Skip to main content

oxi_agent/
tools.rs

1#![allow(unused_doc_comments)]
2/// Agent tools system
3/// This module provides the tool abstraction layer and built-in tools.
4use crate::types::ToolDefinition;
5use async_trait::async_trait;
6use serde_json::Value;
7use std::fmt;
8use std::future::Future;
9use std::path::{Path, PathBuf};
10use std::pin::Pin;
11use std::sync::Arc;
12use tokio::sync::oneshot;
13
14// ═══════════════════════════════════════════════════════════════════════════
15// Capability traits — lightweight interfaces tools need, implemented by the
16// composition root (oxi-cli) bridging to SDK ports. oxi-agent does NOT depend
17// on oxi-sdk, so these are defined here.
18// ═══════════════════════════════════════════════════════════════════════════
19
20/// A single memory item returned by [`MemoryBackend`].
21#[derive(Debug, Clone, serde::Serialize)]
22pub struct MemoryItem {
23    /// Unique identifier.
24    pub id: String,
25    /// Memory kind: "fact", "preference", "context", "summary".
26    pub kind: String,
27    /// The memory content text.
28    pub content: String,
29    /// Project/scope identifier.
30    pub subject: String,
31}
32
33/// Memory backend for the `memory_*` tools. The composition root implements
34/// this, bridging to `oxi_sdk::ports::MemoryStore` + `EmbeddingProvider`.
35pub trait MemoryBackend: Send + Sync + std::fmt::Debug + 'static {
36    /// Store a memory item, returning its new ID.
37    fn put<'a>(
38        &'a self,
39        content: &'a str,
40        kind: &'a str,
41        subject: &'a str,
42    ) -> Pin<Box<dyn Future<Output = Result<String, ToolError>> + Send + 'a>>;
43    /// Semantic-search stored memories, returning up to `k` matches.
44    fn search<'a>(
45        &'a self,
46        query: &'a str,
47        k: usize,
48    ) -> Pin<Box<dyn Future<Output = Result<Vec<MemoryItem>, ToolError>> + Send + 'a>>;
49    /// List memory items for the given subject.
50    fn list<'a>(
51        &'a self,
52        subject: &'a str,
53    ) -> Pin<Box<dyn Future<Output = Result<Vec<MemoryItem>, ToolError>> + Send + 'a>>;
54    /// Delete the memory item with the given ID.
55    fn delete<'a>(
56        &'a self,
57        id: &'a str,
58    ) -> Pin<Box<dyn Future<Output = Result<(), ToolError>> + Send + 'a>>;
59
60    /// Human-readable memory status (None if not supported by this backend).
61    fn memory_info(&self) -> Option<String> {
62        None
63    }
64    /// Trigger sleep consolidation, returning a status message.
65    fn trigger_consolidation(&self) -> Option<String> {
66        None
67    }
68    /// Trigger SHMR harmonization, returning a status message.
69    fn trigger_harmonize(&self) -> Option<String> {
70        None
71    }
72
73    /// Delete every memory item in the backend. Returns the number of
74    /// items removed.
75    ///
76    /// The default implementation is **unsupported**: it returns an
77    /// honest error rather than silently deleting zero rows. Backends
78    /// that can perform a true bulk erase must override this method;
79    /// backends without a list-all subject primitive must surface
80    /// that limitation rather than guess.
81    ///
82    /// This is a destructive operation; the `/memory clear` slash
83    /// command requires an explicit confirmation flag before invoking.
84    fn clear_all<'a>(
85        &'a self,
86    ) -> Pin<Box<dyn Future<Output = Result<usize, ToolError>> + Send + 'a>> {
87        Box::pin(async move {
88            Err(
89                "clear_all not supported by this backend (no list-all subject primitive)"
90                    .to_string(),
91            )
92        })
93    }
94
95    /// Force a consolidation (rebuild) job. Returns a status message
96    /// describing what was dispatched, or `Err` when the backend has
97    /// no in-process capability to enqueue work.
98    ///
99    /// Default: `Err` ("not supported"). Backends that expose
100    /// consolidation via the engine (e.g. Mnemopi sleep) should
101    /// override and run the real operation.
102    fn enqueue_consolidation<'a>(
103        &'a self,
104    ) -> Pin<Box<dyn Future<Output = Result<String, ToolError>> + Send + 'a>> {
105        Box::pin(
106            async move { Err("enqueue_consolidation not supported by this backend".to_string()) },
107        )
108    }
109}
110
111/// Content resolved from an internal protocol URL (e.g. `skill://`, `issue://`).
112pub struct ResolvedContent {
113    /// The resolved text content.
114    pub content: String,
115    /// MIME type: "text/markdown", "application/json", "text/plain".
