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

@setoelkahfi / sigit

claude/sigit-cloud-agent-expansion-reox0a

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 3,223 lines 135 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(&session_id.to_string(), &tc.name) {
1377	permissions::Decision::Allow => {
1378	tools::execute_tool(&tc.name, &tc.arguments).await
1379	}
1380	permissions::Decision::Deny(reason) => {
1381	log::info!(" ✗ {} denied by policy", tc.name);
1382	reason
1383	}
1384	permissions::Decision::Ask => {
1385	match self
1386	.request_tool_permission(cx, &session_id, &tc.name, &tc.arguments)
1387	.await
1388	{
1389	PermissionVerdict::Approved => {
1390	tools::execute_tool(&tc.name, &tc.arguments).await
1391	}
1392	PermissionVerdict::Denied(reason) => {
1393	log::info!(" ✗ {} denied by user", tc.name);
1394	reason
1395	}
1396	PermissionVerdict::TurnCancelled => {
1397	log::info!("prompt({}) cancelled at permission gate", session_id);
1398	// The assistant message carrying these tool
1399	// calls is already in the backend history;
1400	// leaving any of them unanswered makes strict
1401	// OpenAI-compatible endpoints reject every
1402	// later request in the session. Close out this
1403	// call and the ones this round never reached.
1404	for pending in &result.tool_calls[call_index..] {
1405	tool_results.push(BackendToolResult {
1406	tool_call_id: pending.id.clone(),
1407	content: format!(
1408	"`{}` was not executed: the user cancelled the turn \
1409	at the permission prompt.",
1410	pending.name
1411	),
1412	});
1413	}
1414	backend.record_cancelled_tool_results(tool_results).await;
1415	return Ok(PromptResponse::new(StopReason::Cancelled));
1416	}
1417	}
1418	}
1419	};
1420
1421	log::info!(" ← {} chars", output.len());
1422
1423	tool_results.push(BackendToolResult {
1424	tool_call_id: tc.id.clone(),
1425	content: output,
1426	});
1427	}
1428
1429	let next_tools = if round < MAX_TOOL_ROUNDS {
1430	Some(tools.as_slice())
1431	} else {
1432	None // last round: force text
1433	};
1434
1435	result = self
1436	.drain_turn(
1437	cx,
1438	&session_id,
1439	backend.send_tool_results(tool_results, next_tools, Some(&sink)),
1440	&mut sink_rx,
1441	&mut assembled,
1442	&mut sent,
1443	&mut streamed_any,
1444	)
1445	.await
1446	.map_err(\|e\| agent_client_protocol::Error::new(-32603, e.to_string()))?;
1447	}
1448
1449	// ── Final text response ───────────────────────────────────────────
1450	// If anything streamed, the visible reply is already on the wire; only
1451	// send a trailing block for the non-streamed path (e.g. on-device direct
1452	// answers, which onde can't stream while tools are on offer).
1453	if !streamed_any {
1454	let reply_text = result.text.trim().to_string();
1455	let final_text = if reply_text.is_empty() {
1456	if round > 0 {
1457	log::warn!(
1458	"prompt({}) — model returned empty reply after {} tool round(s)",
1459	session_id,
1460	round
1461	);
1462	"Something went wrong — the edits didn't go through. Try rephrasing what you need, or point me at the specific lines.".to_string()
1463	} else {
1464	log::warn!(
1465	"prompt({}) — model returned empty reply (no tool rounds)",
1466	session_id
1467	);
1468	String::new()
1469	}
1470	} else {
1471	// strip <think> blocks so reasoning tokens stay hidden
1472	let (_think, visible) = chat::strip_think_blocks(&reply_text);
1473	visible
1474	};
1475
1476	if !final_text.is_empty() {
1477	self.send_assistant_message(cx, session_id.clone(), final_text)
1478	.ok();
1479	}
1480	}
1481
1482	// Persist the completed turn so a restart (or session/load) can pick
1483	// the conversation back up.
1484	let snapshot = backend.history_snapshot().await;
1485	if let Err(error) = session_store::save(&session_id.to_string(), &snapshot) {
1486	log::warn!("prompt({}) session save failed: {error}", session_id);
1487	}
1488
1489	log::info!("prompt({}) complete — {} tool round(s)", session_id, round);
1490	Ok(PromptResponse::new(StopReason::EndTurn))
1491	}
1492
1493	/// Ask the ACP client for permission to run one tool call. Presents
1494	/// allow-once / allow-for-session / deny; an "always allow" choice is
1495	/// recorded via [`permissions::grant_for_session`]. Only safe to call from
1496	/// a spawned task (see the handler registration in `run_acp_server`): the
1497	/// dispatch loop must be free to route the client's answer back to us.
1498	async fn request_tool_permission(
1499	&self,
1500	cx: &ConnectionTo<Client>,
1501	session_id: &SessionId,
1502	tool_name: &str,
1503	arguments: &str,
1504	) -> PermissionVerdict {
1505	// The user decides from this dialog, so show the arguments with any
1506	// truncation flagged (a silently clipped command could hide its tail
1507	// from the person approving it). The full arguments also travel as
1508	// `raw_input` for clients that render it.
1509	let args_preview = permissions::approval_preview(arguments);
1510	let title = if args_preview.is_empty() {
1511	tool_name.to_string()
1512	} else {
1513	format!("{tool_name}({args_preview})")
1514	};
1515	let raw_input: serde_json::Value = serde_json::from_str(arguments)
1516	.unwrap_or_else(\|_\| serde_json::Value::String(arguments.to_string()));
1517
1518	let request = RequestPermissionRequest::new(
1519	session_id.clone(),
1520	ToolCallUpdate::new(
1521	format!("perm-{}", uuid::Uuid::new_v4()),
1522	ToolCallUpdateFields::new()
1523	.title(title)
1524	.kind(tool_kind_for(tool_name))
1525	.status(ToolCallStatus::Pending)
1526	.raw_input(raw_input),
1527	),
1528	vec![
1529	PermissionOption::new("allow_once", "Allow once", PermissionOptionKind::AllowOnce),
1530	PermissionOption::new(
1531	"allow_session",
1532	"Allow for this session",
1533	PermissionOptionKind::AllowAlways,
1534	),
1535	PermissionOption::new("reject_once", "Deny", PermissionOptionKind::RejectOnce),
1536	],
1537	);
1538
1539	match cx.send_request(request).block_task().await {
1540	Ok(response) => match response.outcome {
1541	RequestPermissionOutcome::Selected(selected) => {
1542	match selected.option_id.0.as_ref() {
1543	"allow_once" => PermissionVerdict::Approved,
1544	"allow_session" => {
1545	permissions::grant_for_session(&session_id.to_string(), tool_name);
1546	PermissionVerdict::Approved
1547	}
1548	_ => PermissionVerdict::Denied(permissions::user_denial(tool_name)),
1549	}
1550	}
1551	RequestPermissionOutcome::Cancelled => PermissionVerdict::TurnCancelled,
1552	// The outcome enum is non_exhaustive; treat anything unknown as
1553	// a denial rather than running a mutating tool unapproved.
1554	_ => PermissionVerdict::Denied(permissions::user_denial(tool_name)),
1555	},
1556	Err(error) => {
1557	log::warn!("permission request for `{tool_name}` failed: {error}");
1558	PermissionVerdict::Denied(format!(
1559	"`{tool_name}` was not executed: this client could not answer the \
1560	permission request ({error}). The user can pre-approve tools in \
1561	settings.toml under [permissions], or set SIGIT_PERMISSIONS=allow \
1562	for clients without permission support."
1563	))
1564	}
1565	}
1566	}
1567
1568	async fn handle_cancel(&self, args: CancelNotification) -> agent_client_protocol::Result<()> {
1569	log::info!("cancel requested for session {}", args.session_id);
1570	Ok(())
1571	}
1572
1573	/// Swap the active backend to a siGit Code Cloud tier and reflect it as the
1574	/// current model so the picker shows it selected. Returns the tier's display
1575	/// name on success, or `None` when no account is signed in (caller prompts
1576	/// for login). Shared by the panel picker and the `/models` slash command.
1577	async fn switch_to_cloud_tier(&self, tier: &str) -> Option<String> {
1578	let cfg = crate::provider::cloud_tier_provider(tier)?;
1579	let mut system_prompt = system_prompt_for_model(true).to_string();
1580	// Mirror the cwd guidance and project instruction files the local engine
1581	// gets at session load, so the cloud model shares the same project context.
1582	if let Some(cwd) = self.session_cwd.lock().ok().and_then(\|g\| g.clone()) {
1583	system_prompt.push_str("\n\n");
1584	system_prompt.push_str(&session_context_message(&cwd));
1585	}
1586	let cloud_backend: Arc<dyn InferenceBackend> = Arc::new(OpenAiBackend::new(
1587	cfg.base_url,
1588	cfg.api_key,
1589	cfg.model,
1590	Some(system_prompt),
1591	));
1592	*self.backend.lock().await = cloud_backend;
1593
1594	let cloud_config = GgufModelConfig {
1595	model_id: format!("sigit-cloud:{tier}"),
1596	files: Vec::new(),
1597	tok_model_id: None,
1598	display_name: cfg.display_name.clone(),
1599	approx_memory: "Cloud".to_string(),
1600	chat_template: None,
1601	};
1602	{
1603	let mut guard = self.current_model.lock().unwrap();
1604	*guard = cloud_config;
1605	}
1606
1607	// Explicitly choosing a cloud tier puts us in cloud mode.
