## Summary

- Fix `fatal: a branch named 'main' already exists` error when reviewing
fork PRs where the contributor's head branch is `main` (their default
branch)
- Fix corrupted diff base — after redirecting origin to the fork,
`origin/$BASE_BRANCH` gets overwritten, so save the base SHA before any
URL changes
- Use fixed branch name `pr-head` with `-B` (force-create) instead of
`-b $HEAD_REF` to avoid name collisions

## Root Cause

When a fork PR uses `main` as its head branch ([PR
#684](#684)):
1. `actions/checkout` creates local `main` branch from the base repo
2. `git checkout -b main origin/main` fails because `main` already
exists
3. Even if that succeeded, `git fetch origin main` (from fork) would
overwrite `origin/main`, making the diff compare fork-vs-fork instead of
base-vs-fork

Failed run:
https://github.com/amd/gaia/actions/runs/23859641383/job/69562473226

## Test plan

- [ ] Trigger Claude review on [PR
#684](#684) (fork PR with `main` as head
branch) — should succeed
- [ ] Trigger Claude review on a non-fork PR — should still work as
before
- [ ] Verify the generated diff in logs shows correct base-vs-fork
comparison

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

## Summary

- Consolidate `publish_installer.yml`, `publish-npm-ui.yml`, and PyPI
publishing into a single `publish.yml` workflow triggered by `v*` tag
pushes
- Single approval gate via GitHub "publish" environment — approve once,
publish both PyPI and npm in parallel
- `pypi.yml` reduced to CI-only build check (verifies package builds on
every push/PR)
- npm publish now uses the validated build artifact from the build step
instead of rebuilding from source

## Workflow flow

```
validate → build (pypi + npm + electron, parallel)
  → approve (single manual gate)
    → publish (pypi + npm, parallel)
      → github release (sigstore + electron artifacts)
```

## One-time setup required

1. Create a **"publish"** GitHub environment with required reviewers
(this is the single approval gate)
2. Keep the **"pypi"** environment for OIDC trusted publishing, but
remove its required reviewers
3. Update PyPI trusted publisher config to reference `publish.yml`
instead of `pypi.yml`
4. npm auth is handled via npm's validated workflow mechanism (manual
config)

## Test plan

- [ ] Verify `pypi.yml` runs on push/PR and skips tag pushes
- [ ] Verify `publish.yml` triggers on `v*` tag push
- [ ] Verify validation catches version mismatches
- [ ] Verify single approval gate blocks until reviewer approves
- [ ] Verify npm publish uses pre-built artifact, not a fresh rebuild
- [ ] Test with `workflow_dispatch` for dry-run validation

## Summary

- **AgentRegistry** discovers agents at boot from built-in Python
agents, custom Python modules (`~/.gaia/agents/*/agent.py`), and YAML
manifests (`~/.gaia/agents/*/agent.yaml`)
- **Per-session agent routing** — sessions store `agent_type`; the
correct agent is instantiated on cache miss
- **Agent selector** in the chat footer and WelcomeScreen — switching
agents mid-conversation creates a new session; switching on an empty
session updates in place
- **Agent name in message header** — shows "GAIA {Agent Name} |
Thinking" during streaming and "GAIA {Agent Name} | 3m ago" after
completion
- **YAML manifest agents** support MCP servers, model preferences, and
conversation starters — drop a folder in `~/.gaia/agents/` and it
appears in the UI after a server restart
- **Builder Agent** — "+" button opens a conversational agent creator
that generates `agent.py` (with optional MCP scaffolding) and
auto-refreshes the selector without a server restart
- **DispatchQueue** boot orchestration — `_check_lemonade` and
`_import_modules` run in parallel at startup; `_load_model` runs after
`_check_lemonade` completes; all tasks report progress to the frontend
- New guide: `docs/guides/custom-agent.mdx`