116    pub content_type: String,
117    /// True if the content is uneditable (suppresses hashline anchors).
118    pub immutable: bool,
119}
120
121/// URL resolver for internal protocol schemes. The composition root
122/// implements this, bridging to `oxi_sdk::ports::InternalUrlRouter`.
123pub trait UrlResolver: Send + Sync + std::fmt::Debug {
124    /// Whether this resolver handles the given input URI.
125    fn can_resolve(&self, input: &str) -> bool;
126    /// Resolve an internal URI to its content, asynchronously.
127    fn resolve<'a>(
128        &'a self,
129        uri: &'a str,
130    ) -> Pin<Box<dyn Future<Output = Result<ResolvedContent, ToolError>> + Send + 'a>>;
131}
132
133/// Todo state access capability. Implemented by the composition root
134/// (oxi-cli) bridging to the session-scoped todo state. Used by the
135/// `todo` agent tool and the TUI sticky panel.
136pub trait TodoStateProvider: Send + Sync + std::fmt::Debug {
137    /// Return a snapshot of the current phase list (read-only, for TUI).
138    fn get_phases(&self) -> Vec<crate::tools::todo::TodoPhase>;
139
140    /// Apply a sequence of todo ops, returning the updated state, the
141    /// newly-completed transitions (for strikethrough animation), and
142    /// any error messages from ambiguous op references.
143    fn apply_ops<'a>(
144        &'a self,
145        ops: Vec<crate::tools::todo::TodoOp>,
146    ) -> Pin<
147        Box<dyn Future<Output = Result<crate::tools::todo::TodoUpdateResult, String>> + Send + 'a>,
148    >;
149}
150
151// ── Agent Hub capability (⑥) ──────────────────────────────────────────
152
153/// Agent kind for Hub display.
154#[derive(Debug, Clone, Copy, PartialEq, Eq)]
155pub enum AgentKind {
156    /// Main conversation agent.
157    Main,
158    /// Task-spawned sub-agent.
159    Task,
160    /// Observation-only advisor.
161    Advisor,
162}
163
164/// Hub display status.
165#[derive(Debug, Clone, Copy, PartialEq, Eq)]
166pub enum AgentHubStatus {
167    /// Currently executing.
168    Running,
169    /// Finished, idle.
170    Idle,
171    /// Parked (memory retained, not running).
172    Parked,
173    /// Abnormal termination.
174    Aborted,
175}
176
177/// Read-only agent info for Hub display.
178#[derive(Debug, Clone)]
179pub struct AgentInfo {
180    /// Unique identifier.
181    pub id: String,
182    /// Display name.
183    pub display_name: String,
184    /// Agent kind.
185    pub kind: AgentKind,
186    /// Current status.
187    pub status: AgentHubStatus,
188    /// Current task description (if any).
189    pub current_task: Option<String>,
190}
191
192/// Agent pool access capability. Implemented by the composition root
193/// to expose live sub-agent info to the Hub overlay and todo matching.
194pub trait AgentPoolProvider: Send + Sync + std::fmt::Debug {
195    /// List all known agents (main + sub-agents).
196    fn list_agents(&self) -> Vec<AgentInfo>;
197    /// Get a specific agent by ID.
198    fn get_agent(&self, id: &str) -> Option<AgentInfo>;
199}
200
201// ── LSP capability (⑧) ────────────────────────────────────────────────
202
203/// Aggregated diagnostics across one or more files (returned by
204/// [`LspProvider::drain_diagnostics`]). Counts severity buckets so callers
205/// can surface a quick "0 errors / 3 warnings" summary without scanning
206/// every diagnostic.
207#[derive(Debug, Clone, Default)]
208pub struct DiagnosticsSummary {
209    /// Total number of fresh diagnostics after filtering.
210    pub count: usize,
211    /// Number of error-severity diagnostics.
212    pub errors: usize,
213    /// Number of warning-severity diagnostics.
214    pub warnings: usize,
215    /// Per-file entries; empty when no diagnostics have arrived yet.
216    pub entries: Vec<FileDiagnosticEntry>,
217}
218
219/// Diagnostics for one file. Path is the LSP document URI as the server
220/// reported it (may be `file://`-prefixed); `diagnostics` is the raw payload
221/// from `textDocument/publishDiagnostics`.
222#[derive(Debug, Clone)]
223pub struct FileDiagnosticEntry {
224    /// Document URI (typically `file://<absolute path>`).
225    pub uri: String,
226    /// Path-relative display of the file (best effort).
227    pub path: String,
228    /// Diagnostics reported by the server for this file.
229    pub diagnostics: serde_json::Value,
230}
231
232/// LSP action enum — the operations the `lsp` tool supports.