1608	let _ = settings::set_local_inference(false);
1609
1610	log::info!("switched to cloud tier {tier}");
1611	Some(cfg.display_name)
1612	}
1613
1614	/// Apply the persisted Local Inference mode at session start. When local
1615	/// inference is off and an account is signed in, route to a cloud tier so the
1616	/// on-device model is never loaded; otherwise leave the on-device backend in
1617	/// place. Call after the session cwd is set so the cloud system prompt picks
1618	/// it up. Does not flip the stored setting on the not-signed-in fallback.
1619	async fn apply_startup_inference_mode(&self) {
1620	if settings::local_inference_enabled() {
1621	return;
1622	}
1623	if self.switch_to_cloud_tier("balanced").await.is_some() {
1624	log::info!("startup: local inference off; routing inference to siGit Code Cloud");
1625	} else {
1626	log::warn!(
1627	"local inference is off but no account is signed in; staying on-device. \
1628	Run /login or set Local Inference on."
1629	);
1630	}
1631	}
1632
1633	/// Route inference back on-device. Used after leaving a cloud tier for a
1634	/// local model. The `LocalBackend` reads the live `engine`, so this just
1635	/// repoints the active backend.
1636	async fn reset_to_local_backend(&self) {
1637	let local_backend: Arc<dyn InferenceBackend> =
1638	Arc::new(LocalBackend::new(Arc::clone(&self.engine)));
1639	*self.backend.lock().await = local_backend;
1640	}
1641
1642	/// Re-attempt the lazy startup model load if the previous attempt failed.
1643	/// Clears the one-shot guard so the next load runs; a healthy load is left
1644	/// untouched so `/reload` doesn't needlessly reload a working model.
1645	fn retry_startup_model_load_if_failed(&self) {
1646	let had_error = self
1647	.model_load_error
1648	.lock()
1649	.map(\|guard\| guard.is_some())
1650	.unwrap_or(false);
1651	if had_error {
1652	self.startup_model_load_started
1653	.store(false, Ordering::Release);
1654	self.start_startup_model_load_if_needed();
1655	}
1656	}
1657
1658	/// Re-sync session state in place — no new session needed. Re-applies the
1659	/// active backend from current credentials (so a fresh `/login` token is
1660	/// picked up), retries a failed model load, and pushes refreshed commands +
1661	/// picker so the editor's UI reflects the current state.
1662	async fn handle_reload(&self, cx: &ConnectionTo<Client>, session_id: SessionId) {
1663	let signed_in = account::status_line().await;
1664
1665	let on_cloud_tier = {
1666	let guard = self.current_model.lock().unwrap();
1667	guard
1668	.model_id
1669	.strip_prefix("sigit-cloud:")
1670	.map(str::to_string)
1671	};
1672
1673	let backend_note = match on_cloud_tier {
1674	Some(tier) => match self.switch_to_cloud_tier(&tier).await {
1675	Some(name) => format!("Active: {name}."),
1676	None => {
1677	self.reset_to_local_backend().await;
1678	"Signed out — back to on-device. Pick a model with /models.".to_string()
1679	}
1680	},
1681	None => {
1682	self.reset_to_local_backend().await;
1683	self.retry_startup_model_load_if_failed();
1684	let guard = self.current_model.lock().unwrap();
1685	format!("Active: {}.", guard.display_name)
1686	}
1687	};
1688
1689	// Push refreshed picker + commands so the editor reflects current state.
1690	let config_options = {
1691	let guard = self.current_model.lock().unwrap();
1692	build_model_config_options(&guard)
1693	};
1694	self.send_tool_call_update(
1695	cx,
1696	session_id.clone(),
1697	SessionUpdate::ConfigOptionUpdate(ConfigOptionUpdate::new(config_options)),
1698	)
1699	.ok();
1700	self.advertise_commands(cx, session_id.clone());
1701
1702	self.send_assistant_message(
1703	cx,
1704	session_id,
1705	format!("Reloaded. {signed_in} {backend_note}"),
1706	)
1707	.ok();
1708	}
1709
1710	async fn handle_set_session_config_option(
1711	&self,
1712	cx: &ConnectionTo<Client>,
1713	args: SetSessionConfigOptionRequest,
1714	) -> agent_client_protocol::Result<SetSessionConfigOptionResponse> {
1715	log::info!(
1716	"set_session_config_option: config_id={}, value={:?}",
1717	args.config_id,
1718	args.value
1719	);
1720
1721	// ── Local Inference toggle ──────────────────────────────────────────
1722	if args.config_id.0.as_ref() == LOCAL_INFERENCE_CONFIG_ID {
1723	let enabled = match args.value.0.as_ref() {
1724	LOCAL_INFERENCE_ON => true,
1725	LOCAL_INFERENCE_OFF => false,
1726	other => {
1727	return Err(agent_client_protocol::Error::new(
1728	-32602,
1729	format!("unknown Local Inference value: {other}"),
1730	));
1731	}
1732	};
1733	if let Err(error) = settings::set_local_inference(enabled) {
1734	return Err(agent_client_protocol::Error::new(
1735	-32603,
1736	format!("could not save Local Inference setting: {error}"),
1737	));
1738	}
1739	let message = if enabled {
1740	"Local inference is on. On-device models are highlighted; pick one from Model."
1741	} else {
1742	"Local inference is off. siGit Code Cloud tiers are highlighted; pick one from Model."
1743	};
1744	self.send_assistant_message(cx, args.session_id.clone(), format!("\n\n{message}"))
1745	.ok();
1746	// Rebuild so the Model picker reflects the new emphasis/order.
1747	let current = self.current_model.lock().unwrap().clone();
1748	let config_options = build_model_config_options(&current);
1749	return Ok(SetSessionConfigOptionResponse::new(config_options));
1750	}
1751
1752	if args.config_id.0.as_ref() != MODEL_CONFIG_ID {
1753	return Err(agent_client_protocol::Error::new(
1754	-32602,
1755	format!("unknown config option: {}", args.config_id.0),
1756	));
1757	}
1758
1759	let model_id = args.value.0.as_ref();
1760
1761	// can't switch while the startup model is still loading — the old
1762	// weights are in GPU memory and the new load gets "does not fit"
1763	if self.startup_model_load_started.load(Ordering::Acquire)
1764	&& !self.model_ready.load(Ordering::Acquire)
1765	{
1766	log::info!("set_session_config_option: waiting for startup model to finish loading");
1767	while !self.model_ready.load(Ordering::Acquire) {
1768	tokio::time::sleep(std::time::Duration::from_millis(200)).await;
1769	}
1770	}
1771
1772	// Zed re-fires the last selection when a thread opens. That re-fire must
1773	// not load anything: on-device models are loaded only on an explicit
1774	// request (`/load`, or actively picking a different model below), so a
1775	// re-fire of the already-current selection is a no-op. Otherwise opening a
1776	// new thread would silently load the local model — exactly what we avoid.
1777	{
1778	let current = self.current_model.lock().unwrap();
1779	if current.model_id == model_id {
1780	log::info!(
1781	"set_session_config_option: {} is already the active selection, skipping",
1782	current.display_name
1783	);
1784	let config_options = build_model_config_options(&current);
1785	return Ok(SetSessionConfigOptionResponse::new(config_options));
1786	}
1787	}
1788
1789	// ── siGit Code Cloud tier: no local load; sign-in gated ─────────────
1790	if let Some(tier) = model_id.strip_prefix("sigit-cloud:") {
1791	let message = match self.switch_to_cloud_tier(tier).await {
1792	Some(display_name) => format!("Switched to {display_name}."),
1793	None => CLOUD_LOGIN_PROMPT.to_string(),
1794	};
1795	// Start on a fresh line: ACP clients concatenate consecutive
1796	// agent-message chunks into one block, so without this the switch
1797	// confirmation runs onto the end of the previous assistant message.
