fix(gaia): keep a loaded skill alive across turns; stop a control resize from wrecking the screen #192

Workflow file for this run

.github/workflows/build_tui.yml at 2983806

	# Copyright(C) 2025-2026 Advanced Micro Devices, Inc. All rights reserved.
	# SPDX-License-Identifier: MIT

	# This workflow builds and tests the Go TUI binary in tui/
	# Produces cross-compiled binaries for 6 platform targets:
	# linux/amd64, linux/arm64, darwin/amd64, darwin/arm64, windows/amd64, windows/arm64

	name: TUI Build & Test

	# NOTE: GitHub Actions does not support YAML anchors, so the two `paths` lists
	# below are duplicated on purpose. Keep them in sync.
	#
	# `src/gaia/daemon/**` is the Python daemon the TUI drives over loopback
	# HTTP/SSE — the actual contract boundary, and narrow on purpose (triggering on
	# all of src/gaia/** would be noise). Be aware of what this does and does not
	# buy: it makes the suite RUN on a daemon-only change, but the TUI's tests fake
	# the daemon with an httptest server (test/fakedaemon_test.go), so wire drift
	# still passes. Detecting drift needs a contract test that exercises the real
	# daemon; until that exists this trigger is early warning, not detection.
	on:
	push:
	branches: [ main ]
	paths:
	- 'tui/**'
	- 'src/gaia/daemon/**'
	# The Python side of the email-scope drift guard (#2730 D2a) — a
	# one-sided edit to the source-of-truth scope constants (or the shared
	# fixture both sides diff against) must still run the Go test that
	# checks connectScopes against them, or drift ships silently.
	- 'hub/agents/email/python/gaia_agent_email/scopes.py'
	- 'hub/agents/email/python/gaia_agent_email/outlook_scopes.py'
	- 'tests/fixtures/connectors/email_scopes.json'
	- '.github/workflows/build_tui.yml'
	pull_request:
	# NO `branches:` filter on purpose. It used to be `branches: [main]`, which
	# silences every stacked PR — one opened against another feature branch
	# rather than main. That is not a corner case here: #2494 never ran this
	# workflow once, and #2496/#2492 last ran several commits behind their heads,
	# because all three are stacked on sibling claudia/* branches. The newest and
	# most safety-critical TUI work had the least CI as a direct result.
	#
	# Dropping the filter runs the suite on every PR regardless of base; the
	# `paths` filter below is what keeps it scoped, and it still applies.
	types: [opened, synchronize, reopened, ready_for_review]
	paths:
	- 'tui/**'
	- 'src/gaia/daemon/**'
	# The Python side of the email-scope drift guard (#2730 D2a) — a
	# one-sided edit to the source-of-truth scope constants (or the shared
	# fixture both sides diff against) must still run the Go test that
	# checks connectScopes against them, or drift ships silently.
	- 'hub/agents/email/python/gaia_agent_email/scopes.py'
	- 'hub/agents/email/python/gaia_agent_email/outlook_scopes.py'
	- 'tests/fixtures/connectors/email_scopes.json'
	- '.github/workflows/build_tui.yml'
	merge_group:
	workflow_dispatch:

	concurrency:
	group: ${{ github.workflow }}-${{ github.head_ref \|\| github.ref }}
	cancel-in-progress: true

	permissions:
	contents: read

	jobs:
	test:
	name: Test (${{ matrix.os }})
	runs-on: ${{ matrix.os }}
	strategy:
	# Report all three platforms on every run. A first-time Windows failure
	# must not hide a real Linux regression behind a cancelled job.
	fail-fast: false
	matrix:
	os: [ubuntu-latest, macos-latest]
	include:
	# Advisory for now. These tests have NEVER run on Windows, and what is
	# most likely to break there is environmental, not logical: 15 test
	# files use t.TempDir() (cleanup fails on Windows if a handle is still
	# open) and 4 compile and spawn real child processes. -race also needs
	# a C toolchain on windows/amd64, which is unverified here.
	#
	# This is additive: ubuntu and macos block as before, nothing is
	# weakened. Flip `advisory` off once a green run proves it.
	- os: windows-latest
	advisory: true
	continue-on-error: ${{ matrix.advisory == true }}
	defaults:
	run:
	# bash on all three so one command works everywhere (Windows defaults to pwsh).
	shell: bash
	working-directory: tui
	steps:
	- uses: actions/checkout@v7

	- uses: actions/setup-go@v7
	with:
	go-version-file: tui/go.mod
	cache-dependency-path: tui/go.sum

	# -race is the point of this job. The TUI is concurrent throughout — SSE
	# streaming, the loopback control API, daemon polling — so this is where
	# its real bugs are. Verified green on all 12 packages locally.
	- name: Run tests (race)
	# Windows runs the 13 unit packages but not ./test. That suite compiles
	# and spawns real child processes and binds ports; two of its cases
	# (TestHubCatalogLoadSurvivesTheChatView, TestASubprocessAgentIsNotGated)
	# do not reach their view there. Naming the gap rather than leaving the
	# job red forever — the unit packages are the coverage worth having on a
	# third OS, and porting the integration harness is its own change.
	run: \|
	if [ "${{ runner.os }}" = "Windows" ]; then
	go test $(go list ./... \| grep -v '/tui/test$') -v -count=1 -race -timeout=15m
	else
	go test ./... -v -count=1 -race -timeout=15m
	fi

	- name: Vet
	run: go vet ./...

	lint:
	name: Lint
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@v7

	- uses: actions/setup-go@v7
	with:
	go-version-file: tui/go.mod
	cache-dependency-path: tui/go.sum

