返回 CodeWhale
ops.rs
根目录 / crates / tui / src / core / ops.rs
1 //! Operations submitted by the UI to the core engine.
2 //!
3 //! These operations flow from the TUI to the engine via a channel,
4 //! allowing the UI to remain responsive while the engine processes requests.
5
6 use crate::compaction::CompactionConfig;
7 use crate::config::ApiProvider;
8 use crate::route_runtime::ResolvedRuntimeRoute;
9 use crate::tools::goal::GoalStatus;
10 use codewhale_config::AppMode;
11 use codewhale_execpolicy::ApprovalMode;
12 use codewhale_models::{Message, SystemPrompt};
13 use codewhale_protocol::runtime::DynamicToolSpec;
14 use std::path::PathBuf;
15
16 /// Prefix used for tool-call ids created by local composer shell shortcuts.
17 pub const USER_SHELL_TOOL_ID_PREFIX: &str = "user_shell_";
18
19 /// Snapshot of session state for saving to disk.
20 /// Returned by `Op::GetSessionSnapshot` via a oneshot channel.
21 #[derive(Debug, Clone)]
22 pub struct SessionSnapshot {
23 pub messages: Vec<Message>,
24 pub total_tokens: u64,
25 pub model: String,
26 /// Generic provider kind retained for serialized compatibility.
27 pub model_provider: String,
28 /// Exact non-secret configured provider key.
29 pub model_provider_id: Option<String>,
30 pub workspace: PathBuf,
31 pub system_prompt: Option<SystemPrompt>,
32 pub mode: String,
33 }
34
35 /// Live context-window posture for one thread, computed where the session
36 /// state actually lives. Returned by `Op::GetContextBudget` via a oneshot
37 /// channel so HTTP clients (GPUI usage panel) never re-derive the engine's
38 /// token math or route limits at the API boundary.
39 #[derive(Debug, Clone)]
40 pub struct SessionContextBudget {
41 /// Total context window for the active route (input + output), in tokens.
42 pub window_tokens: u64,
43 /// Estimated input tokens on the same basis the visible context meter
44 /// uses (`estimate_input_tokens_conservative`, including its safety
45 /// inflation). This is the number a "context filling up" indicator shows.
46 pub input_tokens: u64,
47 /// Provider-billed prompt tokens from the most recent parent-route
48 /// request that still describes the live message list. `None` when no
49 /// provider count exists yet (fresh session) — never a fabricated zero.
50 pub billed_input_tokens: Option<u64>,
51 /// Output tokens reserved for the turn after route clamps.
52 pub output_cap_tokens: u64,
53 /// Spendable input ceiling (`window - output_cap - headroom`, intersected
54 /// with any provider-published hard input limit).
55 pub input_budget_ceiling: u64,
56 /// Input tokens still available before the reserved boundary.
57 pub available_input_tokens: u64,
58 /// Input level at which compaction is suggested.
59 pub compaction_trigger_tokens: u64,
60 /// `input_tokens / window_tokens` as a percentage (0..=100).
61 pub usage_percent: f64,
62 /// Coarse pressure label (`low`/`moderate`/`high`/`critical`).
63 pub pressure: &'static str,
64 /// Route identity the budget was computed for.
65 pub model: String,
66 pub provider: String,
67 pub model_provider_id: Option<String>,
68 }
69
70 /// Provider request runtime state surfaced by `/provider`.
71 /// Returned by `Op::GetProviderRuntimeStatus` via a oneshot channel.
72 #[derive(Debug, Clone, PartialEq, Eq)]
73 pub struct ProviderRuntimeStatus {
74 pub provider: ApiProvider,
75 pub request_concurrency_limit: Option<usize>,
76 pub active_provider_requests: usize,
77 }
78
79 /// Idle Engine snapshot used by a one-shot host before shutting down.
80 /// The completion inbox belongs to the Engine; a terminal worker alone does
81 /// not prove that its parent has consumed the handback.
82 #[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
83 pub struct SubAgentSettlement {
84 pub running_children: usize,
85 /// A Workflow can still be coordinating between child phases.
86 pub running_workflows: usize,
87 pub pending_completions: usize,
88 }
89
90 impl SubAgentSettlement {
91 #[must_use]
92 pub fn is_settled(self) -> bool {
93 self.running_children == 0 && self.running_workflows == 0 && self.pending_completions == 0
94 }
95 }
96
97 /// Engine-owned MCP snapshot plus the exact event generation it supersedes.
98 /// The TUI uses the receipt to reject already-queued boot events even when it
99 /// had not rendered that generation before the direct `/mcp` action.
100 #[derive(Debug, Clone)]
101 pub struct McpManagerUpdate {
102 pub snapshot: crate::mcp::McpManagerSnapshot,
103 pub generation: u64,
104 }
105
106 /// Result of rebuilding the engine-owned MCP pool in process.