Closes #612, #713

## Test plan

- [ ] `python -m pytest tests/unit/agents/ tests/unit/chat/ui/` — all
tests pass
- [ ] Agent selector visible in chat footer and WelcomeScreen; switching
mid-conversation creates new session
- [ ] Message header shows agent name during and after streaming
- [ ] Builder creates a working agent at
`~/.gaia/agents/<name>/agent.py`; appears in selector immediately
- [ ] YAML manifest in `~/.gaia/agents/<name>/agent.yaml` appears after
server restart
- [ ] Boot progress banner shows task names; send button disabled until
ready

---------

Co-authored-by: Tomasz Iniewicz <tomasz.iniewicz@amd.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

)

## Summary

  - Add `GAIA_CACHE_V1` format header and SHA-256 checksum verification
    to cache files before deserializing with `pickle.loads()`
  - Reject cache files exceeding 500 MB size limit
  - Delete corrupted or outdated cache files and re-index automatically
- Old-format caches (pre-header) are detected and rebuilt on first load

  ## Context

  `pickle.load()` was called on cached RAG data without any validation
  (#447). A corrupted cache file would crash the RAG system permanently
  with no recovery path. This adds integrity verification at read time
  and writes a self-verifying format at cache time.

  Note: the diff includes an indentation change (~80 lines dedented)
because data is now read from the file and validated before processing,
  so the processing logic moved outside the `with open()` block.

  ## Test plan

  - [x] Corrupted cache files are detected and deleted automatically
  - [x] Cache load failures trigger clean re-index, not crash
  - [x] Checksum mismatch is detected and triggers re-index
  - [x] Oversized cache files are rejected
  - [x] Old-format (pre-header) cache files are migrated on first load
  - [x] All 33 tests pass (`python -m pytest tests/test_rag.py`)
  - [x] All lint checks pass (`python util/lint.py --all`)

Fixes #728.

## What was broken

Drag-and-drop document upload failed with *"Indexing error: Not found.
The file or folder may have been moved or deleted."* in any browser, and
in the packaged Electron desktop app (Electron 40).

**Root cause:**
`src/gaia/apps/webui/src/components/DocumentLibrary.tsx:265` relied on
`(file as any).path`, an Electron-only augmentation that's `undefined`
in every browser and was [removed in Electron
32+](https://www.electronjs.org/blog/electron-32-0). The fallback
`file.name` (bare filename, no directory) was POSTed to
`/api/documents/upload-path`, which couldn't resolve it → HTTP 404 →
misleading frontend error.

## Fix

New `POST /api/documents/upload` endpoint that accepts the file content
as multipart form data, streams it to
`~/.gaia/documents/<shortid>_<original_name>.<ext>` with a 20 MB cap,
dedupes by SHA-256 hash, and indexes via the standard RAG pipeline.
Drag-and-drop in both DocumentLibrary and ChatView now use this
endpoint. Delete cleans up server-owned files; user-owned files from
`upload-path` are left alone.

Also fixes `getFriendlyError()` in `api.ts` to prefer the backend's
`detail` over the canned 404 string, so future upload errors surface
meaningful messages.

### Details worth knowing
- **On-disk filename embeds the original stem**
(`<12-hex-id>_<sanitized_name>.<ext>`) so the agent LLM and RAG tools
can recognize the file by its human-readable name. A pure-UUID filename
caused the LLM to hallucinate paths (e.g.
`C:\Users\14255\.gaia\documents\<uuid>.pdf` →
`/Users/14255/Documents/<uuid>.pdf`) and fail tool lookups.
- **Filename sanitizer** strips Windows-illegal chars (`<>:"/\|?*`),
control chars, leading/trailing whitespace and dots, and caps length at
80 chars — keeps total path well under Windows MAX_PATH.
- **12-hex-char ID** instead of full UUID. We dedup by SHA-256 first, so
identical-content collisions are impossible; the only theoretical
collision is concurrent uploads of *different* content drawing the same
ID, with probability ~1 in 1.4e14 even at 1M concurrent uploads. Saves
24 chars per filename.
- **Reliability**: streaming write with abort-on-overflow, try/finally
partial cleanup (survives ClientDisconnect), atomic `os.replace`, orphan
cleanup on indexing/db/race failures.
- **Chat-view drop UX**: dropping onto the chat area indexes quietly
without forcing the Document Library side panel open. Dropping onto the
library panel itself still opens it (correct for users intentionally
managing their library).

