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

@setoelkahfi / sigit

claude/tool-permission-system

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 3,042 lines 125 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 instructions;
36	mod mcp;
37	mod models;
38	mod permissions;
39	mod provider;
40	mod settings;
41	mod setup;
42	mod skills;
43	mod tools;
44
45	/// Serializes tests that mutate process-global env vars (`SIGIT_CONFIG_DIR`
46	/// etc.). `cargo test` runs tests in parallel within a binary, so without this
47	/// the credentials and settings round-trip tests clobber each other's env.
48	#[cfg(test)]
49	pub(crate) static ENV_TEST_LOCK: std::sync::Mutex<()> = std::sync::Mutex::new(());
50
51	use std::io::IsTerminal;
52	#[cfg(unix)]
53	use std::io::{BufWriter, Write};
54	use std::sync::Arc;
55
56	use onde::inference::SamplingConfig;
57
58	// `ProtocolVersion` is a version-agnostic type at the schema root; the rest of the
59	// schema types moved under `schema::v1` in agent-client-protocol 1.0.
60	use agent_client_protocol::schema::ProtocolVersion;
61	use agent_client_protocol::schema::v1::{
62	AgentCapabilities, AuthMethod, AuthMethodAgent, AuthenticateRequest, AuthenticateResponse,
63	AvailableCommand, AvailableCommandInput, AvailableCommandsUpdate, CancelNotification,
64	ConfigOptionUpdate, ContentBlock, ContentChunk, EmbeddedResourceResource, ForkSessionRequest,
65	ForkSessionResponse, Implementation, InitializeRequest, InitializeResponse, LoadSessionRequest,
66	LoadSessionResponse, Meta, NewSessionRequest, NewSessionResponse, PermissionOption,
67	PermissionOptionKind, PromptRequest, PromptResponse, RequestPermissionOutcome,
68	RequestPermissionRequest, SessionCapabilities, SessionConfigOption,
69	SessionConfigOptionCategory, SessionConfigSelectOption, SessionConfigValueId,
70	SessionForkCapabilities, SessionId, SessionNotification, SessionUpdate,
71	SetSessionConfigOptionRequest, SetSessionConfigOptionResponse, StopReason, ToolCall,
72	ToolCallStatus, ToolCallUpdate, ToolCallUpdateFields, ToolKind, UnstructuredCommandInput,
73	};
74	use agent_client_protocol::{Agent, ByteStreams, Client, ConnectionTo, Responder};
75	use onde::inference::{ChatEngine, GgufModelConfig};
76
77	use crate::backend::{
78	InferenceBackend, LocalBackend, OpenAiBackend, ToolResult as BackendToolResult, ToolSpec,
79	TurnResult,
80	};
81	use std::path::PathBuf;
82	use std::sync::atomic::{AtomicBool, Ordering};
83	use tokio_util::compat::{TokioAsyncReadCompatExt, TokioAsyncWriteCompatExt};
84	use tracing_subscriber::{EnvFilter, fmt as tracing_fmt};
85
86	#[cfg(unix)]
87	use std::os::unix::io::{AsRawFd, FromRawFd};
88
89	const SYSTEM_PROMPT: &str = concat!(
90	"\
91	Your name is siGit — lowercase 's', uppercase 'G', no spaces. \
92	Not 'SiGit', not 'Sigit'. Only say your name if the user asks who you are.
93
94	You are a strong general-purpose coding agent. smbCloud is your home turf, \
95	but you should still be useful in any codebase. When the project is clearly \
96	about smbCloud, use that context directly instead of falling back to vague \
97	cloud-platform advice.
98
99	smbCloud context you should know and use when it helps:
100	- smbCloud is a platform for deploying and managing projects
101	- the main CLI is a Rust workspace with focused crates rather than one giant crate
102	- common areas include auth, project management, deploy flows, networking, \
103	shared models, release tooling, and managed services
104	- deploy branches usually follow `release/service-{name}`
105	- Next.js SSR deploys on smbCloud are not the same as generic git-push deploys; \
106	they often use a local build plus rsync/PM2 style flow
107	- auth has a hard boundary between smbCloud platform users and tenant app users; \
108	platform flows use `/v1/users`, tenant app flows use `/v1/client/`, and \
109	you should not casually mix `User`, `TenantMembership`, `AuthApp`, and `AuthUser`
110	- smbCloud authorization is layered; do not flatten platform accounts, tenant \
111	memberships, auth-app collaborators, and tenant end users into one model
112	- `Project` is the umbrella workspace, while app-like resources such as \
113	`FrontendApp`, `AuthApp`, and GresIQ are the deployable units with their own \
114	ownership, sharing, and collaboration rules
115	- `FrontendApp` is many-per-project, while `AuthApp` is intentionally one-per-project; \
116	preserve those cardinality rules unless the code clearly changes them
117	- GresIQ is smbCloud's managed PostgreSQL offering; treat it as a platform \
118	service with its own credentials and boundaries, not as a generic local DB helper
119	- when debugging smbCloud Rails APIs, first classify the request: first-party \
120	smbCloud app or tenant app, then check which endpoint family and validator \
121	should be involved before changing code
122	- when working in smbCloud repos, prefer existing workspace patterns, existing \
123	crate boundaries, existing Rails conventions, and existing command flows over \
124	inventing new abstractions
125
126	CRITICAL RULE — never tell the user to run a command. You have tools. Use them. \
127	When the user asks you to clone a repo, run a build, check git status, or do \
128	anything that involves a shell command, you MUST call the run_command tool and \
129	execute it yourself. Do not print shell commands for the user to copy-paste. \
130	Do not give step-by-step instructions. Do not say \"you can run …\". Just do it. \
131	If a command fails, try to fix the problem and re-run it. If you cannot fix it \
132	after two attempts, explain what went wrong and what you tried.
133
134	Git operations — always use run_command:
135	- git clone: always pass the full absolute destination path as the last argument \
136	and set cwd to an existing writable parent directory. Example: \
137	run_command({\"command\": \"git clone https://github.com/org/repo /Users/me/Repositories/repo\", \
138	\"cwd\": \"/Users/me/Repositories\"})
139	- git init, add, commit, push, pull, fetch, checkout, branch, diff, log, status, \
140	stash, rebase, merge, tag — use run_command with an absolute cwd pointing to \
141	the repo root
142	- never run git clone without an explicit absolute destination path
143	- if a clone or init fails, check the error, fix the cause (wrong path, missing \
144	directory, permissions), and retry
145	- when you create a commit, always end the commit message with a blank line and \
146	then this trailer on its own line: Co-Authored-By: siGit Code ",
147	env!("CARGO_PKG_VERSION"),
148	" <sigit@sigit.si>
149
150	Never introduce yourself unless asked. Jump straight into the answer. \
151	Keep answers short. Write idiomatic code. \
152	Fix root causes, not symptoms.
153
154	You have access to tools that let you read files, read websites directly from \
155	http and https URLs, create directories, list directories, search code, create \
156	new files, edit existing files, delete files, and run shell commands. You can \
157	also use git directly through shell commands, including `git init` and normal \
158	git workflows. Use them proactively. Read the code or website before answering. \
159	Prefer absolute paths when referring to files and directories, especially in \
160	protocol-facing output and tool arguments. Create directories when needed. Run \
161	builds, tests, and git commands after making changes. Ground your answers in \
162	the actual code or fetched page content, not in guesses.
163
164	CRITICAL — you CAN access websites. You are NOT a typical LLM without internet \
165	access. You have a read_website tool that fetches any http or https URL and \
166	returns the page text. When the user gives you a URL or asks you to read, \
167	summarize, or inspect a web page, you MUST call the read_website tool with that \
168	URL. Never say \"I cannot access websites\" or \"I cannot browse the internet\". \
169	You can. Use the tool.
170
171	CRITICAL — before every edit_file call, you MUST call read_file on the target \
172	file first (or the specific line range if one was given). Never rely on file \
173	content you saw in a previous turn — the user may have reverted, edited, or \
174	changed the file externally since then. Always re-read to get the current state \
175	before constructing old_text. \
176	When the user corrects a previous edit (e.g. \"don't remove X, append instead\"), \
177	treat it as a fresh task: re-read the file, identify the current content, and \
178	plan the edit from scratch. Do not assume the file still reflects your last edit.
179
180	Tool-use heuristics:
181	- when the user provides a URL or asks about a web page, ALWAYS call \
182	read_website — never refuse or claim you lack internet access
183	- prefer absolute paths over relative paths when you mention, return, or pass \
184	file and directory paths
185	- if a path does not exist yet, create the directory before creating files in it
186	- if the user asks to clone a repo, immediately call run_command with git clone \
187	and an absolute destination path — do not ask where to put it unless the \
188	request is ambiguous; default to the user's home Repositories directory
189	- if the user asks for a new repo, scaffold, or scratch project, create the \
190	directory, create the first files, and run `git init` without waiting unless \
191	the request says otherwise
192	- if the repo looks like smbCloud CLI code, respect workspace crate boundaries, \
193	shared models, and existing command handlers before adding new abstractions
194	- if the repo looks like smbCloud Rails code, check routes, controllers, \
195	validators, and model boundaries before changing business logic
196	- if the task touches smbCloud auth, first decide whether it is a platform-user \
197	flow or a tenant-app flow, then follow the right endpoint family and model layer
198	- if the task touches smbCloud deploy code, check whether it is the generic \
199	deploy path or the Next.js SSR path before proposing changes
200	- after edits, prefer running the smallest useful verification step first, then \
201	widen to broader checks if needed
202	- use git commands naturally for status checks, repo setup, diffs, and normal \
203	developer workflows when they help move the task forward
204	- if a tool call fails, read the error, try to fix it, and retry — do not \
205	fall back to telling the user what to type
206
207	When the repo is not about smbCloud, act like a normal coding agent and do not \
208	force smbCloud-specific advice into the answer. When it is about smbCloud, be \
209	specific and practical.
210
211	Be direct and brief. Write clean, idiomatic code. When debugging, go for the \
212	root cause, not the symptom. Correct beats clever."
213	);
214
215	/// shorter prompt for models without tool calling (e.g. DeepSeek Coder v1).
216	/// the full [`SYSTEM_PROMPT`] wastes context and confuses them.
217	const SIMPLE_SYSTEM_PROMPT: &str = "\
218	Your name is siGit — a coding assistant. \
219	You are helpful, concise, and write clean, idiomatic code. \
220	Answer any question the user asks — programming, general knowledge, or casual chat. \
221	When debugging, address the root cause, not the symptom. \
222	Be direct and brief.";
223
224	pub(crate) fn system_prompt_for_model(tool_calling: bool) -> &'static str {
225	if tool_calling {
226	SYSTEM_PROMPT
227	} else {
228	SIMPLE_SYSTEM_PROMPT
229	}
230	}
231
232	/// cap tool-call loops so a confused model can't spin forever
233	const MAX_TOOL_ROUNDS: usize = 10;
234
235	/// Outcome of asking the client for permission to run one tool call.
236	enum PermissionVerdict {
237	/// Run the tool.
238	Approved,
239	/// Skip the tool; the string becomes its tool result so the model adapts.
240	Denied(String),
241	/// The client cancelled the turn while the request was pending; stop the
242	/// whole prompt with `StopReason::Cancelled` instead of burning rounds.
243	TurnCancelled,
244	}
245
246	/// Display kind for the permission dialog, so editors can show a fitting icon.
247	fn tool_kind_for(tool_name: &str) -> ToolKind {
248	match tool_name {
249	"edit_file" \| "multi_edit" \| "create_file" \| "create_directory" \| "remember" => {
250	ToolKind::Edit
251	}
252	"delete_file" => ToolKind::Delete,
253	"run_command" => ToolKind::Execute,
254	"read_file" \| "list_directory" => ToolKind::Read,
255	"search_files" \| "glob" => ToolKind::Search,
256	"read_website" => ToolKind::Fetch,
257	"write_todos" => ToolKind::Think,
258	_ => ToolKind::Other,
259	}
260	}
261
262	/// Shown when a siGit Code Cloud tier is selected without a signed-in account.
263	const CLOUD_LOGIN_PROMPT: &str = "siGit Code Cloud needs an account. Sign in with \
264	`/login <email> <password>` (or the Authenticate button), then pick the tier again. \
265	Create an account at https://sigit.si.";
266
267	/// The per-session context system message: cwd guidance plus any project
268	/// instruction files (`AGENTS.md` / `CLAUDE.md`) found for that directory. Used
269	/// by every session entry point so on-device and cloud backends get the same
270	/// always-on project context.
271	fn session_context_message(cwd: &std::path::Path) -> String {
272	let mut message = format!(
273	"The user's project working directory is {}. \
274	Always use absolute paths under this directory for all file \
275	and directory operations. This is the root of the project \
276	the user has open in their editor.",
277	cwd.display()
278	);
279	if let Some(project_instructions) = instructions::load_project_instructions(cwd) {
280	message.push_str("\n\n");
281	message.push_str(&project_instructions);
282	}
283	message
284	}
285
286	fn agent_tools_as_specs() -> Vec<ToolSpec> {
287	let mut specs: Vec<ToolSpec> = tools::all_tools()
288	.into_iter()
289	.map(\|t\| ToolSpec {
290	name: t.name.to_string(),
291	description: t.description.to_string(),
292	parameters_schema: t.parameters_schema.to_string(),
293	})
294	.collect();
295
296	// Advertise the `skill` tool only when skills are present, so models without
297	// any skills installed don't see a dangling capability (Agent Skills format,
298	// https://agentskills.io). Discovery metadata lives in the tool description.
299	let discovered = skills::discover_skills();
300	if !discovered.is_empty() {
301	specs.push(ToolSpec {
302	name: skills::SKILL_TOOL_NAME.to_string(),
303	description: skills::skill_tool_description(&discovered),
304	parameters_schema: skills::skill_tool_schema().to_string(),
305	});
306	}
307
308	// Tools discovered from configured MCP servers (incl. the official one).
