nooa-acp 0.0.10__tar.gz

nooa_acp-0.0.10/.gitignore

@@ -0,0 +1,406 @@
+# Rust toolchain (installed for nemo_relay PyO3 build)
+.cargo/
+.rustup/
+
+# Node.js dependencies (never commit)
+node_modules/
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+!packages/nemo-oo-agents-viewer/frontend-react/dist/
+!src/nooa/viewer/frontend-react/dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+# Exception: allow util/*/lib/ directories (our code, not build artifacts)
+!util/*/lib/
+!util/**/lib/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.env.*
+!.env.example
+!.env.template
+.venv
+.venv-macos
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# uv cache directory
+.uv-cache/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# IDEs
+.vscode/*
+.vscode-server/*
+.idea/*
+*.swp
+*.swo
+
+# OS
+.DS_Store
+Thumbs.db
+
+# Temporal
+temporal/
+.temporal/
+
+# Custom
+secrets.env
+local_config.yaml
+
+# Security: JWT tokens and API keys
+**/test-nvauth.py
+**/debug-auth*.py
+# JWT token patterns (eyJ...)
+eyJ*
+
+# Third-party dependencies
+3P/
+3p/
+
+# Trash folder for temporary files
+trash/
+
+# Tools directory (generated traces) - only root level
+/tools/
+
+# Agent trace files and results (root level)
+/trace/
+/traces/
+/results/
+trace*.json
+!trace_viewer_config.json
+/tmp/
+
+# NVIDIA OO Agents generated trace and evaluation files (anywhere)
+*.noo-eval.jsonl
+*.006summary.json
+# Exception: Allow trace fixtures for testing/eval
+!evaluation/fixtures/**/*.noo-eval.jsonl
+
+# Interface output markdown (experiment results)
+docs/interface_output/
+
+# Trace viewer config backups
+**/trace_viewer_config.json.backup
+
+# Local config overrides (user-specific, not tracked)
+**/*.local.json
+
+# Bridge watchpoint file (runtime job monitoring, not tracked)
+.bridge_watch.json
+
+# Slurm job output files (runtime artifacts, not tracked)
+*.err
+*.out
+
+# Cursor IDE rules (keep local but don't track)
+.cursorrules
+
+# PDFs
+*.pdf
+packages/.venv/
+
+# Experimental implementations (on experimental-implementations branch)
+composer-packages/
+gpt-packages/
+trace_tui/
+NEW_AGENT_TUI.md
+RUN_NEW_AGENT_TUI.md
+SETUP_QWEN_TUI.md
+TUI_INTEGRATION_SUMMARY.md
+run-new-tui.sh
+test-tui-debug.py
+test-tui.sh
+scratch/ALIGNMENT_SUMMARY.md
+scratch/DESIGN_ALIGNMENT_REPORT.md
+packages/nemo_oo_agents-core/src/nemo_oo_agents/cli.py
+packages/nemo_oo_agents-tui/src/nemo_oo_agents_tui/cli.py
+
+# Prompt optimization experiment results
+util/prompt-optimization/results/*.json
+util/prompt-optimization/results/*.jsonl
+util/prompt-optimization/results/*/*.json
+util/prompt-optimization/results/*/*.jsonl
+util/prompt-optimization/traces/
+util/prompt-optimization/traces_test/
+
+# Eval pipeline experiments and data
+util/e2e_optimization/experiments/
+util/e2e_optimization/src/e2e_optimization/examples/**/*.jsonl
+util/e2e_optimization/src/e2e_optimization/examples/**/results/
+# Exception: allow data_*.jsonl (test datasets)
+!util/e2e_optimization/src/e2e_optimization/examples/**/data_*.jsonl
+util/eval_pipeline/experiments/
+
+# Experiment test result directories
+experiments/tau_bench_test*/
+experiments/**/results/
+
+traces.db
+traces.db-wal
+traces.db-shm
+# Large result files
+experiments/**/results/
+
+# Test files
+test_arxiv_fetch.py
+trigger_librarian_fetch.py
+experiments/evaluation-ablations/*.db
+experiments/evaluation-ablations/test_case_*/
+experiments/evaluation-ablations/test.csv
+experiments/evaluation-ablations/test.json
+agents/**/traces/
+
+# Claude Code settings (user-specific)
+.claude/settings.json
+
+# Claude Code runtime directories (cache, sessions, IDE state, plugins, skills)
+.claude/backups/
+.claude/history.jsonl
+.claude/cache/