1798	self.send_assistant_message(cx, args.session_id.clone(), format!("\n\n{message}"))
1799	.ok();
1800
1801	let current = self.current_model.lock().unwrap().clone();
1802	let config_options = build_model_config_options(&current);
1803	return Ok(SetSessionConfigOptionResponse::new(config_options));
1804	}
1805
1806	let needs_download = models::local_picker_items()
1807	.into_iter()
1808	.find(\|item\| item.config.model_id == model_id)
1809	.map(\|item\| item.cache_health == setup::ModelCacheHealth::NotDownloaded)
1810	.unwrap_or(false);
1811
1812	// tells the progress poller to stop
1813	let stop_flag = Arc::new(AtomicBool::new(false));
1814
1815	let tool_call_id = format!("model-switch-{}", uuid::Uuid::new_v4());
1816
1817	if needs_download {
1818	let model_id_owned = model_id.to_string();
1819	let expected_bytes = onde::inference::models::SUPPORTED_MODEL_INFO
1820	.iter()
1821	.find(\|m\| m.id == model_id_owned)
1822	.map(\|m\| m.expected_size_bytes)
1823	.unwrap_or(0);
1824
1825	let display_name = models::local_picker_items()
1826	.into_iter()
1827	.find(\|item\| item.config.model_id == model_id_owned)
1828	.map(\|item\| item.display_name.clone())
1829	.unwrap_or_else(\|\| model_id_owned.clone());
1830
1831	let size_hint = if expected_bytes > 0 {
1832	format!(" (~{})", format_size_human(expected_bytes))
1833	} else {
1834	String::new()
1835	};
1836
1837	self.send_tool_call_update(
1838	cx,
1839	args.session_id.clone(),
1840	SessionUpdate::ToolCall(
1841	ToolCall::new(
1842	tool_call_id.clone(),
1843	format!("⏬ Downloading {display_name}{size_hint}"),
1844	)
1845	.kind(ToolKind::Think)
1846	.status(ToolCallStatus::InProgress)
1847	.content(vec![
1848	format!(
1849	"Preparing download for {display_name}. This may take a few minutes."
1850	)
1851	.into(),
1852	]),
1853	),
1854	)
1855	.ok();
1856
1857	// poll download progress and update the spinner in Zed
1858	let cx_for_poller = cx.clone();
1859	let poller_session = args.session_id.clone();
1860	let poller_model_id = model_id_owned.clone();
1861	let poller_stop = Arc::clone(&stop_flag);
1862	let poller_tool_call_id = tool_call_id.clone();
1863
1864	cx.spawn(async move {
1865	const SPINNER: &[char] = &['⠋', '⠙', '⠹', '⠸', '⠼', '⠴', '⠦', '⠧', '⠇', '⠏'];
1866	let cache_path = onde::hf_cache::model_cache_path(&poller_model_id);
1867	let mut tick: usize = 0;
1868	let mut interval = tokio::time::interval(std::time::Duration::from_secs(1));
1869	interval.tick().await; // consume the immediate first tick
1870
1871	while !poller_stop.load(Ordering::Relaxed) {
1872	interval.tick().await;
1873
1874	if poller_stop.load(Ordering::Relaxed) {
1875	break;
1876	}
1877
1878	let downloaded = cache_path
1879	.as_ref()
1880	.filter(\|p\| p.exists())
1881	.map(\|p\| dir_size_recursive(p))
1882	.unwrap_or(0);
1883
1884	let frame = SPINNER[tick % SPINNER.len()];
1885	tick += 1;
1886
1887	let title = if expected_bytes > 0 {
1888	let pct =
1889	((downloaded as f64 / expected_bytes as f64) * 100.0).min(99.0) as u8;
1890	format!("{frame} Downloading {display_name}{size_hint} ({pct}%)")
1891	} else {
1892	format!("{frame} Downloading {display_name}{size_hint}")
1893	};
1894
1895	let msg = if expected_bytes > 0 {
1896	let pct =
1897	((downloaded as f64 / expected_bytes as f64) * 100.0).min(99.0) as u8;
1898	let bar = progress_bar(pct, 20);
1899	format!(
1900	"{display_name} — {bar} {pct}% ({} / {})",
1901	format_size_human(downloaded),
1902	format_size_human(expected_bytes),
1903	)
1904	} else {
1905	format!(
1906	"{display_name} — {} downloaded…",
1907	format_size_human(downloaded)
1908	)
1909	};
1910
1911	let notification = SessionNotification::new(
1912	poller_session.clone(),
1913	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
1914	poller_tool_call_id.clone(),
1915	ToolCallUpdateFields::new()
1916	.title(title)
1917	.status(ToolCallStatus::InProgress)
1918	.content(vec![msg.into()]),
1919	)),
1920	);
1921	if cx_for_poller.send_notification(notification).is_err() {
1922	break;
1923	}
1924	}
1925	Ok(())
1926	})
1927	.ok();
1928	}
1929
1930	// cached models still take 10-30s to load weights; show a spinner
1931	if !needs_download {
1932	let cached_display_name = models::local_picker_items()
1933	.into_iter()
1934	.find(\|item\| item.config.model_id == model_id)
1935	.map(\|item\| item.display_name.clone())
1936	.unwrap_or_else(\|\| model_id.to_string());
1937
1938	self.send_tool_call_update(
1939	cx,
1940	args.session_id.clone(),
1941	SessionUpdate::ToolCall(
1942	ToolCall::new(
1943	tool_call_id.clone(),
1944	format!("Loading {cached_display_name}"),
1945	)
1946	.kind(ToolKind::Think)
1947	.status(ToolCallStatus::InProgress)
1948	.content(vec![format!("Loading {cached_display_name}…").into()]),
1949	),
1950	)
1951	.ok();
1952
1953	// tick every 5s so the user knows we haven't frozen
1954	let cx_for_spinner = cx.clone();
1955	let spinner_session = args.session_id.clone();
1956	let spinner_name = cached_display_name.clone();
1957	let spinner_stop = Arc::clone(&stop_flag);
1958	let spinner_tool_call_id = tool_call_id.clone();
1959	let load_start = std::time::Instant::now();
1960
1961	cx.spawn(async move {
1962	const SPINNER: &[char] = &['⠋', '⠙', '⠹', '⠸', '⠼', '⠴', '⠦', '⠧', '⠇', '⠏'];
1963	let mut tick: usize = 0;
1964	let mut interval = tokio::time::interval(std::time::Duration::from_secs(5));
1965	interval.tick().await; // consume the immediate first tick
1966
1967	while !spinner_stop.load(Ordering::Relaxed) {
1968	interval.tick().await;
1969
1970	if spinner_stop.load(Ordering::Relaxed) {
1971	break;
1972	}
1973
1974	let elapsed = load_start.elapsed();
1975	let elapsed_str = if elapsed.as_secs() >= 60 {
1976	format!("{}m {:02}s", elapsed.as_secs() / 60, elapsed.as_secs() % 60)
1977	} else {
1978	format!("{}s", elapsed.as_secs())
1979	};
1980	let frame = SPINNER[tick % SPINNER.len()];
1981	tick += 1;
1982
1983	let msg = format!("{frame} Loading {spinner_name}… ({elapsed_str})");
1984	let notification = SessionNotification::new(
1985	spinner_session.clone(),
1986	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
1987	spinner_tool_call_id.clone(),
1988	ToolCallUpdateFields::new()
1989	.status(ToolCallStatus::InProgress)
1990	.content(vec![msg.into()]),
1991	)),
1992	);
1993	if cx_for_spinner.send_notification(notification).is_err() {
1994	break;
1995	}
1996	}
1997	Ok(())
1998	})
1999	.ok();
2000	}
2001
2002	let switch_result = self.switch_model_by_id(model_id).await;
2003
2004	stop_flag.store(true, Ordering::Relaxed);
2005
2006	match switch_result {
2007	Ok(new_config) => {
2008	// Route inference back on-device (in case we were on a cloud tier).
2009	self.reset_to_local_backend().await;
2010	// Selecting an on-device model puts us in local mode.
2011	let _ = settings::set_local_inference(true);
2012
2013	let completion_title = if needs_download {
2014	format!("✓ {} downloaded and loaded", new_config.display_name)
2015	} else {
2016	format!("✓ Switched to {}", new_config.display_name)
2017	};
2018	let completion_body = if needs_download {
2019	format!("✓ {} downloaded and loaded.", new_config.display_name)
2020	} else {
2021	format!("✓ Switched to {}.", new_config.display_name)
2022	};
2023
2024	self.send_tool_call_update(
2025	cx,
2026	args.session_id.clone(),
2027	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
2028	tool_call_id,
2029	ToolCallUpdateFields::new()
2030	.title(completion_title)
2031	.status(ToolCallStatus::Completed)
2032	.content(vec![completion_body.into()]),
2033	)),
2034	)
2035	.ok();
2036
2037	let config_options = {
2038	let guard = self.current_model.lock().unwrap();
2039	build_model_config_options(&guard)
2040	};
2041
2042	log::info!("model switch complete");
2043	Ok(SetSessionConfigOptionResponse::new(config_options))
2044	}
2045	Err(err) => {
2046	self.send_tool_call_update(
2047	cx,
2048	args.session_id.clone(),
2049	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
2050	tool_call_id,
2051	ToolCallUpdateFields::new()
2052	.title("Model switch failed".to_string())
2053	.status(ToolCallStatus::Failed)
2054	.content(vec![format!("error loading model: {}", err.message).into()]),
2055	)),
2056	)
2057	.ok();
2058
2059	Err(err)
2060	}
2061	}
2062	}
2063	}
2064
2065	// ── Config option helpers ─────────────────────────────────────────────────────
2066
2067	/// config option ID for the model picker in Zed's agent panel
2068	const MODEL_CONFIG_ID: &str = "sigit-model";
2069
2070	/// config option ID for the Local Inference on/off toggle. Surfaced as a
2071	/// two-option `select` so ACP clients without slash-command support (e.g. Xcode)
2072	/// can still flip the mode from the agent panel.
2073	const LOCAL_INFERENCE_CONFIG_ID: &str = "sigit-local-inference";
2074
2075	/// `select` value ids for the Local Inference toggle.