309	specs.extend(mcp::tool_specs());
310
311	specs
312	}
313
314	fn initialize_meta() -> Meta {
315	let startup_selection = setup::startup_model_selection();
316
317	let active_model_name = startup_selection
318	.as_ref()
319	.map(\|selection\| selection.display_name.clone())
320	.unwrap_or_else(\|\| GgufModelConfig::qwen25_3b().display_name);
321
322	let active_model_id = startup_selection
323	.as_ref()
324	.and_then(\|selection\| selection.selected_model.as_ref())
325	.map(\|selected\| selected.model_id.clone())
326	.unwrap_or_else(\|\| GgufModelConfig::qwen25_3b().model_id);
327
328	let active_model_file = startup_selection
329	.as_ref()
330	.and_then(\|selection\| selection.selected_model.as_ref())
331	.map(\|selected\| selected.gguf_file.clone())
332	.unwrap_or_else(\|\| {
333	GgufModelConfig::qwen25_3b()
334	.files
335	.first()
336	.cloned()
337	.unwrap_or_default()
338	});
339
340	let mut model = serde_json::Map::new();
341	model.insert(
342	"display_name".to_string(),
343	serde_json::Value::String(active_model_name),
344	);
345	model.insert(
346	"model_id".to_string(),
347	serde_json::Value::String(active_model_id),
348	);
349	model.insert(
350	"gguf_file".to_string(),
351	serde_json::Value::String(active_model_file),
352	);
353
354	let mut sigit = serde_json::Map::new();
355	sigit.insert("active_model".to_string(), serde_json::Value::Object(model));
356
357	let mut meta = Meta::new();
358	meta.insert("sigit".to_string(), serde_json::Value::Object(sigit));
359	meta
360	}
361
362	struct SiGitAgent {
363	engine: Arc<ChatEngine>,
364	/// The active inference backend. `LocalBackend` by default; swapped to an
365	/// `OpenAiBackend` when the user selects a siGit Code Cloud tier in the panel.
366	backend: tokio::sync::Mutex<Arc<dyn InferenceBackend>>,
367	/// cwd from the editor — tool calls run here, not where the process started
368	session_cwd: std::sync::Mutex<Option<PathBuf>>,
369	current_model: std::sync::Mutex<GgufModelConfig>,
370	/// flipped once the startup model finishes (success or failure)
371	model_ready: Arc<AtomicBool>,
372	/// guards the one-time lazy startup load for ACP mode
373	startup_model_load_started: Arc<AtomicBool>,
374	/// set if the startup load failed
375	model_load_error: Arc<std::sync::Mutex<Option<String>>>,
376	/// true when the startup model isn't cached yet
377	startup_needs_download: bool,
378	/// for progress UI
379	startup_model_name: String,
380	/// for download-progress polling
381	startup_model_id: String,
382	/// Serializes turn-affecting handlers (prompt, session lifecycle, config
383	/// changes). They run in `cx.spawn`ed tasks so the JSON-RPC dispatch loop
384	/// stays free to route client responses (e.g. permission answers) mid-turn;
385	/// this lock reproduces the strict ordering the dispatch loop used to give
386	/// them for free.
387	turn_lock: Arc<tokio::sync::Mutex<()>>,
388	}
389
390	impl SiGitAgent {
391	fn new(
392	engine: Arc<ChatEngine>,
393	initial_model: GgufModelConfig,
394	model_ready: Arc<AtomicBool>,
395	startup_model_load_started: Arc<AtomicBool>,
396	model_load_error: Arc<std::sync::Mutex<Option<String>>>,
397	startup_needs_download: bool,
398	) -> Self {
399	let startup_model_name = initial_model.display_name.clone();
400	let startup_model_id = initial_model.model_id.clone();
401	let backend: Arc<dyn InferenceBackend> = Arc::new(LocalBackend::new(Arc::clone(&engine)));
402	Self {
403	engine,
404	backend: tokio::sync::Mutex::new(backend),
405	session_cwd: std::sync::Mutex::new(None),
406	current_model: std::sync::Mutex::new(initial_model),
407	model_ready,
408	startup_model_load_started,
409	model_load_error,
410	startup_needs_download,
411	startup_model_name,
412	startup_model_id,
413	turn_lock: Arc::new(tokio::sync::Mutex::new(())),
414	}
415	}
416
417	fn start_startup_model_load_if_needed(&self) {
418	if self
419	.startup_model_load_started
420	.compare_exchange(false, true, Ordering::AcqRel, Ordering::Acquire)
421	.is_err()
422	{
423	return;
424	}
425
426	self.model_ready.store(false, Ordering::Release);
427	if let Ok(mut guard) = self.model_load_error.lock() {
428	*guard = None;
429	}
430
431	let startup_config = self.current_model.lock().unwrap().clone();
432	let (max_tokens, tool_calling) = models::local_picker_items()
433	.into_iter()
434	.find(\|item\| {
435	item.config.model_id == startup_config.model_id
436	&& item
437	.config
438	.files
439	.first()
440	.zip(startup_config.files.first())
441	.map(\|(left, right)\| left == right)
442	.unwrap_or(false)
443	})
444	.map(\|item\| (item.max_tokens, item.tool_calling))
445	.unwrap_or((4096, false));
446
447	let sampling = SamplingConfig {
448	max_tokens: Some(max_tokens),
449	..SamplingConfig::default()
450	};
451
452	let loader_engine = Arc::clone(&self.engine);
453	let loader_system_prompt = system_prompt_for_model(tool_calling).to_string();
454	let model_ready = Arc::clone(&self.model_ready);
455	let model_load_error = Arc::clone(&self.model_load_error);
456
457	std::thread::spawn(move \|\| {
458	let result = tokio::runtime::Runtime::new()
459	.map_err(\|error\| error.to_string())
460	.and_then(\|rt\| {
461	rt.block_on(loader_engine.load_gguf_model(
462	startup_config,
463	Some(loader_system_prompt),
464	Some(sampling),
465	))
466	.map(\|_\| ())
467	.map_err(\|error\| error.to_string())
468	});
469
470	if let Ok(mut guard) = model_load_error.lock() {
471	*guard = result.err();
472	}
473	model_ready.store(true, Ordering::Release);
474	});
475	}
476
477	/// block until the startup model is ready, showing progress in the session.
478	async fn await_model_ready(
479	&self,
480	cx: &ConnectionTo<Client>,
481	session_id: &SessionId,
482	) -> agent_client_protocol::Result<()> {
483	if self.model_ready.load(Ordering::Acquire) {
484	// already done — might be a stored error from earlier
485	if let Some(err) = self.model_load_error.lock().unwrap().as_ref() {
486	return Err(agent_client_protocol::Error::new(
487	-32603,
488	format!("model load failed: {err}"),
489	));
490	}
491	return Ok(());
492	}
493
494	const SPINNER: &[char] = &['⠋', '⠙', '⠹', '⠸', '⠼', '⠴', '⠦', '⠧', '⠇', '⠏'];
495
496	let tool_call_id = format!("startup-load-{}", uuid::Uuid::new_v4());
497	let title = if self.startup_needs_download {
498	format!("Downloading {}", self.startup_model_name)
499	} else {
500	format!("Loading {}", self.startup_model_name)
501	};
502
503	self.send_tool_call_update(
504	cx,
505	session_id.clone(),
506	SessionUpdate::ToolCall(
507	ToolCall::new(tool_call_id.clone(), &title)
508	.kind(ToolKind::Think)
509	.status(ToolCallStatus::InProgress)
510	.content(vec![format!("{}…", title).into()]),
511	),
512	)
513	.ok();
514
515	let expected_bytes = if self.startup_needs_download {
516	onde::inference::models::SUPPORTED_MODEL_INFO
517	.iter()
518	.find(\|m\| m.id == self.startup_model_id)
519	.map(\|m\| m.expected_size_bytes)
520	.unwrap_or(0)
521	} else {
522	0
523	};
524
525	let load_start = std::time::Instant::now();
526	let mut tick: usize = 0;
527	let mut interval = tokio::time::interval(std::time::Duration::from_secs(1));
528	interval.tick().await;
529
530	loop {
531	interval.tick().await;
532	tick += 1;
533
534	if self.model_ready.load(Ordering::Acquire) {
535	break;
536	}
537
538	let frame = SPINNER[tick % SPINNER.len()];
539	let elapsed = load_start.elapsed();
540	let elapsed_str = if elapsed.as_secs() >= 60 {
541	format!("{}m {:02}s", elapsed.as_secs() / 60, elapsed.as_secs() % 60)
542	} else {
543	format!("{}s", elapsed.as_secs())
544	};
545
546	let (update_title, update_content) =
547	if self.startup_needs_download && expected_bytes > 0 {
548	let cache_path = onde::hf_cache::model_cache_path(&self.startup_model_id);
549	let downloaded = cache_path
550	.as_ref()
551	.filter(\|p\| p.exists())
552	.map(\|p\| dir_size_recursive(p))
553	.unwrap_or(0);
554	let pct = ((downloaded as f64 / expected_bytes as f64) * 100.0).min(99.0) as u8;
555	let bar = progress_bar(pct, 20);
556	let size_hint = format!(" (~{})", format_size_human(expected_bytes));
557	(
558	format!(
559	"{frame} Downloading {}{size_hint} ({pct}%)",
560	self.startup_model_name
561	),
562	format!(
563	"{} — {bar} {pct}% ({} / {})",
564	self.startup_model_name,
565	format_size_human(downloaded),
566	format_size_human(expected_bytes),
567	),
568	)
569	} else if self.startup_needs_download {
570	let cache_path = onde::hf_cache::model_cache_path(&self.startup_model_id);
571	let downloaded = cache_path
572	.as_ref()
573	.filter(\|p\| p.exists())
574	.map(\|p\| dir_size_recursive(p))
575	.unwrap_or(0);
576	(
577	format!("{frame} Downloading {}", self.startup_model_name),
578	format!(
579	"{} — {} downloaded… ({elapsed_str})",
580	self.startup_model_name,
581	format_size_human(downloaded),
582	),
583	)
584	} else {
585	(
586	format!("{frame} Loading {}", self.startup_model_name),
587	format!(
588	"{frame} Loading {}… ({elapsed_str})",
589	self.startup_model_name
590	),
591	)
592	};
593
594	self.send_tool_call_update(
595	cx,
596	session_id.clone(),
597	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
598	tool_call_id.clone(),
599	ToolCallUpdateFields::new()
600	.title(update_title)
601	.status(ToolCallStatus::InProgress)
602	.content(vec![update_content.into()]),
603	)),
604	)
605	.ok();
606	}
607
608	// done — check if it blew up
609	let load_error = self.model_load_error.lock().unwrap().clone();
610	if let Some(err) = load_error {
611	self.send_tool_call_update(
612	cx,
613	session_id.clone(),
614	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
615	tool_call_id,
616	ToolCallUpdateFields::new()
617	.title("Model load failed".to_string())
618	.status(ToolCallStatus::Failed)
619	.content(vec![format!("error: {err}").into()]),
620	)),
621	)
622	.ok();
623
624	return Err(agent_client_protocol::Error::new(
625	-32603,
626	format!("model load failed: {err}"),
627	));
628	}
629
630	let done_title = if self.startup_needs_download {
631	format!("✓ {} downloaded and loaded", self.startup_model_name)
632	} else {
633	format!("✓ {} loaded", self.startup_model_name)
634	};
635
636	self.send_tool_call_update(
637	cx,
638	session_id.clone(),
639	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
640	tool_call_id,
641	ToolCallUpdateFields::new()
642	.title(done_title)
643	.status(ToolCallStatus::Completed),
644	)),
645	)
646	.ok();
647
648	Ok(())
649	}
650
651	fn send_assistant_message(
652	&self,
653	cx: &ConnectionTo<Client>,
654	session_id: SessionId,
655	text: impl Into<String>,
656	) -> agent_client_protocol::Result<()> {
657	cx.send_notification(SessionNotification::new(
658	session_id,
659	SessionUpdate::AgentMessageChunk(ContentChunk::new(ContentBlock::from(text.into()))),
660	))
661	}
662
663	/// Run one inference turn (`fut`) while concurrently forwarding any streamed
664	/// tokens to the editor. The sink receiver is drained as the future runs, so
665	/// chunks reach the client live rather than all at once when it resolves.
666	///
667	/// `assembled`/`sent`/`streamed_any` persist across the turns of a single
668	/// prompt so reasoning is stripped consistently and we never re-send text.
669	#[allow(clippy::too_many_arguments)]
670	async fn drain_turn<F>(
671	&self,
672	cx: &ConnectionTo<Client>,
673	session_id: &SessionId,
674	fut: F,
675	sink_rx: &mut tokio::sync::mpsc::UnboundedReceiver<String>,
676	assembled: &mut String,
677	sent: &mut String,
678	streamed_any: &mut bool,
679	) -> Result<TurnResult, backend::BackendError>
680	where
681	F: std::future::Future<Output = Result<TurnResult, backend::BackendError>>,
682	{
683	tokio::pin!(fut);
684	let result = loop {
685	tokio::select! {
686	done = &mut fut => break done,
687	Some(piece) = sink_rx.recv() => {
688	self.emit_visible_chunk(cx, session_id, &piece, assembled, sent, streamed_any);
689	}
690	}
691	};
692	// Flush tokens that landed between the last poll and the future resolving.
693	while let Ok(piece) = sink_rx.try_recv() {
694	self.emit_visible_chunk(cx, session_id, &piece, assembled, sent, streamed_any);
695	}
696	result
697	}
698
699	/// Append a streamed fragment, strip `<think>` reasoning from the running
700	/// text, and send only the newly revealed visible suffix as a chunk. Tracking
701	/// the assembled text (not just deltas) keeps think-block stripping correct
702	/// even when a tag spans chunk boundaries.
703	fn emit_visible_chunk(
704	&self,
705	cx: &ConnectionTo<Client>,
706	session_id: &SessionId,
707	piece: &str,
708	assembled: &mut String,
709	sent: &mut String,
710	streamed_any: &mut bool,
711	) {
712	assembled.push_str(piece);
713	let (_think, visible) = chat::strip_think_blocks(assembled);
714	match visible.strip_prefix(sent.as_str()) {
715	Some(extra) if !extra.is_empty() => {
716	let extra = extra.to_string();
717	*sent = visible;
718	*streamed_any = true;
719	self.send_assistant_message(cx, session_id.clone(), extra)
720	.ok();
721	}
722	// No new visible text, or the visible prefix changed retroactively
723	// (rare, e.g. a late-closing think tag): just resync without
724	// resending what's already on the wire.
725	_ => *sent = visible,
726	}
727	}
728
729	fn send_tool_call_update(
730	&self,
731	cx: &ConnectionTo<Client>,
732	session_id: SessionId,
733	update: SessionUpdate,
734	) -> agent_client_protocol::Result<()> {
735	cx.send_notification(SessionNotification::new(session_id, update))
736	}
737
738	/// Advertise siGit's slash commands to the client. Editors like Zed parse
739	/// `/`-prefixed input and only forward commands they've been told about, so
740	/// without this `/login`, `/models`, etc. are rejected client-side.