233#[derive(Debug, Clone)]
234pub enum LspAction {
235    /// Get diagnostics for a file.
236    Diagnostics {
237        /// Path to the file to inspect.
238        file: String,
239    },
240    /// Go to definition.
241    Definition {
242        /// Path to the file containing the symbol.
243        file: String,
244        /// 1-based line number of the symbol.
245        line: u32,
246        /// Optional symbol text to resolve (for disambiguation).
247        symbol: Option<String>,
248    },
249    /// Find references.
250    References {
251        /// Path to the file containing the symbol.
252        file: String,
253        /// 1-based line number of the symbol.
254        line: u32,
255        /// Optional symbol text to find references for.
256        symbol: Option<String>,
257    },
258    /// Hover info.
259    Hover {
260        /// Path to the file containing the symbol.
261        file: String,
262        /// 1-based line number of the symbol.
263        line: u32,
264        /// Optional symbol text to hover.
265        symbol: Option<String>,
266    },
267    /// Rename symbol.
268    Rename {
269        /// Path to the file containing the symbol.
270        file: String,
271        /// 1-based line number of the symbol.
272        line: u32,
273        /// Symbol text to rename.
274        symbol: String,
275        /// New name for the symbol.
276        new_name: String,
277        /// If true, apply the rename; otherwise just preview.
278        apply: bool,
279    },
280    /// Get workspace/document symbols.
281    Symbols {
282        /// Path to the file to inspect (workspace symbols if query-only).
283        file: String,
284        /// Optional filter query for symbols.
285        query: Option<String>,
286    },
287    /// Get server status.
288    Status,
289    /// Available code actions at a position.
290    CodeActions {
291        /// Path to the file containing the position.
292        file: String,
293        /// 1-based line number.
294        line: u32,
295        /// Optional symbol hint for disambiguation.
296        symbol: Option<String>,
297    },
298    /// Go to type definition.
299    TypeDefinition {
300        /// Path to the file containing the symbol.
301        file: String,
302        /// 1-based line number.
303        line: u32,
304        /// Optional symbol hint.
305        symbol: Option<String>,
306    },
307    /// Go to implementation.
308    Implementation {
309        /// Path to the file containing the symbol.
310        file: String,
311        /// 1-based line number.
312        line: u32,
313        /// Optional symbol hint.
314        symbol: Option<String>,
315    },
316    /// Rename a file (workspace/willRenameFiles + applyWorkspaceEdit).
317    FileRename {
318        /// Current path on disk.
319        old_path: String,
320        /// Target path on disk.
321        new_path: String,
322        /// If true, apply the rename; otherwise just preview.
323        apply: bool,
324    },
325}
326
327/// LSP access capability. Implemented by an `oxi-lsp` crate (feature-gated)
328/// or stubbed with `None` when LSP is disabled.
329pub trait LspProvider: Send + Sync + std::fmt::Debug {
330    /// Kick off background initialisation (servers start but `ensure_ready`
331    /// isn't awaited). Idempotent.
332    fn ensure_started_background<'a>(&'a self) -> Pin<Box<dyn Future<Output = ()> + Send + 'a>>;
333
334    /// Block until at least the configured LSP servers have finished their
335    /// `initialize` handshake (or the operation times out per the
336    /// provider's internal budget).
337    fn ensure_ready<'a>(&'a self) -> Pin<Box<dyn Future<Output = Result<(), String>> + Send + 'a>>;
338
339    /// Drain the most recent batch of diagnostics that arrived via
340    /// `textDocument/publishDiagnostics`. Returns `None` when nothing
341    /// fresh has arrived within `timeout`.
342    fn drain_diagnostics<'a>(
343        &'a self,
344        timeout: std::time::Duration,
345    ) -> Pin<Box<dyn Future<Output = Option<DiagnosticsSummary>> + Send + 'a>>;
346
347    /// Read the most recent cached diagnostics for the given file paths
348    /// (zero-copy snapshot — no waiting). Paths that have no fresh
349    /// diagnostics are omitted from the returned vec.
350    fn read_diagnostics<'a>(
351        &'a self,
352        paths: &'a [std::path::PathBuf],
353    ) -> Pin<Box<dyn Future<Output = Vec<FileDiagnosticEntry>> + Send + 'a>>;
354
355    /// Notify the LSP manager that the contents of `path` changed. The
356    /// manager is responsible for forwarding a `workspace/didChange` to
357    /// every server that owns the file. Default implementation is a no-op
358    /// so lightweight providers (e.g. test stubs) don't have to wire it.
359    fn notify_file_changed<'a>(
360        &'a self,
361        _path: &'a std::path::Path,
362        _content: &'a str,
363    ) -> Pin<Box<dyn Future<Output = ()> + Send + 'a>> {
364        Box::pin(async {})
365    }
366
367    /// Execute an LSP action and return formatted text output.
