main.rs · setoelkahfi/sigit · siGit Code & Deploy

@setoelkahfi / sigit

development

claude/code-feature-parity-q003hm claude/elegant-carson-l1menh claude/sigit-acp-local-chat-cx6380 claude/sigit-cloud-agent-expansion-reox0a claude/tool-permission-system claude/zen-feynman-0u78dk development feature/agent-tools-multiedit-glob-todos-remember feature/background-commands feature/commit-coauthor-attribution feature/headless-mode feature/init-command feature/load-local-model-explicitly feature/session-persistence-compaction feature/sigit-code-cloud feature/subagent-tool feature/tool-permission-system feature/tui-repo-tabs feature/tui-tabs main release/v1.3.1

rs 3,252 lines 136 KB

Raw

1	//! siGit Code — local coding agent on Onde Inference.
2	//!
3	//! In TTY mode, all output (log crate, tracing, stray printlns) redirects to
4	//! `$TMPDIR/sigit.log`. Ratatui holds a separate fd to the real terminal so
5	//! the TUI stays clean.
6	//!
7	//! Two modes:
8	//! - ACP over stdio (editor integration, e.g. Zed)
9	//! - interactive terminal (direct TTY)
10	//!
11	//! Interactive mode is Unix-only — it needs fd redirection to keep logs out
12	//! of the TUI. Windows only gets ACP mode for now.
13	//!
14	//! On macOS the HF cache lives in the App Group container shared with the
15	//! siGit desktop app. See [`setup`].
16	//!
17	//! # Zed setup
18	//!
19	//! Add to `~/.config/zed/settings.json`:
20	//! ```json
21	//! {
22	//! "agent_servers": {
23	//! "siGit Code": {
24	//! "type": "custom",
25	//! "command": "/absolute/path/to/target/release/sigit"
26	//! }
27	//! }
28	//! }
29	//! ```
30
31	mod account;
32	mod backend;
33	mod chat;
34	mod credentials;
35	mod headless;
36	mod instructions;
37	mod mcp;
38	mod models;
39	mod permissions;
40	mod provider;
41	mod session_store;
42	mod settings;
43	mod setup;
44	mod skills;
45	mod tools;
46
47	/// Serializes tests that mutate process-global env vars (`SIGIT_CONFIG_DIR`
48	/// etc.). `cargo test` runs tests in parallel within a binary, so without this
49	/// the credentials and settings round-trip tests clobber each other's env.
50	#[cfg(test)]
51	pub(crate) static ENV_TEST_LOCK: std::sync::Mutex<()> = std::sync::Mutex::new(());
52
53	use std::io::IsTerminal;
54	#[cfg(unix)]
55	use std::io::{BufWriter, Write};
56	use std::sync::Arc;
57
58	use onde::inference::SamplingConfig;
59
60	// `ProtocolVersion` is a version-agnostic type at the schema root; the rest of the
61	// schema types moved under `schema::v1` in agent-client-protocol 1.0.
62	use agent_client_protocol::schema::ProtocolVersion;
63	use agent_client_protocol::schema::v1::{
64	AgentCapabilities, AuthMethod, AuthMethodAgent, AuthenticateRequest, AuthenticateResponse,
65	AvailableCommand, AvailableCommandInput, AvailableCommandsUpdate, CancelNotification,
66	ConfigOptionUpdate, ContentBlock, ContentChunk, EmbeddedResourceResource, ForkSessionRequest,
67	ForkSessionResponse, Implementation, InitializeRequest, InitializeResponse, LoadSessionRequest,
68	LoadSessionResponse, Meta, NewSessionRequest, NewSessionResponse, PermissionOption,
69	PermissionOptionKind, PromptRequest, PromptResponse, RequestPermissionOutcome,
70	RequestPermissionRequest, SessionCapabilities, SessionConfigOption,
71	SessionConfigOptionCategory, SessionConfigSelectOption, SessionConfigValueId,
72	SessionForkCapabilities, SessionId, SessionNotification, SessionUpdate,
73	SetSessionConfigOptionRequest, SetSessionConfigOptionResponse, StopReason, ToolCall,
74	ToolCallStatus, ToolCallUpdate, ToolCallUpdateFields, ToolKind, UnstructuredCommandInput,
75	};
76	use agent_client_protocol::{Agent, ByteStreams, Client, ConnectionTo, Responder};
77	use onde::inference::{ChatEngine, GgufModelConfig};
78
79	use crate::backend::{
80	InferenceBackend, LocalBackend, OpenAiBackend, ToolResult as BackendToolResult, ToolSpec,
81	TurnResult,
82	};
83	use std::path::PathBuf;
84	use std::sync::atomic::{AtomicBool, Ordering};
85	use tokio_util::compat::{TokioAsyncReadCompatExt, TokioAsyncWriteCompatExt};
86	use tracing_subscriber::{EnvFilter, fmt as tracing_fmt};
87
88	#[cfg(unix)]
89	use std::os::unix::io::{AsRawFd, FromRawFd};
90
91	const SYSTEM_PROMPT: &str = "\
92	Your name is siGit — lowercase 's', uppercase 'G', no spaces. \
93	Not 'SiGit', not 'Sigit'. Only say your name if the user asks who you are.
94
95	You are a strong general-purpose coding agent. smbCloud is your home turf, \
96	but you should still be useful in any codebase. When the project is clearly \
97	about smbCloud, use that context directly instead of falling back to vague \
98	cloud-platform advice.
99
100	smbCloud context you should know and use when it helps:
101	- smbCloud is a platform for deploying and managing projects
102	- the main CLI is a Rust workspace with focused crates rather than one giant crate
103	- common areas include auth, project management, deploy flows, networking, \
104	shared models, release tooling, and managed services
105	- deploy branches usually follow `release/service-{name}`
106	- Next.js SSR deploys on smbCloud are not the same as generic git-push deploys; \
107	they often use a local build plus rsync/PM2 style flow
108	- auth has a hard boundary between smbCloud platform users and tenant app users; \
109	platform flows use `/v1/users`, tenant app flows use `/v1/client/`, and \
110	you should not casually mix `User`, `TenantMembership`, `AuthApp`, and `AuthUser`
111	- smbCloud authorization is layered; do not flatten platform accounts, tenant \
112	memberships, auth-app collaborators, and tenant end users into one model
113	- `Project` is the umbrella workspace, while app-like resources such as \
114	`FrontendApp`, `AuthApp`, and GresIQ are the deployable units with their own \
115	ownership, sharing, and collaboration rules
116	- `FrontendApp` is many-per-project, while `AuthApp` is intentionally one-per-project; \
117	preserve those cardinality rules unless the code clearly changes them
118	- GresIQ is smbCloud's managed PostgreSQL offering; treat it as a platform \
119	service with its own credentials and boundaries, not as a generic local DB helper
120	- when debugging smbCloud Rails APIs, first classify the request: first-party \
121	smbCloud app or tenant app, then check which endpoint family and validator \
122	should be involved before changing code
123	- when working in smbCloud repos, prefer existing workspace patterns, existing \
124	crate boundaries, existing Rails conventions, and existing command flows over \
125	inventing new abstractions
126
127	CRITICAL RULE — never tell the user to run a command. You have tools. Use them. \
128	When the user asks you to clone a repo, run a build, check git status, or do \
129	anything that involves a shell command, you MUST call the run_command tool and \
130	execute it yourself. Do not print shell commands for the user to copy-paste. \
131	Do not give step-by-step instructions. Do not say \"you can run …\". Just do it. \
132	If a command fails, try to fix the problem and re-run it. If you cannot fix it \
133	after two attempts, explain what went wrong and what you tried.
134
135	Git operations — always use run_command:
136	- git clone: always pass the full absolute destination path as the last argument \
137	and set cwd to an existing writable parent directory. Example: \
138	run_command({\"command\": \"git clone https://github.com/org/repo /Users/me/Repositories/repo\", \
139	\"cwd\": \"/Users/me/Repositories\"})
140	- git init, add, commit, push, pull, fetch, checkout, branch, diff, log, status, \
141	stash, rebase, merge, tag — use run_command with an absolute cwd pointing to \
142	the repo root
143	- never run git clone without an explicit absolute destination path
144	- if a clone or init fails, check the error, fix the cause (wrong path, missing \
145	directory, permissions), and retry
146	- when you create a commit, always end the commit message with a blank line and \
147	then this trailer on its own line: Co-Authored-By: siGit Code <sigit@sigit.si> \
148	— GitHub reads that exact format and credits siGit as co-author. If a commit \
149	lands without it, siGit Code amends the trailer in automatically and the tool \
150	output says so; do not amend again yourself.
151
152	Never introduce yourself unless asked. Jump straight into the answer. \
153	Keep answers short. Write idiomatic code. \
154	Fix root causes, not symptoms.
155
156	You have access to tools that let you read files, read websites directly from \
157	http and https URLs, create directories, list directories, search code, create \
158	new files, edit existing files, delete files, and run shell commands. You can \
159	also use git directly through shell commands, including `git init` and normal \
160	git workflows. Use them proactively. Read the code or website before answering. \
161	Prefer absolute paths when referring to files and directories, especially in \
162	protocol-facing output and tool arguments. Create directories when needed. Run \
163	builds, tests, and git commands after making changes. Ground your answers in \
164	the actual code or fetched page content, not in guesses.
165
166	CRITICAL — you CAN access websites. You are NOT a typical LLM without internet \
167	access. You have a read_website tool that fetches any http or https URL and \
168	returns the page text. When the user gives you a URL or asks you to read, \
169	summarize, or inspect a web page, you MUST call the read_website tool with that \
170	URL. Never say \"I cannot access websites\" or \"I cannot browse the internet\". \
171	You can. Use the tool.
172
173	CRITICAL — before every edit_file call, you MUST call read_file on the target \
174	file first (or the specific line range if one was given). Never rely on file \
175	content you saw in a previous turn — the user may have reverted, edited, or \
176	changed the file externally since then. Always re-read to get the current state \
177	before constructing old_text. \
178	When the user corrects a previous edit (e.g. \"don't remove X, append instead\"), \
179	treat it as a fresh task: re-read the file, identify the current content, and \
180	plan the edit from scratch. Do not assume the file still reflects your last edit.
181
182	Tool-use heuristics:
183	- when the user provides a URL or asks about a web page, ALWAYS call \
184	read_website — never refuse or claim you lack internet access
185	- prefer absolute paths over relative paths when you mention, return, or pass \
186	file and directory paths
187	- if a path does not exist yet, create the directory before creating files in it
188	- if the user asks to clone a repo, immediately call run_command with git clone \
189	and an absolute destination path — do not ask where to put it unless the \
190	request is ambiguous; default to the user's home Repositories directory
191	- if the user asks for a new repo, scaffold, or scratch project, create the \
192	directory, create the first files, and run `git init` without waiting unless \
193	the request says otherwise
194	- if the repo looks like smbCloud CLI code, respect workspace crate boundaries, \
195	shared models, and existing command handlers before adding new abstractions
196	- if the repo looks like smbCloud Rails code, check routes, controllers, \
197	validators, and model boundaries before changing business logic
198	- if the task touches smbCloud auth, first decide whether it is a platform-user \
199	flow or a tenant-app flow, then follow the right endpoint family and model layer
200	- if the task touches smbCloud deploy code, check whether it is the generic \
201	deploy path or the Next.js SSR path before proposing changes
202	- after edits, prefer running the smallest useful verification step first, then \
203	widen to broader checks if needed
204	- use git commands naturally for status checks, repo setup, diffs, and normal \
205	developer workflows when they help move the task forward
206	- if a tool call fails, read the error, try to fix it, and retry — do not \
207	fall back to telling the user what to type
208
209	When the repo is not about smbCloud, act like a normal coding agent and do not \
210	force smbCloud-specific advice into the answer. When it is about smbCloud, be \
211	specific and practical.
212
213	Be direct and brief. Write clean, idiomatic code. When debugging, go for the \
214	root cause, not the symptom. Correct beats clever.";
215
216	/// shorter prompt for models without tool calling (e.g. DeepSeek Coder v1).
217	/// the full [`SYSTEM_PROMPT`] wastes context and confuses them.
218	const SIMPLE_SYSTEM_PROMPT: &str = "\
219	Your name is siGit — a coding assistant. \
220	You are helpful, concise, and write clean, idiomatic code. \
221	Answer any question the user asks — programming, general knowledge, or casual chat. \
222	When debugging, address the root cause, not the symptom. \
223	Be direct and brief.";
224
225	pub(crate) fn system_prompt_for_model(tool_calling: bool) -> &'static str {
226	if tool_calling {
227	SYSTEM_PROMPT
228	} else {
229	SIMPLE_SYSTEM_PROMPT
230	}
231	}
232
233	/// cap tool-call loops so a confused model can't spin forever; auto-compaction
234	/// (see [`backend::DEFAULT_CONTEXT_TOKEN_BUDGET`]) keeps long runs inside the
235	/// context window, so the cap can afford to be generous
236	const MAX_TOOL_ROUNDS: usize = 24;
237
238	/// Outcome of asking the client for permission to run one tool call.
239	enum PermissionVerdict {
240	/// Run the tool.
241	Approved,
242	/// Skip the tool; the string becomes its tool result so the model adapts.
243	Denied(String),
244	/// The client cancelled the turn while the request was pending; stop the
245	/// whole prompt with `StopReason::Cancelled` instead of burning rounds.
246	TurnCancelled,
247	}
248
249	/// Display kind for the permission dialog, so editors can show a fitting icon.
250	fn tool_kind_for(tool_name: &str) -> ToolKind {
251	match tool_name {
252	"edit_file" \| "multi_edit" \| "create_file" \| "create_directory" \| "remember" => {
253	ToolKind::Edit
254	}
255	"delete_file" => ToolKind::Delete,
256	"run_command" \| "kill_command" => ToolKind::Execute,
257	"read_file" \| "list_directory" \| "command_output" => ToolKind::Read,
258	"search_files" \| "glob" => ToolKind::Search,
259	"read_website" => ToolKind::Fetch,
260	"write_todos" => ToolKind::Think,
261	_ => ToolKind::Other,
262	}
263	}
264
265	/// Shown when a siGit Code Cloud tier is selected without a signed-in account.
266	const CLOUD_LOGIN_PROMPT: &str = "siGit Code Cloud needs an account. Sign in with \
267	`/login <email> <password>` (or the Authenticate button), then pick the tier again. \
268	Create an account at https://sigit.si.";
269
270	/// The per-session context system message: cwd guidance plus any project
271	/// instruction files (`AGENTS.md` / `CLAUDE.md`) found for that directory. Used
272	/// by every session entry point so on-device and cloud backends get the same
273	/// always-on project context.
274	fn session_context_message(cwd: &std::path::Path) -> String {
275	let mut message = format!(
276	"The user's project working directory is {}. \
277	Always use absolute paths under this directory for all file \
278	and directory operations. This is the root of the project \
279	the user has open in their editor.",
280	cwd.display()
281	);
282	if let Some(project_instructions) = instructions::load_project_instructions(cwd) {
283	message.push_str("\n\n");
284	message.push_str(&project_instructions);
285	}
286	message
287	}
288
289	fn agent_tools_as_specs() -> Vec<ToolSpec> {
290	let mut specs: Vec<ToolSpec> = tools::all_tools()
291	.into_iter()
292	.map(\|t\| ToolSpec {
293	name: t.name.to_string(),
294	description: t.description.to_string(),
295	parameters_schema: t.parameters_schema.to_string(),
296	})
297	.collect();
298
299	// Advertise the `skill` tool only when skills are present, so models without
300	// any skills installed don't see a dangling capability (Agent Skills format,
301	// https://agentskills.io). Discovery metadata lives in the tool description.
302	let discovered = skills::discover_skills();
303	if !discovered.is_empty() {
304	specs.push(ToolSpec {
305	name: skills::SKILL_TOOL_NAME.to_string(),
306	description: skills::skill_tool_description(&discovered),
307	parameters_schema: skills::skill_tool_schema().to_string(),
308	});
309	}
310
311	// Delegated research (`task`) is offered only when a subagent backend can
312	// actually be built — same conditional pattern as the `skill` tool above.