107 pub type McpReloadResult = Result<McpManagerUpdate, String>;
108
109 /// Result of the one-shot boot connection pass for the engine-owned MCP pool.
110 ///
111 /// This shares the reload result shape while remaining a separate operation:
112 /// boot may fill an empty live pool, but it must not force a config reload or
113 /// invalidate already-ready connections.
114 pub type McpBootstrapResult = Result<McpManagerUpdate, String>;
115
116 /// Origin of text being introduced as a user-role turn.
117 ///
118 /// Chat providers force several runtime/control-plane signals through
119 /// `role = "user"` for compatibility, so role alone is not authority.
120 #[allow(dead_code)] // Some origins are reserved for ingestion sites landing after the first gate.
121 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
122 pub enum UserInputProvenance {
123 /// Text typed or submitted through the active UI/API input boundary.
124 ExternalUser,
125 /// Runtime-generated continuation, diagnostic, or tool feedback.
126 Runtime,
127 /// Completion/event text from a child worker or sub-agent handoff.
128 SubAgentHandoff,
129 /// A bounded, typed Agent Mail envelope delivered by the durable runtime.
130 /// Provider protocols still receive a user-role projection, but this
131 /// provenance can never inherit external-user authority.
132 AgentMail,
133 /// Text restored from a saved/imported transcript.
134 ImportedTranscript,
135 /// Text recalled from memory or another persisted source.
136 MemoryRecall,
137 /// Assistant-authored text that is shaped like a user response.
138 AssistantGenerated,
139 }
140
141 impl UserInputProvenance {
142 pub fn as_str(self) -> &'static str {
143 match self {
144 Self::ExternalUser => "external_user",
145 Self::Runtime => "runtime",
146 Self::SubAgentHandoff => "subagent_handoff",
147 Self::AgentMail => "agent_mail",
148 Self::ImportedTranscript => "imported_transcript",
149 Self::MemoryRecall => "memory_recall",
150 Self::AssistantGenerated => "assistant_generated",
151 }
152 }
153
154 pub fn can_authorize_work(self) -> bool {
155 matches!(self, Self::ExternalUser)
156 }
157 }
158
159 /// Per-turn authority payload carried by [`Op::SendMessage`]. Extracted from
160 /// the enum arm so new per-turn fields accrete here instead of widening the
161 /// variant; the serializable twin is `codewhale_protocol::op::TurnSpec`.
162 #[derive(Debug)]
163 pub struct TurnSpec {
164 /// Admitted allowance for this turn only; never changes session settings.
165 pub max_output_tokens: Option<std::num::NonZeroU32>,
166 pub content: String,
167 /// Inline image bytes validated by Runtime admission; no file references.
168 pub images: Vec<codewhale_protocol::runtime::RuntimeImageInput>,
169 pub mode: AppMode,
170 /// Exact, structurally resolved route authority for this turn. The
171 /// engine activates its client before mutating turn state; injected
172 /// engines may use their already-supplied client with the same receipt.
173 pub route: Box<ResolvedRuntimeRoute>,
174 /// Compaction policy derived from the same provider route. Carrying it
175 /// atomically avoids a model/limit mismatch before `SendMessage`.
176 pub compaction: Box<CompactionConfig>,
177 /// Auxiliary provider calls completed while planning this exact turn
178 /// (currently Auto's classifier), bounded and paired with their own
179 /// immutable routes. The engine folds their tokens into total usage
180 /// only; they never enter the parent route's billing aggregate.
181 pub initial_routed_usage: Box<crate::cost_status::RuntimeUsageBatch>,
182 pub goal_objective: Option<String>,
183 pub goal_token_budget: Option<u32>,
184 pub goal_status: GoalStatus,
185 /// Reasoning-effort tier: `"off" | "low" | "medium" | "high" | "max"`.
186 /// `None` lets the provider apply its default.
187 pub reasoning_effort: Option<String>,
188 /// True when the user selected auto thinking, even though the UI sends
189 /// a concrete per-turn value to the model API.
190 pub reasoning_effort_auto: bool,
191 /// True when the user selected auto model routing.
192 pub auto_model: bool,
193 pub allow_shell: bool,
194 pub trust_mode: bool,
195 pub auto_approve: bool,
196 pub approval_mode: ApprovalMode,
197 pub translation_enabled: bool,
198 /// Tool restriction from custom slash command frontmatter.
199 /// `None` means the current turn may use the normal tool set.
200 pub allowed_tools: Option<Vec<String>>,
201 /// Runtime-supplied tools available only for this turn.
202 pub dynamic_tools: Vec<DynamicToolSpec>,
203 /// Hook executor for control-plane hooks.
