feat(bots): cross-platform bot creation + desktop connector reliability

[haowei2000]

Follow-on work on the same branch after #265 merged. Five commits.

What's in it

Commit	Change
02a5fe17	connector: resolve the opencode adapter before minting a token
e2d99bac	bots: make creating a bot work the same way on web, macOS and iOS
fd5b7bf7	ios: unstick touch input, tame auto-scroll, finish channel management
27872bfb	desktop: carry --config on every connector restart
4209abdf	desktop: split-button "Open in …" with app-tile icons

02a5fe17 — the stranded-bot bug

Onboarding a bot as "opencode" left it permanently unable to sign in (4401 invalid or revoked token).

The gateway preset named opencode-acp, a binary that does not exist — OpenCode exposes ACP as a subcommand of its own binary (opencode acp). The comment on that preset said "best-effort name; verify against your build", and it never was.

That alone would only be a failed setup. The damage came from ordering: redeeming an enrollment code mints a new bot token, destructively replacing the old one and kicking any live connector. The desktop app then failed while writing the config, because it could not resolve the adapter — after the token had already rotated. Net effect: the gateway expects a token no host holds, the connector that was working is offline, and the single-use code is spent. Trying a different agent type on an existing bot was enough to break it.

Three parts, and fixing any one alone is insufficient:

• AgentPreset gains args, so opencode renders command = "opencode", args = ["acp"].
• The adapter is pre-flighted before any server call, keeping "not installed" a side-effect-free error.
• The opencode fallback in detect_agents is dropped — it accepted a command absolutize_adapter_command still refused, so detection and the write path disagreed about what "installed" meant. Left in, it would have let the new pre-flight wave the same broken config straight through.

e2d99bac — bot creation across the three clients

Creating a bot had two implementations sharing no code, and iOS had none. Underneath, the only real variable is where the connector runs; this makes that the explicit question.

Safety, since these paths mint irreversible credentials:

• The agent type now travels with the bot. bridge_provider is read instead of re-defaulted — a bot created as codex was getting a claude connector, and nothing downstream catches that, because the connector runs whatever adapter the config names.
• The bot is created when a mode is picked, not on Continue — backing out used to leave an orphan bot with no delete affordance.
• The wizard ends on live bridge_connected instead of a Done button that closed the dialog and confirmed nothing. The step was already called "Connect".

iOS can now create a bot and hand it off, but never hosts a connector — a phone cannot run a long-lived ACP child process. The steps mirror the web wizard; the carrier changes, because a shell one-liner is useless on a phone: the code renders as a QR, configs go through the share sheet.

That needs somewhere to land, so macOS gains "I have a code" — the missing half of the remote-machine story. The desktop could previously only mint and redeem for itself in one shot, so a code from anywhere else had nowhere to go.

Also drops jargon from user-facing copy: "Rust ACP connector", "Agent Bridge token", bot_token_env, "0600", and a reachability warning that told whoever was adding a bot to run kubectl port-forward. DESIGN.md already bans raw field names in UI copy; this pays that down. Code comments keep the precise terms.

27872bfb — the connector fix

The CLI's restart fails with "restart requires --config" when there is no prior daemon.json — exactly the case right after onboarding or after a crash, which is when a revive matters most. Resolves the config path (the daemon's recorded one, else the onboarded ~/.cheers/cheers-daemon.<safe>.toml) and passes it on three paths: connector_restart, connector_add_allowed_roots, and the supervisor's revive loop. Each also calls ensure_workspace_dirs_at first, so starting from a hand-picked .toml is zero-prep too.

4209abdf — the UI change

One icon per installed opener crowded the file header. Now a split button: the remembered app opens on click, the chevron drops down the full list. Glyphs moved to openerIcons.tsx as self-contained SVG app tiles (no network, no asset files).

Verification

Target	Check	Result
Gateway	cargo test --lib	176/176 (incl. 2 new connector_config regression tests)
Web	tsc --noEmit	clean (2 pre-existing PdfViewer baseline errors only)
Web	rebuild → kind deploy → sign in	passes, zero console errors
macOS	cargo check --all-targets	passes
macOS	cargo test	20/20
macOS	full npm run dev build	binary builds and launches
iOS	xcodebuild (simulator)	BUILD SUCCEEDED
iOS	simulator install + launch	passes, login screen renders

The 02a5fe17 root cause was confirmed against production logs (two redeems 33s apart on one bot) and reproduced independently with a raw websocket handshake, so the diagnosis does not rest on the connector's own error reporting.

Not verified — please check before merging

• The bot-creation UI in e2d99bac was never opened. That commit is almost entirely user-visible copy and layout across three clients, and it has only been type-checked and compiled. The web wizard and the new iOS screens both want a visual pass; several strings got materially longer.
• The connector restart paths were not exercised at runtime. cargo test covers the pure workspace_dirs_* helpers; it does not reach the three new --config call sites. Testing them needs a dev Tauri instance, which would have collided with the running production Cheers.app. Worth a local pass on "fresh onboard → restart" before merge.
• iOS was not functionally re-tested for fd5b7bf7 — this session only confirmed it builds and launches.
• Existing broken bots are not repaired. 02a5fe17 prevents the failure; it cannot recover a bot whose token was already overwritten. Those need re-onboarding.

Note

openerIcons.tsx header says the glyphs are "stylized marks, not the exact trademarked logos", but the VS Code glyph is the verbatim simple-icons logo path. Either the comment or the path should change.

🤖 Generated with Claude Code