313	if tools::subagent_available() {
314	specs.push(tools::task_tool_spec());
315	}
316
317	// Tools discovered from configured MCP servers (incl. the official one).
318	specs.extend(mcp::tool_specs());
319
320	specs
321	}
322
323	/// Register the `task` tool's subagent factory for an OpenAI-compatible
324	/// provider: each subagent run gets a FRESH `OpenAiBackend` against the same
325	/// endpoint (its own conversation history), seeded with the subagent system
326	/// prompt. The provider config is captured by clone. Called once at startup by
327	/// whichever surface resolved the provider; on-device paths register a factory
328	/// that returns `None` instead (onde has a single shared history, so a second
329	/// concurrent context is not possible yet).
330	fn register_subagent_factory_for(cfg: &provider::ProviderConfig) {
331	let cfg = cfg.clone();
332	tools::set_subagent_factory(Box::new(move \|\| {
333	Some(Arc::new(OpenAiBackend::new(
334	cfg.base_url.clone(),
335	cfg.api_key.clone(),
336	cfg.model.clone(),
337	Some(tools::SUBAGENT_SYSTEM_PROMPT.to_string()),
338	)) as Arc<dyn InferenceBackend>)
339	}));
340	}
341
342	fn initialize_meta() -> Meta {
343	let startup_selection = setup::startup_model_selection();
344
345	let active_model_name = startup_selection
346	.as_ref()
347	.map(\|selection\| selection.display_name.clone())
348	.unwrap_or_else(\|\| GgufModelConfig::qwen25_3b().display_name);
349
350	let active_model_id = startup_selection
351	.as_ref()
352	.and_then(\|selection\| selection.selected_model.as_ref())
353	.map(\|selected\| selected.model_id.clone())
354	.unwrap_or_else(\|\| GgufModelConfig::qwen25_3b().model_id);
355
356	let active_model_file = startup_selection
357	.as_ref()
358	.and_then(\|selection\| selection.selected_model.as_ref())
359	.map(\|selected\| selected.gguf_file.clone())
360	.unwrap_or_else(\|\| {
361	GgufModelConfig::qwen25_3b()
362	.files
363	.first()
364	.cloned()
365	.unwrap_or_default()
366	});
367
368	let mut model = serde_json::Map::new();
369	model.insert(
370	"display_name".to_string(),
371	serde_json::Value::String(active_model_name),
372	);
373	model.insert(
374	"model_id".to_string(),
375	serde_json::Value::String(active_model_id),
376	);
377	model.insert(
378	"gguf_file".to_string(),
379	serde_json::Value::String(active_model_file),
380	);
381
382	let mut sigit = serde_json::Map::new();
383	sigit.insert("active_model".to_string(), serde_json::Value::Object(model));
384
385	let mut meta = Meta::new();
386	meta.insert("sigit".to_string(), serde_json::Value::Object(sigit));
387	meta
388	}
389
390	struct SiGitAgent {
391	engine: Arc<ChatEngine>,
392	/// The active inference backend. `LocalBackend` by default; swapped to an
393	/// `OpenAiBackend` when the user selects a siGit Code Cloud tier in the panel.
394	backend: tokio::sync::Mutex<Arc<dyn InferenceBackend>>,
395	/// cwd from the editor — tool calls run here, not where the process started
396	session_cwd: std::sync::Mutex<Option<PathBuf>>,
397	current_model: std::sync::Mutex<GgufModelConfig>,
398	/// flipped once the startup model finishes (success or failure)
399	model_ready: Arc<AtomicBool>,
400	/// guards the one-time lazy startup load for ACP mode
401	startup_model_load_started: Arc<AtomicBool>,
402	/// set if the startup load failed
403	model_load_error: Arc<std::sync::Mutex<Option<String>>>,
404	/// true when the startup model isn't cached yet
405	startup_needs_download: bool,
406	/// for progress UI
407	startup_model_name: String,
408	/// for download-progress polling
409	startup_model_id: String,
410	/// Serializes turn-affecting handlers (prompt, session lifecycle, config
411	/// changes). They run in `cx.spawn`ed tasks so the JSON-RPC dispatch loop
412	/// stays free to route client responses (e.g. permission answers) mid-turn;
413	/// this lock reproduces the strict ordering the dispatch loop used to give
414	/// them for free.
415	turn_lock: Arc<tokio::sync::Mutex<()>>,
416	}
417
418	impl SiGitAgent {
419	fn new(
420	engine: Arc<ChatEngine>,
421	initial_model: GgufModelConfig,
422	model_ready: Arc<AtomicBool>,
423	startup_model_load_started: Arc<AtomicBool>,
424	model_load_error: Arc<std::sync::Mutex<Option<String>>>,
425	startup_needs_download: bool,
426	) -> Self {
427	let startup_model_name = initial_model.display_name.clone();
428	let startup_model_id = initial_model.model_id.clone();
429	let backend: Arc<dyn InferenceBackend> = Arc::new(LocalBackend::new(Arc::clone(&engine)));
430	Self {
431	engine,
432	backend: tokio::sync::Mutex::new(backend),
433	session_cwd: std::sync::Mutex::new(None),
434	current_model: std::sync::Mutex::new(initial_model),
435	model_ready,
436	startup_model_load_started,
437	model_load_error,
438	startup_needs_download,
439	startup_model_name,
440	startup_model_id,
441	turn_lock: Arc::new(tokio::sync::Mutex::new(())),
442	}
443	}
444
445	fn start_startup_model_load_if_needed(&self) {
446	if self
447	.startup_model_load_started
448	.compare_exchange(false, true, Ordering::AcqRel, Ordering::Acquire)
449	.is_err()
450	{
451	return;
452	}
453
454	self.model_ready.store(false, Ordering::Release);
455	if let Ok(mut guard) = self.model_load_error.lock() {
456	*guard = None;
457	}
458
459	let startup_config = self.current_model.lock().unwrap().clone();
460	let (max_tokens, tool_calling) = models::local_picker_items()
461	.into_iter()
462	.find(\|item\| {
463	item.config.model_id == startup_config.model_id
464	&& item
465	.config
466	.files
467	.first()
468	.zip(startup_config.files.first())
469	.map(\|(left, right)\| left == right)
470	.unwrap_or(false)
471	})
472	.map(\|item\| (item.max_tokens, item.tool_calling))
473	.unwrap_or((4096, false));
474
475	let sampling = SamplingConfig {
476	max_tokens: Some(max_tokens),
477	..SamplingConfig::default()
478	};
479
480	let loader_engine = Arc::clone(&self.engine);
481	let loader_system_prompt = system_prompt_for_model(tool_calling).to_string();
482	let model_ready = Arc::clone(&self.model_ready);
483	let model_load_error = Arc::clone(&self.model_load_error);
484
485	std::thread::spawn(move \|\| {
486	let result = tokio::runtime::Runtime::new()
487	.map_err(\|error\| error.to_string())
488	.and_then(\|rt\| {
489	rt.block_on(loader_engine.load_gguf_model(
490	startup_config,
491	Some(loader_system_prompt),
492	Some(sampling),
493	))
494	.map(\|_\| ())
495	.map_err(\|error\| error.to_string())
496	});
497
498	if let Ok(mut guard) = model_load_error.lock() {
499	*guard = result.err();
500	}
501	model_ready.store(true, Ordering::Release);
502	});
503	}
504
505	/// block until the startup model is ready, showing progress in the session.
506	async fn await_model_ready(
507	&self,
508	cx: &ConnectionTo<Client>,
509	session_id: &SessionId,
510	) -> agent_client_protocol::Result<()> {
511	if self.model_ready.load(Ordering::Acquire) {
512	// already done — might be a stored error from earlier
513	if let Some(err) = self.model_load_error.lock().unwrap().as_ref() {
514	return Err(agent_client_protocol::Error::new(
515	-32603,
516	format!("model load failed: {err}"),
517	));
518	}
519	return Ok(());
520	}
521
522	const SPINNER: &[char] = &['⠋', '⠙', '⠹', '⠸', '⠼', '⠴', '⠦', '⠧', '⠇', '⠏'];
523
524	let tool_call_id = format!("startup-load-{}", uuid::Uuid::new_v4());
525	let title = if self.startup_needs_download {
526	format!("Downloading {}", self.startup_model_name)
527	} else {
528	format!("Loading {}", self.startup_model_name)
529	};
530
531	self.send_tool_call_update(
532	cx,
533	session_id.clone(),
534	SessionUpdate::ToolCall(
535	ToolCall::new(tool_call_id.clone(), &title)
536	.kind(ToolKind::Think)
537	.status(ToolCallStatus::InProgress)
538	.content(vec![format!("{}…", title).into()]),
539	),
540	)
541	.ok();
542
543	let expected_bytes = if self.startup_needs_download {
544	onde::inference::models::SUPPORTED_MODEL_INFO
545	.iter()
546	.find(\|m\| m.id == self.startup_model_id)
547	.map(\|m\| m.expected_size_bytes)
548	.unwrap_or(0)
549	} else {
550	0
551	};
552
553	let load_start = std::time::Instant::now();
554	let mut tick: usize = 0;
555	let mut interval = tokio::time::interval(std::time::Duration::from_secs(1));
556	interval.tick().await;
557
558	loop {
559	interval.tick().await;
560	tick += 1;
561
562	if self.model_ready.load(Ordering::Acquire) {
563	break;
564	}
565
566	let frame = SPINNER[tick % SPINNER.len()];
567	let elapsed = load_start.elapsed();
568	let elapsed_str = if elapsed.as_secs() >= 60 {
569	format!("{}m {:02}s", elapsed.as_secs() / 60, elapsed.as_secs() % 60)
570	} else {
571	format!("{}s", elapsed.as_secs())
572	};
573
574	let (update_title, update_content) =
575	if self.startup_needs_download && expected_bytes > 0 {
576	let cache_path = onde::hf_cache::model_cache_path(&self.startup_model_id);
577	let downloaded = cache_path
578	.as_ref()
579	.filter(\|p\| p.exists())
580	.map(\|p\| dir_size_recursive(p))
581	.unwrap_or(0);
582	let pct = ((downloaded as f64 / expected_bytes as f64) * 100.0).min(99.0) as u8;
583	let bar = progress_bar(pct, 20);
584	let size_hint = format!(" (~{})", format_size_human(expected_bytes));
585	(
586	format!(
587	"{frame} Downloading {}{size_hint} ({pct}%)",
588	self.startup_model_name
589	),
590	format!(
591	"{} — {bar} {pct}% ({} / {})",
592	self.startup_model_name,
593	format_size_human(downloaded),
594	format_size_human(expected_bytes),
595	),
596	)
597	} else if self.startup_needs_download {
598	let cache_path = onde::hf_cache::model_cache_path(&self.startup_model_id);
599	let downloaded = cache_path
600	.as_ref()
601	.filter(\|p\| p.exists())
602	.map(\|p\| dir_size_recursive(p))
603	.unwrap_or(0);
604	(
605	format!("{frame} Downloading {}", self.startup_model_name),
606	format!(
607	"{} — {} downloaded… ({elapsed_str})",
608	self.startup_model_name,
609	format_size_human(downloaded),
610	),
611	)
612	} else {
613	(
614	format!("{frame} Loading {}", self.startup_model_name),
615	format!(
616	"{frame} Loading {}… ({elapsed_str})",
617	self.startup_model_name
618	),
619	)
620	};
621
622	self.send_tool_call_update(
623	cx,
624	session_id.clone(),
625	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
626	tool_call_id.clone(),
627	ToolCallUpdateFields::new()
628	.title(update_title)
629	.status(ToolCallStatus::InProgress)
630	.content(vec![update_content.into()]),
631	)),
632	)
633	.ok();
634	}
635
636	// done — check if it blew up
637	let load_error = self.model_load_error.lock().unwrap().clone();
638	if let Some(err) = load_error {
639	self.send_tool_call_update(
640	cx,
641	session_id.clone(),
642	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
643	tool_call_id,
644	ToolCallUpdateFields::new()
645	.title("Model load failed".to_string())
646	.status(ToolCallStatus::Failed)
647	.content(vec![format!("error: {err}").into()]),
648	)),
649	)
650	.ok();
651
652	return Err(agent_client_protocol::Error::new(
653	-32603,
654	format!("model load failed: {err}"),
655	));
656	}
657
658	let done_title = if self.startup_needs_download {
659	format!("✓ {} downloaded and loaded", self.startup_model_name)
660	} else {
661	format!("✓ {} loaded", self.startup_model_name)
662	};
663
664	self.send_tool_call_update(
665	cx,
666	session_id.clone(),
667	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
668	tool_call_id,
669	ToolCallUpdateFields::new()
670	.title(done_title)
671	.status(ToolCallStatus::Completed),
672	)),
673	)
674	.ok();
675
676	Ok(())
677	}
678
679	fn send_assistant_message(
680	&self,
681	cx: &ConnectionTo<Client>,
682	session_id: SessionId,
683	text: impl Into<String>,
684	) -> agent_client_protocol::Result<()> {
685	cx.send_notification(SessionNotification::new(
686	session_id,
687	SessionUpdate::AgentMessageChunk(ContentChunk::new(ContentBlock::from(text.into()))),
688	))
689	}
690
691	/// Run one inference turn (`fut`) while concurrently forwarding any streamed
692	/// tokens to the editor. The sink receiver is drained as the future runs, so
693	/// chunks reach the client live rather than all at once when it resolves.
694	///
695	/// `assembled`/`sent`/`streamed_any` persist across the turns of a single
696	/// prompt so reasoning is stripped consistently and we never re-send text.
697	#[allow(clippy::too_many_arguments)]
698	async fn drain_turn<F>(
699	&self,
700	cx: &ConnectionTo<Client>,
701	session_id: &SessionId,
702	fut: F,
703	sink_rx: &mut tokio::sync::mpsc::UnboundedReceiver<String>,
704	assembled: &mut String,
705	sent: &mut String,
706	streamed_any: &mut bool,
707	) -> Result<TurnResult, backend::BackendError>
708	where
709	F: std::future::Future<Output = Result<TurnResult, backend::BackendError>>,
710	{
711	tokio::pin!(fut);
712	let result = loop {
713	tokio::select! {
714	done = &mut fut => break done,
715	Some(piece) = sink_rx.recv() => {
716	self.emit_visible_chunk(cx, session_id, &piece, assembled, sent, streamed_any);
717	}
718	}
719	};
720	// Flush tokens that landed between the last poll and the future resolving.
721	while let Ok(piece) = sink_rx.try_recv() {
722	self.emit_visible_chunk(cx, session_id, &piece, assembled, sent, streamed_any);
723	}
724	result
725	}
726
727	/// Append a streamed fragment, strip `<think>` reasoning from the running
728	/// text, and send only the newly revealed visible suffix as a chunk. Tracking
729	/// the assembled text (not just deltas) keeps think-block stripping correct
730	/// even when a tag spans chunk boundaries.
731	fn emit_visible_chunk(
732	&self,
733	cx: &ConnectionTo<Client>,
734	session_id: &SessionId,
735	piece: &str,
736	assembled: &mut String,
737	sent: &mut String,
738	streamed_any: &mut bool,
739	) {
740	assembled.push_str(piece);
741	let (_think, visible) = chat::strip_think_blocks(assembled);
742	match visible.strip_prefix(sent.as_str()) {
743	Some(extra) if !extra.is_empty() => {
744	let extra = extra.to_string();
745	*sent = visible;
746	*streamed_any = true;
747	self.send_assistant_message(cx, session_id.clone(), extra)
748	.ok();
749	}
750	// No new visible text, or the visible prefix changed retroactively
751	// (rare, e.g. a late-closing think tag): just resync without
752	// resending what's already on the wire.