+.claude/file-history/
+.claude/ide/
+.claude/plugins/
+.claude/policy-limits.json
+.claude/projects/
+.claude/sessions/
+.claude/shell-snapshots/
+.claude.json
+
+# Agent runtime directory
+.agents/
+
+# Shell history
+.bash_history
+
+# User-level XDG config/local dirs (injected by sandbox/environment)
+.config/
+.local/
+
+# Rust toolchain (installed by rustup for nemo-relay build)
+.cargo/
+.rustup/
+
+# VS Code Server (remote development runtime)
+.vscode-server/
+
+debug_http_logs/
+
+# SWE-bench evaluation logs (generated during test runs)
+**/logs/run_evaluation/
+
+# Trace analyzer batch results
+**/analysis_batch/
+
+# Claude Code local settings (personal, not shared)
+.claude/settings.local.json
+.claude/agents
+
+# Claude Code sandbox artifacts (created at runtime, not project files)
+.srt-settings.json
+# Debug handler dumps (written to cwd on SIGUSR2)
+debug_dump_*.txt
+# NeMo OO project directory (libraries, traces, config)
+.nemo_oo/
+.nooa/
+.nemo_oo_agents/
+# Shell config stubs (empty read-only files injected by sandbox)
+.zshrc
+.zprofile
+.bash_profile
+.bashrc
+.profile
+.gitconfig
+.ripgreprc
+.idea
+.vscode
+# Git internal stubs (root-level sandbox artifacts only)
+/HEAD
+/config
+/hooks
+/objects
+/refs
+# pytest temp directories
+pytest-of-*/
+experiments/evaluation-ablations/task_temp/
+
+# DABStep experiment outputs (top-level scratch directory)
+/dabstep/
+
+# Harbor generated task directories (regenerated via `harbor adapter run`)
+util/harbor/tasks/
+
+# OpenSkills workspace files
+.openskills/
+
+# Local lockfile
+prpm.lock
+
+# Git worktrees
+.worktrees/
+worktrees/
+
+# Bridge session ID file — written by slack bridge into each agent worktree root
+.claude_session_id
+
+sft_data/**/*.jsonl
+
+# Git credentials (must not be committed)
+.git-credentials
+# Bridge watchpoint file — per-worktree, never committed
+.bridge_watch.json
+
+.claude/backups
+.claude/file-history
+.claude/projects
+.claude/shell-snapshots
+.claude/telemetry
+./claude/sessions
+./claude/cache
+./claude/plugins
+
+.local/
+.config/
+util/harbor/tasks/terminal_bench/
+
+# WTF skills (generated by wtf install-skills)
+.claude/skills/wtf*
+tests/integration/fixtures/*.db
+tests/integration/fixtures/*.lock
+
+# Worktree directories
+.wt-*
+
+# Throwaway benchmark smoke-test configs
+util/harbor/tb1_opus_smoke.yaml
+util/harbor/tb1_ultra_smoke.yaml
+.claude/scheduled_tasks.lock
+.commit_msg
+
+# LoCoMo benchmark data (downloaded on demand, not vendored)
+examples/memory_bench/data/
+
+# Local scratch/working files (never commit)
+.scratch/
+
+# Claude Code local session dir (never commit)
+.claude/

nooa_acp-0.0.10/PKG-INFO

@@ -0,0 +1,201 @@
+Metadata-Version: 2.5
+Name: nooa-acp
+Version: 0.0.10
+Summary: Agent Client Protocol adapter for the NOOA coding agent
+Project-URL: Homepage, https://github.com/NVIDIA-NeMo/labs-OO-Agents
+Project-URL: Repository, https://github.com/NVIDIA-NeMo/labs-OO-Agents
+Project-URL: Issues, https://github.com/NVIDIA-NeMo/labs-OO-Agents/issues
+License: Apache-2.0
+Requires-Python: <3.14,>=3.12
+Requires-Dist: agent-client-protocol<0.12,>=0.11
+Requires-Dist: click>=8.1.0
+Requires-Dist: litellm!=1.95.0
+Requires-Dist: nooa-cli
+Requires-Dist: nooa[mcp]
+Description-Content-Type: text/markdown
+
+# nooa-acp
+
+**Run the NOOA coding agent inside your editor.** `nooa-acp` is an
+[Agent Client Protocol](https://agentclientprotocol.com) server, so any
+ACP-speaking client — Zed today — can drive the same agent the terminal host
+uses: CodeAct, repository tools, a persistent shell, installed skills, workspace
+slash commands and durable sessions, with file edits and terminal commands
+surfaced as structured activity.