	# Config and enabled set live in tui/.golangci.yml. The set was chosen by
	# measuring: every linter in it reports zero findings on the tree today, so
	# this lands green and only fires on new code.
	- uses: golangci/golangci-lint-action@v9
	with:
	version: v2.12.2
	working-directory: tui

	# Both linters only analyze the host GOOS, so internal/daemon/lock_windows.go
	# and pid_windows.go are invisible to every BLOCKING job — the only leg that
	# compiles them is the advisory Windows one, which cannot fail the run. Vet
	# them here so Windows-only code has enforcement that bites. darwin needs no
	# equivalent: the macos test leg blocks and vets natively.
	- name: Vet Windows-only sources
	working-directory: tui
	env:
	GOOS: windows
	run: go vet ./...

	coverage:
	name: Coverage
	runs-on: ubuntu-latest
	# Genuinely non-blocking, not just labelled that way. `make cover` runs the
	# whole suite to produce a profile, so without this a test failure would
	# paint this job red too — a second, duplicate X for something the `test`
	# job already reports. Nothing `needs:` this job.
	continue-on-error: true
	steps:
	- uses: actions/checkout@v7

	- uses: actions/setup-go@v7
	with:
	go-version-file: tui/go.mod
	cache-dependency-path: tui/go.sum

	# Informational, NOT a gate. A threshold that blocks PRs needs buy-in that
	# a printed number does not, and the naive per-package figure is actively
	# misleading here — internal/ui/hub and internal/ui/root read 0% because
	# their tests live in the test/ package, so a naive gate would fail on two
	# of the better-covered packages. `make cover` uses -coverpkg=./... to
	# attribute coverage back to the package that owns the code.
	#
	# pipefail is set explicitly: a bare `make cover \| tee` reports TEE's exit
	# status, so a failing test run would read as a pass. Piping (rather than
	# redirecting) keeps the full run visible in the log for debugging.
	- name: Run coverage
	working-directory: tui
	run: \|
	set -o pipefail
	make cover 2>&1 \| tee /tmp/cover.txt

	# Runs even when the step above failed, so a partial profile still reports.
	# Every extraction is `\|\| true`: GitHub runs `run:` under `bash -e`, and a
	# grep that matches nothing exits 1 — which would turn this purely
	# informational step into a second, misleading red X on the PR.
	- name: Publish to job summary
	if: always()
	run: \|
	TOTAL=$(grep -E '^total:' /tmp/cover.txt 2>/dev/null \| grep -oE '[0-9]+\.[0-9]+%' \| tail -1 \|\| true)
	ROWS=$(grep -E '^[0-9]+\.[0-9]+%' /tmp/cover.txt 2>/dev/null \
	\| sed 's\|github.com/amd/gaia/tui/\|\|' \
	\| sort -k2 \
	\| sed 's/^/\| /; s/\t/ \| /; s/$/ \|/' \|\| true)
	{
	echo "## TUI coverage"
	echo
	echo "Total: ${TOTAL:-unavailable} (statements, \`-coverpkg=./...\`)"
	echo
	if [ -n "$ROWS" ]; then
	echo "\| Coverage \| Package \|"
	echo "\| --- \| --- \|"
	echo "$ROWS"
	else
	echo "_No per-package data — the coverage run did not complete._"
	fi
	} >> "$GITHUB_STEP_SUMMARY"

	build:
	name: Build (${{ matrix.goos }}/${{ matrix.goarch }})
	needs: test
	runs-on: ubuntu-latest
	strategy:
	matrix:
	include:
	- goos: linux
	goarch: amd64
	- goos: linux
	goarch: arm64
	- goos: darwin
	goarch: amd64
	- goos: darwin
	goarch: arm64
	- goos: windows
	goarch: amd64
	- goos: windows
	goarch: arm64
	defaults:
	run:
	working-directory: tui
	steps:
	- uses: actions/checkout@v7

	- uses: actions/setup-go@v7
	with:
	go-version-file: tui/go.mod
	cache-dependency-path: tui/go.sum

	- name: Build
	env:
	CGO_ENABLED: '0'
	GOOS: ${{ matrix.goos }}
	GOARCH: ${{ matrix.goarch }}
	run: \|
	EXT=""
	if [ "${{ matrix.goos }}" = "windows" ]; then EXT=".exe"; fi
	PKG=github.com/amd/gaia/tui/internal/cli
	go build -ldflags="-s -w -X ${PKG}.version=${{ github.ref_name }} -X ${PKG}.commit=${{ github.sha }} -X ${PKG}.date=$(date -u +%Y-%m-%dT%H:%M:%SZ)" \
	-o ../bin/gaia-${{ matrix.goos }}-${{ matrix.goarch }}${EXT} \
	./cmd/gaia

	- name: Upload artifact
	uses: actions/upload-artifact@v7
	with:
	name: gaia-${{ matrix.goos }}-${{ matrix.goarch }}
	path: bin/gaia-*
	retention-days: 14

	size-check:
	name: Binary size check
	needs: build
	runs-on: ubuntu-latest
	steps:
	- name: Download linux/amd64 artifact
	uses: actions/download-artifact@v8
	with:
	name: gaia-linux-amd64

	- name: Check size
	run: \|
	SIZE=$(stat --printf="%s" gaia-linux-amd64)
	SIZE_MB=$((SIZE / 1024 / 1024))
	echo "Binary size: ${SIZE_MB}MB (${SIZE} bytes)"
	if [ "$SIZE" -gt 15728640 ]; then
	echo "::warning::Binary size ${SIZE_MB}MB exceeds 15MB target"
	fi

Navigation Menu

Uh oh!

fix(gaia): keep a loaded skill alive across turns; stop a control resize from wrecking the screen #192

Workflow file

fix(gaia): keep a loaded skill alive across turns; stop a control resize from wrecking the screen #192

Uh oh!

Workflow file for this run