2076	const LOCAL_INFERENCE_ON: &str = "local-inference-on";
2077	const LOCAL_INFERENCE_OFF: &str = "local-inference-off";
2078
2079	/// Replace non-ASCII chars so a downstream byte-index truncation can't split a
2080	/// multi-byte char. Zed slices the model-picker label at a fixed byte offset
2081	/// (`agent_ui/src/config_options.rs`) and panics — crashing the whole editor —
2082	/// when the cut lands mid-glyph (e.g. inside `☁` or `·`). Mapping to `-` keeps
2083	/// separators readable; ASCII bytes are always char boundaries.
2084	fn ascii_safe(s: &str) -> String {
2085	s.chars()
2086	.map(\|c\| if c.is_ascii() { c } else { '-' })
2087	.collect()
2088	}
2089
2090	fn build_model_config_options(current_model: &GgufModelConfig) -> Vec<SessionConfigOption> {
2091	// The full list, including the siGit Code Cloud tiers, so the panel picker
2092	// mirrors the TUI `/models`. Cloud entries are sign-in gated at selection.
2093	let items = models::build_model_picker_items();
2094	let active_kind = models::active_inference_kind();
2095
2096	let options: Vec<SessionConfigSelectOption> = items
2097	.iter()
2098	.filter(\|item\| item.cache_health != setup::ModelCacheHealth::Incomplete)
2099	.map(\|item\| {
2100	let mut desc_parts = Vec::new();
2101	// Mark options in the inactive mode so the active group reads as the
2102	// recommended set (the list is already ordered active-group-first).
2103	if item.source.kind() != active_kind {
2104	desc_parts.push("inactive mode".to_string());
2105	}
2106	if item.tool_calling {
2107	desc_parts.push("tool calling".to_string());
2108	}
2109	desc_parts.push(item.description.clone());
2110	if item.cache_health == setup::ModelCacheHealth::NotDownloaded {
2111	desc_parts.push("download on select".to_string());
2112	}
2113	// ASCII-only for the same reason as the name (see `ascii_safe`).
2114	let description = ascii_safe(&desc_parts.join(" - "));
2115	// Keep badges ASCII: Zed truncates the picker label at a fixed byte
2116	// offset and panics if the cut splits a multi-byte char. See
2117	// `ascii_safe` below.
2118	let source_badge = if item.cloud_tier.is_some() {
2119	" [siGit Code Cloud]"
2120	} else if item.cache_health == setup::ModelCacheHealth::NotDownloaded {
2121	" [Onde]"
2122	} else {
2123	match item.source_label.as_str() {
2124	"Onde" => " [Onde]",
2125	"HuggingFace" => " [HuggingFace]",
2126	_ => "",
2127	}
2128	};
2129	// For cloud tiers use just the tier title (e.g. "Balanced") so the
2130	// label reads "Balanced [siGit Code Cloud]" instead of repeating the
2131	// brand. The display name can carry non-ASCII (the cloud tier label
2132	// is "siGit Code Cloud · Balanced"), so sanitize the whole label.
2133	let base_name = match &item.cloud_tier {
2134	Some(tier) => crate::provider::tier_title(tier),
2135	None => item.display_name.clone(),
2136	};
2137	let name = ascii_safe(&format!("{base_name}{source_badge}"));
2138	SessionConfigSelectOption::new(
2139	SessionConfigValueId::new(item.config.model_id.as_str()),
2140	name,
2141	)
2142	.description(description)
2143	})
2144	.collect();
2145
2146	// Local Inference on/off toggle, modeled as a two-option select so panel-only
2147	// ACP clients (no slash commands) can flip the mode.
2148	let local_on = settings::local_inference_enabled();
2149	let local_current = SessionConfigValueId::new(if local_on {
2150	LOCAL_INFERENCE_ON
2151	} else {
2152	LOCAL_INFERENCE_OFF
2153	});
2154	let local_options = vec![
2155	SessionConfigSelectOption::new(
2156	SessionConfigValueId::new(LOCAL_INFERENCE_ON),
2157	"On (on-device)".to_string(),
2158	)
2159	.description("Run inference on-device; on-device models are highlighted".to_string()),
2160	SessionConfigSelectOption::new(
2161	SessionConfigValueId::new(LOCAL_INFERENCE_OFF),
2162	"Off (siGit Code Cloud)".to_string(),
2163	)
2164	.description("Use siGit Code Cloud; cloud tiers are highlighted".to_string()),
2165	];
2166	let local_option = SessionConfigOption::select(
2167	LOCAL_INFERENCE_CONFIG_ID,
2168	"Local Inference",
2169	local_current,
2170	local_options,
2171	)
2172	.description("Toggle on-device inference; changes which models are highlighted");
2173
2174	if options.is_empty() {
2175	return vec![local_option];
2176	}
2177
2178	let current_value = SessionConfigValueId::new(current_model.model_id.as_str());
2179
2180	vec![
2181	SessionConfigOption::select(MODEL_CONFIG_ID, "Model", current_value, options)
2182	.category(SessionConfigOptionCategory::Model)
2183	.description("Select an on-device model or a siGit Code Cloud tier"),
2184	local_option,
2185	]
2186	}
2187
2188	/// returns `(config, max_tokens, tool_calling)` for a picker model_id, or None
2189	fn resolve_model_config(model_id: &str) -> Option<(GgufModelConfig, u64, bool)> {
2190	let items = models::local_picker_items();
2191	items
2192	.into_iter()
2193	.find(\|item\| {
2194	item.config.model_id == model_id
2195	&& item.cache_health != setup::ModelCacheHealth::Incomplete
2196	})
2197	.map(\|item\| (item.config, item.max_tokens, item.tool_calling))
2198	}
2199
2200	// ── Slash commands ────────────────────────────────────────────────────────────
2201
2202	#[derive(Debug, Clone)]
2203	enum SlashCommand {
2204	Help,
2205	Clear,
2206	Status,
2207	Models(Option<usize>),
2208	/// toggle on-device inference mode. `Some(true/false)` sets it, `None` flips it.
2209	Local(Option<bool>),
2210	/// List discovered Agent Skills.
2211	Skills,
2212	/// List configured MCP servers and their tools.
2213	Mcp,
2214	/// Explicitly load the selected (or default) on-device model.
2215	Load,
2216	/// `/login <email> <password>` — the raw argument, parsed when executed.
2217	Login(Option<String>),
2218	Logout,
2219	Whoami,
2220	/// Re-sync session state (auth, backend, picker) without a new session.
2221	Reload,
2222	/// Toggle plan mode (read-only research; mutating tools denied with a
2223	/// prompt to present a plan). `Some(true/false)` sets it, `None` flips it.
2224	Plan(Option<bool>),
2225	/// Show the effective permission policy for this session.
2226	Permissions,
2227	/// Summarize-and-shrink the conversation history on demand.
2228	Compact,
2229	Exit,
2230	Unknown(String),
2231	}
2232
2233	fn parse_slash(input: &str) -> Option<SlashCommand> {
2234	let trimmed = input.trim();
2235	if !trimmed.starts_with('/') {
2236	return None;
2237	}
2238	let mut parts = trimmed.splitn(2, char::is_whitespace);
2239	let command = parts.next().unwrap_or("");
2240	let argument = parts.next().map(str::trim);
2241	Some(match command {
2242	"/help" => SlashCommand::Help,
2243	"/clear" => SlashCommand::Clear,
2244	"/status" => SlashCommand::Status,
2245	"/models" => SlashCommand::Models(argument.and_then(\|v\| v.parse::<usize>().ok())),
2246	"/local" => SlashCommand::Local(parse_on_off(argument)),
2247	"/skills" => SlashCommand::Skills,
2248	"/mcp" => SlashCommand::Mcp,
2249	"/load" => SlashCommand::Load,
2250	"/login" => SlashCommand::Login(argument.map(str::to_string)),
2251	"/logout" => SlashCommand::Logout,
2252	"/whoami" => SlashCommand::Whoami,
2253	"/reload" => SlashCommand::Reload,
2254	"/plan" => SlashCommand::Plan(parse_on_off(argument)),
2255	"/permissions" => SlashCommand::Permissions,
2256	"/compact" => SlashCommand::Compact,
2257	"/exit" \| "/quit" \| "/q" => SlashCommand::Exit,
2258	other => SlashCommand::Unknown(other.to_string()),
2259	})
2260	}
2261
2262	/// `on`/`off` (and synonyms) → `Some(bool)`; missing or unrecognized → `None`
2263	/// (meaning "toggle the current value").
