Summary

Adds a new LLM provider — Claude Code CLI (local) — that lets users with a Claude Code subscription pick Claude models in settings without pasting an Anthropic API key. The app spawns the local claude binary as a subprocess and streams the conversation over stdin/stdout as stream-json, so the OAuth credentials already in ~/.claude/ are reused and no key ever touches this app.

The transport sits behind the same StreamCallbacks contract as every HTTP provider, so the chat panel, cancellation (AbortSignal), and tokenized streaming all work identically — the rest of the code doesn't know which transport it's talking to.

What changed

Rust (Tauri backend) — new module src-tauri/src/commands/claude_cli.rs:

• claude_cli_detect — locates claude on PATH, runs claude --version with a 3s timeout, returns {installed, version, path, error}. Surfaces the macOS Gatekeeper quarantine remediation hint (xattr -d com.apple.quarantine …) when that specific failure is detected.
• claude_cli_spawn — spawns claude -p --output-format stream-json --input-format stream-json --verbose --model <model>, writes the serialized history to stdin (closing it so claude starts processing), and forwards each stdout line as a Tauri event claude-cli:{streamId}. stderr is collected and bundled into the final claude-cli:{streamId}:done event so non-zero exits surface the real diagnostic.
• claude_cli_kill — cancels a running subprocess for AbortSignal support.
• ClaudeCliState — HashMap<stream_id, Child> in a tokio Mutex; registered via .manage() in lib.rs.

TS (frontend) — new file src/lib/claude-cli-transport.ts:

• createClaudeCodeStreamParser() — parses stream-json lines into token text. Handles both stream_event (real Anthropic-API deltas, passthrough when --verbose is on) and assistant events (full in-progress message on every emission). When both arrive, deltas are authoritative and the fat assistant events are suppressed to avoid double-render. When only assistant events arrive, emits the novel tail by prefix-diffing against what we've already shown.
• streamClaudeCodeCli() — wires invoke/listen to the Rust commands, enforces the AbortSignal contract, and maps exit codes to user-facing errors. Sampling overrides (temperature/top_p/max_tokens/stop) have no CLI equivalents and are silently ignored with a dev-only console warning so callers don't silently wonder why they don't take effect.

Dispatch — src/lib/llm-client.ts routes provider === "claude-code" to the subprocess transport before getProviderConfig, which explicitly throws for this provider since it has no URL/headers.

Settings UI — llm-provider-section.tsx excludes claude-code from the needsApiKey check (no key field shown) and renders a ClaudeCliStatusPill that calls claude_cli_detect on mount so the user sees immediately whether the binary is installed, quarantined, or missing.

Preset — LLM_PRESETS gets a claude-code-cli entry with defaultModel: "claude-sonnet-4-6" and the full Opus/Sonnet/Haiku lineup as suggestions. 200k context window.

Notable details for reviewers

• content must be an array of blocks on both user and assistant turns. During end-to-end testing I hit W is not an Object. (evaluating '"tool_use_id"in W') the moment assistant history appeared — the CLI iterates content blocks looking for tool_use_id and crashes on a raw string. User-only single-turn tolerates a string (that's the trap), so the Rust code normalizes both roles to [{type:"text",text:...}].
• System messages are inlined into the first user turn rather than passed via --system-prompt / --append-system-prompt, because flag availability varies across claude CLI versions. Inlining works on every version.
• Why tokio::process directly, not tauri-plugin-shell — the plugin's scope model is designed for sidecars or fixed absolute paths; scoping a user-installed PATH binary cleanly is awkward. A hardcoded Rust command that always and only spawns claude provides the same security property (the webview can't call this command to execute anything else) without pulling in another plugin or editing capabilities/*.json.
• Agent-mode state is explicitly out of scope — the CLI's tool use, MCPs, file-edit abilities, and --resume session state are all ignored. We use claude purely as a text-completion engine. Multi-turn history is reconstructed from the messages array on every call, symmetric with every other provider.
• Cargo.toml adds tokio (features: process, io-util, sync, macros, rt), which = "7", and uuid = "1" with v4.

Test plan

• Unit tests — src/lib/__tests__/claude-cli-transport.test.ts covers 9 parser scenarios (delta-only, assistant-only, delta-then-assistant dedupe, multi-part text, unknown event types, malformed JSON). All pass.
• getProviderConfig({provider:"claude-code"}) throws — test added in llm-providers.test.ts.
• tsc --noEmit clean on frontend.
• cargo check clean (only pre-existing warnings in fs.rs).
• Injection test: ran the exact command the Rust side spawns, confirmed single-turn and multi-turn-with-assistant-history both return valid responses ("Mango" → "Pineapple" for a fruit follow-up). This is how the content-must-be-array bug was found.
• Manual smoke in the desktop app: install claude, open Settings, pick "Claude Code CLI (local)", send a message, verify streaming tokens render.
• Manual cancel test: hit Stop mid-response, verify claude_cli_kill terminates the subprocess.
• Manual macOS-quarantine test: if reachable, confirm the status pill surfaces the xattr remediation hint.

🤖 Generated with Claude Code