753	_ => *sent = visible,
754	}
755	}
756
757	fn send_tool_call_update(
758	&self,
759	cx: &ConnectionTo<Client>,
760	session_id: SessionId,
761	update: SessionUpdate,
762	) -> agent_client_protocol::Result<()> {
763	cx.send_notification(SessionNotification::new(session_id, update))
764	}
765
766	/// Advertise siGit's slash commands to the client. Editors like Zed parse
767	/// `/`-prefixed input and only forward commands they've been told about, so
768	/// without this `/login`, `/models`, etc. are rejected client-side.
769	fn advertise_commands(&self, cx: &ConnectionTo<Client>, session_id: SessionId) {
770	let with_hint = \|name: &str, desc: &str, hint: &str\| {
771	AvailableCommand::new(name, desc).input(AvailableCommandInput::Unstructured(
772	UnstructuredCommandInput::new(hint),
773	))
774	};
775	let commands = vec![
776	AvailableCommand::new("help", "Show available commands"),
777	AvailableCommand::new("models", "List available models").input(
778	AvailableCommandInput::Unstructured(UnstructuredCommandInput::new(
779	"model number to switch to (optional)",
780	)),
781	),
782	with_hint(
783	"local",
784	"Toggle on-device inference mode",
785	"on\|off (optional)",
786	),
787	AvailableCommand::new("skills", "List available Agent Skills"),
788	AvailableCommand::new("mcp", "List MCP servers and their tools"),
789	AvailableCommand::new("load", "Load the selected on-device model"),
790	with_hint("login", "Sign in to siGit Code Cloud", "<email> <password>"),
791	AvailableCommand::new("logout", "Sign out of siGit Code Cloud"),
792	AvailableCommand::new("whoami", "Show the signed-in account"),
793	AvailableCommand::new("reload", "Re-sync sign-in and model state"),
794	with_hint(
795	"plan",
796	"Plan mode: research only, no edits or commands",
797	"on\|off (optional)",
798	),
799	AvailableCommand::new("permissions", "Show the tool permission policy"),
800	AvailableCommand::new("compact", "Summarize and shrink the conversation history"),
801	AvailableCommand::new("clear", "Wipe the conversation history"),
802	AvailableCommand::new("status", "Show engine status"),
803	];
804	self.send_tool_call_update(
805	cx,
806	session_id,
807	SessionUpdate::AvailableCommandsUpdate(AvailableCommandsUpdate::new(commands)),
808	)
809	.ok();
810	}
811
812	async fn switch_model_by_id(
813	&self,
814	model_id: &str,
815	) -> agent_client_protocol::Result<GgufModelConfig> {
816	let (new_config, max_tokens, new_tool_calling) = resolve_model_config(model_id)
817	.ok_or_else(\|\| {
818	agent_client_protocol::Error::new(
819	-32602,
820	format!("unknown or unavailable model: {model_id}"),
821	)
822	})?;
823
824	log::info!(
825	"switching model to {} (max_tokens={max_tokens})",
826	new_config.display_name
827	);
828
829	let sampling = SamplingConfig {
830	max_tokens: Some(max_tokens),
831	..SamplingConfig::default()
832	};
833
834	// block_in_place inside spawn_local panics, so run the load on a
835	// dedicated thread with its own runtime (same trick as startup)
836	let (result_tx, result_rx) = tokio::sync::oneshot::channel::<Result<(), String>>();
837	let loader_engine = Arc::clone(&self.engine);
838	let loader_config = new_config.clone();
839	let loader_system_prompt = system_prompt_for_model(new_tool_calling).to_string();
840	let loader_sampling = sampling;
841
842	std::thread::spawn(move \|\| {
843	let rt = tokio::runtime::Runtime::new().expect("failed to create loader runtime");
844	let result = rt.block_on(async move {
845	// load_gguf_model already unloads the old model internally;
846	// calling unload first would leave a gap where prompts fail
847	loader_engine
848	.load_gguf_model(
849	loader_config,
850	Some(loader_system_prompt),
851	Some(loader_sampling),
852	)
853	.await
854	});
855	let _ = result_tx.send(result.map(\|_\| ()).map_err(\|e\| e.to_string()));
856	});
857
858	result_rx
859	.await
860	.map_err(\|_\| agent_client_protocol::Error::new(-32603, "model loader thread crashed"))?
861	.map_err(\|error\| {
862	log::error!("model switch failed: {error}");
863	agent_client_protocol::Error::new(-32603, format!("model switch failed: {error}"))
864	})?;
865
866	self.startup_model_load_started
867	.store(true, Ordering::Release);
868	self.model_ready.store(true, Ordering::Release);
869	if let Ok(mut guard) = self.model_load_error.lock() {
870	*guard = None;
871	}
872
873	if let Some(item) = models::local_picker_items()
874	.iter()
875	.find(\|item\| item.config.model_id == new_config.model_id)
876	&& let Err(err) = setup::save_selected_model(&setup::SelectedModel {
877	model_id: item.config.model_id.clone(),
878	gguf_file: item.config.files.first().cloned().unwrap_or_default(),
879	})
880	{
881	log::warn!("failed to persist model selection: {err}");
882	}
883
884	{
885	let mut guard = self.current_model.lock().unwrap();
886	*guard = new_config.clone();
887	}
888
889	if let Some(cwd) = self.session_cwd.lock().ok().and_then(\|g\| g.clone()) {
890	self.engine
891	.push_history(onde::inference::ChatMessage::system(
892	session_context_message(&cwd),
893	))
894	.await;
895	}
896
897	Ok(new_config)
898	}
899	}
900
901	// ── ACP handler implementations ───────────────────────────────────────────────
902
903	impl SiGitAgent {
904	async fn handle_initialize(
905	&self,
906	_req: InitializeRequest,
907	) -> agent_client_protocol::Result<InitializeResponse> {
908	log::info!("initialize");
909
910	// Agent-handled auth method. We don't use `AuthMethod::Terminal`: editors
911	// like Zed advertise terminal-auth capability but don't actually spawn the
912	// login terminal for custom ACP agents, so the button is a silent no-op.
913	// With an Agent method, clicking calls `authenticate`, which returns either
914	// confirmation (already signed in via `/login`) or a message telling the
915	// user to run `/login <email> <password>` — so the button does something.
916	let auth_methods = vec![AuthMethod::Agent(
917	AuthMethodAgent::new("sigit", "Sign in to siGit Code")
918	.description("Sign in with `/login <email> <password>` in the message box."),
919	)];
920
921	Ok(InitializeResponse::new(ProtocolVersion::V1)
922	.agent_info(
923	Implementation::new("sigit", env!("CARGO_PKG_VERSION"))
924	.title("siGit Code - AI Coding Agent"),
925	)
926	.auth_methods(auth_methods)
927	.agent_capabilities(
928	AgentCapabilities::default()
929	.load_session(true)
930	.session_capabilities(
931	SessionCapabilities::new().fork(SessionForkCapabilities::new()),
932	),
933	)
934	.meta(initialize_meta()))
935	}
936
937	async fn handle_authenticate(
938	&self,
939	req: AuthenticateRequest,
940	) -> agent_client_protocol::Result<AuthenticateResponse> {
941	log::info!("authenticate: method={}", req.method_id.0);
942
943	// Confirm the stored token works. The button can't collect a password,
944	// so an unsigned-in user is pointed at the `/login` slash command; a user
945	// already signed in via `/login` gets the gate cleared.
946	match account::verify_session().await {
947	Ok(email) => {
948	log::info!("authenticate: verified session for {email}");
949	Ok(AuthenticateResponse::default())
950	}
951	Err(reason) => Err(agent_client_protocol::Error::new(
952	-32000,
953	format!(
954	"Not signed in to siGit Code Cloud ({reason}). \
955	Sign in with `/login <email> <password>` in the message box, \
956	or create an account at https://sigit.si."
957	),
958	)),
959	}
960	}
961
962	async fn handle_load_session(
963	&self,
964	cx: &ConnectionTo<Client>,
965	args: LoadSessionRequest,
966	) -> agent_client_protocol::Result<LoadSessionResponse> {
967	log::info!(
968	"load_session: id={}, cwd={}, additional_directories={:?}",
969	args.session_id,
970	args.cwd.display(),
971	args.additional_directories
972	.iter()
973	.map(\|p\| p.display().to_string())
974	.collect::<Vec<_>>()
975	);
976
977	if let Ok(mut guard) = self.session_cwd.lock() {
978	*guard = Some(args.cwd.clone());
979	}
980
981	// A session boundary: grants and plan mode from the previous life of
982	// this session id must not carry over — and since one shared engine
983	// means one live conversation, state for every other id is dead too.
984	permissions::reset_all();
985
986	// tool calls use relative paths, so we need to match the editor's cwd
987	if args.cwd.is_dir()
988	&& let Err(err) = std::env::set_current_dir(&args.cwd)
989	{
990	log::warn!("could not set cwd to {}: {err}", args.cwd.display());
991	}
992
993	// start from a clean slate; a stored session (below) replaces it
994	self.engine.clear_history().await;
995
996	self.engine
997	.push_history(onde::inference::ChatMessage::system(
998	session_context_message(&args.cwd),
999	))
1000	.await;
1001
1002	// Honor the persisted Local Inference toggle (off + signed in → cloud).
1003	self.apply_startup_inference_mode().await;
1004
1005	// Durable sessions: when this session id was saved before, restore its
1006	// history into the active backend. The snapshot includes the system
1007	// messages that were live when it was saved, so restore replaces the
1008	// freshly seeded state wholesale.
1009	if let Some(history) = session_store::load(&args.session_id.to_string()) {
1010	let restored = history.len();
1011	let backend = self.backend.lock().await.clone();
1012	backend.restore_history(history).await;
1013	log::info!(
1014	"load_session: restored {restored} message(s) for {}",
1015	args.session_id
1016	);
1017	}
1018
1019	let config_options = {
1020	let guard = self.current_model.lock().unwrap();
1021	build_model_config_options(&guard)
1022	};
1023
1024	self.advertise_commands(cx, args.session_id.clone());
1025
1026	Ok(LoadSessionResponse::new().config_options(config_options))
1027	}
1028
1029	async fn handle_fork_session(
1030	&self,
1031	cx: &ConnectionTo<Client>,
1032	args: ForkSessionRequest,
1033	) -> agent_client_protocol::Result<ForkSessionResponse> {
1034	let new_id = SessionId::new(uuid::Uuid::new_v4().to_string());
1035	// Session boundary: permission grants and plan mode never cross it
1036	// (see handle_load_session), so a fork starts with a clean slate.
1037	permissions::reset_all();
1038	log::info!(
1039	"fork_session: from={} new={new_id}, cwd={}, additional_directories={:?}",
1040	args.session_id,
1041	args.cwd.display(),
1042	args.additional_directories
1043	.iter()
1044	.map(\|p\| p.display().to_string())
1045	.collect::<Vec<_>>()
1046	);
1047
1048	if let Ok(mut guard) = self.session_cwd.lock() {
1049	*guard = Some(args.cwd.clone());
1050	}
1051	if args.cwd.is_dir()
1052	&& let Err(err) = std::env::set_current_dir(&args.cwd)
1053	{
1054	log::warn!("could not set cwd to {}: {err}", args.cwd.display());
1055	}
1056
1057	// no persistence, so fork == fresh session
1058	self.engine.clear_history().await;
1059
1060	self.engine
1061	.push_history(onde::inference::ChatMessage::system(
1062	session_context_message(&args.cwd),
1063	))
1064	.await;
1065
1066	// Honor the persisted Local Inference toggle (off + signed in → cloud).
1067	self.apply_startup_inference_mode().await;
1068
1069	let config_options = {
1070	let guard = self.current_model.lock().unwrap();
1071	build_model_config_options(&guard)
1072	};
1073
1074	self.advertise_commands(cx, new_id.clone());
1075
1076	Ok(ForkSessionResponse::new(new_id).config_options(config_options))
1077	}
1078
1079	async fn handle_new_session(
1080	&self,
1081	cx: &ConnectionTo<Client>,
1082	args: NewSessionRequest,
1083	) -> agent_client_protocol::Result<NewSessionResponse> {
1084	let session_id = SessionId::new(uuid::Uuid::new_v4().to_string());
1085	// Session boundary: permission grants and plan mode never cross it
1086	// (see handle_load_session), so stale ids stop accumulating state.
1087	permissions::reset_all();
1088	log::info!(
1089	"new_session: id={session_id}, cwd={}, additional_directories={:?}",
1090	args.cwd.display(),
1091	args.additional_directories
1092	.iter()
1093	.map(\|p\| p.display().to_string())
1094	.collect::<Vec<_>>()
1095	);
1096
1097	if let Ok(mut guard) = self.session_cwd.lock() {
1098	*guard = Some(args.cwd.clone());
1099	}
1100	if args.cwd.is_dir()
1101	&& let Err(err) = std::env::set_current_dir(&args.cwd)
1102	{
1103	log::warn!("could not set cwd to {}: {err}", args.cwd.display());
1104	}
1105
1106	self.engine.clear_history().await;
1107
1108	self.engine
1109	.push_history(onde::inference::ChatMessage::system(
1110	session_context_message(&args.cwd),
1111	))
1112	.await;
1113
1114	// Honor the persisted Local Inference toggle (off + signed in → cloud).
1115	self.apply_startup_inference_mode().await;
1116
1117	let config_options = {
1118	let guard = self.current_model.lock().unwrap();
1119	build_model_config_options(&guard)
1120	};
1121
1122	self.advertise_commands(cx, session_id.clone());
1123
1124	Ok(NewSessionResponse::new(session_id).config_options(config_options))
1125	}
1126
1127	async fn handle_prompt(
1128	&self,
1129	cx: &ConnectionTo<Client>,
1130	args: PromptRequest,
1131	) -> agent_client_protocol::Result<PromptResponse> {
1132	let session_id = args.session_id.clone();
1133
1134	// log every block so we can debug @ references and file context
1135	for (i, block) in args.prompt.iter().enumerate() {
1136	match block {
1137	ContentBlock::Text(t) => {
1138	log::info!(
1139	"prompt({}) block[{}]: Text({} chars) = \"{}\"",
1140	session_id,
1141	i,
1142	t.text.len(),
1143	t.text.chars().take(200).collect::<String>()
1144	);
1145	}
1146	ContentBlock::Resource(embedded) => {
1147	log::info!(
1148	"prompt({}) block[{}]: EmbeddedResource = {:?}",
1149	session_id,
1150	i,
1151	match &embedded.resource {
1152	EmbeddedResourceResource::TextResourceContents(t) =>
1153	format!("TextResource(uri={}, {} chars)", t.uri, t.text.len()),
1154	EmbeddedResourceResource::BlobResourceContents(b) =>
1155	format!("BlobResource(uri={})", b.uri),
1156	_ => "Unknown".to_string(),
1157	}
1158	);
1159	}
1160	ContentBlock::ResourceLink(link) => {
1161	log::info!(
1162	"prompt({}) block[{}]: ResourceLink(name={}, uri={}, title={:?}, desc={:?})",
1163	session_id,
1164	i,
1165	link.name,
1166	link.uri,
1167	link.title,
1168	link.description
1169	);
1170	}
1171	other => {
1172	log::info!(
1173	"prompt({}) block[{}]: Other({:?})",
1174	session_id,
1175	i,
1176	std::mem::discriminant(other)
1177	);
1178	}
1179	}
1180	}
1181
1182	let mut parts: Vec<String> = Vec::new();
1183
1184	for block in &args.prompt {
1185	match block {
1186	ContentBlock::Text(t) => {
1187	parts.push(t.text.clone());
1188	}
1189	ContentBlock::Resource(embedded) => {
1190	// editor inlined the file content already
1191	match &embedded.resource {
1192	EmbeddedResourceResource::TextResourceContents(text_resource) => {
1193	parts.push(format!(
1194	"\n--- {} ---\n{}\n--- end {} ---",
1195	text_resource.uri, text_resource.text, text_resource.uri
1196	));
1197	}
1198	EmbeddedResourceResource::BlobResourceContents(blob) => {
1199	parts.push(format!("[binary resource: {}]", blob.uri));
1200	}
1201	_ => {
1202	log::debug!("ignoring unsupported embedded resource variant");
1203	}
1204	}
1205	}
1206	ContentBlock::ResourceLink(link) => {
1207	// reference without content; read the file ourselves
1208	let label = link.name.clone();
1209
1210	if let Some(raw_path) = link.uri.strip_prefix("file://") {
1211	let (file_path, line_range) = if let Some(hash_pos) = raw_path.rfind('#') {
1212	let fragment = &raw_path[hash_pos + 1..];
1213	let path = &raw_path[..hash_pos];
1214	// Parse "L207:219" or "L207-219" → (207, 219)
1215	let range = fragment.strip_prefix('L').and_then(\|rest\| {
1216	let sep = if rest.contains(':') { ':' } else { '-' };
1217	let mut parts = rest.splitn(2, sep);
1218	let start = parts.next()?.parse::<usize>().ok()?;
1219	let end = parts.next()?.parse::<usize>().ok()?;
1220	Some((start, end))
1221	});
1222	(path, range)
1223	} else {
1224	(raw_path, None)
1225	};
1226
1227	match std::fs::read_to_string(file_path) {
1228	Ok(contents) => {
1229	let extracted = if let Some((start, end)) = line_range {
1230	let selected: Vec<&str> = contents
1231	.lines()
1232	.enumerate()
1233	.filter(\|(i, _)\| {
1234	let line_num = i + 1;
1235	line_num >= start && line_num <= end
1236	})
1237	.map(\|(_, line)\| line)
1238	.collect();
1239	format!(
1240	"\n--- {label} ({file_path} lines {start}-{end}) ---\n{}\n--- end {label} ---",
1241	selected.join("\n")
1242	)
1243	} else {
1244	format!(
1245	"\n--- {label} ({file_path}) ---\n{contents}\n--- end {label} ---"
1246	)
1247	};
1248	parts.push(extracted);
1249	}
1250	Err(err) => {
1251	log::warn!("could not read ResourceLink {}: {err}", link.uri);
1252	parts.push(format!("[referenced file: {label} ({file_path})]"));
1253	}
1254	}
1255	} else {
1256	parts.push(format!("[resource link: {label} ({})]", link.uri));
1257	}
1258	}
1259	_ => {
1260	log::debug!("ignoring unsupported content block type in prompt");
1261	}
1262	}
1263	}
1264
1265	let user_text = parts.join("\n");
1266
1267	if user_text.trim().is_empty() {
1268	return Ok(PromptResponse::new(StopReason::EndTurn));
1269	}
1270
1271	if let Some(command) = parse_slash(&user_text) {
1272	return exec_slash_acp(self, cx, session_id, command).await;
1273	}
1274
1275	log::info!(
1276	"prompt({}): \"{}\"",
1277	session_id,
1278	user_text.chars().take(80).collect::<String>()
1279	);
1280
1281	// The active backend drives the turn. Snapshot it once so a mid-turn
1282	// model switch doesn't split the conversation across backends.
