feat(tui): replace Textual with a Go/Bubble Tea interface

[kusonooyasumi]

Summary

Replaces the Textual interactive interface with a Go/Bubble Tea TUI that runs as a sidecar process, and makes it the only interactive interface — no fallback, no opt-out env var.

The split is: Python owns the scan (orchestration, controller state, protocol projection), Go owns the screen (rendering, input, selection, clipboard, images). They talk over a socketpair with a versioned JSON protocol:

python backend ──socketpair(STRIX_TUI_FD)── strix-tui (Go)
   hello/ready handshake (protocol v3, capability set must match)
   backend → sidecar: snapshot + collection bootstrap/delta frames
   sidecar → backend: commands (start, steer, quit, mount answer, ...)

All TUI code lives under strix/interface/tui/: backend/ (controller, projection, server, live view), sidecar.py + runtime.py (process lifecycle), and the Go module (cmd/strix-tui, internal/app, internal/protocol, internal/render).

Notable behavior beyond the port:

• Start screen is a prompt, not a config form. A bare prompt launches a scan; with no target it offers to mount the working directory (confirmed by the user, persisted so resume restores it, kept out of targets_info). Modeled on opencode's home screen.
• Two-tier tool output. Only output-heavy tools collapse (exec_command, write_stdin, view_request, repeat_request, view_sitemap_entry, apply_patch) to a 10-line preview with click-to-expand; everything else always renders in full.
• Real images, or none. view_image renders actual pixels through the kitty graphics protocol using Unicode placeholders so it survives Bubble Tea's scrolling viewport. Support is decided by a runtime handshake (1×1 a=q query bounded by DA1), the same one viuer and yazi use — no env allowlist, and no low-fidelity block-art fallback in terminals that can't do it. The local viewer renders the same payload as an <img>.
• Mouse selection and copy inside the TUI (drag to select, copy on release, selection clears with a toast), with image placeholder cells excluded from the copied text.
• Rendering perf with many images: chat blocks, image placements and the bordered trace pane are memoized per event/width. Post-update frame on a 20-image trace went 33.3ms → 7.7ms (~30 → ~130 FPS); benchmarks are committed.
• Markdown emphasis follows CommonMark flanking rules, so ls *.py *.go, a * b, and snake_case no longer render as italics.

Packaging: every wheel and the frozen build bundle a platform strix-tui; a source checkout runs the sidecar from source with go run.

Two fixes worth calling out because they are invisible in the diff's shape:

• Graphics detection used os.Stdin.SetReadDeadline, which returns file type does not support deadline on a tty — a terminal that never answered the query hung the sidecar forever and the backend failed the handshake, so the packaged binary showed only a generic "setup unavailable" panel. The read now goes through a non-blocking fd bounded by the deadline (POSIX) / is only attempted when deadlines work (Windows), and the CLI surfaces the real launch error instead of a generic one.
• docker-py's exec_start never closes the urllib3 responses behind sandbox terminal streams; their finalizers hit the torn-down socket at interpreter exit and printed ValueError: I/O operation on closed file spam. A narrowly-scoped sys.unraisablehook drops exactly that case and delegates everything else.

Verification

• uv run pytest (764 passed), ruff check, mypy strix, pre-commit across all files
• gofmt, go vet ./..., go test ./...; cross-compiled for windows and darwin
• Built the release binary via scripts/build.sh and confirmed the packaged strix-1.4.1-linux-x86_64 boots into the start screen (this is what caught the detection hang above)

Link to Devin session: https://app.devin.ai/sessions/477b26e9b6e24ca48f260b7a71ed87f1
Requested by: @0xallam

[greptile-apps]

Greptile Summary

Replaces the Textual interface with a Go/Bubble Tea sidecar architecture.

• Adds a versioned socket protocol and Python backend for scan orchestration, projection, and command handling.
• Adds Go-based rendering, setup, input, selection, clipboard, image, and vulnerability-report views.
• Updates CLI setup, resume behavior, packaging, release builds, and tests to support the bundled sidecar.
• Removes the legacy Textual implementation and dependency.

Confidence Score: 5/5

The PR appears safe to merge because no eligible unresolved failure remains in this follow-up review.

No blocking failure remains.