741	fn advertise_commands(&self, cx: &ConnectionTo<Client>, session_id: SessionId) {
742	let with_hint = \|name: &str, desc: &str, hint: &str\| {
743	AvailableCommand::new(name, desc).input(AvailableCommandInput::Unstructured(
744	UnstructuredCommandInput::new(hint),
745	))
746	};
747	let commands = vec![
748	AvailableCommand::new("help", "Show available commands"),
749	AvailableCommand::new("models", "List available models").input(
750	AvailableCommandInput::Unstructured(UnstructuredCommandInput::new(
751	"model number to switch to (optional)",
752	)),
753	),
754	with_hint(
755	"local",
756	"Toggle on-device inference mode",
757	"on\|off (optional)",
758	),
759	AvailableCommand::new("skills", "List available Agent Skills"),
760	AvailableCommand::new("mcp", "List MCP servers and their tools"),
761	AvailableCommand::new("load", "Load the selected on-device model"),
762	with_hint("login", "Sign in to siGit Code Cloud", "<email> <password>"),
763	AvailableCommand::new("logout", "Sign out of siGit Code Cloud"),
764	AvailableCommand::new("whoami", "Show the signed-in account"),
765	AvailableCommand::new("reload", "Re-sync sign-in and model state"),
766	with_hint(
767	"plan",
768	"Plan mode: research only, no edits or commands",
769	"on\|off (optional)",
770	),
771	AvailableCommand::new("permissions", "Show the tool permission policy"),
772	AvailableCommand::new("clear", "Wipe the conversation history"),
773	AvailableCommand::new("status", "Show engine status"),
774	];
775	self.send_tool_call_update(
776	cx,
777	session_id,
778	SessionUpdate::AvailableCommandsUpdate(AvailableCommandsUpdate::new(commands)),
779	)
780	.ok();
781	}
782
783	async fn switch_model_by_id(
784	&self,
785	model_id: &str,
786	) -> agent_client_protocol::Result<GgufModelConfig> {
787	let (new_config, max_tokens, new_tool_calling) = resolve_model_config(model_id)
788	.ok_or_else(\|\| {
789	agent_client_protocol::Error::new(
790	-32602,
791	format!("unknown or unavailable model: {model_id}"),
792	)
793	})?;
794
795	log::info!(
796	"switching model to {} (max_tokens={max_tokens})",
797	new_config.display_name
798	);
799
800	let sampling = SamplingConfig {
801	max_tokens: Some(max_tokens),
802	..SamplingConfig::default()
803	};
804
805	// block_in_place inside spawn_local panics, so run the load on a
806	// dedicated thread with its own runtime (same trick as startup)
807	let (result_tx, result_rx) = tokio::sync::oneshot::channel::<Result<(), String>>();
808	let loader_engine = Arc::clone(&self.engine);
809	let loader_config = new_config.clone();
810	let loader_system_prompt = system_prompt_for_model(new_tool_calling).to_string();
811	let loader_sampling = sampling;
812
813	std::thread::spawn(move \|\| {
814	let rt = tokio::runtime::Runtime::new().expect("failed to create loader runtime");
815	let result = rt.block_on(async move {
816	// load_gguf_model already unloads the old model internally;
817	// calling unload first would leave a gap where prompts fail
818	loader_engine
819	.load_gguf_model(
820	loader_config,
821	Some(loader_system_prompt),
822	Some(loader_sampling),
823	)
824	.await
825	});
826	let _ = result_tx.send(result.map(\|_\| ()).map_err(\|e\| e.to_string()));
827	});
828
829	result_rx
830	.await
831	.map_err(\|_\| agent_client_protocol::Error::new(-32603, "model loader thread crashed"))?
832	.map_err(\|error\| {
833	log::error!("model switch failed: {error}");
834	agent_client_protocol::Error::new(-32603, format!("model switch failed: {error}"))
835	})?;
836
837	self.startup_model_load_started
838	.store(true, Ordering::Release);
839	self.model_ready.store(true, Ordering::Release);
840	if let Ok(mut guard) = self.model_load_error.lock() {
841	*guard = None;
842	}
843
844	if let Some(item) = models::local_picker_items()
845	.iter()
846	.find(\|item\| item.config.model_id == new_config.model_id)
847	&& let Err(err) = setup::save_selected_model(&setup::SelectedModel {
848	model_id: item.config.model_id.clone(),
849	gguf_file: item.config.files.first().cloned().unwrap_or_default(),
850	})
851	{
852	log::warn!("failed to persist model selection: {err}");
853	}
854
855	{
856	let mut guard = self.current_model.lock().unwrap();
857	*guard = new_config.clone();
858	}
859
860	if let Some(cwd) = self.session_cwd.lock().ok().and_then(\|g\| g.clone()) {
861	self.engine
862	.push_history(onde::inference::ChatMessage::system(
863	session_context_message(&cwd),
864	))
865	.await;
866	}
867
868	Ok(new_config)
869	}
870	}
871
872	// ── ACP handler implementations ───────────────────────────────────────────────
873
874	impl SiGitAgent {
875	async fn handle_initialize(
876	&self,
877	_req: InitializeRequest,
878	) -> agent_client_protocol::Result<InitializeResponse> {
879	log::info!("initialize");
880
881	// Agent-handled auth method. We don't use `AuthMethod::Terminal`: editors
882	// like Zed advertise terminal-auth capability but don't actually spawn the
883	// login terminal for custom ACP agents, so the button is a silent no-op.
884	// With an Agent method, clicking calls `authenticate`, which returns either
885	// confirmation (already signed in via `/login`) or a message telling the
886	// user to run `/login <email> <password>` — so the button does something.
887	let auth_methods = vec![AuthMethod::Agent(
888	AuthMethodAgent::new("sigit", "Sign in to siGit Code")
889	.description("Sign in with `/login <email> <password>` in the message box."),
890	)];
891
892	Ok(InitializeResponse::new(ProtocolVersion::V1)
893	.agent_info(
894	Implementation::new("sigit", env!("CARGO_PKG_VERSION"))
895	.title("siGit Code - AI Coding Agent"),
896	)
897	.auth_methods(auth_methods)
898	.agent_capabilities(
899	AgentCapabilities::default()
900	.load_session(true)
901	.session_capabilities(
902	SessionCapabilities::new().fork(SessionForkCapabilities::new()),
903	),
904	)
905	.meta(initialize_meta()))
906	}
907
908	async fn handle_authenticate(
909	&self,
910	req: AuthenticateRequest,
911	) -> agent_client_protocol::Result<AuthenticateResponse> {
912	log::info!("authenticate: method={}", req.method_id.0);
913
914	// Confirm the stored token works. The button can't collect a password,
915	// so an unsigned-in user is pointed at the `/login` slash command; a user
916	// already signed in via `/login` gets the gate cleared.
917	match account::verify_session().await {
918	Ok(email) => {
919	log::info!("authenticate: verified session for {email}");
920	Ok(AuthenticateResponse::default())
921	}
922	Err(reason) => Err(agent_client_protocol::Error::new(
923	-32000,
924	format!(
925	"Not signed in to siGit Code Cloud ({reason}). \
926	Sign in with `/login <email> <password>` in the message box, \
927	or create an account at https://sigit.si."
928	),
929	)),
930	}
931	}
932
933	async fn handle_load_session(
934	&self,
935	cx: &ConnectionTo<Client>,
936	args: LoadSessionRequest,
937	) -> agent_client_protocol::Result<LoadSessionResponse> {
938	log::info!(
939	"load_session: id={}, cwd={}, additional_directories={:?}",
940	args.session_id,
941	args.cwd.display(),
942	args.additional_directories
943	.iter()
944	.map(\|p\| p.display().to_string())
945	.collect::<Vec<_>>()
946	);
947
948	if let Ok(mut guard) = self.session_cwd.lock() {
949	*guard = Some(args.cwd.clone());
950	}
951
952	// A reloaded session starts fresh: grants and plan mode from the
953	// previous life of this session id must not carry over.
954	permissions::reset_session(&args.session_id.to_string());
955
956	// tool calls use relative paths, so we need to match the editor's cwd
957	if args.cwd.is_dir()
958	&& let Err(err) = std::env::set_current_dir(&args.cwd)
959	{
960	log::warn!("could not set cwd to {}: {err}", args.cwd.display());
961	}
962
963	// no session persistence, so "load" just resets
964	self.engine.clear_history().await;
965
966	self.engine
967	.push_history(onde::inference::ChatMessage::system(
968	session_context_message(&args.cwd),
969	))
970	.await;
971
972	// Honor the persisted Local Inference toggle (off + signed in → cloud).
973	self.apply_startup_inference_mode().await;
974
975	let config_options = {
976	let guard = self.current_model.lock().unwrap();
977	build_model_config_options(&guard)
978	};
979
980	self.advertise_commands(cx, args.session_id.clone());
981
982	Ok(LoadSessionResponse::new().config_options(config_options))
983	}
984
985	async fn handle_fork_session(
986	&self,
987	cx: &ConnectionTo<Client>,
988	args: ForkSessionRequest,
989	) -> agent_client_protocol::Result<ForkSessionResponse> {
990	let new_id = SessionId::new(uuid::Uuid::new_v4().to_string());
991	log::info!(
992	"fork_session: from={} new={new_id}, cwd={}, additional_directories={:?}",
993	args.session_id,
994	args.cwd.display(),
995	args.additional_directories
996	.iter()
997	.map(\|p\| p.display().to_string())
998	.collect::<Vec<_>>()
999	);
1000
1001	if let Ok(mut guard) = self.session_cwd.lock() {
1002	*guard = Some(args.cwd.clone());
1003	}
1004	if args.cwd.is_dir()
1005	&& let Err(err) = std::env::set_current_dir(&args.cwd)
1006	{
1007	log::warn!("could not set cwd to {}: {err}", args.cwd.display());
1008	}
1009
1010	// no persistence, so fork == fresh session
1011	self.engine.clear_history().await;
1012
1013	self.engine
1014	.push_history(onde::inference::ChatMessage::system(
1015	session_context_message(&args.cwd),
1016	))
1017	.await;
1018
1019	// Honor the persisted Local Inference toggle (off + signed in → cloud).
1020	self.apply_startup_inference_mode().await;
1021
1022	let config_options = {
1023	let guard = self.current_model.lock().unwrap();
1024	build_model_config_options(&guard)
1025	};
1026
1027	self.advertise_commands(cx, new_id.clone());
1028
1029	Ok(ForkSessionResponse::new(new_id).config_options(config_options))
1030	}
1031
1032	async fn handle_new_session(
1033	&self,
1034	cx: &ConnectionTo<Client>,
1035	args: NewSessionRequest,
1036	) -> agent_client_protocol::Result<NewSessionResponse> {
1037	let session_id = SessionId::new(uuid::Uuid::new_v4().to_string());
1038	log::info!(
1039	"new_session: id={session_id}, cwd={}, additional_directories={:?}",
1040	args.cwd.display(),
1041	args.additional_directories
1042	.iter()
1043	.map(\|p\| p.display().to_string())
1044	.collect::<Vec<_>>()
1045	);
1046
1047	if let Ok(mut guard) = self.session_cwd.lock() {
1048	*guard = Some(args.cwd.clone());
1049	}
1050	if args.cwd.is_dir()
1051	&& let Err(err) = std::env::set_current_dir(&args.cwd)
1052	{
1053	log::warn!("could not set cwd to {}: {err}", args.cwd.display());
1054	}
1055
1056	self.engine.clear_history().await;
1057
1058	self.engine
1059	.push_history(onde::inference::ChatMessage::system(
1060	session_context_message(&args.cwd),
1061	))
1062	.await;
1063
1064	// Honor the persisted Local Inference toggle (off + signed in → cloud).
1065	self.apply_startup_inference_mode().await;
1066
1067	let config_options = {
1068	let guard = self.current_model.lock().unwrap();
1069	build_model_config_options(&guard)
1070	};
1071
1072	self.advertise_commands(cx, session_id.clone());
1073
1074	Ok(NewSessionResponse::new(session_id).config_options(config_options))
1075	}
1076
1077	async fn handle_prompt(
1078	&self,
1079	cx: &ConnectionTo<Client>,
1080	args: PromptRequest,
1081	) -> agent_client_protocol::Result<PromptResponse> {
1082	let session_id = args.session_id.clone();
1083
1084	// log every block so we can debug @ references and file context
1085	for (i, block) in args.prompt.iter().enumerate() {
1086	match block {
1087	ContentBlock::Text(t) => {
1088	log::info!(
1089	"prompt({}) block[{}]: Text({} chars) = \"{}\"",
1090	session_id,
1091	i,
1092	t.text.len(),
1093	t.text.chars().take(200).collect::<String>()
1094	);
1095	}
1096	ContentBlock::Resource(embedded) => {
1097	log::info!(
1098	"prompt({}) block[{}]: EmbeddedResource = {:?}",
1099	session_id,
1100	i,
1101	match &embedded.resource {
1102	EmbeddedResourceResource::TextResourceContents(t) =>
1103	format!("TextResource(uri={}, {} chars)", t.uri, t.text.len()),
1104	EmbeddedResourceResource::BlobResourceContents(b) =>
1105	format!("BlobResource(uri={})", b.uri),
1106	_ => "Unknown".to_string(),
1107	}
1108	);
1109	}
1110	ContentBlock::ResourceLink(link) => {
1111	log::info!(
1112	"prompt({}) block[{}]: ResourceLink(name={}, uri={}, title={:?}, desc={:?})",
1113	session_id,
1114	i,
1115	link.name,
1116	link.uri,
1117	link.title,
1118	link.description
1119	);
1120	}
1121	other => {
1122	log::info!(
1123	"prompt({}) block[{}]: Other({:?})",
1124	session_id,
1125	i,
1126	std::mem::discriminant(other)
1127	);
1128	}
1129	}
1130	}
1131
1132	let mut parts: Vec<String> = Vec::new();
1133
1134	for block in &args.prompt {
1135	match block {
1136	ContentBlock::Text(t) => {
1137	parts.push(t.text.clone());
1138	}
1139	ContentBlock::Resource(embedded) => {
1140	// editor inlined the file content already
1141	match &embedded.resource {
1142	EmbeddedResourceResource::TextResourceContents(text_resource) => {
1143	parts.push(format!(
1144	"\n--- {} ---\n{}\n--- end {} ---",
1145	text_resource.uri, text_resource.text, text_resource.uri
1146	));
1147	}
1148	EmbeddedResourceResource::BlobResourceContents(blob) => {
1149	parts.push(format!("[binary resource: {}]", blob.uri));
1150	}
1151	_ => {
1152	log::debug!("ignoring unsupported embedded resource variant");
1153	}
1154	}
1155	}
1156	ContentBlock::ResourceLink(link) => {
1157	// reference without content; read the file ourselves
1158	let label = link.name.clone();
1159
1160	if let Some(raw_path) = link.uri.strip_prefix("file://") {
1161	let (file_path, line_range) = if let Some(hash_pos) = raw_path.rfind('#') {
1162	let fragment = &raw_path[hash_pos + 1..];
1163	let path = &raw_path[..hash_pos];
1164	// Parse "L207:219" or "L207-219" → (207, 219)
1165	let range = fragment.strip_prefix('L').and_then(\|rest\| {
1166	let sep = if rest.contains(':') { ':' } else { '-' };
1167	let mut parts = rest.splitn(2, sep);
1168	let start = parts.next()?.parse::<usize>().ok()?;
1169	let end = parts.next()?.parse::<usize>().ok()?;
1170	Some((start, end))
1171	});
1172	(path, range)
1173	} else {
1174	(raw_path, None)
1175	};
1176
1177	match std::fs::read_to_string(file_path) {
1178	Ok(contents) => {
1179	let extracted = if let Some((start, end)) = line_range {
1180	let selected: Vec<&str> = contents
1181	.lines()
1182	.enumerate()
1183	.filter(\|(i, _)\| {
1184	let line_num = i + 1;
1185	line_num >= start && line_num <= end
1186	})
1187	.map(\|(_, line)\| line)
1188	.collect();
1189	format!(
1190	"\n--- {label} ({file_path} lines {start}-{end}) ---\n{}\n--- end {label} ---",
1191	selected.join("\n")
1192	)
1193	} else {
1194	format!(
1195	"\n--- {label} ({file_path}) ---\n{contents}\n--- end {label} ---"
1196	)
1197	};
1198	parts.push(extracted);
1199	}
1200	Err(err) => {
1201	log::warn!("could not read ResourceLink {}: {err}", link.uri);
1202	parts.push(format!("[referenced file: {label} ({file_path})]"));
1203	}
1204	}
1205	} else {
1206	parts.push(format!("[resource link: {label} ({})]", link.uri));
1207	}
1208	}
1209	_ => {
1210	log::debug!("ignoring unsupported content block type in prompt");
1211	}
1212	}
1213	}
1214
1215	let user_text = parts.join("\n");
1216
1217	if user_text.trim().is_empty() {
1218	return Ok(PromptResponse::new(StopReason::EndTurn));
1219	}
1220
1221	if let Some(command) = parse_slash(&user_text) {
1222	return exec_slash_acp(self, cx, session_id, command).await;
1223	}
1224
1225	log::info!(
1226	"prompt({}): \"{}\"",
1227	session_id,
1228	user_text.chars().take(80).collect::<String>()
1229	);
1230
1231	// The active backend drives the turn. Snapshot it once so a mid-turn
1232	// model switch doesn't split the conversation across backends.
