Changelog

All notable changes to OpenHarness should be recorded in this file.

The format is based on Keep a Changelog, and this project currently tracks changes in a lightweight, repository-oriented way.

[Unreleased]

Added

• Hooks now support a priority field (default 0). Within an event, hooks run highest-priority first, and hooks sharing a priority keep their registration order. This lets users order, for example, a security-check hook ahead of a logging hook regardless of where each is declared in settings or contributed by plugins.
• edit_file and write_file in the React TUI now preview a unified diff before applying file changes, let users approve once or for the rest of the session, and skip the extra prompt automatically in full_auto mode.

Fixed

• Codex subscription requests now pass reasoning effort separately, enabling gpt-5.5 with xhigh effort instead of treating gpt-5.5 xhigh as an unsupported model name.
• Telegram channel now delivers replies again under ohmo init --no-interactive and other configs that do not write a reply_to_message field. TelegramConfig declares reply_to_message: bool = True so the attribute access in TelegramChannel.send no longer raises AttributeError and outbound progress/tool-hint/final messages are sent as expected. See issue #243.

[0.1.9] - 2026-05-07

Added

• Added a bundled skill-creator skill for creating, improving, and verifying OpenHarness/ohmo skills.
• User-invocable skills can now be triggered directly as slash commands, with support for skill-specific arguments and model override metadata.

Fixed

• oh setup can now update the API key for an already-configured API-key provider profile instead of only changing the model.
• oh provider edit <profile> --api-key <key> can now replace a saved profile API key, and oh provider add ... --api-key <key> can store one during profile creation.

[0.1.8] - 2026-05-06

Added

• Built-in nvidia provider profile so oh setup offers NVIDIA NIM as a first-class OpenAI-compatible provider choice, with NVIDIA_API_KEY auth source, openai/gpt-oss-120b as the default model, and the NVIDIA NIM endpoint.
• Built-in qwen provider profile so oh setup offers Qwen (DashScope) as a first-class provider choice, with dashscope_api_key auth source, qwen-plus as the default model, and the DashScope OpenAI-compatible endpoint.
• Plugin tool discovery: plugins can now provide BaseTool subclasses in a <plugin>/tools/ directory and they are auto-discovered, instantiated, and registered in the tool registry at runtime. Add tools_dir to plugin.json (defaults to "tools").
• oh --dry-run safe preview mode for inspecting resolved runtime settings, auth state, prompt assembly, commands, skills, tools, and configured MCP servers without executing the model or tools.
• Built-in minimax provider profile so oh setup offers MiniMax as a first-class provider choice, with MINIMAX_API_KEY auth source, MiniMax-M2.7 as the default model, and MiniMax-M2.7-highspeed in the model picker.
• Docker as an alternative sandbox backend (sandbox.backend = "docker") for stronger execution isolation with configurable resource limits, network isolation, and automatic image management.
• Built-in gemini provider profile so oh setup offers Google Gemini as a first-class provider choice, with gemini_api_key auth source and gemini-2.5-flash as the default model.
• diagnose skill: trace agent run failures and regressions using structured evidence from run artifacts.
• OpenAI-compatible API client (--api-format openai) supporting any provider that implements the OpenAI /v1/chat/completions format, including Alibaba DashScope, DeepSeek, GitHub Models, Groq, Together AI, Ollama, and more.
• OPENHARNESS_API_FORMAT environment variable for selecting the API format.
• OPENAI_API_KEY fallback when using OpenAI-format providers.
• GitHub Actions CI workflow for Python linting, tests, and frontend TypeScript checks.
• CONTRIBUTING.md with local setup, validation commands, and PR expectations.
• docs/SHOWCASE.md with concrete OpenHarness usage patterns and demo commands.
• GitHub issue templates and a pull request template.
• React TUI assistant messages now render structured Markdown blocks, including headings, lists, code fences, blockquotes, links, and tables.
• Built-in codex output style for compact, low-noise transcript rendering in React TUI.

Fixed

• Subprocess teammate spawn (agent tool, task_create) now works on Windows under Git Bash. subprocess_backend.spawn builds a direct-exec argv list and passes it through new argv= and env= kwargs on BackgroundTaskManager.create_agent_task / create_shell_task; _start_process then runs the executable via asyncio.create_subprocess_exec(*argv) with no shell in between. Previously the spawn command was a single string interpreted by bash -lc, which on Windows could not reliably exec a Windows-pathed Python interpreter (e.g. C:\Users\...\python.exe) — Git Bash's escape parser consumed the backslashes from the embedded env-prefix and, even with proper quoting, bash launched via asyncio.create_subprocess_exec returned command not found for Windows-pathed binaries that worked perfectly when invoked interactively. Bypassing the shell sidesteps the entire class of cross-platform quoting and path-translation hazard. The legacy shell-evaluated command= path is preserved for callers (e.g. BashTool) that legitimately want shell semantics. See issue #230.

• Bundled skill loader now uses yaml.safe_load for SKILL.md frontmatter, matching the user-skill loader. The shared parser is extracted to openharness.skills._frontmatter so bundled and user skills handle YAML block scalars (>, |), quoted values, and other standard YAML constructs the same way.

• Compaction now detects llama.cpp/OpenAI-compatible context overflow errors, accounts for image blocks in auto-compact token estimates, and strips image payloads from summarizer-only compaction requests.