1283	let backend = self.backend.lock().await.clone();
1284
1285	// Only on-device inference needs a local model in memory. Cloud tiers run
1286	// over the network, so they never need a local model. We never load the
1287	// on-device model implicitly: the user loads it explicitly with `/load`
1288	// (or by picking one in `/models`). If a prompt arrives before that, guide
1289	// them rather than blocking on a multi-minute download/load.
1290	if !backend.is_remote()
1291	&& self.engine.info().await.status == onde::inference::EngineStatus::Unloaded
1292	{
1293	self.send_assistant_message(
1294	cx,
1295	session_id,
1296	"No on-device model is loaded. Run `/load` to load the selected model, \
1297	or `/models` to choose one.",
1298	)
1299	.ok();
1300	return Ok(PromptResponse::new(StopReason::EndTurn));
1301	}
1302
1303	// ── tool-calling loop ────────────────────────────────────────────
1304	// send message → execute any tool calls → feed results back
1305	// repeat up to MAX_TOOL_ROUNDS, then force a text reply
1306
1307	let tools = agent_tools_as_specs();
1308
1309	// Token sink: backends stream assistant text through this while a turn
1310	// runs. We forward the visible portion to the editor as agent-message
1311	// chunks live (see `drain_turn` / `emit_visible_chunk`). The sink stays
1312	// alive for the whole prompt so `recv()` only ends when a turn future
1313	// resolves, never because every sender was dropped.
1314	let (sink, mut sink_rx) = tokio::sync::mpsc::unbounded_channel::<String>();
1315	let mut assembled = String::new();
1316	let mut sent = String::new();
1317	let mut streamed_any = false;
1318
1319	let mut result = self
1320	.drain_turn(
1321	cx,
1322	&session_id,
1323	backend.send_message_with_tools(&user_text, &tools, Some(&sink)),
1324	&mut sink_rx,
1325	&mut assembled,
1326	&mut sent,
1327	&mut streamed_any,
1328	)
1329	.await
1330	.map_err(\|error\| {
1331	log::error!("send_message_with_tools failed: {error}");
1332	agent_client_protocol::Error::new(-32603, format!("inference failed: {error}"))
1333	})?;
1334
1335	let mut round = 0;
1336
1337	while !result.tool_calls.is_empty() && round < MAX_TOOL_ROUNDS {
1338	round += 1;
1339	log::info!(
1340	"prompt({}) tool round {} — {} call(s)",
1341	session_id,
1342	round,
1343	result.tool_calls.len()
1344	);
1345
1346	// Auto-compaction: long tool runs grow history fast; fold it into
1347	// a summary before the next round rather than blowing the window.
1348	let estimate = backend::estimate_tokens(&backend.history_snapshot().await);
1349	if estimate > backend::DEFAULT_CONTEXT_TOKEN_BUDGET {
1350	log::info!(
1351	"prompt({}) history ≈{} tokens exceeds budget {} — compacting",
1352	session_id,
1353	estimate,
1354	backend::DEFAULT_CONTEXT_TOKEN_BUDGET
1355	);
1356	match backend.compact_history(backend::COMPACT_KEEP_LAST).await {
1357	Ok(()) => {
1358	let after = backend::estimate_tokens(&backend.history_snapshot().await);
1359	log::info!("prompt({}) compacted to ≈{} tokens", session_id, after);
1360	}
1361	Err(error) => log::warn!("prompt({}) compaction failed: {error}", session_id),
1362	}
1363	}
1364
1365	let mut tool_results = Vec::new();
1366
1367	for (call_index, tc) in result.tool_calls.iter().enumerate() {
1368	log::info!(
1369	" → {}({})",
1370	tc.name,
1371	tc.arguments.chars().take(120).collect::<String>()
1372	);
1373
1374	// Permission gate: read-only tools pass straight through; a
1375	// mutating tool consults policy and may ask the client.
1376	let output = match permissions::decision_for(
1377	&session_id.to_string(),
1378	&tc.name,
1379	&tc.arguments,
1380	) {
1381	permissions::Decision::Allow => {
1382	tools::execute_tool(&tc.name, &tc.arguments).await
1383	}
1384	permissions::Decision::Deny(reason) => {
1385	log::info!(" ✗ {} denied by policy", tc.name);
1386	reason
1387	}
1388	permissions::Decision::Ask => {
1389	match self
1390	.request_tool_permission(cx, &session_id, &tc.name, &tc.arguments)
1391	.await
1392	{
1393	PermissionVerdict::Approved => {
1394	tools::execute_tool(&tc.name, &tc.arguments).await
1395	}
1396	PermissionVerdict::Denied(reason) => {
1397	log::info!(" ✗ {} denied by user", tc.name);
1398	reason
1399	}
1400	PermissionVerdict::TurnCancelled => {
1401	log::info!("prompt({}) cancelled at permission gate", session_id);
1402	// The assistant message carrying these tool
1403	// calls is already in the backend history;
1404	// leaving any of them unanswered makes strict
1405	// OpenAI-compatible endpoints reject every
1406	// later request in the session. Close out this
1407	// call and the ones this round never reached.
1408	for pending in &result.tool_calls[call_index..] {
1409	tool_results.push(BackendToolResult {
1410	tool_call_id: pending.id.clone(),
1411	content: format!(
1412	"`{}` was not executed: the user cancelled the turn \
1413	at the permission prompt.",
1414	pending.name
1415	),
1416	});
1417	}
1418	backend.record_cancelled_tool_results(tool_results).await;
1419	return Ok(PromptResponse::new(StopReason::Cancelled));
1420	}
1421	}
1422	}
1423	};
1424
1425	log::info!(" ← {} chars", output.len());
1426
1427	tool_results.push(BackendToolResult {
1428	tool_call_id: tc.id.clone(),
1429	content: output,
1430	});
1431	}
1432
1433	let next_tools = if round < MAX_TOOL_ROUNDS {
1434	Some(tools.as_slice())
1435	} else {
1436	None // last round: force text
1437	};
1438
1439	result = self
1440	.drain_turn(
1441	cx,
1442	&session_id,
1443	backend.send_tool_results(tool_results, next_tools, Some(&sink)),
1444	&mut sink_rx,
1445	&mut assembled,
1446	&mut sent,
1447	&mut streamed_any,
1448	)
1449	.await
1450	.map_err(\|e\| agent_client_protocol::Error::new(-32603, e.to_string()))?;
1451	}
1452
1453	// ── Final text response ───────────────────────────────────────────
1454	// If anything streamed, the visible reply is already on the wire; only
1455	// send a trailing block for the non-streamed path (e.g. on-device direct
1456	// answers, which onde can't stream while tools are on offer).
1457	if !streamed_any {
1458	let reply_text = result.text.trim().to_string();
1459	let final_text = if reply_text.is_empty() {
1460	if round > 0 {
1461	log::warn!(
1462	"prompt({}) — model returned empty reply after {} tool round(s)",
1463	session_id,
1464	round
1465	);
1466	"Something went wrong — the edits didn't go through. Try rephrasing what you need, or point me at the specific lines.".to_string()
1467	} else {
1468	log::warn!(
1469	"prompt({}) — model returned empty reply (no tool rounds)",
1470	session_id
1471	);
1472	String::new()
1473	}
1474	} else {
1475	// strip <think> blocks so reasoning tokens stay hidden
1476	let (_think, visible) = chat::strip_think_blocks(&reply_text);
1477	visible
1478	};
1479
1480	if !final_text.is_empty() {
1481	self.send_assistant_message(cx, session_id.clone(), final_text)
1482	.ok();
1483	}
1484	}
1485
1486	// Persist the completed turn so a restart (or session/load) can pick
1487	// the conversation back up.
1488	let snapshot = backend.history_snapshot().await;
1489	if let Err(error) = session_store::save(&session_id.to_string(), &snapshot) {
1490	log::warn!("prompt({}) session save failed: {error}", session_id);
1491	}
1492
1493	log::info!("prompt({}) complete — {} tool round(s)", session_id, round);
1494	Ok(PromptResponse::new(StopReason::EndTurn))
1495	}
1496
1497	/// Ask the ACP client for permission to run one tool call. Presents
1498	/// allow-once / allow-for-session / deny; an "always allow" choice is
1499	/// recorded via [`permissions::grant_for_session`]. Only safe to call from
1500	/// a spawned task (see the handler registration in `run_acp_server`): the
1501	/// dispatch loop must be free to route the client's answer back to us.
1502	async fn request_tool_permission(
1503	&self,
1504	cx: &ConnectionTo<Client>,
1505	session_id: &SessionId,
1506	tool_name: &str,
1507	arguments: &str,
1508	) -> PermissionVerdict {
1509	// The user decides from this dialog, so show the arguments with any
1510	// truncation flagged (a silently clipped command could hide its tail
1511	// from the person approving it). The full arguments also travel as
1512	// `raw_input` for clients that render it.
1513	let args_preview = permissions::approval_preview(arguments);
1514	let title = if args_preview.is_empty() {
1515	tool_name.to_string()
1516	} else {
1517	format!("{tool_name}({args_preview})")
1518	};
1519	let raw_input: serde_json::Value = serde_json::from_str(arguments)
1520	.unwrap_or_else(\|_\| serde_json::Value::String(arguments.to_string()));
1521
1522	let request = RequestPermissionRequest::new(
1523	session_id.clone(),
1524	ToolCallUpdate::new(
1525	format!("perm-{}", uuid::Uuid::new_v4()),
1526	ToolCallUpdateFields::new()
1527	.title(title)
1528	.kind(tool_kind_for(tool_name))
1529	.status(ToolCallStatus::Pending)
1530	.raw_input(raw_input),
1531	),
1532	vec![
1533	PermissionOption::new("allow_once", "Allow once", PermissionOptionKind::AllowOnce),
1534	PermissionOption::new(
1535	"allow_session",
1536	"Allow for this session",
1537	PermissionOptionKind::AllowAlways,
1538	),
1539	PermissionOption::new("reject_once", "Deny", PermissionOptionKind::RejectOnce),
1540	],
1541	);
1542
1543	match cx.send_request(request).block_task().await {
1544	Ok(response) => match response.outcome {
1545	RequestPermissionOutcome::Selected(selected) => {
1546	match selected.option_id.0.as_ref() {
1547	"allow_once" => PermissionVerdict::Approved,
1548	"allow_session" => {
1549	permissions::grant_for_session(
1550	&session_id.to_string(),
1551	tool_name,
1552	arguments,
1553	);
1554	PermissionVerdict::Approved
1555	}
1556	_ => PermissionVerdict::Denied(permissions::user_denial(tool_name)),
1557	}
1558	}
1559	RequestPermissionOutcome::Cancelled => PermissionVerdict::TurnCancelled,
1560	// The outcome enum is non_exhaustive; treat anything unknown as
1561	// a denial rather than running a mutating tool unapproved.
1562	_ => PermissionVerdict::Denied(permissions::user_denial(tool_name)),
1563	},
1564	Err(error) => {
1565	log::warn!("permission request for `{tool_name}` failed: {error}");
1566	PermissionVerdict::Denied(format!(
1567	"`{tool_name}` was not executed: this client could not answer the \
1568	permission request ({error}). The user can pre-approve tools in \
1569	settings.toml under [permissions], or set SIGIT_PERMISSIONS=allow \
1570	for clients without permission support."
1571	))
1572	}
1573	}
1574	}
1575
1576	async fn handle_cancel(&self, args: CancelNotification) -> agent_client_protocol::Result<()> {
1577	log::info!("cancel requested for session {}", args.session_id);
1578	Ok(())
1579	}
1580
1581	/// Swap the active backend to a siGit Code Cloud tier and reflect it as the
1582	/// current model so the picker shows it selected. Returns the tier's display
1583	/// name on success, or `None` when no account is signed in (caller prompts
1584	/// for login). Shared by the panel picker and the `/models` slash command.
1585	async fn switch_to_cloud_tier(&self, tier: &str) -> Option<String> {
1586	let cfg = crate::provider::cloud_tier_provider(tier)?;
1587	let mut system_prompt = system_prompt_for_model(true).to_string();
1588	// Mirror the cwd guidance and project instruction files the local engine
1589	// gets at session load, so the cloud model shares the same project context.
1590	if let Some(cwd) = self.session_cwd.lock().ok().and_then(\|g\| g.clone()) {
1591	system_prompt.push_str("\n\n");
1592	system_prompt.push_str(&session_context_message(&cwd));
1593	}
1594	let cloud_backend: Arc<dyn InferenceBackend> = Arc::new(OpenAiBackend::new(
1595	cfg.base_url,
1596	cfg.api_key,
1597	cfg.model,
1598	Some(system_prompt),
1599	));
1600	*self.backend.lock().await = cloud_backend;
1601
1602	let cloud_config = GgufModelConfig {
1603	model_id: format!("sigit-cloud:{tier}"),
1604	files: Vec::new(),
1605	tok_model_id: None,
1606	display_name: cfg.display_name.clone(),
1607	approx_memory: "Cloud".to_string(),
1608	chat_template: None,
1609	};
1610	{
1611	let mut guard = self.current_model.lock().unwrap();
1612	*guard = cloud_config;
1613	}
1614
1615	// Explicitly choosing a cloud tier puts us in cloud mode.