1233	let backend = self.backend.lock().await.clone();
1234
1235	// Only on-device inference needs a local model in memory. Cloud tiers run
1236	// over the network, so they never need a local model. We never load the
1237	// on-device model implicitly: the user loads it explicitly with `/load`
1238	// (or by picking one in `/models`). If a prompt arrives before that, guide
1239	// them rather than blocking on a multi-minute download/load.
1240	if !backend.is_remote()
1241	&& self.engine.info().await.status == onde::inference::EngineStatus::Unloaded
1242	{
1243	self.send_assistant_message(
1244	cx,
1245	session_id,
1246	"No on-device model is loaded. Run `/load` to load the selected model, \
1247	or `/models` to choose one.",
1248	)
1249	.ok();
1250	return Ok(PromptResponse::new(StopReason::EndTurn));
1251	}
1252
1253	// ── tool-calling loop ────────────────────────────────────────────
1254	// send message → execute any tool calls → feed results back
1255	// repeat up to MAX_TOOL_ROUNDS, then force a text reply
1256
1257	let tools = agent_tools_as_specs();
1258
1259	// Token sink: backends stream assistant text through this while a turn
1260	// runs. We forward the visible portion to the editor as agent-message
1261	// chunks live (see `drain_turn` / `emit_visible_chunk`). The sink stays
1262	// alive for the whole prompt so `recv()` only ends when a turn future
1263	// resolves, never because every sender was dropped.
1264	let (sink, mut sink_rx) = tokio::sync::mpsc::unbounded_channel::<String>();
1265	let mut assembled = String::new();
1266	let mut sent = String::new();
1267	let mut streamed_any = false;
1268
1269	let mut result = self
1270	.drain_turn(
1271	cx,
1272	&session_id,
1273	backend.send_message_with_tools(&user_text, &tools, Some(&sink)),
1274	&mut sink_rx,
1275	&mut assembled,
1276	&mut sent,
1277	&mut streamed_any,
1278	)
1279	.await
1280	.map_err(\|error\| {
1281	log::error!("send_message_with_tools failed: {error}");
1282	agent_client_protocol::Error::new(-32603, format!("inference failed: {error}"))
1283	})?;
1284
1285	let mut round = 0;
1286
1287	while !result.tool_calls.is_empty() && round < MAX_TOOL_ROUNDS {
1288	round += 1;
1289	log::info!(
1290	"prompt({}) tool round {} — {} call(s)",
1291	session_id,
1292	round,
1293	result.tool_calls.len()
1294	);
1295
1296	let mut tool_results = Vec::new();
1297
1298	for tc in &result.tool_calls {
1299	log::info!(
1300	" → {}({})",
1301	tc.name,
1302	tc.arguments.chars().take(120).collect::<String>()
1303	);
1304
1305	// Permission gate: read-only tools pass straight through; a
1306	// mutating tool consults policy and may ask the client.
1307	let output = match permissions::decision_for(&session_id.to_string(), &tc.name) {
1308	permissions::Decision::Allow => {
1309	tools::execute_tool(&tc.name, &tc.arguments).await
1310	}
1311	permissions::Decision::Deny(reason) => {
1312	log::info!(" ✗ {} denied by policy", tc.name);
1313	reason
1314	}
1315	permissions::Decision::Ask => {
1316	match self
1317	.request_tool_permission(cx, &session_id, &tc.name, &tc.arguments)
1318	.await
1319	{
1320	PermissionVerdict::Approved => {
1321	tools::execute_tool(&tc.name, &tc.arguments).await
1322	}
1323	PermissionVerdict::Denied(reason) => {
1324	log::info!(" ✗ {} denied by user", tc.name);
1325	reason
1326	}
1327	PermissionVerdict::TurnCancelled => {
1328	log::info!("prompt({}) cancelled at permission gate", session_id);
1329	return Ok(PromptResponse::new(StopReason::Cancelled));
1330	}
1331	}
1332	}
1333	};
1334
1335	log::info!(" ← {} chars", output.len());
1336
1337	tool_results.push(BackendToolResult {
1338	tool_call_id: tc.id.clone(),
1339	content: output,
1340	});
1341	}
1342
1343	let next_tools = if round < MAX_TOOL_ROUNDS {
1344	Some(tools.as_slice())
1345	} else {
1346	None // last round: force text
1347	};
1348
1349	result = self
1350	.drain_turn(
1351	cx,
1352	&session_id,
1353	backend.send_tool_results(tool_results, next_tools, Some(&sink)),
1354	&mut sink_rx,
1355	&mut assembled,
1356	&mut sent,
1357	&mut streamed_any,
1358	)
1359	.await
1360	.map_err(\|e\| agent_client_protocol::Error::new(-32603, e.to_string()))?;
1361	}
1362
1363	// ── Final text response ───────────────────────────────────────────
1364	// If anything streamed, the visible reply is already on the wire; only
1365	// send a trailing block for the non-streamed path (e.g. on-device direct
1366	// answers, which onde can't stream while tools are on offer).
1367	if !streamed_any {
1368	let reply_text = result.text.trim().to_string();
1369	let final_text = if reply_text.is_empty() {
1370	if round > 0 {
1371	log::warn!(
1372	"prompt({}) — model returned empty reply after {} tool round(s)",
1373	session_id,
1374	round
1375	);
1376	"Something went wrong — the edits didn't go through. Try rephrasing what you need, or point me at the specific lines.".to_string()
1377	} else {
1378	log::warn!(
1379	"prompt({}) — model returned empty reply (no tool rounds)",
1380	session_id
1381	);
1382	String::new()
1383	}
1384	} else {
1385	// strip <think> blocks so reasoning tokens stay hidden
1386	let (_think, visible) = chat::strip_think_blocks(&reply_text);
1387	visible
1388	};
1389
1390	if !final_text.is_empty() {
1391	self.send_assistant_message(cx, session_id.clone(), final_text)
1392	.ok();
1393	}
1394	}
1395
1396	log::info!("prompt({}) complete — {} tool round(s)", session_id, round);
1397	Ok(PromptResponse::new(StopReason::EndTurn))
1398	}
1399
1400	/// Ask the ACP client for permission to run one tool call. Presents
1401	/// allow-once / allow-for-session / deny; an "always allow" choice is
1402	/// recorded via [`permissions::grant_for_session`]. Only safe to call from
1403	/// a spawned task (see the handler registration in `run_acp_server`): the
1404	/// dispatch loop must be free to route the client's answer back to us.
1405	async fn request_tool_permission(
1406	&self,
1407	cx: &ConnectionTo<Client>,
1408	session_id: &SessionId,
1409	tool_name: &str,
1410	arguments: &str,
1411	) -> PermissionVerdict {
1412	let args_preview: String = arguments.chars().take(120).collect();
1413	let title = if args_preview.is_empty() {
1414	tool_name.to_string()
1415	} else {
1416	format!("{tool_name}({args_preview})")
1417	};
1418
1419	let request = RequestPermissionRequest::new(
1420	session_id.clone(),
1421	ToolCallUpdate::new(
1422	format!("perm-{}", uuid::Uuid::new_v4()),
1423	ToolCallUpdateFields::new()
1424	.title(title)
1425	.kind(tool_kind_for(tool_name))
1426	.status(ToolCallStatus::Pending),
1427	),
1428	vec![
1429	PermissionOption::new("allow_once", "Allow once", PermissionOptionKind::AllowOnce),
1430	PermissionOption::new(
1431	"allow_session",
1432	"Allow for this session",
1433	PermissionOptionKind::AllowAlways,
1434	),
1435	PermissionOption::new("reject_once", "Deny", PermissionOptionKind::RejectOnce),
1436	],
1437	);
1438
1439	match cx.send_request(request).block_task().await {
1440	Ok(response) => match response.outcome {
1441	RequestPermissionOutcome::Selected(selected) => {
1442	match selected.option_id.0.as_ref() {
1443	"allow_once" => PermissionVerdict::Approved,
1444	"allow_session" => {
1445	permissions::grant_for_session(&session_id.to_string(), tool_name);
1446	PermissionVerdict::Approved
1447	}
1448	_ => PermissionVerdict::Denied(permissions::user_denial(tool_name)),
1449	}
1450	}
1451	RequestPermissionOutcome::Cancelled => PermissionVerdict::TurnCancelled,
1452	// The outcome enum is non_exhaustive; treat anything unknown as
1453	// a denial rather than running a mutating tool unapproved.
1454	_ => PermissionVerdict::Denied(permissions::user_denial(tool_name)),
1455	},
1456	Err(error) => {
1457	log::warn!("permission request for `{tool_name}` failed: {error}");
1458	PermissionVerdict::Denied(format!(
1459	"`{tool_name}` was not executed: this client could not answer the \
1460	permission request ({error}). The user can pre-approve tools in \
1461	settings.toml under [permissions], or set SIGIT_PERMISSIONS=allow \
1462	for clients without permission support."
1463	))
1464	}
1465	}
1466	}
1467
1468	async fn handle_cancel(&self, args: CancelNotification) -> agent_client_protocol::Result<()> {
1469	log::info!("cancel requested for session {}", args.session_id);
1470	Ok(())
1471	}
1472
1473	/// Swap the active backend to a siGit Code Cloud tier and reflect it as the
1474	/// current model so the picker shows it selected. Returns the tier's display
1475	/// name on success, or `None` when no account is signed in (caller prompts
1476	/// for login). Shared by the panel picker and the `/models` slash command.
1477	async fn switch_to_cloud_tier(&self, tier: &str) -> Option<String> {
1478	let cfg = crate::provider::cloud_tier_provider(tier)?;
1479	let mut system_prompt = system_prompt_for_model(true).to_string();
1480	// Mirror the cwd guidance and project instruction files the local engine
1481	// gets at session load, so the cloud model shares the same project context.
1482	if let Some(cwd) = self.session_cwd.lock().ok().and_then(\|g\| g.clone()) {
1483	system_prompt.push_str("\n\n");
1484	system_prompt.push_str(&session_context_message(&cwd));
1485	}
1486	let cloud_backend: Arc<dyn InferenceBackend> = Arc::new(OpenAiBackend::new(
1487	cfg.base_url,
1488	cfg.api_key,
1489	cfg.model,
1490	Some(system_prompt),
1491	));
1492	*self.backend.lock().await = cloud_backend;
1493
1494	let cloud_config = GgufModelConfig {
1495	model_id: format!("sigit-cloud:{tier}"),
1496	files: Vec::new(),
1497	tok_model_id: None,
1498	display_name: cfg.display_name.clone(),
1499	approx_memory: "Cloud".to_string(),
1500	chat_template: None,
1501	};
1502	{
1503	let mut guard = self.current_model.lock().unwrap();
1504	*guard = cloud_config;
1505	}
1506
1507	// Explicitly choosing a cloud tier puts us in cloud mode.
1508	let _ = settings::set_local_inference(false);
1509
1510	log::info!("switched to cloud tier {tier}");
1511	Some(cfg.display_name)
1512	}
1513
1514	/// Apply the persisted Local Inference mode at session start. When local
1515	/// inference is off and an account is signed in, route to a cloud tier so the
1516	/// on-device model is never loaded; otherwise leave the on-device backend in
1517	/// place. Call after the session cwd is set so the cloud system prompt picks
1518	/// it up. Does not flip the stored setting on the not-signed-in fallback.
1519	async fn apply_startup_inference_mode(&self) {
1520	if settings::local_inference_enabled() {
1521	return;
1522	}
1523	if self.switch_to_cloud_tier("balanced").await.is_some() {
1524	log::info!("startup: local inference off; routing inference to siGit Code Cloud");
1525	} else {
1526	log::warn!(
1527	"local inference is off but no account is signed in; staying on-device. \
1528	Run /login or set Local Inference on."
1529	);
1530	}
1531	}
1532
1533	/// Route inference back on-device. Used after leaving a cloud tier for a
1534	/// local model. The `LocalBackend` reads the live `engine`, so this just
1535	/// repoints the active backend.