Important Files Changed

Filename	Overview
strix/interface/tui/backend/server.py	Implements the Python side of the sidecar protocol, including handshake, snapshots, collection updates, and command dispatch.
strix/interface/tui/backend/controller.py	Coordinates scan setup, lifecycle, resume behavior, and sidecar commands.
strix/interface/tui/runtime.py	Manages sidecar discovery, launch, socket inheritance, handshake, shutdown, and launch-error reporting.
strix/interface/tui/internal/app/update.go	Implements Bubble Tea event handling and interactive state transitions.
strix/interface/tui/internal/app/view.go	Renders the primary Go TUI screens, panes, prompts, status information, and cached trace content.
strix/interface/tui/internal/render/image_detect.go	Adds bounded runtime detection of terminal graphics support.
strix/interface/main.py	Refactors the CLI entrypoint around extracted argument, environment, scan-setup, and interactive-launch modules.
scripts/tui_sidecar_hook.py	Adds a strict Hatchling build hook that compiles and embeds a platform-specific Go sidecar in wheels.
.github/workflows/build-release.yml	Builds platform wheels and frozen releases with the Go sidecar and verifies its inclusion.
strix/interface/viewer/frontend/src/components/live/tool-renderers/ViewImageRenderer.tsx	Updates the local viewer to render image tool payloads as actual images.

_{Reviews (3): Last reviewed commit: "build: drop the ruff exemption for a mod..." | Re-trigger Greptile}

[devin-ai-integration]

Review of the 21 new commits (49a048f..18b134c)

Ran locally on the pulled branch: gofmt clean, go vet clean, go test ./... green, uv run pytest 764 passed, ruff clean, mypy clean (91 files). No dangling references left from the removed provider layer (provider_for_model, providers.list, pickerNone, paged-models/setup-run-controls capabilities are all gone from both sides), and strix/config/loader.py + models.py are now byte-identical to main.

1. Blocking: the working-directory mount bypasses check_mountable_dir

TuiController._start sets pending_workspace_mount = str(Path.cwd()) and, once confirmed, attach_workspace_mount() appends it to local_sources, which build_bind_mounts turns into a writable bind mount (read_only: False, only .git/.agents/.codex remounted read-only).

Every other local-directory path in the codebase goes through check_mountable_dir() first — the guard that refuses /, $HOME, /etc, and credential dirs (.ssh, .aws, .gnupg, .config, …). The new path never calls it, so cd ~ && strix → confirm → the agent gets the user's whole home directory mounted writable inside the sandbox. tests/test_local_sources.py::test_workspace_mount_is_mounted_without_becoming_a_target currently encodes that behavior with $HOME as the mount.

Suggested fix: run check_mountable_dir(Path.cwd()) in _start before setting pending_workspace_mount, and surface the refusal as a setup-log error ("run Strix from your project directory, or give it a target"). Keeping the resume path exempt is fine — that directory was already confirmed.

2. ctrl+c does nothing while a modal is open

updateModal handles esc/arrows/enter only, so ctrl+c is swallowed by the quit, stop and the new mount-confirm prompts. Given the recent "quit must be immediate" changes, the mount prompt in particular should treat ctrl+c as decline-and-quit rather than ignoring it.

3. syncMountPrompt can steal an open modal

case m.snapshot.PendingMount != "" && m.modal != modalConfirmMount: m.openModal(modalConfirmMount) replaces whatever modal is up (help/vulnerability detail). In practice the prompt only appears at launch, so this is a nit — but re-raising it on every snapshot while another modal is open would also reset modalChoice.

4. composerHeight measures through a by-value copy of the textarea

probe := input; probe.SetValue(line) copies textarea.Model, which holds a *memoization.MemoCache pointer — the probe's SetValue/LineInfo writes into the real composer's cache. Correctness looks fine (the value slice is reallocated), but it pollutes the memo cache for the actual composer on every keystroke on a wrapping line. Measuring with ansi.StringWidth(line)/width arithmetic, or a package-level scratch textarea, avoids that.

Non-blocking observations

• The PR description is now badly stale: it still advertises the Textual fallback, STRIX_TEXTUAL_TUI, provider/model setup, /budget, /turns, /scope, /mount, /target-list, /prompt-file, setup-run-controls, and "889 passed" — all removed. Worth rewriting before merge.
• _is_internal_agent_turn matches exact prefixes taken from strix.core.*; those strings now have a second consumer with no test-level coupling to their sources. A shared constant (or a test that asserts the producer strings still start with these prefixes) would keep them from silently drifting.
• setupLogAppend de-duplicates by moving an identical line to the end, which reorders the log rather than showing a repeat count; intentional, just noting the trade-off.

[devin-ai-integration]

@greptile

[0xallam]

@greptile