• Large tool results are now bounded in conversation history: oversized outputs are saved under tool_artifacts, old MCP results become microcompactable, and context collapse trims stale tool-result payloads.

• ohmo now keeps personal memory isolated from OpenHarness project memory: /memory in ohmo sessions targets the ohmo workspace memory store, and ohmo runtime prompt refreshes no longer inject project memory unless explicitly requested.

• Fixed glob and grep tools hanging indefinitely when the rg subprocess produced enough stderr output to fill the OS pipe buffer. stderr is now redirected to DEVNULL so it is discarded rather than blocking the child process.

• Fixed bash_tool hanging after a timed-out command when the subprocess stdout stream stayed open. _read_remaining_output now applies a 2-second asyncio.wait_for timeout so the tool always returns promptly.

• Fixed session_runner background task deadlock caused by an unread stderr=PIPE stream. The subprocess now uses stderr=STDOUT so all output merges into the single readable stdout pipe.

• React TUI prompt input now treats the raw DEL byte (0x7f) as backward delete while preserving true forward-delete escape sequences, fixing backspace failures seen in some macOS terminal environments.

• todo_write tool now updates an existing unchecked item in-place when checked=True instead of appending a duplicate [x] line.

• Built-in Explore and claude-code-guide agents no longer hard-code model="haiku", which caused them to fail for users on non-Anthropic providers (OpenAI, Bedrock, custom base URLs, etc.). Both agents now use model="inherit" so they run with whatever model the parent session is using. build_inherited_cli_flags is also fixed to skip the --model flag entirely when the value is "inherit", letting the subprocess correctly inherit the parent model via the OPENHARNESS_MODEL environment variable instead of receiving the literal string "inherit" as a model name.

• React TUI spinner now stays visible throughout the entire agent turn: assistant_complete no longer resets busy state prematurely, and tool_started explicitly sets busy=true so the status bar remains active even when tool calls follow an assistant message. line_complete is the sole signal that ends the turn and clears the spinner.

• Skill loader now uses yaml.safe_load to parse SKILL.md frontmatter, correctly handling YAML block scalars (>, |), quoted values, and other standard YAML constructs instead of naive line-by-line splitting.

• BackendHostConfig was missing the cwd field, causing AttributeError: 'BackendHostConfig' object has no attribute 'cwd' on startup when oh was run after the runtime refactor that added cwd support to build_runtime.

• Shell-escape $ARGUMENTS substitution in command hooks to prevent shell injection from payload values containing metacharacters like $(...) or backticks.

• Swarm _READ_ONLY_TOOLS now uses actual registered tool names (snake_case) instead of PascalCase, fixing read-only auto-approval in handle_permission_request.

• Memory scanner now parses YAML frontmatter (name, description, type) instead of returning raw --- as description.

• Memory search matches against body content in addition to metadata, with metadata weighted higher for relevance.

• Memory search tokenizer handles Han characters for multilingual queries.

• Fixed duplicate response in React TUI caused by double Enter key submission in the input handler.

• Fixed concurrent permission modals overwriting each other in TUI default mode when the LLM returns multiple tool calls in one response; _ask_permission now serialises callers via an asyncio.Lock so each modal is shown and resolved before the next one is emitted.

• Fixed React TUI Markdown tables to size columns from rendered cell text so inline formatting like code spans and bold text no longer breaks alignment.

• Fixed grep tool crashing with ValueError / LimitOverrunError when ripgrep outputs a line longer than 64 KB (e.g. minified assets or lock files). The asyncio subprocess stream limit is now 8 MB and oversized lines are skipped rather than terminating the session.

• Fixed React TUI exit leaving the shell prompt concatenated with the last TUI line. The terminal cleanup handler now writes a trailing newline (\n) alongside the cursor-show escape sequence so the shell prompt always starts on a fresh line.

• Reduced React TUI redraw pressure when output_style=codex by avoiding token-level assistant buffer flushes during streaming.

Changed

• ohmo Feishu group routing now supports managed group creation, gateway-scoped provider/model commands, and stricter group mention handling so group conversations only wake ohmo when explicitly addressed.
• Dry-run output now reports a ready / warning / blocked readiness verdict, concrete next_actions, likely matching skills/tools for normal prompts, and richer slash-command previews for read-only vs stateful command paths.
• React TUI now groups consecutive tool + tool_result transcript rows into a single compound row: success shows the result line count inline (e.g. → 24L), errors show a red icon and up to 5 lines of error detail beneath the tool row. Standalone successful tool results are suppressed to reduce transcript noise; standalone errors are still surfaced.
• README now links to contribution docs, changelog, showcase material, and provider compatibility guidance.
• README quick start now includes a one-command demo and clearer provider compatibility notes.
• README provider compatibility section updated to include OpenAI-format providers.

[0.1.7] - 2026-04-18

Fixed

• Install script now links oh, ohmo, and openharness into ~/.local/bin instead of prepending the virtualenv bin directory to PATH, which avoids overriding Conda-managed shells while preserving global command discovery.
• React TUI prompt now supports Shift+Enter for inserting a newline without submitting the current prompt.
• React TUI busy-state animation is less error-prone on Windows terminals: the extra pseudo-animation line was removed, Windows now uses conservative ASCII spinner frames, and the spinner interval was slightly slowed to reduce flashing.

[0.1.0] - 2026-04-01

Added

• Initial public release of OpenHarness.
• Core agent loop, tool registry, permission system, hooks, skills, plugins, MCP support, and terminal UI.