1616	let _ = settings::set_local_inference(false);
1617
1618	log::info!("switched to cloud tier {tier}");
1619	Some(cfg.display_name)
1620	}
1621
1622	/// Apply the persisted Local Inference mode at session start. When local
1623	/// inference is off and an account is signed in, route to a cloud tier so the
1624	/// on-device model is never loaded; otherwise leave the on-device backend in
1625	/// place. Call after the session cwd is set so the cloud system prompt picks
1626	/// it up. Does not flip the stored setting on the not-signed-in fallback.
1627	async fn apply_startup_inference_mode(&self) {
1628	if settings::local_inference_enabled() {
1629	return;
1630	}
1631	if self.switch_to_cloud_tier("balanced").await.is_some() {
1632	log::info!("startup: local inference off; routing inference to siGit Code Cloud");
1633	} else {
1634	log::warn!(
1635	"local inference is off but no account is signed in; staying on-device. \
1636	Run /login or set Local Inference on."
1637	);
1638	}
1639	}
1640
1641	/// Route inference back on-device. Used after leaving a cloud tier for a
1642	/// local model. The `LocalBackend` reads the live `engine`, so this just
1643	/// repoints the active backend.
1644	async fn reset_to_local_backend(&self) {
1645	let local_backend: Arc<dyn InferenceBackend> =
1646	Arc::new(LocalBackend::new(Arc::clone(&self.engine)));
1647	*self.backend.lock().await = local_backend;
1648	}
1649
1650	/// Re-attempt the lazy startup model load if the previous attempt failed.
1651	/// Clears the one-shot guard so the next load runs; a healthy load is left
1652	/// untouched so `/reload` doesn't needlessly reload a working model.
1653	fn retry_startup_model_load_if_failed(&self) {
1654	let had_error = self
1655	.model_load_error
1656	.lock()
1657	.map(\|guard\| guard.is_some())
1658	.unwrap_or(false);
1659	if had_error {
1660	self.startup_model_load_started
1661	.store(false, Ordering::Release);
1662	self.start_startup_model_load_if_needed();
1663	}
1664	}
1665
1666	/// Re-sync session state in place — no new session needed. Re-applies the
1667	/// active backend from current credentials (so a fresh `/login` token is
1668	/// picked up), retries a failed model load, and pushes refreshed commands +
1669	/// picker so the editor's UI reflects the current state.
1670	async fn handle_reload(&self, cx: &ConnectionTo<Client>, session_id: SessionId) {
1671	let signed_in = account::status_line().await;
1672
1673	let on_cloud_tier = {
1674	let guard = self.current_model.lock().unwrap();
1675	guard
1676	.model_id
1677	.strip_prefix("sigit-cloud:")
1678	.map(str::to_string)
1679	};
1680
1681	let backend_note = match on_cloud_tier {
1682	Some(tier) => match self.switch_to_cloud_tier(&tier).await {
1683	Some(name) => format!("Active: {name}."),
1684	None => {
1685	self.reset_to_local_backend().await;
1686	"Signed out — back to on-device. Pick a model with /models.".to_string()
1687	}
1688	},
1689	None => {
1690	self.reset_to_local_backend().await;
1691	self.retry_startup_model_load_if_failed();
1692	let guard = self.current_model.lock().unwrap();
1693	format!("Active: {}.", guard.display_name)
1694	}
1695	};
1696
1697	// Push refreshed picker + commands so the editor reflects current state.
1698	let config_options = {
1699	let guard = self.current_model.lock().unwrap();
1700	build_model_config_options(&guard)
1701	};
1702	self.send_tool_call_update(
1703	cx,
1704	session_id.clone(),
1705	SessionUpdate::ConfigOptionUpdate(ConfigOptionUpdate::new(config_options)),
1706	)
1707	.ok();
1708	self.advertise_commands(cx, session_id.clone());
1709
1710	self.send_assistant_message(
1711	cx,
1712	session_id,
1713	format!("Reloaded. {signed_in} {backend_note}"),
1714	)
1715	.ok();
1716	}
1717
1718	async fn handle_set_session_config_option(
1719	&self,
1720	cx: &ConnectionTo<Client>,
1721	args: SetSessionConfigOptionRequest,
1722	) -> agent_client_protocol::Result<SetSessionConfigOptionResponse> {
1723	log::info!(
1724	"set_session_config_option: config_id={}, value={:?}",
1725	args.config_id,
1726	args.value
1727	);
1728
1729	// ── Local Inference toggle ──────────────────────────────────────────
1730	if args.config_id.0.as_ref() == LOCAL_INFERENCE_CONFIG_ID {
1731	let enabled = match args.value.0.as_ref() {
1732	LOCAL_INFERENCE_ON => true,
1733	LOCAL_INFERENCE_OFF => false,
1734	other => {
1735	return Err(agent_client_protocol::Error::new(
1736	-32602,
1737	format!("unknown Local Inference value: {other}"),
1738	));
1739	}
1740	};
1741	if let Err(error) = settings::set_local_inference(enabled) {
1742	return Err(agent_client_protocol::Error::new(
1743	-32603,
1744	format!("could not save Local Inference setting: {error}"),
1745	));
1746	}
1747	let message = if enabled {
1748	"Local inference is on. On-device models are highlighted; pick one from Model."
1749	} else {
1750	"Local inference is off. siGit Code Cloud tiers are highlighted; pick one from Model."
1751	};
1752	self.send_assistant_message(cx, args.session_id.clone(), format!("\n\n{message}"))
1753	.ok();
1754	// Rebuild so the Model picker reflects the new emphasis/order.
1755	let current = self.current_model.lock().unwrap().clone();
1756	let config_options = build_model_config_options(&current);
1757	return Ok(SetSessionConfigOptionResponse::new(config_options));
1758	}
1759
1760	if args.config_id.0.as_ref() != MODEL_CONFIG_ID {
1761	return Err(agent_client_protocol::Error::new(
1762	-32602,
1763	format!("unknown config option: {}", args.config_id.0),
1764	));
1765	}
1766
1767	let model_id = args.value.0.as_ref();
1768
1769	// can't switch while the startup model is still loading — the old
1770	// weights are in GPU memory and the new load gets "does not fit"
1771	if self.startup_model_load_started.load(Ordering::Acquire)
1772	&& !self.model_ready.load(Ordering::Acquire)
1773	{
1774	log::info!("set_session_config_option: waiting for startup model to finish loading");
1775	while !self.model_ready.load(Ordering::Acquire) {
1776	tokio::time::sleep(std::time::Duration::from_millis(200)).await;
1777	}
1778	}
1779
1780	// Zed re-fires the last selection when a thread opens. That re-fire must
1781	// not load anything: on-device models are loaded only on an explicit
1782	// request (`/load`, or actively picking a different model below), so a
1783	// re-fire of the already-current selection is a no-op. Otherwise opening a
1784	// new thread would silently load the local model — exactly what we avoid.
1785	{
1786	let current = self.current_model.lock().unwrap();
1787	if current.model_id == model_id {
1788	log::info!(
1789	"set_session_config_option: {} is already the active selection, skipping",
1790	current.display_name
1791	);
1792	let config_options = build_model_config_options(&current);
1793	return Ok(SetSessionConfigOptionResponse::new(config_options));
1794	}
1795	}
1796
1797	// ── siGit Code Cloud tier: no local load; sign-in gated ─────────────
1798	if let Some(tier) = model_id.strip_prefix("sigit-cloud:") {
1799	let message = match self.switch_to_cloud_tier(tier).await {
1800	Some(display_name) => format!("Switched to {display_name}."),
1801	None => CLOUD_LOGIN_PROMPT.to_string(),
1802	};
1803	// Start on a fresh line: ACP clients concatenate consecutive
1804	// agent-message chunks into one block, so without this the switch
1805	// confirmation runs onto the end of the previous assistant message.
1806	self.send_assistant_message(cx, args.session_id.clone(), format!("\n\n{message}"))
1807	.ok();
1808
1809	let current = self.current_model.lock().unwrap().clone();
1810	let config_options = build_model_config_options(&current);
1811	return Ok(SetSessionConfigOptionResponse::new(config_options));
1812	}
1813
1814	let needs_download = models::local_picker_items()
1815	.into_iter()
1816	.find(\|item\| item.config.model_id == model_id)
1817	.map(\|item\| item.cache_health == setup::ModelCacheHealth::NotDownloaded)
1818	.unwrap_or(false);
1819
1820	// tells the progress poller to stop
1821	let stop_flag = Arc::new(AtomicBool::new(false));
1822
1823	let tool_call_id = format!("model-switch-{}", uuid::Uuid::new_v4());
1824
1825	if needs_download {
1826	let model_id_owned = model_id.to_string();
1827	let expected_bytes = onde::inference::models::SUPPORTED_MODEL_INFO
1828	.iter()
1829	.find(\|m\| m.id == model_id_owned)
1830	.map(\|m\| m.expected_size_bytes)
1831	.unwrap_or(0);
1832
1833	let display_name = models::local_picker_items()
1834	.into_iter()
1835	.find(\|item\| item.config.model_id == model_id_owned)
1836	.map(\|item\| item.display_name.clone())
1837	.unwrap_or_else(\|\| model_id_owned.clone());
1838
1839	let size_hint = if expected_bytes > 0 {
1840	format!(" (~{})", format_size_human(expected_bytes))
1841	} else {
1842	String::new()
1843	};
1844
1845	self.send_tool_call_update(
1846	cx,
1847	args.session_id.clone(),
1848	SessionUpdate::ToolCall(
1849	ToolCall::new(
1850	tool_call_id.clone(),
1851	format!("⏬ Downloading {display_name}{size_hint}"),
1852	)
1853	.kind(ToolKind::Think)
1854	.status(ToolCallStatus::InProgress)
1855	.content(vec![
1856	format!(
1857	"Preparing download for {display_name}. This may take a few minutes."
1858	)
1859	.into(),
1860	]),
1861	),
1862	)
1863	.ok();
1864
1865	// poll download progress and update the spinner in Zed
1866	let cx_for_poller = cx.clone();
1867	let poller_session = args.session_id.clone();
1868	let poller_model_id = model_id_owned.clone();
1869	let poller_stop = Arc::clone(&stop_flag);
1870	let poller_tool_call_id = tool_call_id.clone();
1871
1872	cx.spawn(async move {
1873	const SPINNER: &[char] = &['⠋', '⠙', '⠹', '⠸', '⠼', '⠴', '⠦', '⠧', '⠇', '⠏'];
1874	let cache_path = onde::hf_cache::model_cache_path(&poller_model_id);
1875	let mut tick: usize = 0;
1876	let mut interval = tokio::time::interval(std::time::Duration::from_secs(1));
1877	interval.tick().await; // consume the immediate first tick
1878
1879	while !poller_stop.load(Ordering::Relaxed) {
1880	interval.tick().await;
1881
1882	if poller_stop.load(Ordering::Relaxed) {
1883	break;
1884	}
1885
1886	let downloaded = cache_path
1887	.as_ref()
1888	.filter(\|p\| p.exists())
1889	.map(\|p\| dir_size_recursive(p))
1890	.unwrap_or(0);
1891
1892	let frame = SPINNER[tick % SPINNER.len()];
1893	tick += 1;
1894
1895	let title = if expected_bytes > 0 {
1896	let pct =
1897	((downloaded as f64 / expected_bytes as f64) * 100.0).min(99.0) as u8;
1898	format!("{frame} Downloading {display_name}{size_hint} ({pct}%)")
1899	} else {
1900	format!("{frame} Downloading {display_name}{size_hint}")
1901	};
1902
1903	let msg = if expected_bytes > 0 {
1904	let pct =
1905	((downloaded as f64 / expected_bytes as f64) * 100.0).min(99.0) as u8;
1906	let bar = progress_bar(pct, 20);
1907	format!(
1908	"{display_name} — {bar} {pct}% ({} / {})",
1909	format_size_human(downloaded),
1910	format_size_human(expected_bytes),
1911	)
1912	} else {
1913	format!(
1914	"{display_name} — {} downloaded…",
1915	format_size_human(downloaded)
1916	)
1917	};
1918
1919	let notification = SessionNotification::new(
1920	poller_session.clone(),
1921	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
1922	poller_tool_call_id.clone(),
1923	ToolCallUpdateFields::new()
1924	.title(title)
1925	.status(ToolCallStatus::InProgress)
1926	.content(vec![msg.into()]),
1927	)),
1928	);
1929	if cx_for_poller.send_notification(notification).is_err() {
1930	break;
1931	}
1932	}
1933	Ok(())
1934	})
1935	.ok();
1936	}
1937
1938	// cached models still take 10-30s to load weights; show a spinner
1939	if !needs_download {
1940	let cached_display_name = models::local_picker_items()
1941	.into_iter()
1942	.find(\|item\| item.config.model_id == model_id)
1943	.map(\|item\| item.display_name.clone())
1944	.unwrap_or_else(\|\| model_id.to_string());
1945
1946	self.send_tool_call_update(
1947	cx,
1948	args.session_id.clone(),
1949	SessionUpdate::ToolCall(
1950	ToolCall::new(
1951	tool_call_id.clone(),
1952	format!("Loading {cached_display_name}"),
1953	)
1954	.kind(ToolKind::Think)
1955	.status(ToolCallStatus::InProgress)
1956	.content(vec![format!("Loading {cached_display_name}…").into()]),
1957	),
1958	)
1959	.ok();
1960
1961	// tick every 5s so the user knows we haven't frozen
1962	let cx_for_spinner = cx.clone();
1963	let spinner_session = args.session_id.clone();
1964	let spinner_name = cached_display_name.clone();
1965	let spinner_stop = Arc::clone(&stop_flag);
1966	let spinner_tool_call_id = tool_call_id.clone();
1967	let load_start = std::time::Instant::now();
1968
1969	cx.spawn(async move {
1970	const SPINNER: &[char] = &['⠋', '⠙', '⠹', '⠸', '⠼', '⠴', '⠦', '⠧', '⠇', '⠏'];
1971	let mut tick: usize = 0;
1972	let mut interval = tokio::time::interval(std::time::Duration::from_secs(5));
1973	interval.tick().await; // consume the immediate first tick
1974
1975	while !spinner_stop.load(Ordering::Relaxed) {
1976	interval.tick().await;
1977
1978	if spinner_stop.load(Ordering::Relaxed) {
1979	break;
1980	}
1981
1982	let elapsed = load_start.elapsed();
1983	let elapsed_str = if elapsed.as_secs() >= 60 {
1984	format!("{}m {:02}s", elapsed.as_secs() / 60, elapsed.as_secs() % 60)
1985	} else {
1986	format!("{}s", elapsed.as_secs())
1987	};
1988	let frame = SPINNER[tick % SPINNER.len()];
1989	tick += 1;
1990
1991	let msg = format!("{frame} Loading {spinner_name}… ({elapsed_str})");
1992	let notification = SessionNotification::new(
1993	spinner_session.clone(),
1994	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
1995	spinner_tool_call_id.clone(),
1996	ToolCallUpdateFields::new()
1997	.status(ToolCallStatus::InProgress)
1998	.content(vec![msg.into()]),
1999	)),
2000	);
2001	if cx_for_spinner.send_notification(notification).is_err() {
2002	break;
2003	}
2004	}
2005	Ok(())
2006	})
2007	.ok();
2008	}
2009
2010	let switch_result = self.switch_model_by_id(model_id).await;
2011
2012	stop_flag.store(true, Ordering::Relaxed);
2013
2014	match switch_result {
2015	Ok(new_config) => {
2016	// Route inference back on-device (in case we were on a cloud tier).
2017	self.reset_to_local_backend().await;
2018	// Selecting an on-device model puts us in local mode.