1536	async fn reset_to_local_backend(&self) {
1537	let local_backend: Arc<dyn InferenceBackend> =
1538	Arc::new(LocalBackend::new(Arc::clone(&self.engine)));
1539	*self.backend.lock().await = local_backend;
1540	}
1541
1542	/// Re-attempt the lazy startup model load if the previous attempt failed.
1543	/// Clears the one-shot guard so the next load runs; a healthy load is left
1544	/// untouched so `/reload` doesn't needlessly reload a working model.
1545	fn retry_startup_model_load_if_failed(&self) {
1546	let had_error = self
1547	.model_load_error
1548	.lock()
1549	.map(\|guard\| guard.is_some())
1550	.unwrap_or(false);
1551	if had_error {
1552	self.startup_model_load_started
1553	.store(false, Ordering::Release);
1554	self.start_startup_model_load_if_needed();
1555	}
1556	}
1557
1558	/// Re-sync session state in place — no new session needed. Re-applies the
1559	/// active backend from current credentials (so a fresh `/login` token is
1560	/// picked up), retries a failed model load, and pushes refreshed commands +
1561	/// picker so the editor's UI reflects the current state.
1562	async fn handle_reload(&self, cx: &ConnectionTo<Client>, session_id: SessionId) {
1563	let signed_in = account::status_line().await;
1564
1565	let on_cloud_tier = {
1566	let guard = self.current_model.lock().unwrap();
1567	guard
1568	.model_id
1569	.strip_prefix("sigit-cloud:")
1570	.map(str::to_string)
1571	};
1572
1573	let backend_note = match on_cloud_tier {
1574	Some(tier) => match self.switch_to_cloud_tier(&tier).await {
1575	Some(name) => format!("Active: {name}."),
1576	None => {
1577	self.reset_to_local_backend().await;
1578	"Signed out — back to on-device. Pick a model with /models.".to_string()
1579	}
1580	},
1581	None => {
1582	self.reset_to_local_backend().await;
1583	self.retry_startup_model_load_if_failed();
1584	let guard = self.current_model.lock().unwrap();
1585	format!("Active: {}.", guard.display_name)
1586	}
1587	};
1588
1589	// Push refreshed picker + commands so the editor reflects current state.
1590	let config_options = {
1591	let guard = self.current_model.lock().unwrap();
1592	build_model_config_options(&guard)
1593	};
1594	self.send_tool_call_update(
1595	cx,
1596	session_id.clone(),
1597	SessionUpdate::ConfigOptionUpdate(ConfigOptionUpdate::new(config_options)),
1598	)
1599	.ok();
1600	self.advertise_commands(cx, session_id.clone());
1601
1602	self.send_assistant_message(
1603	cx,
1604	session_id,
1605	format!("Reloaded. {signed_in} {backend_note}"),
1606	)
1607	.ok();
1608	}
1609
1610	async fn handle_set_session_config_option(
1611	&self,
1612	cx: &ConnectionTo<Client>,
1613	args: SetSessionConfigOptionRequest,
1614	) -> agent_client_protocol::Result<SetSessionConfigOptionResponse> {
1615	log::info!(
1616	"set_session_config_option: config_id={}, value={:?}",
1617	args.config_id,
1618	args.value
1619	);
1620
1621	// ── Local Inference toggle ──────────────────────────────────────────
1622	if args.config_id.0.as_ref() == LOCAL_INFERENCE_CONFIG_ID {
1623	let enabled = match args.value.0.as_ref() {
1624	LOCAL_INFERENCE_ON => true,
1625	LOCAL_INFERENCE_OFF => false,
1626	other => {
1627	return Err(agent_client_protocol::Error::new(
1628	-32602,
1629	format!("unknown Local Inference value: {other}"),
1630	));
1631	}
1632	};
1633	if let Err(error) = settings::set_local_inference(enabled) {
1634	return Err(agent_client_protocol::Error::new(
1635	-32603,
1636	format!("could not save Local Inference setting: {error}"),
1637	));
1638	}
1639	let message = if enabled {
1640	"Local inference is on. On-device models are highlighted; pick one from Model."
1641	} else {
1642	"Local inference is off. siGit Code Cloud tiers are highlighted; pick one from Model."
1643	};
1644	self.send_assistant_message(cx, args.session_id.clone(), format!("\n\n{message}"))
1645	.ok();
1646	// Rebuild so the Model picker reflects the new emphasis/order.
1647	let current = self.current_model.lock().unwrap().clone();
1648	let config_options = build_model_config_options(&current);
1649	return Ok(SetSessionConfigOptionResponse::new(config_options));
1650	}
1651
1652	if args.config_id.0.as_ref() != MODEL_CONFIG_ID {
1653	return Err(agent_client_protocol::Error::new(
1654	-32602,
1655	format!("unknown config option: {}", args.config_id.0),
1656	));
1657	}
1658
1659	let model_id = args.value.0.as_ref();
1660
1661	// can't switch while the startup model is still loading — the old
1662	// weights are in GPU memory and the new load gets "does not fit"
1663	if self.startup_model_load_started.load(Ordering::Acquire)
1664	&& !self.model_ready.load(Ordering::Acquire)
1665	{
1666	log::info!("set_session_config_option: waiting for startup model to finish loading");
1667	while !self.model_ready.load(Ordering::Acquire) {
1668	tokio::time::sleep(std::time::Duration::from_millis(200)).await;
1669	}
1670	}
1671
1672	// Zed re-fires the last selection when a thread opens. That re-fire must
1673	// not load anything: on-device models are loaded only on an explicit
1674	// request (`/load`, or actively picking a different model below), so a
1675	// re-fire of the already-current selection is a no-op. Otherwise opening a
1676	// new thread would silently load the local model — exactly what we avoid.
1677	{
1678	let current = self.current_model.lock().unwrap();
1679	if current.model_id == model_id {
1680	log::info!(
1681	"set_session_config_option: {} is already the active selection, skipping",
1682	current.display_name
1683	);
1684	let config_options = build_model_config_options(&current);
1685	return Ok(SetSessionConfigOptionResponse::new(config_options));
1686	}
1687	}
1688
1689	// ── siGit Code Cloud tier: no local load; sign-in gated ─────────────
1690	if let Some(tier) = model_id.strip_prefix("sigit-cloud:") {
1691	let message = match self.switch_to_cloud_tier(tier).await {
1692	Some(display_name) => format!("Switched to {display_name}."),
1693	None => CLOUD_LOGIN_PROMPT.to_string(),
1694	};
1695	// Start on a fresh line: ACP clients concatenate consecutive
1696	// agent-message chunks into one block, so without this the switch
1697	// confirmation runs onto the end of the previous assistant message.
1698	self.send_assistant_message(cx, args.session_id.clone(), format!("\n\n{message}"))
1699	.ok();
1700
1701	let current = self.current_model.lock().unwrap().clone();
1702	let config_options = build_model_config_options(&current);
1703	return Ok(SetSessionConfigOptionResponse::new(config_options));
1704	}
1705
1706	let needs_download = models::local_picker_items()
1707	.into_iter()
1708	.find(\|item\| item.config.model_id == model_id)
1709	.map(\|item\| item.cache_health == setup::ModelCacheHealth::NotDownloaded)
1710	.unwrap_or(false);
1711
1712	// tells the progress poller to stop
1713	let stop_flag = Arc::new(AtomicBool::new(false));
1714
1715	let tool_call_id = format!("model-switch-{}", uuid::Uuid::new_v4());
1716
1717	if needs_download {
1718	let model_id_owned = model_id.to_string();
1719	let expected_bytes = onde::inference::models::SUPPORTED_MODEL_INFO
1720	.iter()
1721	.find(\|m\| m.id == model_id_owned)
1722	.map(\|m\| m.expected_size_bytes)
1723	.unwrap_or(0);
1724
1725	let display_name = models::local_picker_items()
1726	.into_iter()
1727	.find(\|item\| item.config.model_id == model_id_owned)
1728	.map(\|item\| item.display_name.clone())
1729	.unwrap_or_else(\|\| model_id_owned.clone());
1730
1731	let size_hint = if expected_bytes > 0 {
1732	format!(" (~{})", format_size_human(expected_bytes))
1733	} else {
1734	String::new()
1735	};
1736
1737	self.send_tool_call_update(
1738	cx,
1739	args.session_id.clone(),
1740	SessionUpdate::ToolCall(
1741	ToolCall::new(
1742	tool_call_id.clone(),
1743	format!("⏬ Downloading {display_name}{size_hint}"),
1744	)
1745	.kind(ToolKind::Think)
1746	.status(ToolCallStatus::InProgress)
1747	.content(vec![
1748	format!(
1749	"Preparing download for {display_name}. This may take a few minutes."
1750	)
1751	.into(),
1752	]),
1753	),
1754	)
1755	.ok();
1756
1757	// poll download progress and update the spinner in Zed
1758	let cx_for_poller = cx.clone();
1759	let poller_session = args.session_id.clone();
1760	let poller_model_id = model_id_owned.clone();
1761	let poller_stop = Arc::clone(&stop_flag);
1762	let poller_tool_call_id = tool_call_id.clone();
1763
1764	cx.spawn(async move {
1765	const SPINNER: &[char] = &['⠋', '⠙', '⠹', '⠸', '⠼', '⠴', '⠦', '⠧', '⠇', '⠏'];
1766	let cache_path = onde::hf_cache::model_cache_path(&poller_model_id);
1767	let mut tick: usize = 0;
1768	let mut interval = tokio::time::interval(std::time::Duration::from_secs(1));
1769	interval.tick().await; // consume the immediate first tick
1770
1771	while !poller_stop.load(Ordering::Relaxed) {
1772	interval.tick().await;
1773
1774	if poller_stop.load(Ordering::Relaxed) {
1775	break;
1776	}
1777
1778	let downloaded = cache_path
1779	.as_ref()
1780	.filter(\|p\| p.exists())
1781	.map(\|p\| dir_size_recursive(p))
1782	.unwrap_or(0);
1783
1784	let frame = SPINNER[tick % SPINNER.len()];
1785	tick += 1;
1786
1787	let title = if expected_bytes > 0 {
1788	let pct =
1789	((downloaded as f64 / expected_bytes as f64) * 100.0).min(99.0) as u8;
1790	format!("{frame} Downloading {display_name}{size_hint} ({pct}%)")
1791	} else {
1792	format!("{frame} Downloading {display_name}{size_hint}")
1793	};
1794
1795	let msg = if expected_bytes > 0 {
1796	let pct =
1797	((downloaded as f64 / expected_bytes as f64) * 100.0).min(99.0) as u8;
1798	let bar = progress_bar(pct, 20);
1799	format!(
1800	"{display_name} — {bar} {pct}% ({} / {})",
1801	format_size_human(downloaded),
1802	format_size_human(expected_bytes),
1803	)
1804	} else {
1805	format!(
1806	"{display_name} — {} downloaded…",
1807	format_size_human(downloaded)
1808	)
1809	};
1810
1811	let notification = SessionNotification::new(
1812	poller_session.clone(),
1813	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
1814	poller_tool_call_id.clone(),
1815	ToolCallUpdateFields::new()
1816	.title(title)
1817	.status(ToolCallStatus::InProgress)
1818	.content(vec![msg.into()]),
1819	)),
1820	);
1821	if cx_for_poller.send_notification(notification).is_err() {
1822	break;
1823	}
1824	}
1825	Ok(())
1826	})
1827	.ok();
1828	}
1829
1830	// cached models still take 10-30s to load weights; show a spinner
1831	if !needs_download {
1832	let cached_display_name = models::local_picker_items()
1833	.into_iter()
1834	.find(\|item\| item.config.model_id == model_id)
1835	.map(\|item\| item.display_name.clone())
1836	.unwrap_or_else(\|\| model_id.to_string());
1837
1838	self.send_tool_call_update(
1839	cx,
1840	args.session_id.clone(),
1841	SessionUpdate::ToolCall(
1842	ToolCall::new(
1843	tool_call_id.clone(),
1844	format!("Loading {cached_display_name}"),
1845	)
1846	.kind(ToolKind::Think)
1847	.status(ToolCallStatus::InProgress)
1848	.content(vec![format!("Loading {cached_display_name}…").into()]),
1849	),
1850	)
1851	.ok();
1852
1853	// tick every 5s so the user knows we haven't frozen
1854	let cx_for_spinner = cx.clone();
1855	let spinner_session = args.session_id.clone();
1856	let spinner_name = cached_display_name.clone();
1857	let spinner_stop = Arc::clone(&stop_flag);
1858	let spinner_tool_call_id = tool_call_id.clone();
1859	let load_start = std::time::Instant::now();
1860
1861	cx.spawn(async move {
1862	const SPINNER: &[char] = &['⠋', '⠙', '⠹', '⠸', '⠼', '⠴', '⠦', '⠧', '⠇', '⠏'];
1863	let mut tick: usize = 0;
1864	let mut interval = tokio::time::interval(std::time::Duration::from_secs(5));
1865	interval.tick().await; // consume the immediate first tick
1866
1867	while !spinner_stop.load(Ordering::Relaxed) {
1868	interval.tick().await;
1869
1870	if spinner_stop.load(Ordering::Relaxed) {
1871	break;
1872	}
1873
1874	let elapsed = load_start.elapsed();
1875	let elapsed_str = if elapsed.as_secs() >= 60 {
1876	format!("{}m {:02}s", elapsed.as_secs() / 60, elapsed.as_secs() % 60)
1877	} else {
1878	format!("{}s", elapsed.as_secs())
1879	};
1880	let frame = SPINNER[tick % SPINNER.len()];
1881	tick += 1;
1882
1883	let msg = format!("{frame} Loading {spinner_name}… ({elapsed_str})");
1884	let notification = SessionNotification::new(
1885	spinner_session.clone(),
1886	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
1887	spinner_tool_call_id.clone(),
1888	ToolCallUpdateFields::new()
1889	.status(ToolCallStatus::InProgress)
1890	.content(vec![msg.into()]),
1891	)),
1892	);
1893	if cx_for_spinner.send_notification(notification).is_err() {
1894	break;
1895	}
1896	}
1897	Ok(())
1898	})
1899	.ok();
1900	}
1901
1902	let switch_result = self.switch_model_by_id(model_id).await;
1903
1904	stop_flag.store(true, Ordering::Relaxed);
1905
1906	match switch_result {
1907	Ok(new_config) => {
1908	// Route inference back on-device (in case we were on a cloud tier).
1909	self.reset_to_local_backend().await;
1910	// Selecting an on-device model puts us in local mode.