### Bundled cleanup
This PR also includes a doc-consistency pass auto-applied by
`util/lint.py --all --fix` (step 8) that aligns the supported-extension
allowlist across backend (`utils.py`, `files.py`), frontend
(`UnsupportedFeature.tsx`), and tests.

### What's NOT changed
- `POST /api/documents/upload-path` is untouched — still used by the
server-side file browser in `FileBrowser.tsx`, where the user explicitly
picks a real filesystem path.
- The Electron preload — deliberately not adding
`webUtils.getPathForFile` since it's [broken for drag-drop in Electron
30-33+](electron/electron#44600) and the blob
path works everywhere.

### Known limitation / follow-up
The agent LLM still occasionally cites the on-disk filename (e.g.
`9c59c981f8e7_report.pdf`) when answering questions about a document,
instead of the user's display name (`report.pdf`). This is because
`src/gaia/agents/chat/tools/rag_tools.py` returns chunk metadata using
raw filepaths from `self.rag.indexed_files` rather than translating them
to the display name stored in `documents.filename`. **This is a
separate, pre-existing concern in the RAG tool layer** and fixing it
requires a focused refactor of `rag_tools.py` (~1800 lines, touches
`query_documents`, `query_specific_file`, `summarize_document`,
`search_indexed_chunks`, `list_indexed_documents`). Tracking it as a
follow-up rather than expanding this PR's scope.

### Conflict notes for reviewers
- **#723** (surface indexing failures) also touches
`src/gaia/ui/routers/documents.py`, but edits the body of
`upload_by_path` — this PR only adds new top-of-file constants, helpers,
and a new `upload_document_blob` function inserted after
`upload_by_path`. Three-way merge should be clean. Semantically, #723's
"raise on failure" change is **more correct** under my new endpoint than
the previous silent-0-chunk behavior.
- **#725** (Ask Agent file indexing) touches `ChatView.tsx` in a
disjoint line range. No conflict.

---

## Steps to reproduce the bug (before this fix)

### One-time setup
```bash
git checkout main
uv venv
uv pip install -e ".[dev,ui]"
cd src/gaia/apps/webui && npm install && npm run build && cd -
```

> **Windows PowerShell users:** quote the extras → `uv pip install -e
".[ui]"`. Without quotes PowerShell silently strips them.

### Start the services
```bash
# Terminal 1
lemonade-server serve

# Terminal 2
gaia chat --ui
```

### Reproduce
1. Open the UI in **Firefox** (most reliable repro; the bug also
reproduces in Chrome and the packaged Electron 40 desktop app).
2. Open DevTools → Network tab, filter on `documents`.
3. Open or create a chat session.
4. Drag any `.pdf`/`.txt`/`.md`/`.docx` from your file manager onto the
Document Library drop zone or the chat area.
5. **Observe:**
- UI: *"Indexing error: Not found. The file or folder may have been
moved or deleted."*
   - Network: `POST /api/documents/upload-path` → **HTTP 404**
- Request body: `{"filepath":"report.pdf"}` (bare filename — the smoking
gun)
   - Backend log: `HTTPException: File not found: report.pdf`

### Verify the fix
```bash
git checkout kalin/fix-728-document-upload
cd src/gaia/apps/webui && npm run build && cd -
# Restart `gaia chat --ui`
```
Repeat the drop. Expected:
- Network: `POST /api/documents/upload` (multipart, **not**
`upload-path`) → **HTTP 200**
- File persisted at `~/.gaia/documents/<shortid>_report.pdf` with the
original name embedded
- Backend log: `Blob upload indexed: report.pdf (<N> bytes, <M> chunks)`
- Drop the same file twice → `Blob upload dedup: report.pdf matches
existing doc <id>` (no second file on disk)
- Ask the agent about the doc → answers correctly using the indexed
content