2019	let _ = settings::set_local_inference(true);
2020
2021	let completion_title = if needs_download {
2022	format!("✓ {} downloaded and loaded", new_config.display_name)
2023	} else {
2024	format!("✓ Switched to {}", new_config.display_name)
2025	};
2026	let completion_body = if needs_download {
2027	format!("✓ {} downloaded and loaded.", new_config.display_name)
2028	} else {
2029	format!("✓ Switched to {}.", new_config.display_name)
2030	};
2031
2032	self.send_tool_call_update(
2033	cx,
2034	args.session_id.clone(),
2035	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
2036	tool_call_id,
2037	ToolCallUpdateFields::new()
2038	.title(completion_title)
2039	.status(ToolCallStatus::Completed)
2040	.content(vec![completion_body.into()]),
2041	)),
2042	)
2043	.ok();
2044
2045	let config_options = {
2046	let guard = self.current_model.lock().unwrap();
2047	build_model_config_options(&guard)
2048	};
2049
2050	log::info!("model switch complete");
2051	Ok(SetSessionConfigOptionResponse::new(config_options))
2052	}
2053	Err(err) => {
2054	self.send_tool_call_update(
2055	cx,
2056	args.session_id.clone(),
2057	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
2058	tool_call_id,
2059	ToolCallUpdateFields::new()
2060	.title("Model switch failed".to_string())
2061	.status(ToolCallStatus::Failed)
2062	.content(vec![format!("error loading model: {}", err.message).into()]),
2063	)),
2064	)
2065	.ok();
2066
2067	Err(err)
2068	}
2069	}
2070	}
2071	}
2072
2073	// ── Config option helpers ─────────────────────────────────────────────────────
2074
2075	/// config option ID for the model picker in Zed's agent panel
2076	const MODEL_CONFIG_ID: &str = "sigit-model";
2077
2078	/// config option ID for the Local Inference on/off toggle. Surfaced as a
2079	/// two-option `select` so ACP clients without slash-command support (e.g. Xcode)
2080	/// can still flip the mode from the agent panel.
2081	const LOCAL_INFERENCE_CONFIG_ID: &str = "sigit-local-inference";
2082
2083	/// `select` value ids for the Local Inference toggle.
2084	const LOCAL_INFERENCE_ON: &str = "local-inference-on";
2085	const LOCAL_INFERENCE_OFF: &str = "local-inference-off";
2086
2087	/// Replace non-ASCII chars so a downstream byte-index truncation can't split a
2088	/// multi-byte char. Zed slices the model-picker label at a fixed byte offset
2089	/// (`agent_ui/src/config_options.rs`) and panics — crashing the whole editor —
2090	/// when the cut lands mid-glyph (e.g. inside `☁` or `·`). Mapping to `-` keeps
2091	/// separators readable; ASCII bytes are always char boundaries.
2092	fn ascii_safe(s: &str) -> String {
2093	s.chars()
2094	.map(\|c\| if c.is_ascii() { c } else { '-' })
2095	.collect()
2096	}
2097
2098	fn build_model_config_options(current_model: &GgufModelConfig) -> Vec<SessionConfigOption> {
2099	// The full list, including the siGit Code Cloud tiers, so the panel picker
2100	// mirrors the TUI `/models`. Cloud entries are sign-in gated at selection.
2101	let items = models::build_model_picker_items();
2102	let active_kind = models::active_inference_kind();
2103
2104	let options: Vec<SessionConfigSelectOption> = items
2105	.iter()
2106	.filter(\|item\| item.cache_health != setup::ModelCacheHealth::Incomplete)
2107	.map(\|item\| {
2108	let mut desc_parts = Vec::new();
2109	// Mark options in the inactive mode so the active group reads as the
2110	// recommended set (the list is already ordered active-group-first).
2111	if item.source.kind() != active_kind {
2112	desc_parts.push("inactive mode".to_string());
2113	}
2114	if item.tool_calling {
2115	desc_parts.push("tool calling".to_string());
2116	}
2117	desc_parts.push(item.description.clone());
2118	if item.cache_health == setup::ModelCacheHealth::NotDownloaded {
2119	desc_parts.push("download on select".to_string());
2120	}
2121	// ASCII-only for the same reason as the name (see `ascii_safe`).
2122	let description = ascii_safe(&desc_parts.join(" - "));
2123	// Keep badges ASCII: Zed truncates the picker label at a fixed byte
2124	// offset and panics if the cut splits a multi-byte char. See
2125	// `ascii_safe` below.
2126	let source_badge = if item.cloud_tier.is_some() {
2127	" [siGit Code Cloud]"
2128	} else if item.cache_health == setup::ModelCacheHealth::NotDownloaded {
2129	" [Onde]"
2130	} else {
2131	match item.source_label.as_str() {
2132	"Onde" => " [Onde]",
2133	"HuggingFace" => " [HuggingFace]",
2134	_ => "",
2135	}
2136	};
2137	// For cloud tiers use just the tier title (e.g. "Balanced") so the
2138	// label reads "Balanced [siGit Code Cloud]" instead of repeating the
2139	// brand. The display name can carry non-ASCII (the cloud tier label
2140	// is "siGit Code Cloud · Balanced"), so sanitize the whole label.
2141	let base_name = match &item.cloud_tier {
2142	Some(tier) => crate::provider::tier_title(tier),
2143	None => item.display_name.clone(),
2144	};
2145	let name = ascii_safe(&format!("{base_name}{source_badge}"));
2146	SessionConfigSelectOption::new(
2147	SessionConfigValueId::new(item.config.model_id.as_str()),
2148	name,
2149	)
2150	.description(description)
2151	})
2152	.collect();
2153
2154	// Local Inference on/off toggle, modeled as a two-option select so panel-only
2155	// ACP clients (no slash commands) can flip the mode.
2156	let local_on = settings::local_inference_enabled();
2157	let local_current = SessionConfigValueId::new(if local_on {
2158	LOCAL_INFERENCE_ON
2159	} else {
2160	LOCAL_INFERENCE_OFF
2161	});
2162	let local_options = vec![
2163	SessionConfigSelectOption::new(
2164	SessionConfigValueId::new(LOCAL_INFERENCE_ON),
2165	"On (on-device)".to_string(),
2166	)
2167	.description("Run inference on-device; on-device models are highlighted".to_string()),
2168	SessionConfigSelectOption::new(
2169	SessionConfigValueId::new(LOCAL_INFERENCE_OFF),
2170	"Off (siGit Code Cloud)".to_string(),
2171	)
2172	.description("Use siGit Code Cloud; cloud tiers are highlighted".to_string()),
2173	];
2174	let local_option = SessionConfigOption::select(
2175	LOCAL_INFERENCE_CONFIG_ID,
2176	"Local Inference",
2177	local_current,
2178	local_options,
2179	)
2180	.description("Toggle on-device inference; changes which models are highlighted");
2181
2182	if options.is_empty() {
2183	return vec![local_option];
2184	}
2185
2186	let current_value = SessionConfigValueId::new(current_model.model_id.as_str());
2187
2188	vec![
2189	SessionConfigOption::select(MODEL_CONFIG_ID, "Model", current_value, options)
2190	.category(SessionConfigOptionCategory::Model)
2191	.description("Select an on-device model or a siGit Code Cloud tier"),
2192	local_option,
2193	]
2194	}
2195
2196	/// returns `(config, max_tokens, tool_calling)` for a picker model_id, or None
2197	fn resolve_model_config(model_id: &str) -> Option<(GgufModelConfig, u64, bool)> {
2198	let items = models::local_picker_items();
2199	items
2200	.into_iter()
2201	.find(\|item\| {
2202	item.config.model_id == model_id
2203	&& item.cache_health != setup::ModelCacheHealth::Incomplete
2204	})
2205	.map(\|item\| (item.config, item.max_tokens, item.tool_calling))
2206	}
2207
2208	// ── Slash commands ────────────────────────────────────────────────────────────
2209
2210	#[derive(Debug, Clone)]
2211	enum SlashCommand {
2212	Help,
2213	Clear,
2214	Status,
2215	Models(Option<usize>),
2216	/// toggle on-device inference mode. `Some(true/false)` sets it, `None` flips it.
2217	Local(Option<bool>),
2218	/// List discovered Agent Skills.
2219	Skills,
2220	/// List configured MCP servers and their tools.
2221	Mcp,
2222	/// Explicitly load the selected (or default) on-device model.
2223	Load,
2224	/// `/login <email> <password>` — the raw argument, parsed when executed.
2225	Login(Option<String>),
2226	Logout,
2227	Whoami,
2228	/// Re-sync session state (auth, backend, picker) without a new session.
2229	Reload,
2230	/// Toggle plan mode (read-only research; mutating tools denied with a
2231	/// prompt to present a plan). `Some(true/false)` sets it, `None` flips it.
2232	Plan(Option<bool>),
2233	/// Show the effective permission policy for this session.
2234	Permissions,
2235	/// Summarize-and-shrink the conversation history on demand.
2236	Compact,
2237	Exit,
2238	Unknown(String),
2239	}
2240
2241	fn parse_slash(input: &str) -> Option<SlashCommand> {
2242	let trimmed = input.trim();
2243	if !trimmed.starts_with('/') {
2244	return None;
2245	}
2246	let mut parts = trimmed.splitn(2, char::is_whitespace);
2247	let command = parts.next().unwrap_or("");
2248	let argument = parts.next().map(str::trim);
2249	Some(match command {
2250	"/help" => SlashCommand::Help,
2251	"/clear" => SlashCommand::Clear,
2252	"/status" => SlashCommand::Status,
2253	"/models" => SlashCommand::Models(argument.and_then(\|v\| v.parse::<usize>().ok())),
2254	"/local" => SlashCommand::Local(parse_on_off(argument)),
2255	"/skills" => SlashCommand::Skills,
2256	"/mcp" => SlashCommand::Mcp,
2257	"/load" => SlashCommand::Load,
2258	"/login" => SlashCommand::Login(argument.map(str::to_string)),
2259	"/logout" => SlashCommand::Logout,
2260	"/whoami" => SlashCommand::Whoami,
2261	"/reload" => SlashCommand::Reload,
2262	"/plan" => SlashCommand::Plan(parse_on_off(argument)),
2263	"/permissions" => SlashCommand::Permissions,
2264	"/compact" => SlashCommand::Compact,
2265	"/exit" \| "/quit" \| "/q" => SlashCommand::Exit,
2266	other => SlashCommand::Unknown(other.to_string()),
2267	})
2268	}
2269
2270	/// `on`/`off` (and synonyms) → `Some(bool)`; missing or unrecognized → `None`
2271	/// (meaning "toggle the current value").
2272	fn parse_on_off(arg: Option<&str>) -> Option<bool> {
2273	match arg.map(\|s\| s.trim().to_ascii_lowercase())?.as_str() {
2274	"on" \| "true" \| "1" \| "yes" => Some(true),
2275	"off" \| "false" \| "0" \| "no" => Some(false),
2276	_ => None,
2277	}
2278	}
2279
2280	fn format_models_list(current_model: &GgufModelConfig) -> String {
2281	let items = models::build_model_picker_items();
2282	if items.is_empty() {
2283	return "No local models found. siGit will use the platform default model.".to_string();
2284	}
2285
2286	let mut lines = vec!["Available models:".to_string()];
2287	let mut last_source: Option<&str> = None;
2288
2289	for (index, item) in items.iter().enumerate() {
2290	let source_key = match item.source_label.as_str() {
2291	"Onde" => "Onde",
2292	"HuggingFace" => "HuggingFace",
2293	"siGit Code Cloud" => "Cloud",
2294	_ => "Fallback",
2295	};
2296
2297	if last_source != Some(source_key) {
2298	if last_source.is_some() {
2299	lines.push(String::new());
2300	}
2301	let section = match source_key {
2302	"Onde" => "Onde Inference",
2303	"HuggingFace" => "Hugging Face cache",
2304	"Cloud" => "siGit Code Cloud",
2305	_ => "Fallback",
2306	};
2307	lines.push(section.to_string());
2308	// Blank line so the following "N." items render as an ordered list.
2309	// CommonMark only lets an ordered list interrupt a paragraph when it
2310	// starts at 1, so without this the cloud section (items 9+) would be
2311	// absorbed into the header paragraph.
2312	lines.push(String::new());
2313	last_source = Some(source_key);
2314	}
2315
2316	let number = index + 1;
2317	let current_badge = if item.config.model_id == current_model.model_id {
2318	" <- current"
2319	} else {
2320	""
2321	};
2322	let tool_badge = if item.tool_calling {
2323	" tool calling"
2324	} else {
2325	""
2326	};
2327	let health_badge = match item.cache_health {
2328	setup::ModelCacheHealth::Complete => "",
2329	setup::ModelCacheHealth::Incomplete => " ! incomplete cache",
2330	setup::ModelCacheHealth::NotDownloaded => " ↓ download on select",
2331	};
2332	let source = match source_key {
2333	"Onde" => " [Onde]",
2334	"HuggingFace" => " [HuggingFace]",
2335	"Cloud" => " [☁ Cloud]",
2336	_ => " [default]",
2337	};
2338
2339	lines.push(format!(
2340	"{number}. {} {}{}{}{}{}",
2341	item.display_name, item.description, tool_badge, health_badge, current_badge, source,
2342	));
2343	}
2344
2345	lines.push(String::new());
2346	lines.push("Use /models N to switch models.".to_string());
2347	lines.join("\n")
2348	}
2349
2350	async fn exec_slash_acp(
2351	agent: &SiGitAgent,
2352	cx: &ConnectionTo<Client>,
2353	session_id: SessionId,
2354	command: SlashCommand,
2355	) -> agent_client_protocol::Result<PromptResponse> {
2356	match command {
2357	SlashCommand::Help => {
2358	agent
2359	.send_assistant_message(
2360	cx,
2361	session_id,
2362	"/help - show this message\n\
2363	/models - list available models\n\
2364	/models N - switch to model N\n\
2365	/local [on\|off]- toggle on-device inference mode\n\
2366	/skills - list available Agent Skills\n\
2367	/mcp - list MCP servers and their tools\n\
2368	/load - load the selected on-device model\n\
2369	/login E P - sign in to siGit Code Cloud\n\
2370	/logout - sign out\n\
2371	/whoami - show the signed-in account\n\
2372	/reload - re-sync sign-in and model state\n\
2373	/plan [on\|off] - plan mode: research only, no edits or commands\n\
2374	/permissions - show the tool permission policy\n\
2375	/compact - summarize and shrink conversation history\n\
2376	/clear - wipe conversation history\n\
2377	/status - show engine status\n\
2378	/exit - end this turn",
2379	)
2380	.ok();
2381	}
2382	SlashCommand::Clear => {
2383	let cleared = agent.engine.clear_history().await;
2384	permissions::reset_session(&session_id.to_string());
2385	// The saved session must not resurrect what the user just wiped.
2386	session_store::delete(&session_id.to_string());
2387	agent
2388	.send_assistant_message(
2389	cx,
2390	session_id,
2391	format!("Cleared {cleared} turn(s). History is empty."),
2392	)
2393	.ok();
2394	}
2395	SlashCommand::Plan(value) => {
2396	let session_key = session_id.to_string();
2397	let enabled = value.unwrap_or_else(\|\| !permissions::plan_mode(&session_key));
2398	permissions::set_plan_mode(&session_key, enabled);
2399	let message = if enabled {
2400	"Plan mode ON — the agent researches with read-only tools and presents a \
2401	plan; edits and commands are blocked until /plan off."
2402	} else {
2403	"Plan mode OFF — the agent may execute tools again (subject to the \
2404	permission policy)."
2405	};
2406	agent.send_assistant_message(cx, session_id, message).ok();
2407	}
2408	SlashCommand::Permissions => {
2409	let summary = permissions::describe(&session_id.to_string());
2410	agent.send_assistant_message(cx, session_id, summary).ok();
2411	}
2412	SlashCommand::Compact => {
2413	let backend = agent.backend.lock().await.clone();
2414	let before = backend::estimate_tokens(&backend.history_snapshot().await);
2415	let message = match backend.compact_history(backend::COMPACT_KEEP_LAST).await {
2416	Ok(()) => {
2417	let snapshot = backend.history_snapshot().await;
2418	let after = backend::estimate_tokens(&snapshot);
2419	// Keep the saved session in step with the compacted state.