1911	let _ = settings::set_local_inference(true);
1912
1913	let completion_title = if needs_download {
1914	format!("✓ {} downloaded and loaded", new_config.display_name)
1915	} else {
1916	format!("✓ Switched to {}", new_config.display_name)
1917	};
1918	let completion_body = if needs_download {
1919	format!("✓ {} downloaded and loaded.", new_config.display_name)
1920	} else {
1921	format!("✓ Switched to {}.", new_config.display_name)
1922	};
1923
1924	self.send_tool_call_update(
1925	cx,
1926	args.session_id.clone(),
1927	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
1928	tool_call_id,
1929	ToolCallUpdateFields::new()
1930	.title(completion_title)
1931	.status(ToolCallStatus::Completed)
1932	.content(vec![completion_body.into()]),
1933	)),
1934	)
1935	.ok();
1936
1937	let config_options = {
1938	let guard = self.current_model.lock().unwrap();
1939	build_model_config_options(&guard)
1940	};
1941
1942	log::info!("model switch complete");
1943	Ok(SetSessionConfigOptionResponse::new(config_options))
1944	}
1945	Err(err) => {
1946	self.send_tool_call_update(
1947	cx,
1948	args.session_id.clone(),
1949	SessionUpdate::ToolCallUpdate(ToolCallUpdate::new(
1950	tool_call_id,
1951	ToolCallUpdateFields::new()
1952	.title("Model switch failed".to_string())
1953	.status(ToolCallStatus::Failed)
1954	.content(vec![format!("error loading model: {}", err.message).into()]),
1955	)),
1956	)
1957	.ok();
1958
1959	Err(err)
1960	}
1961	}
1962	}
1963	}
1964
1965	// ── Config option helpers ─────────────────────────────────────────────────────
1966
1967	/// config option ID for the model picker in Zed's agent panel
1968	const MODEL_CONFIG_ID: &str = "sigit-model";
1969
1970	/// config option ID for the Local Inference on/off toggle. Surfaced as a
1971	/// two-option `select` so ACP clients without slash-command support (e.g. Xcode)
1972	/// can still flip the mode from the agent panel.
1973	const LOCAL_INFERENCE_CONFIG_ID: &str = "sigit-local-inference";
1974
1975	/// `select` value ids for the Local Inference toggle.
1976	const LOCAL_INFERENCE_ON: &str = "local-inference-on";
1977	const LOCAL_INFERENCE_OFF: &str = "local-inference-off";
1978
1979	/// Replace non-ASCII chars so a downstream byte-index truncation can't split a
1980	/// multi-byte char. Zed slices the model-picker label at a fixed byte offset
1981	/// (`agent_ui/src/config_options.rs`) and panics — crashing the whole editor —
1982	/// when the cut lands mid-glyph (e.g. inside `☁` or `·`). Mapping to `-` keeps
1983	/// separators readable; ASCII bytes are always char boundaries.
1984	fn ascii_safe(s: &str) -> String {
1985	s.chars()
1986	.map(\|c\| if c.is_ascii() { c } else { '-' })
1987	.collect()
1988	}
1989
1990	fn build_model_config_options(current_model: &GgufModelConfig) -> Vec<SessionConfigOption> {
1991	// The full list, including the siGit Code Cloud tiers, so the panel picker
1992	// mirrors the TUI `/models`. Cloud entries are sign-in gated at selection.
1993	let items = models::build_model_picker_items();
1994	let active_kind = models::active_inference_kind();
1995
1996	let options: Vec<SessionConfigSelectOption> = items
1997	.iter()
1998	.filter(\|item\| item.cache_health != setup::ModelCacheHealth::Incomplete)
1999	.map(\|item\| {
2000	let mut desc_parts = Vec::new();
2001	// Mark options in the inactive mode so the active group reads as the
2002	// recommended set (the list is already ordered active-group-first).
2003	if item.source.kind() != active_kind {
2004	desc_parts.push("inactive mode".to_string());
2005	}
2006	if item.tool_calling {
2007	desc_parts.push("tool calling".to_string());
2008	}
2009	desc_parts.push(item.description.clone());
2010	if item.cache_health == setup::ModelCacheHealth::NotDownloaded {
2011	desc_parts.push("download on select".to_string());
2012	}
2013	// ASCII-only for the same reason as the name (see `ascii_safe`).
2014	let description = ascii_safe(&desc_parts.join(" - "));
2015	// Keep badges ASCII: Zed truncates the picker label at a fixed byte
2016	// offset and panics if the cut splits a multi-byte char. See
2017	// `ascii_safe` below.
2018	let source_badge = if item.cloud_tier.is_some() {
2019	" [siGit Code Cloud]"
2020	} else if item.cache_health == setup::ModelCacheHealth::NotDownloaded {
2021	" [Onde]"
2022	} else {
2023	match item.source_label.as_str() {
2024	"Onde" => " [Onde]",
2025	"HuggingFace" => " [HuggingFace]",
2026	_ => "",
2027	}
2028	};
2029	// For cloud tiers use just the tier title (e.g. "Balanced") so the
2030	// label reads "Balanced [siGit Code Cloud]" instead of repeating the
2031	// brand. The display name can carry non-ASCII (the cloud tier label
2032	// is "siGit Code Cloud · Balanced"), so sanitize the whole label.
2033	let base_name = match &item.cloud_tier {
2034	Some(tier) => crate::provider::tier_title(tier),
2035	None => item.display_name.clone(),
2036	};
2037	let name = ascii_safe(&format!("{base_name}{source_badge}"));
2038	SessionConfigSelectOption::new(
2039	SessionConfigValueId::new(item.config.model_id.as_str()),
2040	name,
2041	)
2042	.description(description)
2043	})
2044	.collect();
2045
2046	// Local Inference on/off toggle, modeled as a two-option select so panel-only
2047	// ACP clients (no slash commands) can flip the mode.
2048	let local_on = settings::local_inference_enabled();
2049	let local_current = SessionConfigValueId::new(if local_on {
2050	LOCAL_INFERENCE_ON
2051	} else {
2052	LOCAL_INFERENCE_OFF
2053	});
2054	let local_options = vec![
2055	SessionConfigSelectOption::new(
2056	SessionConfigValueId::new(LOCAL_INFERENCE_ON),
2057	"On (on-device)".to_string(),
2058	)
2059	.description("Run inference on-device; on-device models are highlighted".to_string()),
2060	SessionConfigSelectOption::new(
2061	SessionConfigValueId::new(LOCAL_INFERENCE_OFF),
2062	"Off (siGit Code Cloud)".to_string(),
2063	)
2064	.description("Use siGit Code Cloud; cloud tiers are highlighted".to_string()),
2065	];
2066	let local_option = SessionConfigOption::select(
2067	LOCAL_INFERENCE_CONFIG_ID,
2068	"Local Inference",
2069	local_current,
2070	local_options,
2071	)
2072	.description("Toggle on-device inference; changes which models are highlighted");
2073
2074	if options.is_empty() {
2075	return vec![local_option];
2076	}
2077
2078	let current_value = SessionConfigValueId::new(current_model.model_id.as_str());
2079
2080	vec![
2081	SessionConfigOption::select(MODEL_CONFIG_ID, "Model", current_value, options)
2082	.category(SessionConfigOptionCategory::Model)
2083	.description("Select an on-device model or a siGit Code Cloud tier"),
2084	local_option,
2085	]
2086	}
2087
2088	/// returns `(config, max_tokens, tool_calling)` for a picker model_id, or None
2089	fn resolve_model_config(model_id: &str) -> Option<(GgufModelConfig, u64, bool)> {
2090	let items = models::local_picker_items();
2091	items
2092	.into_iter()
2093	.find(\|item\| {
2094	item.config.model_id == model_id
2095	&& item.cache_health != setup::ModelCacheHealth::Incomplete
2096	})
2097	.map(\|item\| (item.config, item.max_tokens, item.tool_calling))
2098	}
2099
2100	// ── Slash commands ────────────────────────────────────────────────────────────
2101
2102	#[derive(Debug, Clone)]
2103	enum SlashCommand {
2104	Help,
2105	Clear,
2106	Status,
2107	Models(Option<usize>),
2108	/// toggle on-device inference mode. `Some(true/false)` sets it, `None` flips it.
2109	Local(Option<bool>),
2110	/// List discovered Agent Skills.
2111	Skills,
2112	/// List configured MCP servers and their tools.
2113	Mcp,
2114	/// Explicitly load the selected (or default) on-device model.
2115	Load,
2116	/// `/login <email> <password>` — the raw argument, parsed when executed.
2117	Login(Option<String>),
2118	Logout,
2119	Whoami,
2120	/// Re-sync session state (auth, backend, picker) without a new session.
2121	Reload,
2122	/// Toggle plan mode (read-only research; mutating tools denied with a
2123	/// prompt to present a plan). `Some(true/false)` sets it, `None` flips it.
2124	Plan(Option<bool>),
2125	/// Show the effective permission policy for this session.
2126	Permissions,
2127	Exit,
2128	Unknown(String),
2129	}
2130
2131	fn parse_slash(input: &str) -> Option<SlashCommand> {
2132	let trimmed = input.trim();
2133	if !trimmed.starts_with('/') {
2134	return None;
2135	}
2136	let mut parts = trimmed.splitn(2, char::is_whitespace);
2137	let command = parts.next().unwrap_or("");
2138	let argument = parts.next().map(str::trim);
2139	Some(match command {
2140	"/help" => SlashCommand::Help,
2141	"/clear" => SlashCommand::Clear,
2142	"/status" => SlashCommand::Status,
2143	"/models" => SlashCommand::Models(argument.and_then(\|v\| v.parse::<usize>().ok())),
2144	"/local" => SlashCommand::Local(parse_on_off(argument)),
2145	"/skills" => SlashCommand::Skills,
2146	"/mcp" => SlashCommand::Mcp,
2147	"/load" => SlashCommand::Load,
2148	"/login" => SlashCommand::Login(argument.map(str::to_string)),
2149	"/logout" => SlashCommand::Logout,
2150	"/whoami" => SlashCommand::Whoami,
2151	"/reload" => SlashCommand::Reload,
2152	"/plan" => SlashCommand::Plan(parse_on_off(argument)),
2153	"/permissions" => SlashCommand::Permissions,
2154	"/exit" \| "/quit" \| "/q" => SlashCommand::Exit,
2155	other => SlashCommand::Unknown(other.to_string()),
2156	})
2157	}
2158
2159	/// `on`/`off` (and synonyms) → `Some(bool)`; missing or unrecognized → `None`
2160	/// (meaning "toggle the current value").
2161	fn parse_on_off(arg: Option<&str>) -> Option<bool> {
2162	match arg.map(\|s\| s.trim().to_ascii_lowercase())?.as_str() {
2163	"on" \| "true" \| "1" \| "yes" => Some(true),
2164	"off" \| "false" \| "0" \| "no" => Some(false),
2165	_ => None,
2166	}
2167	}
2168
2169	fn format_models_list(current_model: &GgufModelConfig) -> String {
2170	let items = models::build_model_picker_items();
2171	if items.is_empty() {
2172	return "No local models found. siGit will use the platform default model.".to_string();
2173	}
2174
2175	let mut lines = vec!["Available models:".to_string()];
2176	let mut last_source: Option<&str> = None;
2177
2178	for (index, item) in items.iter().enumerate() {
2179	let source_key = match item.source_label.as_str() {
2180	"Onde" => "Onde",
2181	"HuggingFace" => "HuggingFace",
2182	"siGit Code Cloud" => "Cloud",
2183	_ => "Fallback",
2184	};
2185
2186	if last_source != Some(source_key) {
2187	if last_source.is_some() {
2188	lines.push(String::new());
2189	}
2190	let section = match source_key {
2191	"Onde" => "Onde Inference",
2192	"HuggingFace" => "Hugging Face cache",
2193	"Cloud" => "siGit Code Cloud",
2194	_ => "Fallback",
2195	};
2196	lines.push(section.to_string());
2197	// Blank line so the following "N." items render as an ordered list.
2198	// CommonMark only lets an ordered list interrupt a paragraph when it
2199	// starts at 1, so without this the cloud section (items 9+) would be
2200	// absorbed into the header paragraph.
2201	lines.push(String::new());
2202	last_source = Some(source_key);
2203	}
2204
2205	let number = index + 1;
2206	let current_badge = if item.config.model_id == current_model.model_id {
2207	" <- current"
2208	} else {
2209	""
2210	};
2211	let tool_badge = if item.tool_calling {
2212	" tool calling"
2213	} else {
2214	""
2215	};
2216	let health_badge = match item.cache_health {
2217	setup::ModelCacheHealth::Complete => "",
2218	setup::ModelCacheHealth::Incomplete => " ! incomplete cache",
2219	setup::ModelCacheHealth::NotDownloaded => " ↓ download on select",
2220	};
2221	let source = match source_key {
2222	"Onde" => " [Onde]",
2223	"HuggingFace" => " [HuggingFace]",
2224	"Cloud" => " [☁ Cloud]",
2225	_ => " [default]",
2226	};
2227
2228	lines.push(format!(
2229	"{number}. {} {}{}{}{}{}",
2230	item.display_name, item.description, tool_badge, health_badge, current_badge, source,
2231	));
2232	}
2233
2234	lines.push(String::new());
2235	lines.push("Use /models N to switch models.".to_string());
2236	lines.join("\n")
2237	}
2238
2239	async fn exec_slash_acp(
2240	agent: &SiGitAgent,
2241	cx: &ConnectionTo<Client>,
2242	session_id: SessionId,
2243	command: SlashCommand,
2244	) -> agent_client_protocol::Result<PromptResponse> {
2245	match command {
2246	SlashCommand::Help => {
2247	agent
2248	.send_assistant_message(
2249	cx,
2250	session_id,
2251	"/help - show this message\n\
2252	/models - list available models\n\
2253	/models N - switch to model N\n\
2254	/local [on\|off]- toggle on-device inference mode\n\
2255	/skills - list available Agent Skills\n\
2256	/mcp - list MCP servers and their tools\n\
2257	/load - load the selected on-device model\n\
2258	/login E P - sign in to siGit Code Cloud\n\
2259	/logout - sign out\n\
2260	/whoami - show the signed-in account\n\
2261	/reload - re-sync sign-in and model state\n\
2262	/plan [on\|off] - plan mode: research only, no edits or commands\n\
2263	/permissions - show the tool permission policy\n\
2264	/clear - wipe conversation history\n\
2265	/status - show engine status\n\
2266	/exit - end this turn",
2267	)
2268	.ok();
2269	}
2270	SlashCommand::Clear => {
2271	let cleared = agent.engine.clear_history().await;
2272	permissions::reset_session(&session_id.to_string());
2273	agent
2274	.send_assistant_message(
2275	cx,
2276	session_id,
2277	format!("Cleared {cleared} turn(s). History is empty."),
2278	)
2279	.ok();
2280	}
2281	SlashCommand::Plan(value) => {
2282	let session_key = session_id.to_string();
2283	let enabled = value.unwrap_or_else(\|\| !permissions::plan_mode(&session_key));
2284	permissions::set_plan_mode(&session_key, enabled);
2285	let message = if enabled {
2286	"Plan mode ON — the agent researches with read-only tools and presents a \
2287	plan; edits and commands are blocked until /plan off."