## Test plan

- [x] 10 new integration tests
(`tests/integration/test_documents_router.py`): happy path, dedup,
empty/oversized/bad-extension/illegal-chars rejection, indexing failure
cleanup, delete unlinks server-owned, delete preserves user-owned, 404
on delete
- [x] Adjacent regression sweep: 186 tests pass (files router, folder
indexing, database, server unit tests)
- [x] `python util/lint.py --all --fix` clean
- [x] `tsc --noEmit` clean (no new errors in changed files)
- [x] End-to-end smoke test via TestClient (upload → list → delete →
file unlinked)
- [x] **Manual (Windows + Firefox)**: drag-and-drop indexes successfully
→ ✅ validated
- [x] **Manual**: dropped doc auto-attaches to chat session → ✅
validated
- [x] **Manual**: dedup short-circuit verified → ✅ validated
- [x] **Manual**: delete cleanup verified for both server-owned and
user-owned files → ✅ validated
- [x] **Manual**: agent answers questions correctly using the indexed
document → ✅ validated (display-name citation polish tracked as
follow-up; see "Known limitation" above)

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Closes #530.

Ships GAIA as a real desktop app with one-click installers — the primary
install path for non-developer end users.

## What users get

**One-click install on every platform.** Download `gaia-agent-ui-*.exe`
(Windows), `.dmg` (macOS), `.deb` or `.AppImage` (Linux) from GitHub
Releases. Double-click and you're done. No Python, no `pip`, no
terminal.

**Automatic backend setup on first launch.** The app shows a progress
dialog while it downloads `uv`, creates a Python venv, installs the GAIA
backend, and starts Lemonade Server. About 5–10 minutes the first time,
instant on every launch after. If anything fails, users see an error
dialog with a Retry button, a "Manual install" link, and a log file path
— no more dead-end *"Install with: pip install amd-gaia"*.

**Auto-updates in the background.** GAIA checks GitHub Releases on
launch and every few hours. When a new version is available, users see a
"Restart to update" prompt — same flow as Slack, VS Code, etc. No manual
reinstalls.

**Native uninstall.** Add or Remove Programs / drag to Trash / `apt
remove` all work as expected. User data in `~/.gaia/` is preserved
unless explicitly purged. A `gaia uninstall` CLI offers tiered cleanup
(`--venv`, `--purge`, `--purge-lemonade`, `--purge-models`,
`--purge-hf-cache`, `--dry-run`) for power users.

**Windows autostart by default.** GAIA starts in the system tray when
you sign in to Windows, matching the Slack/Discord/Zoom convention.
Disable via Task Manager → Startup or the in-app tray menu.

**Trusted installers (when secrets land).** Code-signing scaffolding is
in place for SignPath (Windows) and Apple Developer ID + notarization
(macOS). Activates with zero code changes once #732 and #733 unblock —
no more SmartScreen / Gatekeeper warnings.

## Naming

Display name is **GAIA** (dock, Start Menu, title bar, tray). Electron
binary is `gaia-desktop` (avoids collision with the Python CLI `gaia`).
Download artifacts use `gaia-agent-ui-*` for URL stability.

## Verification

- Windows NSIS installer tested end-to-end on real hardware: install →
first-launch bootstrap (uv + venv + pip + gaia init + Lemonade) → chat
UI loads → functional
- Local macOS DMG verified: mounted, code-signed, entitlements
confirmed, first-launch bootstrap completes
- Cross-platform CI green on Windows + macOS + Linux via the new
`build-installers.yml` workflow (51/51 checks passing)
- 3 automated code review passes: 19 issues found and fixed (name
collisions, stale refs, orphaned registry keys, missing require, cwd
bug, frontend loading bug, API_BASE file:// fix, doc sync)
- 42 unit tests for `gaia uninstall` (containment guard, log handler
cleanup, silent-purge refusal, Lemonade python resolution, exit codes,
GAIA_HOME env var)