2420	if let Err(error) = session_store::save(&session_id.to_string(), &snapshot) {
2421	log::warn!("session save after /compact failed: {error}");
2422	}
2423	format!("Compacted history: ~{before} → ~{after} tokens (estimated).")
2424	}
2425	Err(error) => format!("Compaction failed: {error}"),
2426	};
2427	agent.send_assistant_message(cx, session_id, message).ok();
2428	}
2429	SlashCommand::Status => {
2430	let info = agent.engine.info().await;
2431	let model = info.model_name.as_deref().unwrap_or("(none)");
2432	let memory = info.approx_memory.as_deref().unwrap_or("unknown");
2433	agent
2434	.send_assistant_message(
2435	cx,
2436	session_id,
2437	format!(
2438	"status: {:?} model: {} memory: {} history: {} turns",
2439	info.status, model, memory, info.history_length,
2440	),
2441	)
2442	.ok();
2443	}
2444	SlashCommand::Models(None) => {
2445	let current_model = agent.current_model.lock().unwrap().clone();
2446	agent
2447	.send_assistant_message(cx, session_id, format_models_list(&current_model))
2448	.ok();
2449	}
2450	SlashCommand::Skills => {
2451	agent
2452	.send_assistant_message(cx, session_id, skills::format_skills_list())
2453	.ok();
2454	}
2455	SlashCommand::Mcp => {
2456	agent
2457	.send_assistant_message(cx, session_id, mcp::status_summary())
2458	.ok();
2459	}
2460	SlashCommand::Models(Some(number)) => {
2461	let items = models::build_model_picker_items();
2462	let index = number.saturating_sub(1);
2463	match items.get(index).cloned() {
2464	None => {
2465	agent
2466	.send_assistant_message(
2467	cx,
2468	session_id,
2469	format!("error: no model #{number} - type /models to see the list."),
2470	)
2471	.ok();
2472	}
2473	Some(model) if model.cloud_tier.is_some() => {
2474	// siGit Code Cloud tier: swap backend, sign-in gated.
2475	let tier = model.cloud_tier.clone().unwrap_or_default();
2476	let message = match agent.switch_to_cloud_tier(&tier).await {
2477	Some(display_name) => format!("Switched to {display_name}."),
2478	None => CLOUD_LOGIN_PROMPT.to_string(),
2479	};
2480	agent.send_assistant_message(cx, session_id, message).ok();
2481	}
2482	Some(model) => {
2483	if model.cache_health == setup::ModelCacheHealth::Incomplete {
2484	agent
2485	.send_assistant_message(
2486	cx,
2487	session_id,
2488	format!(
2489	"error: {} has an incomplete local cache and cannot be selected yet.",
2490	model.display_name
2491	),
2492	)
2493	.ok();
2494	} else if model.cache_health == setup::ModelCacheHealth::NotDownloaded {
2495	agent
2496	.send_assistant_message(
2497	cx,
2498	session_id.clone(),
2499	format!(
2500	"Downloading and loading {} ({})… this may take a few minutes.",
2501	model.display_name, model.description
2502	),
2503	)
2504	.ok();
2505
2506	match agent.switch_model_by_id(&model.config.model_id).await {
2507	Ok(new_config) => {
2508	agent.reset_to_local_backend().await;
2509	let _ = settings::set_local_inference(true);
2510	agent.engine.clear_history().await;
2511	agent
2512	.send_assistant_message(
2513	cx,
2514	session_id,
2515	format!(
2516	"✓ Downloaded and switched to {}",
2517	new_config.display_name
2518	),
2519	)
2520	.ok();
2521	}
2522	Err(err) => {
2523	agent
2524	.send_assistant_message(
2525	cx,
2526	session_id,
2527	format!("error downloading model: {}", err.message),
2528	)
2529	.ok();
2530	}
2531	}
2532	} else {
2533	agent
2534	.send_assistant_message(
2535	cx,
2536	session_id.clone(),
2537	format!("Loading {}...", model.display_name),
2538	)
2539	.ok();
2540
2541	let switched = agent.switch_model_by_id(&model.config.model_id).await?;
2542	agent.reset_to_local_backend().await;
2543	let _ = settings::set_local_inference(true);
2544	agent.engine.clear_history().await;
2545
2546	agent
2547	.send_assistant_message(
2548	cx,
2549	session_id,
2550	format!("Switched to {}.", switched.display_name),
2551	)
2552	.ok();
2553	}
2554	}
2555	}
2556	}
2557	SlashCommand::Local(value) => {
2558	let enabled = value.unwrap_or(!settings::local_inference_enabled());
2559	let message = match settings::set_local_inference(enabled) {
2560	Ok(()) if enabled => "Local inference is on. On-device models are highlighted; \
2561	pick one with /models."
2562	.to_string(),
2563	Ok(()) => "Local inference is off. siGit Code Cloud tiers are highlighted; \
2564	pick one with /models."
2565	.to_string(),
2566	Err(error) => format!("error: could not save local inference setting: {error}"),
2567	};
2568	agent
2569	.send_assistant_message(cx, session_id.clone(), message)
2570	.ok();
2571	// Refresh the panel so the Model picker reflects the new emphasis.
2572	let config_options = {
2573	let current = agent.current_model.lock().unwrap();
2574	build_model_config_options(&current)
2575	};
2576	agent
2577	.send_tool_call_update(
2578	cx,
2579	session_id,
2580	SessionUpdate::ConfigOptionUpdate(ConfigOptionUpdate::new(config_options)),
2581	)
2582	.ok();
2583	}
2584	SlashCommand::Load => {
2585	// Explicitly load the on-device model. This is the only path that
2586	// brings a local model into memory; prompts never do it implicitly.
2587	// If a cloud tier is active, fall back to a local default so we don't
2588	// try to load the (file-less) cloud config as GGUF.
2589	let on_cloud = {
2590	let guard = agent.current_model.lock().unwrap();
2591	guard.model_id.starts_with("sigit-cloud:")
2592	};
2593	if on_cloud {
2594	let default_config = default_local_model_config();
2595	*agent.current_model.lock().unwrap() = default_config;
2596	agent.reset_to_local_backend().await;
2597	}
2598	// Loading an on-device model puts us in local inference mode.
2599	let _ = settings::set_local_inference(true);
2600	// `await_model_ready` drives the download/load progress UI and reports
2601	// success or failure to the editor.
2602	agent.start_startup_model_load_if_needed();
2603	agent.await_model_ready(cx, &session_id).await?;
2604	}
2605	SlashCommand::Login(argument) => {
2606	let message = match argument.as_deref().and_then(account::parse_login_args) {
2607	Some((email, password)) => match account::authenticate(&email, &password).await {
2608	Ok(email) => format!(
2609	"Signed in as {email}. Pick a siGit Code Cloud tier in /models to use it."
2610	),
2611	Err(error) => format!("Login failed: {error}"),
2612	},
2613	None => "usage: /login <email> <password>".to_string(),
2614	};
2615	agent.send_assistant_message(cx, session_id, message).ok();
2616	}
2617	SlashCommand::Logout => {
2618	// If we're on a cloud tier, drop back to local — the token is gone.
2619	let on_cloud = {
2620	let guard = agent.current_model.lock().unwrap();
2621	guard.model_id.starts_with("sigit-cloud:")
2622	};
2623	let message = account::end_session().await;
2624	if on_cloud {
2625	agent.reset_to_local_backend().await;
2626	}
2627	agent.send_assistant_message(cx, session_id, message).ok();
2628	}
2629	SlashCommand::Whoami => {
2630	let message = account::status_line().await;
2631	agent.send_assistant_message(cx, session_id, message).ok();
2632	}
2633	SlashCommand::Reload => {
2634	agent.handle_reload(cx, session_id).await;
2635	}
2636	SlashCommand::Exit => {
2637	agent
2638	.send_assistant_message(
2639	cx,
2640	session_id,
2641	"Use the panel controls to close or switch threads.",
2642	)
2643	.ok();
2644	}
2645	SlashCommand::Unknown(command) => {
2646	agent
2647	.send_assistant_message(cx, session_id, format!("unknown command: {command}"))
2648	.ok();
2649	}
2650	}
2651
2652	Ok(PromptResponse::new(StopReason::EndTurn))
2653	}
2654
2655	// ── Request dispatch helper ───────────────────────────────────────────────────
2656
2657	fn handle_response<T: agent_client_protocol::JsonRpcResponse>(
2658	responder: Responder<T>,
2659	result: agent_client_protocol::Result<T>,
2660	) -> agent_client_protocol::Result<()> {
2661	match result {
2662	Ok(resp) => responder.respond(resp),
2663	Err(err) => responder.respond_with_error(err),
2664	}
2665	}
2666
2667	// ── Download progress helpers ─────────────────────────────────────────────────
2668
2669	/// total bytes on disk under `path`. needed because hf-hub uses staging
2670	/// names during download, so we can't just stat the final blobs.
2671	fn dir_size_recursive(path: &std::path::Path) -> u64 {
2672	let mut total: u64 = 0;
2673	let Ok(entries) = std::fs::read_dir(path) else {
2674	return 0;
2675	};
2676	for entry in entries.flatten() {
2677	let entry_path = entry.path();
2678	if entry_path.is_dir() {
2679	total += dir_size_recursive(&entry_path);
2680	} else if let Ok(meta) = entry_path.metadata() {
2681	total += meta.len();
2682	}
2683	}
2684	total
2685	}
2686
2687	fn format_size_human(bytes: u64) -> String {
2688	const GB: u64 = 1_073_741_824;
2689	const MB: u64 = 1_048_576;
2690	const KB: u64 = 1_024;
2691	if bytes >= GB {
2692	format!("{:.2} GB", bytes as f64 / GB as f64)
2693	} else if bytes >= MB {
2694	format!("{:.1} MB", bytes as f64 / MB as f64)
2695	} else if bytes >= KB {
2696	format!("{:.0} KB", bytes as f64 / KB as f64)
2697	} else {
2698	format!("{bytes} B")
2699	}
2700	}
2701
2702	fn progress_bar(pct: u8, width: usize) -> String {
2703	let filled = ((pct as usize) * width) / 100;
2704	let empty = width.saturating_sub(filled);
2705	format!("[{}{}]", "█".repeat(filled), "░".repeat(empty))
2706	}
2707
2708	// ── Output capture ────────────────────────────────────────────────────────────
2709
2710	/// redirect stdout+stderr to `$TMPDIR/sigit.log` at the fd level so
2711	/// mistralrs/tracing noise never hits the terminal. returns two dup'd
2712	/// fds to the real tty: one for ratatui, one for cleanup (ratatui 0.29
2713	/// doesn't expose `writer_mut()`).
2714	#[cfg(unix)]
2715	fn redirect_output_to_log() -> anyhow::Result<(std::fs::File, std::fs::File)> {
2716	let log_path = std::env::temp_dir().join("sigit.log");
2717	let log_file = std::fs::File::create(&log_path)?;
2718	let log_fd = log_file.as_raw_fd();
2719
2720	// two copies: ratatui needs one, cleanup needs another
2721	let saved_tui = unsafe { libc::dup(libc::STDOUT_FILENO) };
2722	anyhow::ensure!(
2723	saved_tui >= 0,
2724	"dup(stdout) for tui failed: {}",
2725	std::io::Error::last_os_error()
2726	);
2727	let saved_cleanup = unsafe { libc::dup(libc::STDOUT_FILENO) };
2728	anyhow::ensure!(
2729	saved_cleanup >= 0,
2730	"dup(stdout) for cleanup failed: {}",
2731	std::io::Error::last_os_error()
2732	);
2733
2734	unsafe {
2735	libc::dup2(log_fd, libc::STDOUT_FILENO);
2736	libc::dup2(log_fd, libc::STDERR_FILENO);
2737	}
2738
2739	// safe to drop log_file; dup2 keeps the fd alive via stdout/stderr
2740
2741	Ok((unsafe { std::fs::File::from_raw_fd(saved_tui) }, unsafe {
2742	std::fs::File::from_raw_fd(saved_cleanup)
2743	}))
2744	}
2745
2746	// ── Logging ───────────────────────────────────────────────────────────────────
2747
2748	/// in TUI mode stderr is the log file (redirected earlier);
2749	/// in ACP mode it's real stderr. either way, write there.
2750	fn init_logging(is_tty: bool) {
2751	let filter = EnvFilter::try_from_default_env().unwrap_or_else(\|_\| EnvFilter::new("info"));
2752	let _ = tracing_fmt::Subscriber::builder()
2753	.with_env_filter(filter)
2754	.with_writer(std::io::stderr)
2755	.with_ansi(!is_tty)
2756	.try_init();
2757	}
2758
2759	// ── Interactive TUI mode ──────────────────────────────────────────────────────
2760
2761	/// boot the TUI and load the model on a background thread.
2762	/// `tty` goes to ratatui; `cleanup_tty` is a separate fd for
2763	/// LeaveAlternateScreen (ratatui 0.29 hides `writer_mut()`).
2764	#[cfg(unix)]
2765	async fn run_interactive(tty: std::fs::File, mut cleanup_tty: std::fs::File) -> anyhow::Result<()> {
2766	let engine = Arc::new(ChatEngine::new());
2767
2768	let startup_selection = setup::startup_model_selection();
2769	let startup_model_name = startup_selection
2770	.as_ref()
2771	.map(\|selection\| selection.display_name.clone())
2772	.unwrap_or_else(\|\| GgufModelConfig::qwen25_3b().display_name);
2773
2774	// Signals the loading phase to finish. On-device models are no longer loaded
2775	// at startup, so this resolves immediately for both backends; it stays a
2776	// channel so the loading-phase plumbing in `chat::run_with` is unchanged.
2777	let (load_tx, load_rx) = std::sync::mpsc::channel::<Result<(), String>>();
2778
2779	// Project instruction files (AGENTS.md / CLAUDE.md) for the launch directory,
2780	// injected into the system prompt so the TUI shares the same always-on
2781	// project context the ACP sessions get.
2782	let project_instructions = std::env::current_dir()
2783	.ok()
2784	.and_then(\|cwd\| instructions::load_project_instructions(&cwd));
2785	let with_instructions = \|base: String\| match &project_instructions {
2786	Some(extra) => format!("{base}\n\n{extra}"),
2787	None => base,
2788	};
2789
2790	// Pick the inference backend: a configured provider if present, else on-device.
2791	let (inference_backend, startup_model_name): (Arc<dyn InferenceBackend>, String) =
2792	match provider::active_provider() {
2793	Some(provider) => {
2794	log::info!(
2795	"inference: using {} (model {}) at {}",
2796	provider.display_name,
2797	provider.model,
2798	provider.base_url
2799	);
2800	// No local model to load; the endpoint is ready immediately.
2801	let _ = load_tx.send(Ok(()));
2802	let label = provider.display_name.clone();
2803	register_subagent_factory_for(&provider);
2804	let backend = Arc::new(OpenAiBackend::new(
2805	provider.base_url,
2806	provider.api_key,
2807	provider.model,
2808	Some(with_instructions(SYSTEM_PROMPT.to_string())),
2809	)) as Arc<dyn InferenceBackend>;
2810	(backend, label)
2811	}
2812	None => {
2813	// Honor the Local Inference toggle: when off and signed in, start
2814	// on a cloud tier. Otherwise bring up on-device WITHOUT loading a
2815	// model — the user loads it explicitly with /load (or /models), so
2816	// the UI comes up immediately. Project instructions are injected at
2817	// load time in `chat.rs`.
2818	let cloud_when_off = if settings::local_inference_enabled() {
2819	None
2820	} else {
2821	provider::cloud_tier_provider("balanced")
2822	};
2823
2824	match cloud_when_off {
2825	Some(provider) => {
2826	log::info!(
2827	"inference: local inference off; using {} (model {})",
2828	provider.display_name,
2829	provider.model
2830	);
2831	let _ = load_tx.send(Ok(()));
2832	let label = provider.display_name.clone();
2833	register_subagent_factory_for(&provider);
2834	let backend = Arc::new(OpenAiBackend::new(
2835	provider.base_url,
2836	provider.api_key,
2837	provider.model,
2838	Some(with_instructions(SYSTEM_PROMPT.to_string())),
2839	)) as Arc<dyn InferenceBackend>;
2840	(backend, label)
2841	}
2842	None => {
2843	if !settings::local_inference_enabled() {
2844	log::warn!(
2845	"local inference is off but no account is signed in; \
2846	bringing up on-device. Run /login or /local on."
