The people building Daintree.

Creator of Daintree. Builds the desktop habitat, the agent integration layer, and the review-first workflow the rest of the app is arranged around. Most weeks that means working across the whole stack rather than inside any one system.

The last 30 days ran to 625 merged PRs, and plugins took the largest share. A project can load its own plugins from .daintree/plugins, each with one named MCP server. Plugins can now read declared databases, ship their own SVG icons, give their views an identity and header actions, list agents across every open project, read a terminal's screen behind a terminal:read permission, and track the actions they have running.

The last week brought dictation reliability fixes and new assistant powers over MCP: deleting worktrees, closing a pane it launched once the handback arrives, and changing Daintree's confirmation settings. Rendered Markdown now draws Mermaid diagrams, and the E2E suite was overhauled. Before that came a performance pass on re-renders and git spawns, and a consistency sweep that put UI controls onto shared primitives.

Works on test reliability and platform QA. Most of the pull requests land in the E2E suite and the Windows-specific paths that only break on CI. Most of the issues come from running Daintree on Windows and Linux and writing down exactly where it breaks: 51 of them so far, more than anyone else, each one fixed by Greg.

The reports read like a platform matrix. On Windows: a PTY with pid 0 starting the process detector and spamming invalid PID warnings, Smart App Control blocking the unsigned installer (and the follow-up request for Authenticode signing), a system requirements check failing on REG_EXPAND_SZ registry values, a better-sqlite3 ABI mismatch. On Linux: an AppImage write EIO on startup, the terminal opening blank from a fit-and-resize race, the CLI unable to open projects from the deb or AppImage. Across both: polling loops retrying a dead git path after a project directory was deleted, a terminal that would not clear its Paused backpressure state, a window restoring collapsed when an external monitor disconnected.

The pull requests are the same territory from the other side: coverage for untested Windows kill and shell paths and the write-error throttle in the pty layer, a fix for duplicate agent spawns when Enter arrives at the trust prompt, Windows CI failures traced to path separators and shell quoting, and a run of E2E repairs covering selector precision and adaptive timeouts.

Works on dev preview and the MCP server, with 12 merged pull requests and 35 resolved issues. Lately that means driving Daintree from the outside: running an orchestrator against the MCP server and reporting each place it couldn't do what an orchestrator needs.

Thirteen issues in September came from that work. There was no way to discover workspace ids, panel ownership didn't survive a client reconnect, a bound session couldn't keep its workspace resident, there was no interrupt short of destroying the panel, a closed terminal read the same as an idle agent, and a client couldn't confirm that a send was delivered. Each became an MCP feature: workspace.list, a session-scoped interrupt, client metadata on terminals, submission confirmation. The latest asks that a RESOURCE_NOT_OWNED refusal tell the agent how the user can hand the terminal over.

The pull requests are mostly test infrastructure around MCP: an end-to-end test covering bound-view eviction protection and a Playwright-free harness for the CDP renderer freeze. Earlier work added the per-worktree dev server port registry, the agent flavor system with CCR discovery, and the per-worktree remote compute lifecycle. A request for GitLab support alongside GitHub also shipped.

Uses Daintree the way a user does and files what turns up: 41 issues, each one fixed by Greg. Most came during the early releases in March and April 2026, and together they are a usability pass over the app, from the notes panel to the project switcher to what happens after a crash.

Two more came in September. The welcome screen didn't scale its layout or type up on large displays, and file paths that an agent's terminal UI hard-wrapped at the margin weren't linked correctly. Both were fixed within a day of being filed.

The earlier specifics: a saved note not editable in the modal view, custom session names not persisting across projects, dark mode toggle and button contrast, cursor states on modal close icons, tooltips missing from the refresh, plus, copy and menu icons, keyboard navigation in the projects list, a way out of a stuck loading screen, Start Fresh after a crash removing every existing project, a close warning, CMD W to close an instance, and Option Enter to insert a newline in the command input. The one pull request restored the fontFamily in the CodeMirror editor theme.

Extended Daintree's built-in agent roster. Added Kiro CLI as a first-class agent, covering setup wizard integration, auth detection, and panel glue. Kiro users don't have to hand-roll a custom terminal recipe anymore.

Followed up with AWS SSO token support in the auth check, so Kiro users authenticating through an AWS identity center get detected correctly instead of bouncing through the setup flow on every launch. Also added Kiro to the README agent list and install section so newcomers evaluating the project can see the full agent lineup from the landing doc. Quiet, infrastructural work that makes agent onboarding invisible for a whole class of users.

Reported a Windows 11 bug where the launch dialog kept saying the Claude Code binary wasn't found, even though the CLI ran fine from a shell and a restart cleared it for a while. The report traced it to two separate causes. On a native install, probeNativePaths checked the path without appending .exe, so claude.exe could never match. And one 10 second timeout covered the checks for all 18 or so agents; when it ran out, every agent was marked missing and the error was discarded. Greg fixed both the next day.

Found a terminal input ordering bug. When a submit timed out, its trailing Enter keypress was interleaved with the body of the next submit, so the second command reached the terminal out of order (issue #11875). Fixed in pull request #11882 on 20 August 2026.