## Test plan

- [x] Smoke test on real Windows: download → install → first-run → chat
UI loads
- [ ] Smoke test on real macOS: mount DMG → drag to Applications →
first-run backend install → single-instance lock → uninstall
- [ ] Smoke test on real Linux: `.deb` + `.AppImage`
- [ ] Re-verify after SignPath / Apple Developer ID secrets land (#732,
#733)

## Follow-ups

- #732 — SignPath OSS application for Windows code signing
- #733 — AMD Apple Developer ID for macOS notarization
- #734 — R2 / amd-gaia.ai distribution mirror

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

)

Fixes two critical bugs found during end-to-end install testing on
Ubuntu 24.04 (amd64), testing against PR #731.

## Bug 1: chrome-sandbox missing SUID bit

**Symptom:** App crashes immediately after `apt install` with:
```
FATAL: The SUID sandbox helper binary was found, but is not configured correctly.
Rather than run without sandboxing I'm aborting now. You need to make sure that
/opt/GAIA/chrome-sandbox is owned by root and has mode 4755.
```

**Root cause:** `electron-builder` packages `chrome-sandbox` with mode
`755`. Chromium's sandbox helper requires `4755` (setuid root) to
function. Without it, no user can launch the app after a clean install.

**Fix:** `installer/debian/postinst` now sets `chown root:root` + `chmod
4755` on `/opt/GAIA/chrome-sandbox` during package configuration, which
runs as root via `apt`.

## Bug 2: Model download silently skipped — app crashes when user types

**Symptom:** Bootstrap completes and reports `ready`, but Qwen3-0.6B is
never downloaded. Lemonade has no model loaded, so the app crashes when
the user sends their first message.

**Root cause:** `backend-installer.cjs` runs `gaia init --profile
minimal` with piped stdio (for log streaming). With no TTY attached,
`gaia init`'s interactive prompts ("Continue? [Y/n]:" and "Run model
verification? [Y/n]:") receive EOF and answer themselves as "Skipping".
The non-zero-ish exit is treated as non-fatal, so state reaches `ready`
despite no model existing.

**Fix:** Pass `--yes` to `gaia init` so prompts are auto-answered. The
flag already exists for this purpose.

## Testing

Both bugs reproduced on Ubuntu 24.04.4 LTS (x86_64, kernel 6.17) with
the `.deb` from CI run
[24255947495](https://github.com/amd/gaia/actions/runs/24255947495).
Verified manually:
- Before fix: immediate crash on launch (Bug 1), silent model skip +
crash on chat (Bug 2)
- After applying patches locally: app launches, bootstrap downloads
model, chat works

Closes part of #530.

## Summary

Closes #712

- MCP tool calls in the Agent Activity panel now show a purple **"via
{server}"** badge identifying the MCP server
- **Latency** (in ms) is displayed in the collapsed tool header,
measured at the SSE handler layer via `time.monotonic()`
- MCP tool cards start **collapsed by default** (native tools still
auto-expand)
- **Filter bar** appears when 2+ tool steps exist — filter by tool name
text or status (All/Success/Error)
- Server name resolved via `_TOOL_REGISTRY` lookup (no fragile prefix
parsing)
- No new SSE event types, no DB migration, no new API endpoints

## Changes

| File | Change |
|------|--------|
| `src/gaia/ui/sse_handler.py` | `mcp_server` on `tool_start`,
`latency_ms` on `tool_result` |
| `src/gaia/ui/_chat_helpers.py` | `mcpServer`/`latencyMs` in captured
steps |
| `src/gaia/ui/models.py` | New fields on `AgentStepResponse` |
| `src/gaia/apps/webui/src/types/index.ts` | Extend `AgentStep` +
`StreamEvent` |
| `src/gaia/apps/webui/src/components/ChatView.tsx` | Forward
`mcp_server`/`latency_ms` in event handlers |
| `src/gaia/apps/webui/src/components/AgentActivity.tsx` | MCP meta,
badge, latency, filter bar, collapsed-by-default |
| `src/gaia/apps/webui/src/components/AgentActivity.css` | Filter bar,
MCP badge, latency styles |
| `tests/unit/chat/ui/test_sse_handler.py` | 7 new MCP visualization
tests |
| `tests/unit/chat/ui/test_models.py` | 4 new AgentStepResponse tests |