204 /// `ToolCallBefore` hooks may deny a tool call with exit code 2.
205 pub hook_executor: Option<std::sync::Arc<crate::hooks::HookExecutor>>,
206 pub verbosity: Option<String>,
207 /// Structural input origin. This gates whether the turn may inherit
208 /// YOLO/auto-approval authority; user-shaped text is not enough.
209 pub provenance: UserInputProvenance,
210 }
211
212 /// Operations that can be submitted to the engine.
213 #[derive(Debug)]
214 pub enum Op {
215 /// Send a message to the AI
216 SendMessage(TurnSpec),
217
218 /// Re-check and dispatch an interactive goal continuation when this
219 /// operation reaches the front of the engine queue. Keeping this distinct
220 /// from `SendMessage` prevents a queued `/goal pause` or `/goal clear`
221 /// from being overwritten by a stale synthetic Active snapshot.
222 ContinueGoal {
223 /// Runtime-supplied tools remain available across the synthetic turn
224 /// that continues the same logical goal run.
225 dynamic_tools: Vec<DynamicToolSpec>,
226 /// Opaque identity for an engine-owned synthetic continuation. Direct
227 /// callers use `None`; the engine uses `Some` to coalesce one token
228 /// across capacity-waiting, enqueued, and running-adjacent states.
229 engine_schedule_id: Option<u64>,
230 },
231
232 /// Execute a user-submitted composer shell command (`! <command>`) without
233 /// sending a model turn. This still routes through `exec_shell`, approval,
234 /// sandbox, and command-safety handling.
235 RunShellCommand {
236 command: String,
237 mode: AppMode,
238 allow_shell: bool,
239 trust_mode: bool,
240 auto_approve: bool,
241 approval_mode: ApprovalMode,
242 },
243
244 /// Set the runtime goal status without dispatching a model turn. Used by
245 /// `/goal pause`, `/goal resume`, `/goal clear`, etc. so the engine's
246 /// `SharedGoalState` learns the new status immediately and a queued
247 /// continuation doesn't overwrite it back to Active.
248 SetGoalStatus {
249 status: GoalStatus,
250 /// When `true`, clear the objective entirely (`/goal clear`).
251 clear: bool,
252 /// Accepted control revision; None lets direct callers mint it.
253 goal_id: Option<String>,
254 },
255
256 /// Set (or replace) the active goal objective and immediately start goal
257 /// work through the runtime's continuation steering. `/goal <objective>`
258 /// is the caller; the objective is never echoed as a raw user message.
259 SetGoalObjective {
260 objective: String,
261 token_budget: Option<u32>,
262 /// Accepted control revision; None lets direct callers mint it.
263 goal_id: Option<String>,
264 },
265
266 /// Describe the exact request the next turn would send, without
267 /// sending it (`/preview-request`, #1004).
268 ///
269 /// Handled by the engine because only the engine can rebuild the current
270 /// tool catalog, MCP state, mode, gates, permission posture, and resolved
271 /// route. Pure inspection: it adds no message, no turn, and no tool call.
272 PreviewOutboundRequest {
273 inputs: Box<crate::core::engine::preview::PreviewRequestInputs>,
274 /// Render the manifest as JSON instead of the human-readable table.
275 json: bool,
276 /// Explicit disclosure of the base prompt only; effective system text
277 /// remains protected behind hashes.
278 base_prompt_only: bool,
279 },
280
281 /// List current sub-agents and their status
282 ListSubAgents,
283
284 /// Inspect child settlement at the Engine's idle operation boundary.
285 /// This does not drain the completion inbox or dispatch a second loop.
286 GetSubAgentSettlement {
287 tx: std::sync::Arc<
288 std::sync::Mutex<Option<tokio::sync::oneshot::Sender<SubAgentSettlement>>>,
289 >,
290 },
291
292 /// Cancel a running sub-agent by id or session name.
293 CancelSubAgent { agent_id: String },
294
295 /// Deliver an operator follow-up to one child on its own fork: live
296 /// delivery to a running child, or a checkpoint continuation (new agent
297 /// id) for an interrupted or completed child. Terminal failed/cancelled
298 /// children answer with a receipt explaining why they cannot continue.
299 FollowUpSubAgent { agent_id: String, text: String },
300
301 /// Change the operating mode
302 ChangeMode {
303 mode: AppMode,
304 allow_shell: bool,
305 trust_mode: bool,
306 auto_approve: bool,
307 approval_mode: ApprovalMode,
308 configured_sandbox_mode: Option<String>,
309 },
310
311 /// Update the model being used and refresh stable prompt context.
312 SetModel {
313 model: String,
314 mode: AppMode,
315 route_limits: Option<codewhale_config::route::RouteLimits>,
316 },
317
318 /// Update auto-compaction settings
319 SetCompaction { config: CompactionConfig },
320
321 /// Update the SSE idle timeout used for subsequent streamed turns.