2264	fn parse_on_off(arg: Option<&str>) -> Option<bool> {
2265	match arg.map(\|s\| s.trim().to_ascii_lowercase())?.as_str() {
2266	"on" \| "true" \| "1" \| "yes" => Some(true),
2267	"off" \| "false" \| "0" \| "no" => Some(false),
2268	_ => None,
2269	}
2270	}
2271
2272	fn format_models_list(current_model: &GgufModelConfig) -> String {
2273	let items = models::build_model_picker_items();
2274	if items.is_empty() {
2275	return "No local models found. siGit will use the platform default model.".to_string();
2276	}
2277
2278	let mut lines = vec!["Available models:".to_string()];
2279	let mut last_source: Option<&str> = None;
2280
2281	for (index, item) in items.iter().enumerate() {
2282	let source_key = match item.source_label.as_str() {
2283	"Onde" => "Onde",
2284	"HuggingFace" => "HuggingFace",
2285	"siGit Code Cloud" => "Cloud",
2286	_ => "Fallback",
2287	};
2288
2289	if last_source != Some(source_key) {
2290	if last_source.is_some() {
2291	lines.push(String::new());
2292	}
2293	let section = match source_key {
2294	"Onde" => "Onde Inference",
2295	"HuggingFace" => "Hugging Face cache",
2296	"Cloud" => "siGit Code Cloud",
2297	_ => "Fallback",
2298	};
2299	lines.push(section.to_string());
2300	// Blank line so the following "N." items render as an ordered list.
2301	// CommonMark only lets an ordered list interrupt a paragraph when it
2302	// starts at 1, so without this the cloud section (items 9+) would be
2303	// absorbed into the header paragraph.
2304	lines.push(String::new());
2305	last_source = Some(source_key);
2306	}
2307
2308	let number = index + 1;
2309	let current_badge = if item.config.model_id == current_model.model_id {
2310	" <- current"
2311	} else {
2312	""
2313	};
2314	let tool_badge = if item.tool_calling {
2315	" tool calling"
2316	} else {
2317	""
2318	};
2319	let health_badge = match item.cache_health {
2320	setup::ModelCacheHealth::Complete => "",
2321	setup::ModelCacheHealth::Incomplete => " ! incomplete cache",
2322	setup::ModelCacheHealth::NotDownloaded => " ↓ download on select",
2323	};
2324	let source = match source_key {
2325	"Onde" => " [Onde]",
2326	"HuggingFace" => " [HuggingFace]",
2327	"Cloud" => " [☁ Cloud]",
2328	_ => " [default]",
2329	};
2330
2331	lines.push(format!(
2332	"{number}. {} {}{}{}{}{}",
2333	item.display_name, item.description, tool_badge, health_badge, current_badge, source,
2334	));
2335	}
2336
2337	lines.push(String::new());
2338	lines.push("Use /models N to switch models.".to_string());
2339	lines.join("\n")
2340	}
2341
2342	async fn exec_slash_acp(
2343	agent: &SiGitAgent,
2344	cx: &ConnectionTo<Client>,
2345	session_id: SessionId,
2346	command: SlashCommand,
2347	) -> agent_client_protocol::Result<PromptResponse> {
2348	match command {
2349	SlashCommand::Help => {
2350	agent
2351	.send_assistant_message(
2352	cx,
2353	session_id,
2354	"/help - show this message\n\
2355	/models - list available models\n\
2356	/models N - switch to model N\n\
2357	/local [on\|off]- toggle on-device inference mode\n\
2358	/skills - list available Agent Skills\n\
2359	/mcp - list MCP servers and their tools\n\
2360	/load - load the selected on-device model\n\
2361	/login E P - sign in to siGit Code Cloud\n\
2362	/logout - sign out\n\
2363	/whoami - show the signed-in account\n\
2364	/reload - re-sync sign-in and model state\n\
2365	/plan [on\|off] - plan mode: research only, no edits or commands\n\
2366	/permissions - show the tool permission policy\n\
2367	/compact - summarize and shrink conversation history\n\
2368	/clear - wipe conversation history\n\
2369	/status - show engine status\n\
2370	/exit - end this turn",
2371	)
2372	.ok();
2373	}
2374	SlashCommand::Clear => {
2375	let cleared = agent.engine.clear_history().await;
2376	permissions::reset_session(&session_id.to_string());
2377	// The saved session must not resurrect what the user just wiped.
2378	session_store::delete(&session_id.to_string());
2379	agent
2380	.send_assistant_message(
2381	cx,
2382	session_id,
2383	format!("Cleared {cleared} turn(s). History is empty."),
2384	)
2385	.ok();
2386	}
2387	SlashCommand::Plan(value) => {
2388	let session_key = session_id.to_string();
2389	let enabled = value.unwrap_or_else(\|\| !permissions::plan_mode(&session_key));
2390	permissions::set_plan_mode(&session_key, enabled);
2391	let message = if enabled {
2392	"Plan mode ON — the agent researches with read-only tools and presents a \
2393	plan; edits and commands are blocked until /plan off."
2394	} else {
2395	"Plan mode OFF — the agent may execute tools again (subject to the \
2396	permission policy)."
2397	};
2398	agent.send_assistant_message(cx, session_id, message).ok();
2399	}
2400	SlashCommand::Permissions => {
2401	let summary = permissions::describe(&session_id.to_string());
2402	agent.send_assistant_message(cx, session_id, summary).ok();
2403	}
2404	SlashCommand::Compact => {
2405	let backend = agent.backend.lock().await.clone();
2406	let before = backend::estimate_tokens(&backend.history_snapshot().await);
2407	let message = match backend.compact_history(backend::COMPACT_KEEP_LAST).await {
2408	Ok(()) => {
2409	let snapshot = backend.history_snapshot().await;
2410	let after = backend::estimate_tokens(&snapshot);
2411	// Keep the saved session in step with the compacted state.
2412	if let Err(error) = session_store::save(&session_id.to_string(), &snapshot) {
2413	log::warn!("session save after /compact failed: {error}");
2414	}
2415	format!("Compacted history: ~{before} → ~{after} tokens (estimated).")
2416	}
2417	Err(error) => format!("Compaction failed: {error}"),
2418	};
2419	agent.send_assistant_message(cx, session_id, message).ok();
2420	}
2421	SlashCommand::Status => {
2422	let info = agent.engine.info().await;
2423	let model = info.model_name.as_deref().unwrap_or("(none)");
2424	let memory = info.approx_memory.as_deref().unwrap_or("unknown");
2425	agent
2426	.send_assistant_message(
2427	cx,
2428	session_id,
2429	format!(
2430	"status: {:?} model: {} memory: {} history: {} turns",
2431	info.status, model, memory, info.history_length,
2432	),
2433	)
2434	.ok();
2435	}
2436	SlashCommand::Models(None) => {
2437	let current_model = agent.current_model.lock().unwrap().clone();
2438	agent
2439	.send_assistant_message(cx, session_id, format_models_list(&current_model))
2440	.ok();
2441	}
2442	SlashCommand::Skills => {
2443	agent
2444	.send_assistant_message(cx, session_id, skills::format_skills_list())
2445	.ok();
2446	}
2447	SlashCommand::Mcp => {
2448	agent
2449	.send_assistant_message(cx, session_id, mcp::status_summary())
2450	.ok();
2451	}
2452	SlashCommand::Models(Some(number)) => {
2453	let items = models::build_model_picker_items();
2454	let index = number.saturating_sub(1);
2455	match items.get(index).cloned() {
2456	None => {
2457	agent
2458	.send_assistant_message(
2459	cx,
2460	session_id,
2461	format!("error: no model #{number} - type /models to see the list."),
2462	)
2463	.ok();
2464	}
2465	Some(model) if model.cloud_tier.is_some() => {
2466	// siGit Code Cloud tier: swap backend, sign-in gated.
2467	let tier = model.cloud_tier.clone().unwrap_or_default();
2468	let message = match agent.switch_to_cloud_tier(&tier).await {
2469	Some(display_name) => format!("Switched to {display_name}."),
2470	None => CLOUD_LOGIN_PROMPT.to_string(),
2471	};
2472	agent.send_assistant_message(cx, session_id, message).ok();
2473	}
2474	Some(model) => {
2475	if model.cache_health == setup::ModelCacheHealth::Incomplete {
2476	agent
2477	.send_assistant_message(
2478	cx,
2479	session_id,
2480	format!(
2481	"error: {} has an incomplete local cache and cannot be selected yet.",
2482	model.display_name
2483	),
2484	)
2485	.ok();
2486	} else if model.cache_health == setup::ModelCacheHealth::NotDownloaded {
2487	agent
2488	.send_assistant_message(
2489	cx,
2490	session_id.clone(),
2491	format!(
2492	"Downloading and loading {} ({})… this may take a few minutes.",
2493	model.display_name, model.description
2494	),
2495	)
2496	.ok();
2497
2498	match agent.switch_model_by_id(&model.config.model_id).await {
2499	Ok(new_config) => {
2500	agent.reset_to_local_backend().await;
2501	let _ = settings::set_local_inference(true);
2502	agent.engine.clear_history().await;
2503	agent
2504	.send_assistant_message(
2505	cx,
2506	session_id,
2507	format!(
2508	"✓ Downloaded and switched to {}",
2509	new_config.display_name
2510	),
2511	)
2512	.ok();
2513	}
2514	Err(err) => {
2515	agent
2516	.send_assistant_message(
2517	cx,
2518	session_id,
2519	format!("error downloading model: {}", err.message),
2520	)
2521	.ok();
2522	}
2523	}
2524	} else {
2525	agent
2526	.send_assistant_message(
2527	cx,
2528	session_id.clone(),
2529	format!("Loading {}...", model.display_name),
2530	)
2531	.ok();
2532
2533	let switched = agent.switch_model_by_id(&model.config.model_id).await?;
2534	agent.reset_to_local_backend().await;
2535	let _ = settings::set_local_inference(true);
2536	agent.engine.clear_history().await;
2537
2538	agent
2539	.send_assistant_message(
2540	cx,
2541	session_id,
2542	format!("Switched to {}.", switched.display_name),
2543	)
2544	.ok();
2545	}
2546	}
2547	}
2548	}
2549	SlashCommand::Local(value) => {
2550	let enabled = value.unwrap_or(!settings::local_inference_enabled());
2551	let message = match settings::set_local_inference(enabled) {
2552	Ok(()) if enabled => "Local inference is on. On-device models are highlighted; \
2553	pick one with /models."