368    fn execute_action<'a>(
369        &'a self,
370        action: &'a LspAction,
371    ) -> Pin<Box<dyn Future<Output = Result<String, ToolError>> + Send + 'a>>;
372}
373
374// ── Sub-agent delegation (issue #28 gap 3) ─────────────────────────────
375
376/// Result of an in-process isolated sub-agent fork run.
377///
378/// Produced by [`SubagentRunner::run_isolated`]. The sub-agent runs
379/// with a **fresh, empty context** — its conversation history is
380/// completely isolated from the parent agent. Only the final text and
381/// usage statistics are returned, keeping the parent's context small.
382///
383/// This is the library-native alternative to shelling out to the `oxi`
384/// CLI binary. Library consumers (e.g. Oxios) that embed `oxi-agent`
385/// without an `oxi` subprocess implement this trait so the `subagent`
386/// tool works in-process.
387#[derive(Debug, Clone, Default)]
388pub struct ForkResult {
389    /// Final response text from the sub-agent.
390    pub text: String,
391    /// Input tokens consumed (last reported turn).
392    pub input_tokens: usize,
393    /// Output tokens consumed (last reported turn).
394    pub output_tokens: usize,
395    /// Number of agent turns executed.
396    pub turns: u32,
397    /// Model ID used by the sub-agent.
398    pub model: Option<String>,
399    /// Error message if the run failed.
400    pub error: Option<String>,
401}
402
403/// In-process sub-agent runner — the library-native delegation backend.
404///
405/// When wired into [`ToolContext`] via
406/// [`ToolContext::with_subagent_runner`], the `subagent` tool prefers
407/// this in-process path over shelling out to the `oxi` CLI binary.
408/// This is essential for library consumers (Oxios) that embed
409/// `oxi-agent` as a kernel without an `oxi` subprocess.
410///
411/// The SDK provides a ready-made implementation
412/// (`oxi_sdk::SdkSubagentRunner`) that wraps an `Oxi` instance and
413/// creates a fresh `Agent` for each invocation.
414#[async_trait::async_trait]
415#[allow(clippy::too_many_arguments)]
416pub trait SubagentRunner: Send + Sync + std::fmt::Debug {
417    /// Run a single agent task with an isolated (empty) context.
418    ///
419    /// # Arguments
420    /// * `agent_name` — Agent definition name (for logging / display).
421    /// * `task` — The task prompt to execute.
422    /// * `system_prompt` — Optional system prompt override.
423    /// * `model` — Optional model ID override (e.g. `"anthropic/claude-...`).
424    /// * `tools` — Optional tool whitelist (empty = all registered tools).
425    /// * `cwd` — Working directory for file tools.
426    /// * `depth` — Current sub-agent nesting depth. The runner sets
427    ///   the forked agent's `subagent_depth` to `depth + 1` so the
428    ///   fork's own subagent tool can enforce a recursion cap without
429    ///   env vars (issue #28 gap 3 — concurrent `set_var` is UB).
430    async fn run_isolated(
431        &self,
432        agent_name: &str,
433        task: &str,
434        system_prompt: Option<&str>,
435        model: Option<&str>,
436        tools: &[String],
437        cwd: &Path,
438        depth: u8,
439    ) -> anyhow::Result<ForkResult>;
440}
441
442/// Context passed to tools at execution time.
443///
444/// This allows tools to operate on a specific workspace without being
445/// rebuilt. When `root_dir` is `Some`, tools use it as their base directory.
446/// When `None`, tools should fall back to `workspace_dir`.
447#[derive(Clone)]
448pub struct ToolContext {
449    /// Primary workspace directory (used when root_dir is None).
450    pub workspace_dir: PathBuf,
451    /// Optional explicit root directory for file tools.
452    /// Takes priority over workspace_dir if present.
453    pub root_dir: Option<PathBuf>,
454    /// Session identifier for logging/tracing.
455    pub session_id: Option<String>,
456    /// Snapshot store for hashline tag emission/validation.
457    /// When `None`, hashline edit mode is unavailable.
458    pub snapshot_store: Option<Arc<dyn oxi_hashline::SnapshotStore>>,
459    /// Memory backend for `memory_*` tools.
460    /// When `None`, memory tools return an error.
461    pub memory: Option<Arc<dyn MemoryBackend>>,
462    /// URL resolver for internal protocol schemes (`issue://`, `pr://`, etc.).
463    /// When `None`, URL-prefixed paths are treated as regular file paths.
464    pub url_resolver: Option<Arc<dyn UrlResolver>>,
465    /// Todo state for the `todo` agent tool.