## Test plan

- [x] 216 unit tests pass (`python -m pytest tests/unit/chat/ui/`)
- [x] Frontend builds clean (`npm run build`)
- [x] Lint passes (`python util/lint.py --all --fix`)
- [x] Manual: `gaia chat --ui` → configure MCP server → trigger tool
calls → verify badge/latency/filter/collapse
- [x] Manual: Refresh page → verify tool traces persist in session

## Images
<img width="813" height="751" alt="image"
src="https://github.com/user-attachments/assets/82f6d6d1-a272-4c30-a5cc-9aebfe64b92f"
/>

---------

Co-authored-by: Tomasz Iniewicz <infancy_shred.0d@icloud.com>
Co-authored-by: Tomasz Iniewicz <itomek@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Release notes for v0.17.2. See
[docs/releases/v0.17.2.mdx](https://github.com/amd/gaia/blob/v0.17.2-release/docs/releases/v0.17.2.mdx)
for the full body.

**Headline features:**
- One-click desktop installers for Windows, macOS, and Linux (#731 +
#743)
- Customizable agents and Builder Agent (#720)
- MCP tool execution visibility in Agent Activity panel (#717)

**Security:** RAG cache integrity verification (#722, closes #447)

**Bug fixes:** drag-and-drop upload in any browser (#729), Linux install
fixes (#743)

**Release & CI:** unified publish workflow (#715), Claude review on fork
PRs (#685), npm OIDC fix (#683), webui package.json sync (#682)

Full changelog: v0.17.1...v0.17.2

---------

Co-authored-by: Tomasz Iniewicz <tomasz@iniewicz.com>

## Summary

- The `Publish Release` workflow failed with `startup_failure` (0 jobs,
instant failure) on the v0.17.2 tag push:
https://github.com/amd/gaia/actions/runs/24267844380
- **Root cause:** The heredoc body in the `github-release` job's
"Extract release notes" step was at column 0 (no indentation), which
broke out of the `run: |` YAML block scalar. The YAML parser interpreted
`- **Windows**` as a sequence item with `*Windows` as an alias
reference, causing a parse error before any job could start.
- **Fix:** Indent the heredoc body to 10 spaces to match the block
scalar's indentation level. YAML strips the leading spaces
automatically, so bash still receives the content at column 0 — no
change to the generated release body.

## Test plan

- [x] `yaml-lint` passes on `publish.yml` (was failing before the fix)
- [ ] Re-tag v0.17.2 (or use `workflow_dispatch`) to verify the publish
workflow starts successfully

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…748)

## Summary

- Fixes the `startup_failure` on the Publish Release workflow:
https://github.com/amd/gaia/actions/runs/24278877956
- **Root cause:** `build-electron-apps.yml` declares `permissions:
contents: write` at its top level. When called as a reusable workflow
from `publish.yml`, the calling `build-electron` job only had `contents:
read` (inherited from publish.yml's workflow-level permissions). GitHub
Actions validates this at startup and rejects the run before any jobs
start.
- **Fix:** Add explicit `permissions: contents: write` to the
`build-electron` job in publish.yml so it matches what the reusable
workflow declares.

> **Note:** The previous PR #747 fixed a YAML block scalar issue
(heredoc indentation) that was also causing a parse error. This PR fixes
the second startup_failure — the permissions mismatch discovered from
the GitHub Actions error annotation.

## Test plan

- [ ] Merge, move v0.17.2 tag, verify Publish Release workflow starts
and reaches the Validate step

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

## Summary

- Fixes the macOS DMG build failure in the Publish Release workflow:
https://github.com/amd/gaia/actions/runs/24278980001
- **Root cause:** electron-builder auto-skips code signing for PR builds
(`Current build is a part of pull request, code signing will be
skipped`) but NOT for tag pushes. Without valid signing secrets, it
tries to sign with empty credentials and fails with `⨯
.../src/gaia/apps/webui not a file`.
- **Fix:** Pass `--config.mac.identity=null` when `APPLE_ID` is not set,
explicitly telling electron-builder to skip signing. Windows and Linux
builds are unaffected (they don't attempt signing without their
respective secrets).