2554	.to_string(),
2555	Ok(()) => "Local inference is off. siGit Code Cloud tiers are highlighted; \
2556	pick one with /models."
2557	.to_string(),
2558	Err(error) => format!("error: could not save local inference setting: {error}"),
2559	};
2560	agent
2561	.send_assistant_message(cx, session_id.clone(), message)
2562	.ok();
2563	// Refresh the panel so the Model picker reflects the new emphasis.
2564	let config_options = {
2565	let current = agent.current_model.lock().unwrap();
2566	build_model_config_options(&current)
2567	};
2568	agent
2569	.send_tool_call_update(
2570	cx,
2571	session_id,
2572	SessionUpdate::ConfigOptionUpdate(ConfigOptionUpdate::new(config_options)),
2573	)
2574	.ok();
2575	}
2576	SlashCommand::Load => {
2577	// Explicitly load the on-device model. This is the only path that
2578	// brings a local model into memory; prompts never do it implicitly.
2579	// If a cloud tier is active, fall back to a local default so we don't
2580	// try to load the (file-less) cloud config as GGUF.
2581	let on_cloud = {
2582	let guard = agent.current_model.lock().unwrap();
2583	guard.model_id.starts_with("sigit-cloud:")
2584	};
2585	if on_cloud {
2586	let default_config = default_local_model_config();
2587	*agent.current_model.lock().unwrap() = default_config;
2588	agent.reset_to_local_backend().await;
2589	}
2590	// Loading an on-device model puts us in local inference mode.
2591	let _ = settings::set_local_inference(true);
2592	// `await_model_ready` drives the download/load progress UI and reports
2593	// success or failure to the editor.
2594	agent.start_startup_model_load_if_needed();
2595	agent.await_model_ready(cx, &session_id).await?;
2596	}
2597	SlashCommand::Login(argument) => {
2598	let message = match argument.as_deref().and_then(account::parse_login_args) {
2599	Some((email, password)) => match account::authenticate(&email, &password).await {
2600	Ok(email) => format!(
2601	"Signed in as {email}. Pick a siGit Code Cloud tier in /models to use it."
2602	),
2603	Err(error) => format!("Login failed: {error}"),
2604	},
2605	None => "usage: /login <email> <password>".to_string(),
2606	};
2607	agent.send_assistant_message(cx, session_id, message).ok();
2608	}
2609	SlashCommand::Logout => {
2610	// If we're on a cloud tier, drop back to local — the token is gone.
2611	let on_cloud = {
2612	let guard = agent.current_model.lock().unwrap();
2613	guard.model_id.starts_with("sigit-cloud:")
2614	};
2615	let message = account::end_session().await;
2616	if on_cloud {
2617	agent.reset_to_local_backend().await;
2618	}
2619	agent.send_assistant_message(cx, session_id, message).ok();
2620	}
2621	SlashCommand::Whoami => {
2622	let message = account::status_line().await;
2623	agent.send_assistant_message(cx, session_id, message).ok();
2624	}
2625	SlashCommand::Reload => {
2626	agent.handle_reload(cx, session_id).await;
2627	}
2628	SlashCommand::Exit => {
2629	agent
2630	.send_assistant_message(
2631	cx,
2632	session_id,
2633	"Use the panel controls to close or switch threads.",
2634	)
2635	.ok();
2636	}
2637	SlashCommand::Unknown(command) => {
2638	agent
2639	.send_assistant_message(cx, session_id, format!("unknown command: {command}"))
2640	.ok();
2641	}
2642	}
2643
2644	Ok(PromptResponse::new(StopReason::EndTurn))
2645	}
2646
2647	// ── Request dispatch helper ───────────────────────────────────────────────────
2648
2649	fn handle_response<T: agent_client_protocol::JsonRpcResponse>(
2650	responder: Responder<T>,
2651	result: agent_client_protocol::Result<T>,
2652	) -> agent_client_protocol::Result<()> {
2653	match result {
2654	Ok(resp) => responder.respond(resp),
2655	Err(err) => responder.respond_with_error(err),
2656	}
2657	}
2658
2659	// ── Download progress helpers ─────────────────────────────────────────────────
2660
2661	/// total bytes on disk under `path`. needed because hf-hub uses staging
2662	/// names during download, so we can't just stat the final blobs.
2663	fn dir_size_recursive(path: &std::path::Path) -> u64 {
2664	let mut total: u64 = 0;
2665	let Ok(entries) = std::fs::read_dir(path) else {
2666	return 0;
2667	};
2668	for entry in entries.flatten() {
2669	let entry_path = entry.path();
2670	if entry_path.is_dir() {
2671	total += dir_size_recursive(&entry_path);
2672	} else if let Ok(meta) = entry_path.metadata() {
2673	total += meta.len();
2674	}
2675	}
2676	total
2677	}
2678
2679	fn format_size_human(bytes: u64) -> String {
2680	const GB: u64 = 1_073_741_824;
2681	const MB: u64 = 1_048_576;
2682	const KB: u64 = 1_024;
2683	if bytes >= GB {
2684	format!("{:.2} GB", bytes as f64 / GB as f64)
2685	} else if bytes >= MB {
2686	format!("{:.1} MB", bytes as f64 / MB as f64)
2687	} else if bytes >= KB {
2688	format!("{:.0} KB", bytes as f64 / KB as f64)
2689	} else {
2690	format!("{bytes} B")
2691	}
2692	}
2693
2694	fn progress_bar(pct: u8, width: usize) -> String {
2695	let filled = ((pct as usize) * width) / 100;
2696	let empty = width.saturating_sub(filled);
2697	format!("[{}{}]", "█".repeat(filled), "░".repeat(empty))
2698	}
2699
2700	// ── Output capture ────────────────────────────────────────────────────────────
2701
2702	/// redirect stdout+stderr to `$TMPDIR/sigit.log` at the fd level so
2703	/// mistralrs/tracing noise never hits the terminal. returns two dup'd
2704	/// fds to the real tty: one for ratatui, one for cleanup (ratatui 0.29
2705	/// doesn't expose `writer_mut()`).
2706	#[cfg(unix)]
2707	fn redirect_output_to_log() -> anyhow::Result<(std::fs::File, std::fs::File)> {
2708	let log_path = std::env::temp_dir().join("sigit.log");
2709	let log_file = std::fs::File::create(&log_path)?;
2710	let log_fd = log_file.as_raw_fd();
2711
2712	// two copies: ratatui needs one, cleanup needs another
2713	let saved_tui = unsafe { libc::dup(libc::STDOUT_FILENO) };
2714	anyhow::ensure!(
2715	saved_tui >= 0,
2716	"dup(stdout) for tui failed: {}",
2717	std::io::Error::last_os_error()
2718	);
2719	let saved_cleanup = unsafe { libc::dup(libc::STDOUT_FILENO) };
2720	anyhow::ensure!(
2721	saved_cleanup >= 0,
2722	"dup(stdout) for cleanup failed: {}",
2723	std::io::Error::last_os_error()
2724	);
2725
2726	unsafe {
2727	libc::dup2(log_fd, libc::STDOUT_FILENO);
2728	libc::dup2(log_fd, libc::STDERR_FILENO);
2729	}
2730
2731	// safe to drop log_file; dup2 keeps the fd alive via stdout/stderr
2732
2733	Ok((unsafe { std::fs::File::from_raw_fd(saved_tui) }, unsafe {
2734	std::fs::File::from_raw_fd(saved_cleanup)
2735	}))
2736	}
2737
2738	// ── Logging ───────────────────────────────────────────────────────────────────
2739
2740	/// in TUI mode stderr is the log file (redirected earlier);
2741	/// in ACP mode it's real stderr. either way, write there.
2742	fn init_logging(is_tty: bool) {
2743	let filter = EnvFilter::try_from_default_env().unwrap_or_else(\|_\| EnvFilter::new("info"));
2744	let _ = tracing_fmt::Subscriber::builder()
2745	.with_env_filter(filter)
2746	.with_writer(std::io::stderr)
2747	.with_ansi(!is_tty)
2748	.try_init();
2749	}
2750
2751	// ── Interactive TUI mode ──────────────────────────────────────────────────────
2752
2753	/// boot the TUI and load the model on a background thread.
2754	/// `tty` goes to ratatui; `cleanup_tty` is a separate fd for
2755	/// LeaveAlternateScreen (ratatui 0.29 hides `writer_mut()`).