466    /// When `None`, the `todo` tool returns an error.
467    pub todo: Option<Arc<dyn TodoStateProvider>>,
468    /// Agent pool for Hub display and todo sub-agent matching.
469    pub agent_pool: Option<Arc<dyn AgentPoolProvider>>,
470    /// LSP provider for the `lsp` tool.
471    pub lsp: Option<Arc<dyn LspProvider>>,
472    /// In-process sub-agent runner (issue #28 gap 3).
473    /// When `Some`, the `subagent` tool prefers an in-process isolated
474    /// run over shelling out to the CLI binary. Library consumers
475    /// (e.g. Oxios) that embed `oxi-agent` without an `oxi` subprocess
476    /// set this so delegation works. When `None`, the CLI backend is
477    /// used (the default for `oxi-cli`).
478    pub subagent_runner: Option<Arc<dyn SubagentRunner>>,
479    /// Current sub-agent nesting depth for the in-process path
480    /// (issue #28 gap 3). The CLI path uses env vars instead.
481    /// Default 0 (top-level agent).
482    pub subagent_depth: u8,
483    /// Intent trace for the current tool call.
484    /// Set by the agent loop before executing each tool, read by tools
485    /// that surface intent to users (e.g. `ask`).
486    pub intent: Option<String>,
487}
488
489impl fmt::Debug for ToolContext {
490    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
491        f.debug_struct("ToolContext")
492            .field("workspace_dir", &self.workspace_dir)
493            .field("root_dir", &self.root_dir)
494            .field("session_id", &self.session_id)
495            .field(
496                "snapshot_store",
497                &self.snapshot_store.as_ref().map(|_| "<dyn SnapshotStore>"),
498            )
499            .field(
500                "memory",
501                &self.memory.as_ref().map(|_| "<dyn MemoryBackend>"),
502            )
503            .field(
504                "url_resolver",
505                &self.url_resolver.as_ref().map(|_| "<dyn UrlResolver>"),
506            )
507            .finish()
508    }
509}
510
511impl ToolContext {
512    /// Create a new context with the given workspace.
513    pub fn new(workspace_dir: impl Into<PathBuf>) -> Self {
514        Self {
515            workspace_dir: workspace_dir.into(),
516            root_dir: None,
517            session_id: None,
518            snapshot_store: None,
519            memory: None,
520            url_resolver: None,
521            todo: None,
522            agent_pool: None,
523            lsp: None,
524            subagent_runner: None,
525            subagent_depth: 0,
526            intent: None,
527        }
528    }
529
530    /// Get the effective root directory.
531    /// Returns root_dir if set, otherwise workspace_dir.
532    pub fn root(&self) -> &Path {
533        self.root_dir.as_deref().unwrap_or(&self.workspace_dir)
534    }
535
536    /// Set a session ID.
537    pub fn with_session(mut self, session_id: impl Into<String>) -> Self {
538        self.session_id = Some(session_id.into());
539        self
540    }
541
542    /// Set an explicit root directory.
543    pub fn with_root(mut self, root_dir: impl Into<PathBuf>) -> Self {
544        self.root_dir = Some(root_dir.into());
545        self
546    }
547
548    /// Set the snapshot store (enables hashline edit mode).
549    pub fn with_snapshot_store(mut self, store: Arc<dyn oxi_hashline::SnapshotStore>) -> Self {
550        self.snapshot_store = Some(store);
551        self
552    }
553
554    /// Set the memory backend (enables memory tools).
555    pub fn with_memory(mut self, memory: Arc<dyn MemoryBackend>) -> Self {
556        self.memory = Some(memory);
557        self
558    }
559
560    /// Set the URL resolver (enables internal URL dispatch).
561    pub fn with_url_resolver(mut self, resolver: Arc<dyn UrlResolver>) -> Self {
562        self.url_resolver = Some(resolver);
563        self
564    }
565
566    /// Set the todo state (enables the `todo` agent tool).
567    pub fn with_todo(mut self, todo: Arc<dyn TodoStateProvider>) -> Self {
568        self.todo = Some(todo);
569        self
570    }
571
572    /// Set the in-process sub-agent runner (enables library-native
573    /// delegation — issue #28 gap 3).
574    pub fn with_subagent_runner(mut self, runner: Arc<dyn SubagentRunner>) -> Self {
575        self.subagent_runner = Some(runner);
576        self
577    }
578
579    /// Attach an intent trace to this context.