2288	} else {
2289	"Plan mode OFF — the agent may execute tools again (subject to the \
2290	permission policy)."
2291	};
2292	agent.send_assistant_message(cx, session_id, message).ok();
2293	}
2294	SlashCommand::Permissions => {
2295	let summary = permissions::describe(&session_id.to_string());
2296	agent.send_assistant_message(cx, session_id, summary).ok();
2297	}
2298	SlashCommand::Status => {
2299	let info = agent.engine.info().await;
2300	let model = info.model_name.as_deref().unwrap_or("(none)");
2301	let memory = info.approx_memory.as_deref().unwrap_or("unknown");
2302	agent
2303	.send_assistant_message(
2304	cx,
2305	session_id,
2306	format!(
2307	"status: {:?} model: {} memory: {} history: {} turns",
2308	info.status, model, memory, info.history_length,
2309	),
2310	)
2311	.ok();
2312	}
2313	SlashCommand::Models(None) => {
2314	let current_model = agent.current_model.lock().unwrap().clone();
2315	agent
2316	.send_assistant_message(cx, session_id, format_models_list(&current_model))
2317	.ok();
2318	}
2319	SlashCommand::Skills => {
2320	agent
2321	.send_assistant_message(cx, session_id, skills::format_skills_list())
2322	.ok();
2323	}
2324	SlashCommand::Mcp => {
2325	agent
2326	.send_assistant_message(cx, session_id, mcp::status_summary())
2327	.ok();
2328	}
2329	SlashCommand::Models(Some(number)) => {
2330	let items = models::build_model_picker_items();
2331	let index = number.saturating_sub(1);
2332	match items.get(index).cloned() {
2333	None => {
2334	agent
2335	.send_assistant_message(
2336	cx,
2337	session_id,
2338	format!("error: no model #{number} - type /models to see the list."),
2339	)
2340	.ok();
2341	}
2342	Some(model) if model.cloud_tier.is_some() => {
2343	// siGit Code Cloud tier: swap backend, sign-in gated.
2344	let tier = model.cloud_tier.clone().unwrap_or_default();
2345	let message = match agent.switch_to_cloud_tier(&tier).await {
2346	Some(display_name) => format!("Switched to {display_name}."),
2347	None => CLOUD_LOGIN_PROMPT.to_string(),
2348	};
2349	agent.send_assistant_message(cx, session_id, message).ok();
2350	}
2351	Some(model) => {
2352	if model.cache_health == setup::ModelCacheHealth::Incomplete {
2353	agent
2354	.send_assistant_message(
2355	cx,
2356	session_id,
2357	format!(
2358	"error: {} has an incomplete local cache and cannot be selected yet.",
2359	model.display_name
2360	),
2361	)
2362	.ok();
2363	} else if model.cache_health == setup::ModelCacheHealth::NotDownloaded {
2364	agent
2365	.send_assistant_message(
2366	cx,
2367	session_id.clone(),
2368	format!(
2369	"Downloading and loading {} ({})… this may take a few minutes.",
2370	model.display_name, model.description
2371	),
2372	)
2373	.ok();
2374
2375	match agent.switch_model_by_id(&model.config.model_id).await {
2376	Ok(new_config) => {
2377	agent.reset_to_local_backend().await;
2378	let _ = settings::set_local_inference(true);
2379	agent.engine.clear_history().await;
2380	agent
2381	.send_assistant_message(
2382	cx,
2383	session_id,
2384	format!(
2385	"✓ Downloaded and switched to {}",
2386	new_config.display_name
2387	),
2388	)
2389	.ok();
2390	}
2391	Err(err) => {
2392	agent
2393	.send_assistant_message(
2394	cx,
2395	session_id,
2396	format!("error downloading model: {}", err.message),
2397	)
2398	.ok();
2399	}
2400	}
2401	} else {
2402	agent
2403	.send_assistant_message(
2404	cx,
2405	session_id.clone(),
2406	format!("Loading {}...", model.display_name),
2407	)
2408	.ok();
2409
2410	let switched = agent.switch_model_by_id(&model.config.model_id).await?;
2411	agent.reset_to_local_backend().await;
2412	let _ = settings::set_local_inference(true);
2413	agent.engine.clear_history().await;
2414
2415	agent
2416	.send_assistant_message(
2417	cx,
2418	session_id,
2419	format!("Switched to {}.", switched.display_name),
2420	)
2421	.ok();
2422	}
2423	}
2424	}
2425	}
2426	SlashCommand::Local(value) => {
2427	let enabled = value.unwrap_or(!settings::local_inference_enabled());
2428	let message = match settings::set_local_inference(enabled) {
2429	Ok(()) if enabled => "Local inference is on. On-device models are highlighted; \
2430	pick one with /models."
2431	.to_string(),
2432	Ok(()) => "Local inference is off. siGit Code Cloud tiers are highlighted; \
2433	pick one with /models."
2434	.to_string(),
2435	Err(error) => format!("error: could not save local inference setting: {error}"),
2436	};
2437	agent
2438	.send_assistant_message(cx, session_id.clone(), message)
2439	.ok();
2440	// Refresh the panel so the Model picker reflects the new emphasis.
2441	let config_options = {
2442	let current = agent.current_model.lock().unwrap();
2443	build_model_config_options(&current)
2444	};
2445	agent
2446	.send_tool_call_update(
2447	cx,
2448	session_id,
2449	SessionUpdate::ConfigOptionUpdate(ConfigOptionUpdate::new(config_options)),
2450	)
2451	.ok();
2452	}
2453	SlashCommand::Load => {
2454	// Explicitly load the on-device model. This is the only path that
2455	// brings a local model into memory; prompts never do it implicitly.
2456	// If a cloud tier is active, fall back to a local default so we don't
2457	// try to load the (file-less) cloud config as GGUF.
2458	let on_cloud = {
2459	let guard = agent.current_model.lock().unwrap();
2460	guard.model_id.starts_with("sigit-cloud:")
2461	};
2462	if on_cloud {
2463	let default_config = default_local_model_config();
2464	*agent.current_model.lock().unwrap() = default_config;
2465	agent.reset_to_local_backend().await;
2466	}
2467	// Loading an on-device model puts us in local inference mode.
2468	let _ = settings::set_local_inference(true);
2469	// `await_model_ready` drives the download/load progress UI and reports
2470	// success or failure to the editor.
2471	agent.start_startup_model_load_if_needed();
2472	agent.await_model_ready(cx, &session_id).await?;
2473	}
2474	SlashCommand::Login(argument) => {
2475	let message = match argument.as_deref().and_then(account::parse_login_args) {
2476	Some((email, password)) => match account::authenticate(&email, &password).await {
2477	Ok(email) => format!(
2478	"Signed in as {email}. Pick a siGit Code Cloud tier in /models to use it."
2479	),
2480	Err(error) => format!("Login failed: {error}"),
2481	},
2482	None => "usage: /login <email> <password>".to_string(),
2483	};
2484	agent.send_assistant_message(cx, session_id, message).ok();
2485	}
2486	SlashCommand::Logout => {
2487	// If we're on a cloud tier, drop back to local — the token is gone.
2488	let on_cloud = {
2489	let guard = agent.current_model.lock().unwrap();
2490	guard.model_id.starts_with("sigit-cloud:")
2491	};
2492	let message = account::end_session().await;
2493	if on_cloud {
2494	agent.reset_to_local_backend().await;
2495	}
2496	agent.send_assistant_message(cx, session_id, message).ok();
2497	}
2498	SlashCommand::Whoami => {
2499	let message = account::status_line().await;
2500	agent.send_assistant_message(cx, session_id, message).ok();
2501	}
2502	SlashCommand::Reload => {
2503	agent.handle_reload(cx, session_id).await;
2504	}
2505	SlashCommand::Exit => {
2506	agent
2507	.send_assistant_message(
2508	cx,
2509	session_id,
2510	"Use the panel controls to close or switch threads.",
2511	)
2512	.ok();
2513	}
2514	SlashCommand::Unknown(command) => {
2515	agent
2516	.send_assistant_message(cx, session_id, format!("unknown command: {command}"))
2517	.ok();
2518	}
2519	}
2520
2521	Ok(PromptResponse::new(StopReason::EndTurn))
2522	}
2523
2524	// ── Request dispatch helper ───────────────────────────────────────────────────
2525
2526	fn handle_response<T: agent_client_protocol::JsonRpcResponse>(
2527	responder: Responder<T>,
2528	result: agent_client_protocol::Result<T>,
2529	) -> agent_client_protocol::Result<()> {
2530	match result {
2531	Ok(resp) => responder.respond(resp),
2532	Err(err) => responder.respond_with_error(err),
2533	}
2534	}
2535
2536	// ── Download progress helpers ─────────────────────────────────────────────────
2537
2538	/// total bytes on disk under `path`. needed because hf-hub uses staging
2539	/// names during download, so we can't just stat the final blobs.
2540	fn dir_size_recursive(path: &std::path::Path) -> u64 {
2541	let mut total: u64 = 0;
2542	let Ok(entries) = std::fs::read_dir(path) else {
2543	return 0;
2544	};
2545	for entry in entries.flatten() {
2546	let entry_path = entry.path();
2547	if entry_path.is_dir() {
2548	total += dir_size_recursive(&entry_path);
2549	} else if let Ok(meta) = entry_path.metadata() {
2550	total += meta.len();
2551	}
2552	}
2553	total
2554	}
2555
2556	fn format_size_human(bytes: u64) -> String {
2557	const GB: u64 = 1_073_741_824;
2558	const MB: u64 = 1_048_576;
2559	const KB: u64 = 1_024;
2560	if bytes >= GB {
2561	format!("{:.2} GB", bytes as f64 / GB as f64)
2562	} else if bytes >= MB {
2563	format!("{:.1} MB", bytes as f64 / MB as f64)
2564	} else if bytes >= KB {
2565	format!("{:.0} KB", bytes as f64 / KB as f64)
2566	} else {
2567	format!("{bytes} B")
2568	}
2569	}
2570
2571	fn progress_bar(pct: u8, width: usize) -> String {
2572	let filled = ((pct as usize) * width) / 100;
2573	let empty = width.saturating_sub(filled);
2574	format!("[{}{}]", "█".repeat(filled), "░".repeat(empty))
2575	}
2576
2577	// ── Output capture ────────────────────────────────────────────────────────────
2578
2579	/// redirect stdout+stderr to `$TMPDIR/sigit.log` at the fd level so
2580	/// mistralrs/tracing noise never hits the terminal. returns two dup'd
2581	/// fds to the real tty: one for ratatui, one for cleanup (ratatui 0.29
2582	/// doesn't expose `writer_mut()`).
2583	#[cfg(unix)]
2584	fn redirect_output_to_log() -> anyhow::Result<(std::fs::File, std::fs::File)> {
2585	let log_path = std::env::temp_dir().join("sigit.log");
2586	let log_file = std::fs::File::create(&log_path)?;
2587	let log_fd = log_file.as_raw_fd();
2588
2589	// two copies: ratatui needs one, cleanup needs another
2590	let saved_tui = unsafe { libc::dup(libc::STDOUT_FILENO) };
2591	anyhow::ensure!(
2592	saved_tui >= 0,
2593	"dup(stdout) for tui failed: {}",
2594	std::io::Error::last_os_error()
2595	);
2596	let saved_cleanup = unsafe { libc::dup(libc::STDOUT_FILENO) };
2597	anyhow::ensure!(
2598	saved_cleanup >= 0,
2599	"dup(stdout) for cleanup failed: {}",
2600	std::io::Error::last_os_error()
2601	);
2602
2603	unsafe {
2604	libc::dup2(log_fd, libc::STDOUT_FILENO);
2605	libc::dup2(log_fd, libc::STDERR_FILENO);
2606	}
2607
2608	// safe to drop log_file; dup2 keeps the fd alive via stdout/stderr
2609
2610	Ok((unsafe { std::fs::File::from_raw_fd(saved_tui) }, unsafe {
2611	std::fs::File::from_raw_fd(saved_cleanup)
2612	}))
2613	}
2614
2615	// ── Logging ───────────────────────────────────────────────────────────────────
2616
2617	/// in TUI mode stderr is the log file (redirected earlier);
2618	/// in ACP mode it's real stderr. either way, write there.
2619	fn init_logging(is_tty: bool) {
2620	let filter = EnvFilter::try_from_default_env().unwrap_or_else(\|_\| EnvFilter::new("info"));
2621	let _ = tracing_fmt::Subscriber::builder()
2622	.with_env_filter(filter)
2623	.with_writer(std::io::stderr)
2624	.with_ansi(!is_tty)
2625	.try_init();
2626	}
2627
2628	// ── Interactive TUI mode ──────────────────────────────────────────────────────
2629
2630	/// boot the TUI and load the model on a background thread.
2631	/// `tty` goes to ratatui; `cleanup_tty` is a separate fd for
2632	/// LeaveAlternateScreen (ratatui 0.29 hides `writer_mut()`).
2633	#[cfg(unix)]
2634	async fn run_interactive(tty: std::fs::File, mut cleanup_tty: std::fs::File) -> anyhow::Result<()> {
2635	let engine = Arc::new(ChatEngine::new());
2636
2637	let startup_selection = setup::startup_model_selection();
2638	let startup_model_name = startup_selection
2639	.as_ref()
2640	.map(\|selection\| selection.display_name.clone())
2641	.unwrap_or_else(\|\| GgufModelConfig::qwen25_3b().display_name);
2642
2643	// Signals the loading phase to finish. On-device models are no longer loaded
2644	// at startup, so this resolves immediately for both backends; it stays a
2645	// channel so the loading-phase plumbing in `chat::run_with` is unchanged.
2646	let (load_tx, load_rx) = std::sync::mpsc::channel::<Result<(), String>>();
2647
2648	// Project instruction files (AGENTS.md / CLAUDE.md) for the launch directory,
2649	// injected into the system prompt so the TUI shares the same always-on
2650	// project context the ACP sessions get.