2756	#[cfg(unix)]
2757	async fn run_interactive(tty: std::fs::File, mut cleanup_tty: std::fs::File) -> anyhow::Result<()> {
2758	let engine = Arc::new(ChatEngine::new());
2759
2760	let startup_selection = setup::startup_model_selection();
2761	let startup_model_name = startup_selection
2762	.as_ref()
2763	.map(\|selection\| selection.display_name.clone())
2764	.unwrap_or_else(\|\| GgufModelConfig::qwen25_3b().display_name);
2765
2766	// Signals the loading phase to finish. On-device models are no longer loaded
2767	// at startup, so this resolves immediately for both backends; it stays a
2768	// channel so the loading-phase plumbing in `chat::run_with` is unchanged.
2769	let (load_tx, load_rx) = std::sync::mpsc::channel::<Result<(), String>>();
2770
2771	// Project instruction files (AGENTS.md / CLAUDE.md) for the launch directory,
2772	// injected into the system prompt so the TUI shares the same always-on
2773	// project context the ACP sessions get.
2774	let project_instructions = std::env::current_dir()
2775	.ok()
2776	.and_then(\|cwd\| instructions::load_project_instructions(&cwd));
2777	let with_instructions = \|base: String\| match &project_instructions {
2778	Some(extra) => format!("{base}\n\n{extra}"),
2779	None => base,
2780	};
2781
2782	// Pick the inference backend: a configured provider if present, else on-device.
2783	let (inference_backend, startup_model_name): (Arc<dyn InferenceBackend>, String) =
2784	match provider::active_provider() {
2785	Some(provider) => {
2786	log::info!(
2787	"inference: using {} (model {}) at {}",
2788	provider.display_name,
2789	provider.model,
2790	provider.base_url
2791	);
2792	// No local model to load; the endpoint is ready immediately.
2793	let _ = load_tx.send(Ok(()));
2794	let label = provider.display_name.clone();
2795	register_subagent_factory_for(&provider);
2796	let backend = Arc::new(OpenAiBackend::new(
2797	provider.base_url,
2798	provider.api_key,
2799	provider.model,
2800	Some(with_instructions(SYSTEM_PROMPT.to_string())),
2801	)) as Arc<dyn InferenceBackend>;
2802	(backend, label)
2803	}
2804	None => {
2805	// Honor the Local Inference toggle: when off and signed in, start
2806	// on a cloud tier. Otherwise bring up on-device WITHOUT loading a
2807	// model — the user loads it explicitly with /load (or /models), so
2808	// the UI comes up immediately. Project instructions are injected at
2809	// load time in `chat.rs`.
2810	let cloud_when_off = if settings::local_inference_enabled() {
2811	None
2812	} else {
2813	provider::cloud_tier_provider("balanced")
2814	};
2815
2816	match cloud_when_off {
2817	Some(provider) => {
2818	log::info!(
2819	"inference: local inference off; using {} (model {})",
2820	provider.display_name,
2821	provider.model
2822	);
2823	let _ = load_tx.send(Ok(()));
2824	let label = provider.display_name.clone();
2825	register_subagent_factory_for(&provider);
2826	let backend = Arc::new(OpenAiBackend::new(
2827	provider.base_url,
2828	provider.api_key,
2829	provider.model,
2830	Some(with_instructions(SYSTEM_PROMPT.to_string())),
2831	)) as Arc<dyn InferenceBackend>;
2832	(backend, label)
2833	}
2834	None => {
2835	if !settings::local_inference_enabled() {
2836	log::warn!(
2837	"local inference is off but no account is signed in; \
2838	bringing up on-device. Run /login or /local on."
2839	);
2840	}
2841	let _ = load_tx.send(Ok(()));
2842	// On-device inference has a single shared history; no
2843	// subagent context is possible yet.
2844	tools::set_subagent_factory(Box::new(\|\| None));
2845	let backend = Arc::new(LocalBackend::new(Arc::clone(&engine)))
2846	as Arc<dyn InferenceBackend>;
2847	(backend, startup_model_name)
2848	}
2849	}
2850	}
2851	};
2852
2853	crossterm::terminal::enable_raw_mode()?;
2854	let mut tty = BufWriter::new(tty);
2855	crossterm::execute!(tty, crossterm::terminal::EnterAlternateScreen)?;
2856	let term_backend = ratatui::backend::CrosstermBackend::new(tty);
2857	let mut terminal = ratatui::Terminal::new(term_backend)?;
2858
2859	// polls load_rx with try_recv() each tick, no blocking
2860	let chat_result = chat::run_with(
2861	&mut terminal,
2862	engine,
2863	inference_backend,
2864	load_rx,
2865	startup_model_name,
2866	)
2867	.await;
2868
2869	// cleanup fd because backend's writer is private
2870	crossterm::execute!(cleanup_tty, crossterm::terminal::LeaveAlternateScreen)?;
2871	cleanup_tty.flush()?;
2872	crossterm::terminal::disable_raw_mode()?;
2873
2874	// restore real stdout/stderr for post-TUI error output
2875	#[cfg(unix)]
2876	{
2877	let cleanup_fd = cleanup_tty.as_raw_fd();
2878	unsafe {
2879	libc::dup2(cleanup_fd, libc::STDOUT_FILENO);
2880	libc::dup2(cleanup_fd, libc::STDERR_FILENO);
2881	}
2882	}
2883
2884	chat_result
2885	}
2886
2887	// ── ACP server mode ───────────────────────────────────────────────────────────
2888
2889	/// The on-device model `/load` should bring up by default: the persisted
2890	/// selection if it still resolves to a known local model, otherwise the built-in
2891	/// default (`qwen25_3b`).
2892	fn default_local_model_config() -> GgufModelConfig {
2893	setup::startup_model_selection()
2894	.as_ref()
2895	.and_then(\|selection\| {
2896	selection.selected_model.as_ref().and_then(\|selected\| {
2897	models::local_picker_items()
2898	.into_iter()
2899	.find(\|item\| {
2900	item.config.model_id == selected.model_id
2901	&& item
2902	.config
2903	.files
2904	.iter()
2905	.any(\|file\| file == &selected.gguf_file)
2906	})
2907	.map(\|item\| item.config)
2908	})
2909	})
2910	.unwrap_or_else(GgufModelConfig::qwen25_3b)
2911	}
2912
2913	async fn run_acp_server() -> anyhow::Result<()> {
2914	log::info!("ACP mode — starting agent server");
2915
2916	let config = default_local_model_config();
2917
2918	let needs_download = models::local_picker_items()
2919	.iter()
2920	.find(\|item\| item.config.model_id == config.model_id)
2921	.map(\|item\| item.cache_health != setup::ModelCacheHealth::Complete)
2922	.unwrap_or(true);
2923
2924	log::info!(
2925	"ACP startup model selected: {} ({})",
2926	config.display_name,
2927	if needs_download {
2928	"needs download"
2929	} else {
2930	"cached"
2931	}
2932	);
2933
2934	let engine = Arc::new(ChatEngine::new());
2935
2936	// The on-device model is never loaded implicitly; the user loads it with
2937	// `/load` (or by picking one in `/models`). So initialize/session/new stay
2938	// lightweight and `model_ready` starts true (nothing is loading).
2939	let model_ready = Arc::new(AtomicBool::new(true));
2940	let startup_model_load_started = Arc::new(AtomicBool::new(false));
2941	let model_load_error: Arc<std::sync::Mutex<Option<String>>> =
2942	Arc::new(std::sync::Mutex::new(None));
2943
2944	let state = Arc::new(SiGitAgent::new(
2945	engine,
2946	config,
2947	model_ready,
2948	startup_model_load_started,
2949	model_load_error,
2950	needs_download,
2951	));
2952
2953	// Honor the explicit provider override (OPENAI_BASE_URL/OPENAI_API_KEY or
2954	// an active providers.toml profile) in ACP mode too — the interactive
2955	// client already does. Without this the override was silently ignored here
2956	// and prompts insisted on a local model. It is also what lets the ACP
2957	// integration test drive the agent against a scripted endpoint
2958	// (tests/acp_permissions.rs). The model picker still shows the local
2959	// selection; overrides are a power-user escape hatch, not a tier.
2960	if let Some(cfg) = provider::active_provider() {
2961	log::info!(
2962	"inference: using {} (model {}) at {}",
2963	cfg.display_name,
2964	cfg.model,
2965	cfg.base_url
2966	);
2967	register_subagent_factory_for(&cfg);
2968	let override_backend: Arc<dyn InferenceBackend> = Arc::new(OpenAiBackend::new(
2969	cfg.base_url,
2970	cfg.api_key,
2971	cfg.model,
2972	Some(system_prompt_for_model(true).to_string()),
2973	));
2974	*state.backend.lock().await = override_backend;
2975	} else {
2976	// On-device inference has a single shared conversation history, so a
2977	// second concurrent subagent context is not possible yet.