2847	);
2848	}
2849	let _ = load_tx.send(Ok(()));
2850	// On-device inference has a single shared history; no
2851	// subagent context is possible yet.
2852	tools::set_subagent_factory(Box::new(\|\| None));
2853	let backend = Arc::new(LocalBackend::new(Arc::clone(&engine)))
2854	as Arc<dyn InferenceBackend>;
2855	(backend, startup_model_name)
2856	}
2857	}
2858	}
2859	};
2860
2861	crossterm::terminal::enable_raw_mode()?;
2862	let mut tty = BufWriter::new(tty);
2863	crossterm::execute!(tty, crossterm::terminal::EnterAlternateScreen)?;
2864	let term_backend = ratatui::backend::CrosstermBackend::new(tty);
2865	let mut terminal = ratatui::Terminal::new(term_backend)?;
2866
2867	// polls load_rx with try_recv() each tick, no blocking
2868	let chat_result = chat::run_with(
2869	&mut terminal,
2870	engine,
2871	inference_backend,
2872	load_rx,
2873	startup_model_name,
2874	)
2875	.await;
2876
2877	// cleanup fd because backend's writer is private
2878	crossterm::execute!(cleanup_tty, crossterm::terminal::LeaveAlternateScreen)?;
2879	cleanup_tty.flush()?;
2880	crossterm::terminal::disable_raw_mode()?;
2881
2882	// restore real stdout/stderr for post-TUI error output
2883	#[cfg(unix)]
2884	{
2885	let cleanup_fd = cleanup_tty.as_raw_fd();
2886	unsafe {
2887	libc::dup2(cleanup_fd, libc::STDOUT_FILENO);
2888	libc::dup2(cleanup_fd, libc::STDERR_FILENO);
2889	}
2890	}
2891
2892	chat_result
2893	}
2894
2895	// ── ACP server mode ───────────────────────────────────────────────────────────
2896
2897	/// The on-device model `/load` should bring up by default: the persisted
2898	/// selection if it still resolves to a known local model, otherwise the built-in
2899	/// default (`qwen25_3b`).
2900	fn default_local_model_config() -> GgufModelConfig {
2901	setup::startup_model_selection()
2902	.as_ref()
2903	.and_then(\|selection\| {
2904	selection.selected_model.as_ref().and_then(\|selected\| {
2905	models::local_picker_items()
2906	.into_iter()
2907	.find(\|item\| {
2908	item.config.model_id == selected.model_id
2909	&& item
2910	.config
2911	.files
2912	.iter()
2913	.any(\|file\| file == &selected.gguf_file)
2914	})
2915	.map(\|item\| item.config)
2916	})
2917	})
2918	.unwrap_or_else(GgufModelConfig::qwen25_3b)
2919	}
2920
2921	async fn run_acp_server() -> anyhow::Result<()> {
2922	log::info!("ACP mode — starting agent server");
2923
2924	let config = default_local_model_config();
2925
2926	let needs_download = models::local_picker_items()
2927	.iter()
2928	.find(\|item\| item.config.model_id == config.model_id)
2929	.map(\|item\| item.cache_health != setup::ModelCacheHealth::Complete)
2930	.unwrap_or(true);
2931
2932	log::info!(
2933	"ACP startup model selected: {} ({})",
2934	config.display_name,
2935	if needs_download {
2936	"needs download"
2937	} else {
2938	"cached"
2939	}
2940	);
2941
2942	let engine = Arc::new(ChatEngine::new());
2943
2944	// The on-device model is never loaded implicitly; the user loads it with
2945	// `/load` (or by picking one in `/models`). So initialize/session/new stay
2946	// lightweight and `model_ready` starts true (nothing is loading).
2947	let model_ready = Arc::new(AtomicBool::new(true));
2948	let startup_model_load_started = Arc::new(AtomicBool::new(false));
2949	let model_load_error: Arc<std::sync::Mutex<Option<String>>> =
2950	Arc::new(std::sync::Mutex::new(None));
2951
2952	let state = Arc::new(SiGitAgent::new(
2953	engine,
2954	config,
2955	model_ready,
2956	startup_model_load_started,
2957	model_load_error,
2958	needs_download,
2959	));
2960
2961	// Honor the explicit provider override (OPENAI_BASE_URL/OPENAI_API_KEY or
2962	// an active providers.toml profile) in ACP mode too — the interactive
2963	// client already does. Without this the override was silently ignored here
2964	// and prompts insisted on a local model. It is also what lets the ACP
2965	// integration test drive the agent against a scripted endpoint
2966	// (tests/acp_permissions.rs). The model picker still shows the local
2967	// selection; overrides are a power-user escape hatch, not a tier.
2968	if let Some(cfg) = provider::active_provider() {
2969	log::info!(
2970	"inference: using {} (model {}) at {}",
2971	cfg.display_name,
2972	cfg.model,
2973	cfg.base_url
2974	);
2975	register_subagent_factory_for(&cfg);
2976	let override_backend: Arc<dyn InferenceBackend> = Arc::new(OpenAiBackend::new(
2977	cfg.base_url,
2978	cfg.api_key,
2979	cfg.model,
2980	Some(system_prompt_for_model(true).to_string()),
2981	));
2982	*state.backend.lock().await = override_backend;
2983	} else {
2984	// On-device inference has a single shared conversation history, so a
2985	// second concurrent subagent context is not possible yet.
2986	tools::set_subagent_factory(Box::new(\|\| None));
2987	}
2988
2989	let stdin = tokio::io::stdin().compat();
2990	let stdout = tokio::io::stdout().compat_write();
2991	let transport = ByteStreams::new(stdout, stdin);
2992
2993	Agent
2994	.builder()
2995	.on_receive_request(
2996	{
2997	let state = Arc::clone(&state);
2998	async move \|req: InitializeRequest, responder, _cx: ConnectionTo<Client>\| {
2999	handle_response(responder, state.handle_initialize(req).await)
3000	}
3001	},
3002	agent_client_protocol::on_receive_request!(),
3003	)
3004	.on_receive_request(
3005	{
3006	let state = Arc::clone(&state);
3007	async move \|req: AuthenticateRequest, responder, _cx: ConnectionTo<Client>\| {
3008	handle_response(responder, state.handle_authenticate(req).await)
3009	}
3010	},
3011	agent_client_protocol::on_receive_request!(),
3012	)
3013	// Turn-affecting handlers below run in spawned tasks, serialized by
3014	// `turn_lock`, so the dispatch loop stays free to route client
3015	// responses (permission answers) while a turn is in flight. Awaiting a
3016	// client request from inside a handler would deadlock: the dispatch
3017	// loop can't read the response while the handler blocks it.
3018	.on_receive_request(
3019	{
3020	let state = Arc::clone(&state);
3021	async move \|req: LoadSessionRequest, responder, cx: ConnectionTo<Client>\| {
3022	let state = Arc::clone(&state);
3023	let task_cx = cx.clone();
3024	cx.spawn(async move {
3025	let _turn = state.turn_lock.lock().await;
3026	handle_response(responder, state.handle_load_session(&task_cx, req).await)
3027	})
3028	}
3029	},
3030	agent_client_protocol::on_receive_request!(),
3031	)
3032	.on_receive_request(
3033	{
3034	let state = Arc::clone(&state);
3035	async move \|req: ForkSessionRequest, responder, cx: ConnectionTo<Client>\| {
3036	let state = Arc::clone(&state);
3037	let task_cx = cx.clone();
3038	cx.spawn(async move {
3039	let _turn = state.turn_lock.lock().await;
3040	handle_response(responder, state.handle_fork_session(&task_cx, req).await)
3041	})
3042	}
3043	},
3044	agent_client_protocol::on_receive_request!(),
3045	)
3046	.on_receive_request(
3047	{
3048	let state = Arc::clone(&state);
3049	async move \|req: NewSessionRequest, responder, cx: ConnectionTo<Client>\| {
3050	let state = Arc::clone(&state);
3051	let task_cx = cx.clone();
3052	cx.spawn(async move {
3053	let _turn = state.turn_lock.lock().await;
3054	handle_response(responder, state.handle_new_session(&task_cx, req).await)
3055	})
3056	}
3057	},
3058	agent_client_protocol::on_receive_request!(),
3059	)
3060	.on_receive_request(
3061	{
3062	let state = Arc::clone(&state);
3063	async move \|req: PromptRequest, responder, cx: ConnectionTo<Client>\| {
3064	let state = Arc::clone(&state);
3065	let task_cx = cx.clone();
3066	cx.spawn(async move {
3067	let _turn = state.turn_lock.lock().await;
3068	handle_response(responder, state.handle_prompt(&task_cx, req).await)
3069	})
3070	}
3071	},
3072	agent_client_protocol::on_receive_request!(),
3073	)
3074	.on_receive_request(
3075	{
3076	let state = Arc::clone(&state);
3077	async move \|req: SetSessionConfigOptionRequest,
3078	responder,
3079	cx: ConnectionTo<Client>\| {
3080	let state = Arc::clone(&state);
3081	let task_cx = cx.clone();
3082	cx.spawn(async move {
3083	let _turn = state.turn_lock.lock().await;
3084	handle_response(
3085	responder,
3086	state.handle_set_session_config_option(&task_cx, req).await,
3087	)
3088	})
3089	}
3090	},
3091	agent_client_protocol::on_receive_request!(),
3092	)
3093	.on_receive_notification(
3094	{
3095	let state = Arc::clone(&state);
3096	async move \|notif: CancelNotification, _cx: ConnectionTo<Client>\| {
3097	state.handle_cancel(notif).await
3098	}
3099	},
3100	agent_client_protocol::on_receive_notification!(),
3101	)
3102	.connect_to(transport)
3103	.await
3104	.map_err(\|e\| anyhow::anyhow!("ACP connection error: {e}"))?;
3105
3106	log::info!("siGit shutting down");
3107	Ok(())
3108	}
3109
3110	// ── Entry point ──────────────────────────────────────────────────────────────
3111
3112	#[tokio::main]
3113	async fn main() -> anyhow::Result<()> {
3114	// Account subcommands. The editor launches `sigit login` in an embedded
3115	// terminal for ACP terminal-based authentication; the same verbs are handy
3116	// directly from a shell. These must be handled before the TTY/ACP split.
3117	if let Some(verb) = std::env::args().nth(1) {
3118	match verb.as_str() {
3119	"login" => {
3120	init_logging(true);
3121	match account::interactive_login().await {
3122	Ok(email) => {
3123	println!("Signed in to siGit Code Cloud as {email}.");
3124	return Ok(());
3125	}
3126	Err(error) => {
3127	eprintln!("Login failed: {error}");
3128	std::process::exit(1);
3129	}
3130	}
3131	}
3132	"logout" => {
3133	init_logging(true);
3134	println!("{}", account::end_session().await);
3135	return Ok(());
3136	}
3137	"whoami" => {
3138	init_logging(true);
3139	println!("{}", account::status_line().await);
3140	return Ok(());
3141	}
3142	_ => {}
3143	}
3144	}
3145
3146	// Headless programmatic mode: `sigit -p "<prompt>"` runs one prompt and
3147	// exits. Plain stdio and cross-platform (unlike the Unix-only TUI), so it
3148	// is dispatched here, before the TTY/ACP split, like the account
3149	// subcommands above.
3150	let cli_args: Vec<String> = std::env::args().skip(1).collect();
3151	match headless::parse_args(&cli_args) {
3152	Ok(None) => {}
3153	Ok(Some(config)) => {
3154	// Logs to stderr; stdout stays clean for the assistant's answer.
3155	init_logging(false);
3156	// Enter --cwd before anything loads, so instruction files and
3157	// project-local MCP config resolve from it (the headless
3158	// counterpart of the ACP session cwd handling).
3159	if let Some(dir) = &config.cwd
3160	&& let Err(error) = std::env::set_current_dir(dir)
3161	{
3162	eprintln!("sigit: cannot use --cwd {}: {error}", dir.display());
3163	std::process::exit(2);
3164	}
3165	setup::setup_shared_model_cache();
3166	// Best-effort MCP discovery (incl. the official server), like the
3167	// other entry points, so MCP tools are offered to the model.
3168	mcp::init().await;
3169	log::info!(
3170	"siGit v{} starting (headless mode)",
3171	env!("CARGO_PKG_VERSION")
3172	);
3173	std::process::exit(headless::run(config).await);
3174	}
3175	Err(error) => {
3176	eprintln!("sigit: {error}\n{}", headless::USAGE);
3177	std::process::exit(2);
3178	}
3179	}
3180
3181	let is_tty = std::io::stdin().is_terminal();
3182
3183	if is_tty {
3184	// must redirect before any library code touches stdout
3185	#[cfg(unix)]
3186	{
3187	let (tty, cleanup_tty) = redirect_output_to_log()?;
3188	init_logging(true);
3189	setup::setup_shared_model_cache();
3190	// Best-effort: discover MCP servers (incl. the official one) before
3191	// the first turn so their tools are offered to the model.
3192	mcp::init().await;
3193	run_interactive(tty, cleanup_tty).await
3194	}
3195	#[cfg(not(unix))]
3196	{
3197	anyhow::bail!("interactive mode requires Unix (macOS / Linux)");
3198	}
3199	} else {
3200	// ACP mode: keep stdout untouched for protocol JSON only.
3201	// Logs already go to stderr via `init_logging(false)`.
3202	init_logging(false);
3203	setup::setup_shared_model_cache();
3204	// Best-effort MCP discovery (incl. the official server) before serving.
3205	mcp::init().await;
3206	log::info!("siGit v{} starting (ACP mode)", env!("CARGO_PKG_VERSION"));
3207	run_acp_server().await
3208	}
3209	}
3210
3211	#[cfg(test)]
3212	mod tests {
3213	use super::*;
3214
3215	#[test]
3216	fn system_prompt_advertises_the_commit_co_author_trailer() {
3217	// The prompt instructs the model with the exact trailer that
3218	// `tools::ensure_commit_co_author` enforces; if the two drift apart the
3219	// safety net would re-amend commits the model already attributed.
3220	assert!(
3221	SYSTEM_PROMPT.contains(tools::COMMIT_CO_AUTHOR_TRAILER),
3222	"SYSTEM_PROMPT must quote tools::COMMIT_CO_AUTHOR_TRAILER verbatim"
3223	);
3224	}
3225
3226	#[test]
3227	fn ascii_safe_replaces_multibyte_chars() {
3228	// The exact label that crashed Zed: the cloud tier name plus the old
3229	// "[☁ siGit Code Cloud]" badge. After sanitizing it must be pure ASCII so
3230	// Zed's fixed byte-offset truncation can never split a glyph.
3231	let crashing = "siGit Code Cloud · Balanced [☁ siGit Code Cloud]";
3232	let safe = ascii_safe(crashing);
3233	assert!(safe.is_ascii(), "sanitized label must be ASCII: {safe:?}");
3234	assert_eq!(safe, "siGit Code Cloud - Balanced [- siGit Code Cloud]");
3235	}
3236
3237	#[test]
3238	fn ascii_safe_leaves_ascii_untouched() {
3239	let plain = "Qwen 2.5 3B [Onde]";
3240	assert_eq!(ascii_safe(plain), plain);
3241	}
3242
3243	#[test]
3244	fn ascii_safe_output_has_only_char_boundaries() {
3245	// Every byte index in an ASCII string is a valid char boundary, so any
3246	// downstream truncation is panic-free regardless of where it cuts.
3247	let safe = ascii_safe("Onde · ◉ ↓ ☁ ○ test");
3248	for i in 0..=safe.len() {
3249	assert!(safe.is_char_boundary(i));
3250	}
3251	}
3252	}