pi-experiment-ops 0.6.2

Adds pi-condense 2.11.2 with a configurable four-image request budget for attachments and tool results. Original session images remain available; optional LLM summarization is disabled by default.

Includes the expanded Python analysis and visualization dependencies from 0.6.1, aligns installer defaults and examples, and documents the release checklist in AGENTS.md.

The release workflow builds and attaches the versioned npm tarball and SHA-256 checksum after integration and clean-install validation.

v0.6.0

Set up a workspace with the new pi-experiment-ops config wizard. It guides provider, model, and default selection, with optional MCP and skill setup.

• Add MCP connections with pi-experiment-ops config add-mcp, including HTTP, SSE, and local stdio servers.
• Copy a local skill folder with pi-experiment-ops config add-skill /path/to/skill.
• Ask the agent to configure MCP servers or install skills using the new workspace setup skill, installed during initialization. It includes direct file instructions when the CLI is unavailable.
• New workspaces allow skill discovery by default. Existing permission settings and edited skills are preserved; tool permissions remain separate.
• Applications can reuse the configuration commands through the pi-experiment-ops/config export.

The installer defaults and installation examples now target 0.6.0.

v0.5.1

Workspace skills now load from .pi/skills alongside bundled skills. Initialization creates the skill directory and valid empty models.json and auth.json files while preserving existing workspace configuration.

The README documents skill installation and permissions, the installer's default workspace and preset storage, and selecting a provider/model for the Pi TUI and EAA WebUI.

Install with:

curl -fsSL https://raw.githubusercontent.com/AdvancedPhotonSource/pi-experiment-ops/v0.5.1/install.sh | \
  sh -s -- --repo AdvancedPhotonSource/pi-experiment-ops --version 0.5.1

The release workflow validates the package and attaches the installation tarball and SHA-256 checksum.

v0.5.0

• Replace pi-graph with pi-subagents workflow resources and a JavaScript toy workflow.
• Remove piw CLI/export and graph environment variables; use PI_OPS_PYTHON for the provisioned interpreter.
• Workflow retries execute fresh runs. Existing workspace definitions are preserved for manual migration.

v0.4.0

Highlights

pi-experiment-ops 0.4.0 adds interactive approval controls for Pi tools, shell commands, and direct MCP tools using pi-permission-system 0.8.0.

Unmatched tool calls now ask for approval by default:

• Allow once approves the current call.
• Allow for session remembers approval for that tool during the current session.
• Deny rejects the call.

Use /permissions auto, /permissions ask, or /permissions status to control approval behavior in the TUI. Pattern-based policies can be configured in .pi-experiment-ops/agent/pi-permissions.jsonc.

MCP and CodeMode

• Direct MCP tools pass through the shared permission gate.
• Existing approveTools settings in .pi/mcp.json remain an optional, independent MCP-adapter gate.
• MCP adapter script mode is disabled by default, making CodeMode the bundled scripted tool-call interface.
• Permission policy gates the codemode_execute launch. Calls made inside an accepted CodeMode cell are not prompted separately.

Workspace configuration

New workspaces receive:

• .pi-experiment-ops/agent/permission-system.json
• .pi-experiment-ops/agent/pi-permissions.jsonc

Initialization preserves existing files and configuration.

Validation

The release was validated with:

• Real Pi 0.85.1 sessions
• Direct MCP tools
• Allow-once and session approval flows
• Session-only auto-allow behavior
• CodeMode background execution
• Standalone tarball installation and package acceptance

v0.3.0

Added

• Bundle @ian-pascoe/pi-codemode 0.7.4 and its skills with the standalone distribution.
• Expose CodeMode through the codemode resource replacement key for downstream applications.
• Support non-blocking execution of registered Pi and MCP tools through codemode_execute, with later result retrieval through codemode_result.
• Add an end-to-end acceptance test covering a 60-second MCP operation running in the background while the Pi agent performs independent work.
• Add a README overview of bundled extensions, downstream APIs, helper commands, installation paths, and other package components.

Changed

• Advance the package version from 0.2.0 to 0.3.0.
• Extend bundle and standalone-package tests to verify the CodeMode extension, skills, and tools in both source and packed installations.

CI

• Upload release archives and checksums separately, with up to five attempts and exponential backoff for transient GitHub release failures.

v0.2.0

Experiment-ops now exposes Pi's native SDK through pi-experiment-ops/sdk, with unchanged exports and types. Frontends can depend on experiment-ops for both the SDK and CLI instead of maintaining a separate Pi dependency.

The new workspacePaths(workspace) helper resolves the existing native workspace directories without creating files or changing environment variables. Existing CLI commands, extension replacement keys, configuration formats, and session formats are preserved.

Validation: standalone SDK identity and workspace tests, native TUI/provider integration, real graph execution, and installation from the release tarball.

v0.1.1

v0.1.1

A standalone Pi package for experiment operations, combining MCP connectivity, multiagent workflows, background processes, persistent terminals, and searchable conversation history.

This release includes:

• User-local installation: A single installer provisions Node.js, uv, Python 3.12, and package dependencies as needed, without administrator privileges.
• Argo configuration: An installation preset configures Argo using your Argonne username and selected model. Custom OpenAI-compatible endpoints are also supported.
• Workspace-specific configuration: Keep provider settings, credentials, sessions, and workflow outputs in an explicit workspace.
• Agent and execution tools: Bundled extensions provide subagents, background jobs, persistent interactive shells, and plan/build modes.
• Multiagent workflows: Pi-graph integration includes a small generator–reviewer–renderer example demonstrating review gates, failure recovery, and resumable execution.
• Searchable history: SQLite-backed transcript indexing supports searching previous conversations.
• SDK integration: Public helpers allow other applications to use the bundle’s resources and workspace configuration.

Installation supports Linux x86_64 and aarch64. See the installation instructions to get started.

Native execution uses your account’s permissions; plan mode is not an operating-system sandbox. Automated validation uses local model fixtures and simulated tasks. Live Argo inference and physical instrument operation require validation in your deployment.