+
+It hosts `nooa_cli.coding.CodingAgent` directly. Repository instructions
+(`AGENTS.md`), coding tools, summarization, installed `nooa.skills` entry points
+and semantic file and terminal activity therefore have no separate ACP
+implementations — fix something here and the terminal host gets it too.
+
+This is new and we would like it exercised. If something breaks, please say so.
+
+## Install
+
+```bash
+uv add nooa-acp                 # or: uv add "nooa[acp]"
+```
+
+There is no default model. Set `NOOA_MODEL` or pass `--model`, or the command
+exits with a usage error.
+
+## Quick start: Zed
+
+Zed launches ACP agents as "external agents". Add NOOA to `settings.json`
+(`cmd-,`):
+
+```json
+{
+  "agent_servers": {
+    "NOOA": {
+      "type": "custom",
+      "command": "uvx",
+      "args": ["nooa-acp"],
+      "env": {
+        "NOOA_MODEL": "nvidia_nim/nvidia/nemotron-3-super-120b-a12b",
+        "NVIDIA_API_KEY": "nvapi-..."
+      }
+    }
+  }
+}
+```
+
+Open a repository, then pick **NOOA** from the `+` menu in the agent panel. Zed
+runs the command with your worktree as its working directory, so repository
+instructions, project skills and sessions resolve against the open project.
+
+Credentials go in `env` here rather than in Zed's own settings: the agent is a
+separate process and inherits only what Zed passes it. Use a secret-manager
+wrapper as the `command` if you would rather not put a key in `settings.json`.
+
+From a checkout of this repository, point the client at the workspace package
+instead:
+
+```bash
+uv run --project "$PWD" --package nooa-acp -- nooa-acp
+```
+
+### MCP servers do not carry over from Zed
+
+**Remote MCP servers you authenticated inside Zed are not usable from an ACP
+agent.** Zed holds those OAuth tokens itself and does not pass them down, so a
+server showing a green indicator in Zed's own UI arrives at the agent either
+with no tools at all or with nothing but its `authenticate` /
+`__complete_authentication` stubs. Local stdio MCP servers are unaffected.
+
+This is a known Zed limitation, tracked in
+[zed-industries/zed#54410](https://github.com/zed-industries/zed/issues/54410)
+(open, labelled `area:ai/mcp` + `area:ai/acp`). A maintainer has said the
+plumbing largely exists and the work is queued, but as of this writing it is
+unresolved.
+
+Configure the MCP server directly for NOOA instead — through NOOA's own
+`.mcp.json` — and it works normally, because the agent then owns the
+connection and its credentials rather than borrowing Zed's.
+
+## Launching the server yourself
+
+```bash
+nooa-acp --model nvidia_nim/nvidia/nemotron-3-super-120b-a12b
+```
+
+This is a JSON-RPC server, not an interactive program: it speaks ACP on
+stdin/stdout and exits when its input closes, so running it in a terminal
+without a client does nothing. Launch it this way to wire up an ACP client
+other than Zed, or to watch the diagnostics it writes to stderr while a client
+drives it. `--model` accepts any LiteLLM model name or configured NOOA alias.
+
+## Opening a repository runs code from it
+
+**Creating a session imports Python from the workspace, before you send a
+prompt.** This is deliberate — it is how workspace skills work — but it means
+opening a folder is enough to execute code it contains. Treat opening a
+repository with NOOA as equivalent to running its build.