580    pub fn with_intent(mut self, intent: impl Into<String>) -> Self {
581        self.intent = Some(intent.into());
582        self
583    }
584}
585
586impl Default for ToolContext {
587    fn default() -> Self {
588        Self {
589            workspace_dir: std::env::current_dir().unwrap_or_else(|_| PathBuf::from(".")),
590            root_dir: None,
591            session_id: None,
592            snapshot_store: None,
593            memory: None,
594            url_resolver: None,
595            todo: None,
596            agent_pool: None,
597            lsp: None,
598            subagent_runner: None,
599            subagent_depth: 0,
600            intent: None,
601        }
602    }
603}
604
605/// Result type for tool execution
606pub type ToolError = String;
607
608/// Result of tool execution
609#[derive(Debug)]
610pub struct AgentToolResult {
611    /// pub.
612    pub success: bool,
613    /// pub.
614    pub output: String,
615    /// pub.
616    pub metadata: Option<serde_json::Value>,
617    /// Optional content blocks (e.g., image blocks) to include in the tool result message.
618    /// When present, these are used as the content of the ToolResultMessage instead of
619    /// wrapping `output` in a Text block.
620    pub content_blocks: Option<Vec<oxi_ai::ContentBlock>>,
621    /// When `true`, signals that the agent loop should terminate after this batch
622    /// of tool calls completes.  Defaults to `false` so that the loop continues
623    /// unless a tool explicitly opts-in to termination.
624    pub terminate: bool,
625    /// Intent trace — a concise description of what this specific tool call did.
626    /// Set by the agent loop from the tool's static `intent()` or by the tool
627    /// itself for dynamic intent. Included in `ToolExecutionEnd` events.
628    pub intent: Option<String>,
629}
630
631impl AgentToolResult {
632    /// Creates a successful tool result with the given output text.
633    pub fn success(output: impl Into<String>) -> Self {
634        Self {
635            success: true,
636            output: output.into(),
637            metadata: None,
638            content_blocks: None,
639            terminate: false,
640            intent: None,
641        }
642    }
643
644    /// Creates an error tool result with the given error message.
645    pub fn error(output: impl Into<String>) -> Self {
646        Self {
647            success: false,
648            output: output.into(),
649            metadata: None,
650            content_blocks: None,
651            terminate: false,
652            intent: None,
653        }
654    }
655
656    /// Attaches structured metadata (JSON) to this result.
657    pub fn with_metadata(mut self, metadata: serde_json::Value) -> Self {
658        self.metadata = Some(metadata);
659        self
660    }
661
662    /// Attaches rich content blocks (images, code, etc.) to this result.
663    pub fn with_content_blocks(mut self, blocks: Vec<oxi_ai::ContentBlock>) -> Self {
664        self.content_blocks = Some(blocks);
665        self
666    }
667
668    /// Mark this result as requesting agent-loop termination.
669    pub fn with_terminate(mut self) -> Self {
670        self.terminate = true;
671        self
672    }
673
674    /// Attach an intent trace to this result.
675    pub fn with_intent(mut self, intent: impl Into<String>) -> Self {
676        self.intent = Some(intent.into());
677        self
678    }
679}
680
681impl fmt::Display for AgentToolResult {
682    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
683        write!(f, "{}", self.output)
684    }
685}
686
687/// Callback type for progress updates
688pub type ProgressCallback = Arc<dyn Fn(String) + Send + Sync>;
689
690/// Tool execution mode for parallel safety.
691#[derive(Debug, Clone)]
692pub enum ToolExecutionMode {
693    /// Safe to run in parallel with any other tool
694    ParallelSafe,
695    /// Must run sequentially — no parallel execution
696    SequentialOnly,
697    /// Mutates a specific file — file_mutation_queue serializes same-file access
698    MutatesFile(std::path::PathBuf),
699    /// Read-only — always parallel safe
700    ReadOnly,
701}
702
703/// Render output for TUI visualization.
704#[derive(Debug, Clone)]
705pub struct RenderOutput {
706    /// Rendered text content (markdown or plain)
707    pub content: String,
708    /// Whether to show collapsed by default
709    pub collapsed: bool,
710    /// Optional summary text for TUI footer
711    pub summary: Option<String>,
712}
713
714/// Core trait for all agent tools
715/// Risk tier for approval gating.
716///
717/// Determines which approval tiers gate a tool call.
718/// - `Read`  — no side effects (lookup, search, inspection).
719/// - `Write` — mutates data (creates, edits, commits).
720/// - `Exec`  — arbitrary side effects (shell, eval, network).
721#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
722pub enum ToolTier {
723    /// Read-only — inspection, search, lookup.
724    Read,
725    /// Data mutation — create, edit, commit.
726    Write,
727    /// Arbitrary execution — shell, eval, network, subagent.