2978	tools::set_subagent_factory(Box::new(\|\| None));
2979	}
2980
2981	let stdin = tokio::io::stdin().compat();
2982	let stdout = tokio::io::stdout().compat_write();
2983	let transport = ByteStreams::new(stdout, stdin);
2984
2985	Agent
2986	.builder()
2987	.on_receive_request(
2988	{
2989	let state = Arc::clone(&state);
2990	async move \|req: InitializeRequest, responder, _cx: ConnectionTo<Client>\| {
2991	handle_response(responder, state.handle_initialize(req).await)
2992	}
2993	},
2994	agent_client_protocol::on_receive_request!(),
2995	)
2996	.on_receive_request(
2997	{
2998	let state = Arc::clone(&state);
2999	async move \|req: AuthenticateRequest, responder, _cx: ConnectionTo<Client>\| {
3000	handle_response(responder, state.handle_authenticate(req).await)
3001	}
3002	},
3003	agent_client_protocol::on_receive_request!(),
3004	)
3005	// Turn-affecting handlers below run in spawned tasks, serialized by
3006	// `turn_lock`, so the dispatch loop stays free to route client
3007	// responses (permission answers) while a turn is in flight. Awaiting a
3008	// client request from inside a handler would deadlock: the dispatch
3009	// loop can't read the response while the handler blocks it.
3010	.on_receive_request(
3011	{
3012	let state = Arc::clone(&state);
3013	async move \|req: LoadSessionRequest, responder, cx: ConnectionTo<Client>\| {
3014	let state = Arc::clone(&state);
3015	let task_cx = cx.clone();
3016	cx.spawn(async move {
3017	let _turn = state.turn_lock.lock().await;
3018	handle_response(responder, state.handle_load_session(&task_cx, req).await)
3019	})
3020	}
3021	},
3022	agent_client_protocol::on_receive_request!(),
3023	)
3024	.on_receive_request(
3025	{
3026	let state = Arc::clone(&state);
3027	async move \|req: ForkSessionRequest, responder, cx: ConnectionTo<Client>\| {
3028	let state = Arc::clone(&state);
3029	let task_cx = cx.clone();
3030	cx.spawn(async move {
3031	let _turn = state.turn_lock.lock().await;
3032	handle_response(responder, state.handle_fork_session(&task_cx, req).await)
3033	})
3034	}
3035	},
3036	agent_client_protocol::on_receive_request!(),
3037	)
3038	.on_receive_request(
3039	{
3040	let state = Arc::clone(&state);
3041	async move \|req: NewSessionRequest, responder, cx: ConnectionTo<Client>\| {
3042	let state = Arc::clone(&state);
3043	let task_cx = cx.clone();
3044	cx.spawn(async move {
3045	let _turn = state.turn_lock.lock().await;
3046	handle_response(responder, state.handle_new_session(&task_cx, req).await)
3047	})
3048	}
3049	},
3050	agent_client_protocol::on_receive_request!(),
3051	)
3052	.on_receive_request(
3053	{
3054	let state = Arc::clone(&state);
3055	async move \|req: PromptRequest, responder, cx: ConnectionTo<Client>\| {
3056	let state = Arc::clone(&state);
3057	let task_cx = cx.clone();
3058	cx.spawn(async move {
3059	let _turn = state.turn_lock.lock().await;
3060	handle_response(responder, state.handle_prompt(&task_cx, req).await)
3061	})
3062	}
3063	},
3064	agent_client_protocol::on_receive_request!(),
3065	)
3066	.on_receive_request(
3067	{
3068	let state = Arc::clone(&state);
3069	async move \|req: SetSessionConfigOptionRequest,
3070	responder,
3071	cx: ConnectionTo<Client>\| {
3072	let state = Arc::clone(&state);
3073	let task_cx = cx.clone();
3074	cx.spawn(async move {
3075	let _turn = state.turn_lock.lock().await;
3076	handle_response(
3077	responder,
3078	state.handle_set_session_config_option(&task_cx, req).await,
3079	)
3080	})
3081	}
3082	},
3083	agent_client_protocol::on_receive_request!(),
3084	)
3085	.on_receive_notification(
3086	{
3087	let state = Arc::clone(&state);
3088	async move \|notif: CancelNotification, _cx: ConnectionTo<Client>\| {
3089	state.handle_cancel(notif).await
3090	}
3091	},
3092	agent_client_protocol::on_receive_notification!(),
3093	)
3094	.connect_to(transport)
3095	.await
3096	.map_err(\|e\| anyhow::anyhow!("ACP connection error: {e}"))?;
3097
3098	log::info!("siGit shutting down");
3099	Ok(())
3100	}
3101
3102	// ── Entry point ──────────────────────────────────────────────────────────────
3103
3104	#[tokio::main]
3105	async fn main() -> anyhow::Result<()> {
3106	// Account subcommands. The editor launches `sigit login` in an embedded
3107	// terminal for ACP terminal-based authentication; the same verbs are handy
3108	// directly from a shell. These must be handled before the TTY/ACP split.
3109	if let Some(verb) = std::env::args().nth(1) {
3110	match verb.as_str() {
3111	"login" => {
3112	init_logging(true);
3113	match account::interactive_login().await {
3114	Ok(email) => {
3115	println!("Signed in to siGit Code Cloud as {email}.");
3116	return Ok(());
3117	}
3118	Err(error) => {
3119	eprintln!("Login failed: {error}");
3120	std::process::exit(1);
3121	}
3122	}
3123	}
3124	"logout" => {
3125	init_logging(true);
3126	println!("{}", account::end_session().await);
3127	return Ok(());
3128	}
3129	"whoami" => {
3130	init_logging(true);
3131	println!("{}", account::status_line().await);
3132	return Ok(());
3133	}
3134	"run" => {
3135	// Headless one-shot mode: stdout carries JSONL run events, so
3136	// logs go to stderr exactly like ACP mode.
3137	init_logging(false);
3138	setup::setup_shared_model_cache();
3139	// Best-effort MCP discovery so `mcp__*` tools are offered
3140	// (`SIGIT_MCP=off` skips this, the cloud-runner default).
3141	mcp::init().await;
3142	log::info!(
3143	"siGit v{} starting (headless run)",
3144	env!("CARGO_PKG_VERSION")
3145	);
3146	return headless::run(std::env::args().skip(2)).await;
3147	}
3148	_ => {}
3149	}
3150	}
3151
3152	let is_tty = std::io::stdin().is_terminal();
3153
3154	if is_tty {
3155	// must redirect before any library code touches stdout
3156	#[cfg(unix)]
3157	{
3158	let (tty, cleanup_tty) = redirect_output_to_log()?;
3159	init_logging(true);
3160	setup::setup_shared_model_cache();
3161	// Best-effort: discover MCP servers (incl. the official one) before
3162	// the first turn so their tools are offered to the model.
3163	mcp::init().await;
3164	run_interactive(tty, cleanup_tty).await
3165	}
3166	#[cfg(not(unix))]
3167	{
3168	anyhow::bail!("interactive mode requires Unix (macOS / Linux)");
3169	}
3170	} else {
3171	// ACP mode: keep stdout untouched for protocol JSON only.
3172	// Logs already go to stderr via `init_logging(false)`.
3173	init_logging(false);
3174	setup::setup_shared_model_cache();
3175	// Best-effort MCP discovery (incl. the official server) before serving.
3176	mcp::init().await;
3177	log::info!("siGit v{} starting (ACP mode)", env!("CARGO_PKG_VERSION"));
3178	run_acp_server().await
3179	}
3180	}
3181
3182	#[cfg(test)]
3183	mod tests {
3184	use super::*;
3185
3186	#[test]
3187	fn system_prompt_advertises_the_commit_co_author_trailer() {
3188	// The prompt instructs the model with the exact trailer that
3189	// `tools::ensure_commit_co_author` enforces; if the two drift apart the
3190	// safety net would re-amend commits the model already attributed.
3191	assert!(
3192	SYSTEM_PROMPT.contains(tools::COMMIT_CO_AUTHOR_TRAILER),
3193	"SYSTEM_PROMPT must quote tools::COMMIT_CO_AUTHOR_TRAILER verbatim"
3194	);
3195	}
3196
3197	#[test]
3198	fn ascii_safe_replaces_multibyte_chars() {
3199	// The exact label that crashed Zed: the cloud tier name plus the old
3200	// "[☁ siGit Code Cloud]" badge. After sanitizing it must be pure ASCII so
3201	// Zed's fixed byte-offset truncation can never split a glyph.
3202	let crashing = "siGit Code Cloud · Balanced [☁ siGit Code Cloud]";
3203	let safe = ascii_safe(crashing);
3204	assert!(safe.is_ascii(), "sanitized label must be ASCII: {safe:?}");
3205	assert_eq!(safe, "siGit Code Cloud - Balanced [- siGit Code Cloud]");
3206	}
3207
3208	#[test]
3209	fn ascii_safe_leaves_ascii_untouched() {
3210	let plain = "Qwen 2.5 3B [Onde]";
3211	assert_eq!(ascii_safe(plain), plain);
3212	}
3213
3214	#[test]
3215	fn ascii_safe_output_has_only_char_boundaries() {
3216	// Every byte index in an ASCII string is a valid char boundary, so any
3217	// downstream truncation is panic-free regardless of where it cuts.
3218	let safe = ascii_safe("Onde · ◉ ↓ ☁ ○ test");
3219	for i in 0..=safe.len() {
3220	assert!(safe.is_char_boundary(i));
3221	}
3222	}
3223	}