src/openenv/core/harnesses/init.py	Clean module init re-exporting all public types via __all__. No issues.
tests/core/test_harnesses/init.py	Empty test package init with copyright header. No issues.
tests/core/test_harnesses/test_harness_types.py	Thorough test suite covering all types, validation, serialization roundtrips, ABC enforcement, and module exports. Good coverage of edge cases.

Class Diagram

classDiagram
    class Action {
        +Dict metadata
    }
    class HarnessAction {
        +Literal type = "harness_message"
        +str message
    }
    Action <|-- HarnessAction

    class HarnessTransport {
        <<enum>>
        STDIO = "stdio"
        STREAMABLE_HTTP = "http"
        MCP = "mcp"
    }

    class HarnessEventType {
        <<enum>>
        LLM_REQUEST
        LLM_RESPONSE
        LLM_CHUNK
        TOOL_CALL
        TOOL_RESULT
        TEXT_OUTPUT
        ERROR
        TURN_COMPLETE
    }

    class HarnessEvent {
        +HarnessEventType type
        +float timestamp
        +Dict data
    }
    HarnessEvent --> HarnessEventType

    class HarnessResponse {
        +str response
        +List~HarnessEvent~ events
        +bool done
    }
    HarnessResponse --> HarnessEvent

    class HarnessConfig {
        +str name
        +List~str~ command
        +str working_directory
        +Dict env_vars
        +HarnessTransport transport
        +Optional~str~ mcp_config_path
        +float startup_timeout_s
        +float session_timeout_s
        +Optional~str~ model
        +Optional~str~ api_key_env_var
    }
    HarnessConfig --> HarnessTransport

    class HarnessAdapter {
        <<abstract>>
        +HarnessConfig config
        +start(working_directory) None
        +stop() None
        +inject_tools(tools) None
        +send_message(message) HarnessResponse
        +send_message_streaming(message) AsyncIterator~HarnessEvent~
        +is_alive() bool
    }
    HarnessAdapter --> HarnessConfig
    HarnessAdapter --> HarnessResponse
    HarnessAdapter --> HarnessEvent

    class Tool {
        +str name
        +str description
        +Dict input_schema
    }

    class resolve_tool_conflicts {
        <<function>>
        +resolve(env_tools, harness_builtin_tools) List~Tool~
    }
    resolve_tool_conflicts --> Tool

Loading

_{Last reviewed commit: 0c5d074}

[greptile-apps]

_{6 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[arpitg1991]

will this be merged ?