322 SetStreamChunkTimeout { timeout_secs: u64 },
323
324 /// Update sub-agent runtime controls for subsequent turns.
325 SetSubagentRuntimeConfig {
326 enabled: bool,
327 max_subagents: usize,
328 launch_concurrency: usize,
329 max_spawn_depth: u32,
330 api_timeout_secs: u64,
331 heartbeat_timeout_secs: u64,
332 },
333
334 /// Update the web-search backend for subsequent tool calls.
335 SetSearchProvider {
336 provider: crate::config::SearchProvider,
337 },
338
339 /// Replace the engine's merged Fleet roster after the setup wizard saves a
340 /// project or personal profile. Subsequent turns can use the new role
341 /// immediately instead of requiring an application restart.
342 SetFleetRoster {
343 roster: std::sync::Arc<crate::fleet::roster::FleetRoster>,
344 },
345
346 /// Sync engine session state (used for resume/load)
347 SyncSession {
348 session_id: Option<String>,
349 messages: Vec<Message>,
350 system_prompt: Option<SystemPrompt>,
351 system_prompt_override: bool,
352 model: String,
353 workspace: PathBuf,
354 mode: AppMode,
355 },
356
357 /// Run context compaction on one exact, structurally resolved provider
358 /// route with policy derived from that same descriptor.
359 CompactContext {
360 /// Stable request identity allocated before the operation enters the
361 /// bounded mailbox. Cancellation uses this id even when the provider
362 /// future has not started yet.
363 id: String,
364 route: Box<ResolvedRuntimeRoute>,
365 compaction: Box<CompactionConfig>,
366 },
367
368 /// Cancel one exact queued or running context-compaction request.
369 CancelCompaction { id: String },
370
371 /// Get a snapshot of the current session state (messages, tokens, etc.)
372 /// for saving to disk. Returns the result via the oneshot sender so
373 /// the caller doesn't have to compete with the SSE event stream.
374 GetSessionSnapshot {
375 tx: std::sync::Arc<std::sync::Mutex<Option<tokio::sync::oneshot::Sender<SessionSnapshot>>>>,
376 },
377
378 /// Get the live context-window budget for this session's route. Computed
379 /// on the engine so `active_route_limits`, the memoized token estimate,
380 /// and the last billed prompt size all come from one authority.
381 GetContextBudget {
382 tx: std::sync::Arc<
383 std::sync::Mutex<Option<tokio::sync::oneshot::Sender<Option<SessionContextBudget>>>>,
384 >,
385 },
386
387 /// Get active provider request concurrency state for readiness surfaces.
388 GetProviderRuntimeStatus {
389 tx: std::sync::Arc<
390 std::sync::Mutex<Option<tokio::sync::oneshot::Sender<ProviderRuntimeStatus>>>,
391 >,
392 },
393
394 /// Populate the engine-owned MCP pool once at UI boot and return a
395 /// snapshot from that exact pool. This is not a config reload and never
396 /// constructs a UI-owned discovery pool. Optional servers never block
397 /// the first model turn: that turn snapshots currently-ready tools.
398 BootstrapMcp {
399 tx: std::sync::Arc<
400 std::sync::Mutex<Option<tokio::sync::oneshot::Sender<McpBootstrapResult>>>,
401 >,
402 },
403
404 /// Retry one failed MCP server on the existing engine pool and return a
405 /// full snapshot. Ready siblings are never invalidated or reconnected.
406 RetryMcpServer {
407 name: String,
408 tx: std::sync::Arc<
409 std::sync::Mutex<Option<tokio::sync::oneshot::Sender<McpBootstrapResult>>>,
410 >,
411 },
412
413 /// Force the engine-owned MCP config/catalog to reload and reconnect.
414 /// The returned snapshot is taken from that same live pool.
415 ReloadMcp {
416 config_path: PathBuf,
417 tx: std::sync::Arc<std::sync::Mutex<Option<tokio::sync::oneshot::Sender<McpReloadResult>>>>,
418 },
419
420 /// Run agent-driven context purging.
421 PurgeContext,
422
423 /// Edit the last user message: remove the last user+assistant exchange
424 /// from the session, then re-send with the new content.
425 #[cfg_attr(not(test), expect(dead_code))]
426 EditLastTurn { new_message: String },
427
428 /// Enable or disable the background advisor watcher for this session.
429 /// When enabled, a fire-and-forget background task runs after each turn
430 /// that contained tool calls and emits an `Event::AdvisoryNote` with
431 /// concise observations. (#3982)
432 SetAdvisorEnabled { enabled: bool },
433
434 /// Shutdown the engine
435 Shutdown,
436 }
437
437 lines RUST