> This is the third fix for the v0.17.2 publish pipeline, after #747
(YAML heredoc indentation) and #748 (reusable workflow permissions).

## Test plan

- [x] The PR itself triggers `build-installers.yml` (via `paths:`
filter) — the macOS job will validate the fix
- [ ] After merge, move v0.17.2 tag and re-trigger publish

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

## Summary

- Fixes TypeScript compilation failure in the `Publish to npm` job:
https://github.com/amd/gaia/actions/runs/24279178148/job/70898457497
- `terminalStore.ts` callbacks (`setPaused`, `setActiveTab`,
`openTerminal`, `setShowTerminal`) relied on Zustand type inference for
parameter types, but `tsc` flagged them as implicit `any` during the
build
- Added explicit parameter type annotations matching the `TerminalState`
interface

## Test plan

- [x] `npx tsc --noEmit` passes locally with zero errors
- [ ] Merge, move tag, verify npm publish step passes

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

## Summary

- Fixes the `Publish to npm` job failure:
https://github.com/amd/gaia/actions/runs/24279365243/job/70899004478
- **Root cause:** `npm publish` triggers `prepublishOnly` → `tsc && vite
build`, but the job installs only production deps (`--omit=dev`). `tsc`
fails because `@types/react`, `typescript`, etc. are missing.
- **Fix:** Add `--ignore-scripts` to `npm publish`. The dist/ is already
pre-built in the `build-npm` job and downloaded as an artifact — the
prepublishOnly rebuild is redundant.

> Fix #5 for the v0.17.2 publish pipeline, after #747 (YAML indent),
#748 (permissions), #749 (macOS signing), #750 (terminalStore types).

## Test plan

- [ ] Merge, move tag, verify npm publish succeeds

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

## Summary

- Fixes `ENEEDAUTH` in the `Publish to npm` job:
https://github.com/amd/gaia/actions/runs/24279785278/job/70899991808
- **Root cause:** The old `publish-npm-ui.yml` workflow used
`environment: npm` which carries the OIDC claims npm needs for trusted
publishing. The new unified `publish.yml` dropped this block, so npm
rejected auth.
- **Fix:** Restore the `environment: npm` block and the `npm install -g
npm@latest` upgrade step, both matching the old working workflow.

## Test plan

- [ ] Merge, move tag, verify npm publish authenticates and publishes
successfully

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

## Summary

- Fixes `Cannot find module 'promise-retry'` in the npm publish job:
https://github.com/amd/gaia/actions/runs/24280075274/job/70900655694
- **Root cause:** `npm install -g npm@latest` replaces the bundled npm
with a version incompatible with the runner's Node 22 layout, corrupting
npm's own dependencies.
- **Fix:** Remove the upgrade step. Node 22 ships with npm 10+ which
already supports OIDC trusted publishing.

## Test plan

- [ ] Merge, move tag, verify npm publish works

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

## Summary

- Fixes `ENEEDAUTH` in npm publish:
https://github.com/amd/gaia/actions/runs/24280576961
- **Root cause:** `setup-node` without `registry-url` never creates the
`.npmrc` that maps `NODE_AUTH_TOKEN` to the npm registry. npm's OIDC
trusted publishing needs this to authenticate.
- **Note:** The old `publish-npm-ui.yml` was never actually run (zero
successful runs in history) — v0.17.0 and v0.17.1 were published outside
CI. So the old workflow's lack of `registry-url` was an untested bug
too.
- **Fix:** Add `registry-url: 'https://registry.npmjs.org'` to
`setup-node`.

## Test plan

- [ ] Merge, move tag, verify npm publish authenticates and succeeds

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>