Repository navigation
Research Pilot 0.7.0
Research Pilot 0.7.0 brings work modes, a plan-and-review workflow for research, and rebuilt context management for longer PI Agent sessions. It also updates the product documentation around how to use the application, with diagrams and a separate installation guide.
Work modes across agents
- Choose a workspace's default way of working across Claude Code, Codex and PI Agent. Collaborator is the general mode; experimental choices include Ideator, Builder, Teacher, and Policy & Records.
- Inspect a mode, copy it, or create a custom mode. Saving and applying are separate actions. Switching previews the foreground sessions it will end and starts fresh sessions with the new configuration.
- Mode editing now reports unsupported choices and failed switches accurately, protects shared modes from unsafe deletion, and asks before discarding unsaved edits.
- The phone companion exposes work mode under Settings. PI model and reasoning choices are remembered for the workspace after restarting the app.
Plan and follow research in one place
- Describe a research question on the Research page, generate a readable plan, request changes, and select Approve & start for the version you reviewed. Planning uses its own session and does not occupy ordinary PI chat.
- Draft inputs, revision requests and plan versions are saved. Generating or editing a plan does not start a run. A stopped or outdated planning response cannot replace the current plan.
- New runs can compare independent proposals, critique them, choose a direction, refine it and produce a readable design. They can also refine an existing direction directly. The coordinator retains responsibility for the next action.
- Design outputs and their dependencies are recorded, so replacing an upstream proposal marks affected downstream work for reconsideration. Earlier versions remain inspectable.
- Coordinators can retrieve detailed history and evidence as needed instead of receiving every record on every turn. The Research view presents the current narrative with supporting detail in a side panel.
These changes support recorded, reviewable investigations. They do not establish that a generated design is novel or that the design workflow improves scientific outcomes. The run's model-cost limit starts after launch; planning calls and external compute or service costs are separate.
Longer PI Agent sessions
- Context is checked at the model-request boundary, including tool definitions and the provider's reported usage.
- Large tool results can be stored by the local data service and replaced in working context with excerpts and retrieval references. Their original text remains available for follow-up reads.
- Compatible tool changes preserve compressed working context. Restoring a session and switching providers follow explicit compatibility rules.
- Codex subscription sessions reserve 32k tokens for local budget calculations and compact later. This does not add a response-length cap to the provider request. Anthropic sessions use a 300k working-window ceiling to bound context growth.
- The interface shows compression progress and elapsed time. Approval waits, queued follow-up instructions, file links and conversation history have received related fixes.
- Provider request handling includes fixes for cache-control block limits and diagnostic response-body handling.
Memory grounded in sources and time
- Memory records carry their originating user statement and relationships such as updates, historical facts and equivalence.
- A difference in wording alone no longer becomes a startup conflict warning. Disagreements are considered with their source, time and relevance to the current task.
- The online memory experiment has been removed. The full source-and-time relation behavior is the default when memory is enabled; work-mode settings and “restart without memory” remain available.
Documentation and release fixes
- The README now explains the product, first use, agent and mode choices, papers and memory, research planning, phone access, permissions and upgrades.
- Two diagrams show how project work becomes reusable knowledge and how a research plan reaches a reviewed result. Detailed setup and troubleshooting live in a separate guide.
- Settings loads when opened, reducing the startup JavaScript from about 484 KiB to 410 KiB and keeping it within the existing release budget.
- The Linux search-error test now accepts GNU grep's diagnostic wording and checks that malformed expressions remain visible errors.
Install or upgrade
Finish active work and quit Research Pilot before replacing the app. Research data remains separate under ~/.pipilot by default. Back up that directory and your project folders; do not remove them during an upgrade. Reopen the application so the new desktop and bundled data service are used together.
| Platform | Recommended | Alternative |
|---|---|---|
| macOS · Apple Silicon | DMG | ZIP |
| Ubuntu/Debian · Intel / AMD 64-bit | deb | AppImage |
| Ubuntu/Debian · ARM 64-bit | deb | AppImage |
Windows and Intel Mac installers are not included. Prefer the deb on Ubuntu. AppImages on Ubuntu 24.04 need libfuse2t64; arm64 also needs zlib1g-dev. Older Ubuntu releases use libfuse2. See the guide for the Chromium --no-sandbox fallback when AppImage startup is blocked by sandbox setup.
For Claude or Codex sessions, install the corresponding CLI and sign in. PI Agent's runtime is bundled and needs a configured model and authentication. Research Pilot itself needs no separate Node.js installation. Electron remains at 31.3.1.
Installation, configuration, phone setup and troubleshooting.
Agent permissions
Desktop and phone launch Claude Code with --dangerously-skip-permissions and Codex with --dangerously-bypass-approvals-and-sandbox. Agents can run commands, edit or delete files, and access the network with your OS account's permissions without per-action approval. Codex's own sandbox is disabled. This does not grant administrator privileges or bypass OS access controls. Use trusted workspaces and tasks, and keep backups.
Download a named application package above. GitHub's automatic Source code (zip) and Source code (tar.gz) entries contain public documentation, not the application source or an installer. The named mac-arm64.zip is the packaged Mac application.
Release validation
- Automated checks: 1,323 tests passed locally. GitHub CI passed the Node 18 test job and the Node 20 production release gate. The gate includes Electron typechecking, the production build, size and dependency budgets, and the daemon/MCP demo.
- Settings loading: the production renderer was exercised with isolated API fixtures. Settings was absent from startup requests, loaded on demand, displayed its Memory tab, and closed and reopened successfully.
- macOS packages: the application and DMG passed Developer ID signature, Apple notarization, stapled-ticket and Gatekeeper checks. Applications inside both the DMG and ZIP report 0.7.0 and pass signature and ticket validation. Their application archives match. The DMG application launched its actual bundled interface and opened Settings in an isolated profile.
- Linux packages: both debs installed in fresh Ubuntu 24.04 test containers and set the expected sandbox permissions. Both packaged interfaces displayed the workspace picker under a virtual display and started their bundled 0.7.0 data service.
- Bundled runtime: the Mac application, both installed Linux debs and both extracted AppImage payloads loaded the PI provider modules, spawned a native terminal, started the data service, rejected an unauthenticated API request, and completed a Wiki search through the packaged MCP adapter.
Linux graphical checks used a virtual display and Chromium's --no-sandbox flag inside containers. The x86_64 checks ran under emulation on Apple Silicon, not physical x86_64 hardware. The ARM AppImage extracted through its own launcher; the x86_64 payload was extracted directly because its launcher is not supported in this emulation setup. These packaging checks did not run a new authenticated model conversation or re-evaluate research-answer quality.