728    #[default]
729    Exec,
730}
731
732/// Core trait for all agent tools
733#[async_trait]
734pub trait AgentTool: Send + Sync {
735    /// Tool name (used in function calls)
736    fn name(&self) -> &str;
737
738    /// Human-readable label
739    fn label(&self) -> &str;
740
741    /// Description for the model
742    fn description(&self) -> &str;
743
744    /// JSON Schema for parameters
745    fn parameters_schema(&self) -> Value;
746
747    /// Whether this tool is essential (cannot be disabled).
748    /// Essential tools: read, write, edit, bash, grep, find, ls
749    /// Optional tools: web_search, github, subagent, etc.
750    fn essential(&self) -> bool {
751        false
752    }
753
754    /// Execute the tool with the given tool call ID and parameters.
755    ///
756    /// The `ctx` parameter provides workspace information. File tools should
757    /// use `ctx.root()` to get the effective directory. Custom tools can use
758    /// `ctx.workspace_dir` for workspace-relative operations.
759    ///
760    /// # Examples
761    ///
762    /// ```ignore
763    /// use oxi_agent::{AgentTool, AgentToolResult, ToolContext};
764    /// use serde_json::json;
765    /// struct MyTool;
766    ///
767    /// #[async_trait]
768    /// impl AgentTool for MyTool {
769    ///     fn name(&self) -> &str { "my_tool" }
770    ///     fn label(&self) -> &str { "My Tool" }
771    ///     fn description(&self) -> &str { "A custom tool" }
772    ///     fn parameters_schema(&self) -> Value { json!({
773    ///         "type": "object",
774    ///         "properties": {}
775    ///     }) }
776    ///
777    ///     async fn execute(&self, tool_call_id: &str, params: Value, _signal: Option<oneshot::Receiver<()>>, ctx: &ToolContext) -> Result<AgentToolResult, String> {
778    ///         println!("Tool '{}' called with params: {:?}, workspace: {:?}", tool_call_id, params, ctx.workspace_dir);
779    ///         Ok(AgentToolResult::success("Done!"))
780    ///     }
781    /// }
782    /// ```
783    async fn execute(
784        &self,
785        tool_call_id: &str,
786        params: Value,
787        signal: Option<oneshot::Receiver<()>>,
788        ctx: &ToolContext,
789    ) -> Result<AgentToolResult, ToolError>;
790
791    /// Called with progress updates during execution.
792    /// Tools can override this to emit streaming updates.
793    fn on_progress(&self, _callback: ProgressCallback) {
794        // Default no-op
795    }
796
797    /// Structured browse progress callback for browser tool context enrichment.
798    /// Default implementation is no-op. Only browse tools override this to
799    /// register a callback that enriches `ToolCallContext` with structured
800    /// data from `BrowseProgress` events.
801    fn on_browse_progress(&self, _callback: crate::tools::browse::BrowseProgressCallback) {}
802
803    /// Custom rendering for tool call (TUI visualization).
804    /// Return None to use the default tool_renderer.rs formatter.
805    fn render_call(&self, _params: &serde_json::Value) -> Option<RenderOutput> {
806        None
807    }
808
809    /// Custom rendering for tool result (TUI visualization).
810    /// Return None to use the default tool_renderer.rs formatter.
811    fn render_result(&self, _result: &AgentToolResult) -> Option<RenderOutput> {
812        None
813    }
814
815    /// Intent trace — a concise description of what this tool does.
816    /// Returned value is included in `ToolExecutionStart` / `ToolExecutionEnd`
817    /// events so the agent loop can surface intent to users or telemetry.
818    /// Default `None` (no intent tracing).
819    fn intent(&self) -> Option<&str> {
820        None
821    }
822
823    /// Execution mode for parallel safety.
824    /// Defaults to ParallelSafe. Override for file-mutating or sequential tools.
825    fn execution_mode(&self) -> ToolExecutionMode {
826        ToolExecutionMode::ParallelSafe
827    }
828
829    /// Risk tier for approval gating.
830    ///
831    /// - `Read`  — no side effects (lookup, search, inspection).
832    /// - `Write` — mutates data (creates, edits, commits).
833    /// - `Exec`  — arbitrary side effects (shell, eval, network).
834    ///
835    /// Default: `Exec` (safest default — requires explicit opt-down).
836    fn tool_tier(&self) -> ToolTier {
837        ToolTier::Exec
838    }
839
840    /// Return the current active tab ID, if this tool manages browser tabs.
841    /// Defaults to `None`. Browser tools override this to return the tab ID
842    /// of the currently-open tab during execution, so the agent loop can
843    /// populate `ToolExecutionUpdate.tab_id`.
844    fn current_tab_id(&self) -> Option<uuid::Uuid> {
845        None
846    }
847
848    /// Receive a shared slot where the tool can write the current tab ID.
