Skip to content

Research Pilot 0.8.4

Choose a tag to compare

@daidong daidong released this 27 Oct 14:49

Research Pilot 0.8.4

A bug-fix release focused on long PI sessions, saved images, subscription model identity, and reliable launch instructions.

What's changed

  • An unusable compaction result no longer ends an otherwise usable PI turn. Empty, truncated, or non-shrinking summaries leave the previous working context intact. PI continues if that context still fits the input budget; if it does not fit, PI reports the capacity problem without replacing the original history. Authentication, transport, cancellation, and persistence errors are still errors, not silently ignored failures.
  • PI session images are stored separately from checkpoint JSON. Conversation history, working context, and pending inputs share image files and restore their original encoding when loaded. Missing or corrupt image data is detected rather than sent to the model as an unresolved reference. This avoids repeatedly embedding the same image data in large checkpoints.
  • Subscription routes retain model-specific behavior. Codex context windows are selected per model rather than uniformly using the channel fallback. Anthropic subscription authentication is separated from the SDK model identity, so restored signed thinking blocks keep the provider identity expected by the SDK.
  • PiPilot identity and rules travel with the session launch. Claude and Codex no longer depend on PiPilot-written project guidance for those instructions. Workspace sync retires older PiPilot-managed blocks in AGENTS.md and CLAUDE.md, preserving user content. A separately launched CLI does not receive PiPilot's launch identity or rules; shared skills and delegate definitions remain visible by design.
  • Builder instructions now ask for evidence before completion claims. Agents are asked to check claims against results from the current session after the relevant edits and state what remains unverified. This is model guidance, not an enforced guarantee of correctness.
  • UI verification scripts were brought back into line with the current interface, with documented checks for each feature area.

Downloads and installation

Download a named package below, not GitHub's automatic “Source code” archives. This public repository distributes application packages and documentation, not the application source.

Platform Recommended Alternative
macOS Apple Silicon Research-Pilot-0.8.4-mac-arm64.dmg Research-Pilot-0.8.4-mac-arm64.zip
Linux x86_64 Research-Pilot-0.8.4-linux-amd64.deb Research-Pilot-0.8.4-linux-x86_64.AppImage
Linux arm64 Research-Pilot-0.8.4-linux-arm64.deb Research-Pilot-0.8.4-linux-arm64.AppImage
  • macOS: open the DMG and drag Research Pilot into Applications. The application and DMG are signed, Apple-notarized, and stapled.
  • Ubuntu/Debian: install the deb with sudo apt install ./Research-Pilot-0.8.4-linux-amd64.deb, substituting arm64 on ARM. The deb is recommended.
  • AppImage: make the file executable. Ubuntu 24.04 needs libfuse2t64; older Ubuntu versions use libfuse2. The arm64 AppImage also needs zlib1g-dev. If Chromium's sandbox prevents startup, prefer the deb; --no-sandbox is a fallback that disables that sandbox for the launch.
  • Agent setup: install and sign in to Claude Code and/or Codex CLI for those tabs. The built-in PI Agent does not need a separate CLI; configure its model and authentication in Settings. Packaged use does not need Node.js or Docker.

See the installation guide.

Upgrade and permissions

Finish active work, back up the entire ~/.pipilot directory and your project folders, quit the old app, replace it, and reopen. Start a new agent session to receive the new launch instructions. Refresh the phone companion after upgrading. Do not delete research data to upgrade. Keep PI checkpoint files and their sibling image directories together; older releases cannot read the new image-reference checkpoint format.

Desktop and phone sessions launch Claude Code with --dangerously-skip-permissions and Codex with --dangerously-bypass-approvals-and-sandbox. Agents can edit or delete files, run commands, 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. This is fixed launch behavior, not a selectable mode. Use trusted tasks and keep backups.

Verification

  • Local release gate, Node.js 20.20.2: the final source passed npm run verify:release twice. Each run reported 2,456 tests passed, 8 skipped, and 0 failed or cancelled. Electron typechecking, production builds, release size/dependency checks, and the end-to-end MCP/daemon demo passed. Initial runs exposed two test timing issues; the fixtures now wait for a stopped Compute worker to exit and keep periodic heartbeats from racing the lease-expiry test's simulated clock. These changes do not alter production behavior.
  • macOS Apple Silicon: code signatures, Apple notarization, stapled tickets, and Gatekeeper assessments passed. The app launched from the mounted DMG with an isolated profile. Both DMG and ZIP payloads passed PI runtime imports, native SQLite and PTY checks, daemon health/protocol checks, authenticated Teams API access, rejection of unauthenticated access, and packaged MCP initialization, Wiki search, and knowledge search.
  • Ubuntu 24.04 arm64 and x86_64: installed deb packages and extracted AppImage payloads passed start-screen, PI imports, SQLite, native PTY, authenticated daemon/Teams, and packaged MCP search checks. The installed deb sandbox helper has root ownership and mode 4755.
  • Package consistency: all three targets report version 0.8.4. Their daemon, MCP, Builder instructions, and phone JavaScript bytes match. Builder revision 8 is present in each package. The ASAR integrity check also passed, including detection of a modified archive. All six uploaded assets match the tested local files by size and SHA-256.

Limits: GitHub-hosted CI did not start because of the account billing/spending limit; it is not a passing CI run. Linux checks used containers and a virtual display with --no-sandbox; x86_64 ran under emulation on Apple Silicon. The x86_64 AppImage payload was extracted with unsquashfs, not its bootstrap. These checks do not establish native x86_64 launcher, desktop sandbox, or FUSE-mount behavior. No live model calls or physical-phone tests were made for this release. The installed macOS application was not replaced, and scripted process cleanup is not a graceful-quit test.

SHA-256 checksums

6379bdf6b881fb24fb3786917217757400a6b3f83101bc1bd43ff9263f245574  Research-Pilot-0.8.4-linux-amd64.deb
d15fa45c829eaf2ed856e667d152b044abdebcd354be43a097748860c1972798  Research-Pilot-0.8.4-linux-arm64.AppImage
cc0fca8463e5b611e2aee9c6a2d83f84e0b3bb984f83a7c04b85085fd517c843  Research-Pilot-0.8.4-linux-arm64.deb
4d9e01fc169d3897a37bbeaf6b4d24edcd9999955f38fdefb08b912f2164516a  Research-Pilot-0.8.4-linux-x86_64.AppImage
66bd99520c0544a87056fcc1317d9f7fff29e44a0c2aeb44f549ecf82ecab2a4  Research-Pilot-0.8.4-mac-arm64.dmg
48655161d96a770fb07ee55c1093775f772b6e958ea4d172aeb028ce8a8be30a  Research-Pilot-0.8.4-mac-arm64.zip