+
+Three paths load workspace code at `session/new` and `session/load`:
+
+- **Skill roots.** Every `.py` file under `.agents/skills`, `.cursor/skills`,
+  `.claude/skills`, or `.claude/commands` is imported. Module-level code runs
+  during import, before anything checks whether the file defines a skill, so the
+  contents are irrelevant.
+- **Workspace settings.** `<workspace>/.nooa/settings.yaml` and the legacy
+  `.nooa/config.toml` may name *additional* skill roots. Those paths are not
+  confined to the workspace: a relative path escaping it, an absolute path, or a
+  symlink is accepted as written.
+- **Libraries.** `<workspace>/.nooa/libs/<package>/` is imported and its
+  directory is prepended to `sys.path` for the life of the process. One ACP
+  server serves several workspaces, so a package name there can shadow the same
+  import for later sessions on other workspaces.
+
+The agent runs as you, in a process holding your model credentials. There is no
+consent prompt on these paths.
+
+**Open repositories you would run.** For anything else, use an OS-level sandbox,
+or start a separate server per workspace with credentials scoped to that task.
+
+## How it behaves
+
+ACP uses standard input and output for JSON-RPC. Diagnostics are written to
+standard error. The agent can execute generated Python and shell commands, so
+use an OS-level sandbox for untrusted tasks. Generated code shares the agent's
+process environment, including model credentials; launch it with only the
+credentials and network access that the session may use.
+Cancellation stops cooperative local work immediately. An in-flight provider
+request may finish in the background when its client does not support
+transport-level aborts. Slash commands run on the agent's event loop so they
+have the same semantics as the native TUI and can safely start agent jobs. An
+async command is cooperatively cancellable; a synchronous command that blocks
+that loop cannot be preempted by the current in-process adapter. The planned
+one-process-per-agent boundary is the safe kill mechanism for that case.
+
+## Sessions and skills
+
+Each ACP session has an independent live agent and allows one foreground prompt
+at a time. Sessions are stored in `<workspace>/.nooa/sessions`, where the TUI
+and ACP adapter can share list and replay metadata. These files are inside the
+workspace trust boundary: a repository can supply session records that appear
+in `session/list` and are replayed as conversation history by `session/load`.
+Open only repositories whose code and conversation history you trust. The
+adapter also advertises session close; closing a live session preserves its
+durable history.
+
+The current stdio adapter hosts those live agents in its own process. That is
+an adapter-private implementation detail rather than part of the durable
+session API: the live-session registry is isolated inside `nooa-acp` so it can
+later be replaced by handles to an agent daemon without changing stored
+sessions, the shared coding agent, or the ACP protocol surface.
+
+Python skill packages use the interpreter's normal import machinery. Multiple
+sessions may use distinct skill package names, but two workspaces must not load
+different checkouts under the same top-level Python package name in one ACP
+server process. Launch a separate stdio server for those workspaces. A future
+one-process-per-agent daemon will make that isolation an OS process boundary.
+
+Installed `nooa.skills` entry points are loaded into the shared skill registry
+but remain opt-in. The agent can activate a relevant skill with
+`self.skills.activate(["name"])`. Stdio MCP servers supplied by an ACP client
+are registered and activated as `mcp.<name>` skills for that session.
+
+Workspace and user skill roots are shared with the terminal host through
+layered `settings.yaml`. New configuration should use:
+
+```yaml
+coding:
+  additional_skills_dirs:
+    - ../nemo-oo-skills
+```
+
+The existing `tui.additional_skills_dirs` key remains supported during the
+migration, as does the older project-local `.nooa/config.toml` key
+`[tui].libs_dirs`. Packaged libraries declared through `nooa.skills`, `SKILL.md`
+skills, and standalone Python skills are discovered from each configured root.
+Loaded `@slash_command` methods are advertised through ACP and matching
+`/command arguments` prompts are dispatched through the shared typed command
+router. Command discovery is refreshed when loaded skills change.
+
+The current adapter accepts text and resource-link prompts plus stdio, HTTP,
+and SSE MCP servers forwarded by an ACP client. ACP-transport MCP proxies,
+additional workspace directories, images, and embedded resources are not
+advertised yet. An unavailable, duplicate, or unsupported MCP server is skipped
+with a session warning so it cannot prevent a new or restored NOOA session from
+opening.