2651	let project_instructions = std::env::current_dir()
2652	.ok()
2653	.and_then(\|cwd\| instructions::load_project_instructions(&cwd));
2654	let with_instructions = \|base: String\| match &project_instructions {
2655	Some(extra) => format!("{base}\n\n{extra}"),
2656	None => base,
2657	};
2658
2659	// Pick the inference backend: a configured provider if present, else on-device.
2660	let (inference_backend, startup_model_name): (Arc<dyn InferenceBackend>, String) =
2661	match provider::active_provider() {
2662	Some(provider) => {
2663	log::info!(
2664	"inference: using {} (model {}) at {}",
2665	provider.display_name,
2666	provider.model,
2667	provider.base_url
2668	);
2669	// No local model to load; the endpoint is ready immediately.
2670	let _ = load_tx.send(Ok(()));
2671	let label = provider.display_name.clone();
2672	let backend = Arc::new(OpenAiBackend::new(
2673	provider.base_url,
2674	provider.api_key,
2675	provider.model,
2676	Some(with_instructions(SYSTEM_PROMPT.to_string())),
2677	)) as Arc<dyn InferenceBackend>;
2678	(backend, label)
2679	}
2680	None => {
2681	// Honor the Local Inference toggle: when off and signed in, start
2682	// on a cloud tier. Otherwise bring up on-device WITHOUT loading a
2683	// model — the user loads it explicitly with /load (or /models), so
2684	// the UI comes up immediately. Project instructions are injected at
2685	// load time in `chat.rs`.
2686	let cloud_when_off = if settings::local_inference_enabled() {
2687	None
2688	} else {
2689	provider::cloud_tier_provider("balanced")
2690	};
2691
2692	match cloud_when_off {
2693	Some(provider) => {
2694	log::info!(
2695	"inference: local inference off; using {} (model {})",
2696	provider.display_name,
2697	provider.model
2698	);
2699	let _ = load_tx.send(Ok(()));
2700	let label = provider.display_name.clone();
2701	let backend = Arc::new(OpenAiBackend::new(
2702	provider.base_url,
2703	provider.api_key,
2704	provider.model,
2705	Some(with_instructions(SYSTEM_PROMPT.to_string())),
2706	)) as Arc<dyn InferenceBackend>;
2707	(backend, label)
2708	}
2709	None => {
2710	if !settings::local_inference_enabled() {
2711	log::warn!(
2712	"local inference is off but no account is signed in; \
2713	bringing up on-device. Run /login or /local on."
2714	);
2715	}
2716	let _ = load_tx.send(Ok(()));
2717	let backend = Arc::new(LocalBackend::new(Arc::clone(&engine)))
2718	as Arc<dyn InferenceBackend>;
2719	(backend, startup_model_name)
2720	}
2721	}
2722	}
2723	};
2724
2725	crossterm::terminal::enable_raw_mode()?;
2726	let mut tty = BufWriter::new(tty);
2727	crossterm::execute!(tty, crossterm::terminal::EnterAlternateScreen)?;
2728	let term_backend = ratatui::backend::CrosstermBackend::new(tty);
2729	let mut terminal = ratatui::Terminal::new(term_backend)?;
2730
2731	// polls load_rx with try_recv() each tick, no blocking
2732	let chat_result = chat::run_with(
2733	&mut terminal,
2734	engine,
2735	inference_backend,
2736	load_rx,
2737	startup_model_name,
2738	)
2739	.await;
2740
2741	// cleanup fd because backend's writer is private
2742	crossterm::execute!(cleanup_tty, crossterm::terminal::LeaveAlternateScreen)?;
2743	cleanup_tty.flush()?;
2744	crossterm::terminal::disable_raw_mode()?;
2745
2746	// restore real stdout/stderr for post-TUI error output
2747	#[cfg(unix)]
2748	{
2749	let cleanup_fd = cleanup_tty.as_raw_fd();
2750	unsafe {
2751	libc::dup2(cleanup_fd, libc::STDOUT_FILENO);
2752	libc::dup2(cleanup_fd, libc::STDERR_FILENO);
2753	}
2754	}
2755
2756	chat_result
2757	}
2758
2759	// ── ACP server mode ───────────────────────────────────────────────────────────
2760
2761	/// The on-device model `/load` should bring up by default: the persisted
2762	/// selection if it still resolves to a known local model, otherwise the built-in
2763	/// default (`qwen25_3b`).
2764	fn default_local_model_config() -> GgufModelConfig {
2765	setup::startup_model_selection()
2766	.as_ref()
2767	.and_then(\|selection\| {
2768	selection.selected_model.as_ref().and_then(\|selected\| {
2769	models::local_picker_items()
2770	.into_iter()
2771	.find(\|item\| {
2772	item.config.model_id == selected.model_id
2773	&& item
2774	.config
2775	.files
2776	.iter()
2777	.any(\|file\| file == &selected.gguf_file)
2778	})
2779	.map(\|item\| item.config)
2780	})
2781	})
2782	.unwrap_or_else(GgufModelConfig::qwen25_3b)
2783	}
2784
2785	async fn run_acp_server() -> anyhow::Result<()> {
2786	log::info!("ACP mode — starting agent server");
2787
2788	let config = default_local_model_config();
2789
2790	let needs_download = models::local_picker_items()
2791	.iter()
2792	.find(\|item\| item.config.model_id == config.model_id)
2793	.map(\|item\| item.cache_health != setup::ModelCacheHealth::Complete)
2794	.unwrap_or(true);
2795
2796	log::info!(
2797	"ACP startup model selected: {} ({})",
2798	config.display_name,
2799	if needs_download {
2800	"needs download"
2801	} else {
2802	"cached"
2803	}
2804	);
2805
2806	let engine = Arc::new(ChatEngine::new());
2807
2808	// The on-device model is never loaded implicitly; the user loads it with
2809	// `/load` (or by picking one in `/models`). So initialize/session/new stay
2810	// lightweight and `model_ready` starts true (nothing is loading).
2811	let model_ready = Arc::new(AtomicBool::new(true));
2812	let startup_model_load_started = Arc::new(AtomicBool::new(false));
2813	let model_load_error: Arc<std::sync::Mutex<Option<String>>> =
2814	Arc::new(std::sync::Mutex::new(None));
2815
2816	let state = Arc::new(SiGitAgent::new(
2817	engine,
2818	config,
2819	model_ready,
2820	startup_model_load_started,
2821	model_load_error,
2822	needs_download,
2823	));
2824
2825	let stdin = tokio::io::stdin().compat();
2826	let stdout = tokio::io::stdout().compat_write();
2827	let transport = ByteStreams::new(stdout, stdin);
2828
2829	Agent
2830	.builder()
2831	.on_receive_request(
2832	{
2833	let state = Arc::clone(&state);
2834	async move \|req: InitializeRequest, responder, _cx: ConnectionTo<Client>\| {
2835	handle_response(responder, state.handle_initialize(req).await)
2836	}
2837	},
2838	agent_client_protocol::on_receive_request!(),
2839	)
2840	.on_receive_request(
2841	{
2842	let state = Arc::clone(&state);
2843	async move \|req: AuthenticateRequest, responder, _cx: ConnectionTo<Client>\| {
2844	handle_response(responder, state.handle_authenticate(req).await)
2845	}
2846	},
2847	agent_client_protocol::on_receive_request!(),
2848	)
2849	// Turn-affecting handlers below run in spawned tasks, serialized by
2850	// `turn_lock`, so the dispatch loop stays free to route client
2851	// responses (permission answers) while a turn is in flight. Awaiting a
2852	// client request from inside a handler would deadlock: the dispatch
2853	// loop can't read the response while the handler blocks it.
2854	.on_receive_request(
2855	{
2856	let state = Arc::clone(&state);
2857	async move \|req: LoadSessionRequest, responder, cx: ConnectionTo<Client>\| {
2858	let state = Arc::clone(&state);
2859	let task_cx = cx.clone();
2860	cx.spawn(async move {
2861	let _turn = state.turn_lock.lock().await;
2862	handle_response(responder, state.handle_load_session(&task_cx, req).await)
2863	})
2864	}
2865	},
2866	agent_client_protocol::on_receive_request!(),
2867	)
2868	.on_receive_request(
2869	{
2870	let state = Arc::clone(&state);
2871	async move \|req: ForkSessionRequest, responder, cx: ConnectionTo<Client>\| {
2872	let state = Arc::clone(&state);
2873	let task_cx = cx.clone();
2874	cx.spawn(async move {
2875	let _turn = state.turn_lock.lock().await;
2876	handle_response(responder, state.handle_fork_session(&task_cx, req).await)
2877	})
2878	}
2879	},
2880	agent_client_protocol::on_receive_request!(),
2881	)
2882	.on_receive_request(
2883	{
2884	let state = Arc::clone(&state);
2885	async move \|req: NewSessionRequest, responder, cx: ConnectionTo<Client>\| {
2886	let state = Arc::clone(&state);
2887	let task_cx = cx.clone();
2888	cx.spawn(async move {
2889	let _turn = state.turn_lock.lock().await;
2890	handle_response(responder, state.handle_new_session(&task_cx, req).await)
2891	})
2892	}
2893	},
2894	agent_client_protocol::on_receive_request!(),
2895	)
2896	.on_receive_request(
2897	{
2898	let state = Arc::clone(&state);
2899	async move \|req: PromptRequest, responder, cx: ConnectionTo<Client>\| {
2900	let state = Arc::clone(&state);
2901	let task_cx = cx.clone();
2902	cx.spawn(async move {
2903	let _turn = state.turn_lock.lock().await;
2904	handle_response(responder, state.handle_prompt(&task_cx, req).await)
2905	})
2906	}
2907	},
2908	agent_client_protocol::on_receive_request!(),
2909	)
2910	.on_receive_request(
2911	{
2912	let state = Arc::clone(&state);
2913	async move \|req: SetSessionConfigOptionRequest,
2914	responder,
2915	cx: ConnectionTo<Client>\| {
2916	let state = Arc::clone(&state);
2917	let task_cx = cx.clone();
2918	cx.spawn(async move {
2919	let _turn = state.turn_lock.lock().await;
2920	handle_response(
2921	responder,
2922	state.handle_set_session_config_option(&task_cx, req).await,
2923	)
2924	})
2925	}
2926	},
2927	agent_client_protocol::on_receive_request!(),
2928	)
2929	.on_receive_notification(
2930	{
2931	let state = Arc::clone(&state);
2932	async move \|notif: CancelNotification, _cx: ConnectionTo<Client>\| {
2933	state.handle_cancel(notif).await
2934	}
2935	},
2936	agent_client_protocol::on_receive_notification!(),
2937	)
2938	.connect_to(transport)
2939	.await
2940	.map_err(\|e\| anyhow::anyhow!("ACP connection error: {e}"))?;
2941
2942	log::info!("siGit shutting down");
2943	Ok(())
2944	}
2945
2946	// ── Entry point ──────────────────────────────────────────────────────────────
2947
2948	#[tokio::main]
2949	async fn main() -> anyhow::Result<()> {
2950	// Account subcommands. The editor launches `sigit login` in an embedded
2951	// terminal for ACP terminal-based authentication; the same verbs are handy
2952	// directly from a shell. These must be handled before the TTY/ACP split.
2953	if let Some(verb) = std::env::args().nth(1) {
2954	match verb.as_str() {
2955	"login" => {
2956	init_logging(true);
2957	match account::interactive_login().await {
2958	Ok(email) => {
2959	println!("Signed in to siGit Code Cloud as {email}.");
2960	return Ok(());
2961	}
2962	Err(error) => {
2963	eprintln!("Login failed: {error}");
2964	std::process::exit(1);
2965	}
2966	}
2967	}
2968	"logout" => {
2969	init_logging(true);
2970	println!("{}", account::end_session().await);
2971	return Ok(());
2972	}
2973	"whoami" => {
2974	init_logging(true);
2975	println!("{}", account::status_line().await);
2976	return Ok(());
2977	}
2978	_ => {}
2979	}
2980	}
2981
2982	let is_tty = std::io::stdin().is_terminal();
2983
2984	if is_tty {
2985	// must redirect before any library code touches stdout
2986	#[cfg(unix)]
2987	{
2988	let (tty, cleanup_tty) = redirect_output_to_log()?;
2989	init_logging(true);
2990	setup::setup_shared_model_cache();
2991	// Best-effort: discover MCP servers (incl. the official one) before
2992	// the first turn so their tools are offered to the model.
2993	mcp::init().await;
2994	run_interactive(tty, cleanup_tty).await
2995	}
2996	#[cfg(not(unix))]
2997	{
2998	anyhow::bail!("interactive mode requires Unix (macOS / Linux)");
2999	}
3000	} else {
3001	// ACP mode: keep stdout untouched for protocol JSON only.
3002	// Logs already go to stderr via `init_logging(false)`.
3003	init_logging(false);
3004	setup::setup_shared_model_cache();
3005	// Best-effort MCP discovery (incl. the official server) before serving.
3006	mcp::init().await;
3007	log::info!("siGit v{} starting (ACP mode)", env!("CARGO_PKG_VERSION"));
3008	run_acp_server().await
3009	}
3010	}
3011
3012	#[cfg(test)]
3013	mod tests {
3014	use super::*;
3015
3016	#[test]
3017	fn ascii_safe_replaces_multibyte_chars() {
3018	// The exact label that crashed Zed: the cloud tier name plus the old
3019	// "[☁ siGit Code Cloud]" badge. After sanitizing it must be pure ASCII so
3020	// Zed's fixed byte-offset truncation can never split a glyph.
3021	let crashing = "siGit Code Cloud · Balanced [☁ siGit Code Cloud]";
3022	let safe = ascii_safe(crashing);
3023	assert!(safe.is_ascii(), "sanitized label must be ASCII: {safe:?}");
3024	assert_eq!(safe, "siGit Code Cloud - Balanced [- siGit Code Cloud]");
3025	}
3026
3027	#[test]
3028	fn ascii_safe_leaves_ascii_untouched() {
3029	let plain = "Qwen 2.5 3B [Onde]";
3030	assert_eq!(ascii_safe(plain), plain);
3031	}
3032
3033	#[test]
3034	fn ascii_safe_output_has_only_char_boundaries() {
3035	// Every byte index in an ASCII string is a valid char boundary, so any
3036	// downstream truncation is panic-free regardless of where it cuts.
3037	let safe = ascii_safe("Onde · ◉ ↓ ☁ ○ test");
3038	for i in 0..=safe.len() {
3039	assert!(safe.is_char_boundary(i));
3040	}
3041	}
3042	}