849    /// The agent loop creates the slot and passes it before `on_progress`;
850    /// the tool writes `Some(tab_id)` when it opens a tab and `None` when
851    /// it closes it. Defaults to a no-op — only tab-aware tools override.
852    fn set_tab_id_slot(&self, _slot: Arc<parking_lot::Mutex<Option<uuid::Uuid>>>) {}
853
854    /// Convert to ToolDefinition
855    fn to_definition(&self) -> ToolDefinition {
856        ToolDefinition {
857            name: self.name().to_string(),
858            description: self.description().to_string(),
859            input_schema: serde_json::from_value(self.parameters_schema()).unwrap_or_default(),
860        }
861    }
862}
863
864// Built-in tools
865/// Ask tool — ask the user one or more clarifying questions via the TUI overlay.
866pub mod ask;
867/// AST-aware structural code rewriting tool (ast-grep backed).
868pub mod ast_edit;
869/// AST structural search tool (wraps the `sg` CLI).
870pub mod ast_grep;
871/// Bash shell execution tool.
872pub mod bash;
873/// Browser tools (engine abstraction always compiled).
874pub mod browse;
875/// Checkpoint and Rewind tools — save/restore investigation state.
876pub mod checkpoint_tool;
877/// Conventional-commit tool (deterministic scope + LLM analysis).
878pub mod commit;
879/// Computer tool — computer control using Vision AI.
880pub mod computer_tool;
881/// Context7 documentation tools.
882pub mod context7;
883/// Debug tool — DAP-backed debugger integration (scaffold).
884pub mod debug_tool;
885/// In-place file edit tool.
886pub mod edit;
887/// Diff-based edit helpers.
888pub mod edit_diff;
889/// Eval tool — persistent-kernel code execution (scaffold).
890pub mod eval_tool;
891/// Serialised file-mutation queue.
892pub mod file_mutation_queue;
893/// File-fsystem find tool.
894pub mod find;
895/// Image generation tool (OpenRouter API).
896pub mod generate_image;
897/// GitHub integration tool (gh CLI-based).
898pub mod github;
899/// GitHub repository search tool (legacy REST API).
900pub mod github_search;
901/// Goal tool — manage investigation goals with token budgets.
902pub mod goal_tool;
903/// Content search (grep) tool.
904pub mod grep;
905/// TokioHashlineFs — tokio::fs-backed HashlineFs implementation.
906pub mod hashline_fs;
907/// Shared HTTP client singleton.
908pub mod http_client;
909/// Hub tool — agent coordination for peer messaging and job management.
910pub mod hub_tool;
911/// Inspect Image tool — analyze images using Vision LLM capabilities.
912pub mod inspect_image_tool;
913/// Learn tool — capture a reusable lesson to memory and optionally create a managed skill.
914pub mod learn_tool;
915/// Directory listing tool.
916pub mod ls;
917/// LSP tool (requires LspProvider capability).
918pub mod lsp;
919/// Manage Skill tool — create, update, or delete isolated managed SKILL.md files.
920pub mod manage_skill_tool;
921/// Memory edit tool — update or delete a memory item.
922pub mod memory_edit;
923/// Memory recall tool — semantic search over stored memories.
924pub mod memory_recall;
925/// Memory reflect tool — persist a session summary to memory.
926pub mod memory_reflect;
927/// Memory retain tool — persist a memory item to the backend.
928pub mod memory_retain;
929/// Path security (traversal protection).
930pub mod path_security;
931/// Path manipulation utilities.
932pub mod path_utils;
933/// File reading tool.
934pub mod read;
935/// Rendering utilities for tool output.
936pub mod render_utils;
937/// Review tool — request code review with focus areas and priorities.
938pub mod review_tool;
939/// Search result cache and get_search_results tool.
940pub mod search_cache;
941/// Sub-agent delegation tool.
942pub mod subagent;
943/// Phased todo tool (init/start/done/drop/rm/append/view).
944pub mod todo;
945/// Tool definition wrapper helpers.
946pub mod tool_definition_wrapper;
947/// Output truncation helpers.
948pub mod truncate;
949/// TTS tool — text-to-speech synthesis.
950pub mod tts_tool;
951/// Vibe tool — manage persistent worker sessions.
952pub mod vibe_tool;
953/// Multi-engine web search tool (oxibrowser search module).
954pub mod web_search;
955/// File writing tool.
956pub mod write;
957/// Yield tool — subagent result submission.
958pub mod yield_tool;
959
960// Re-export for convenience
961pub use bash::BashTool;
962pub use debug_tool::DebugTool;
963pub use edit::EditTool;
964pub use eval_tool::EvalTool;
965pub use find::FindTool;
966pub use grep::GrepTool;
967pub use ls::LsTool;
968pub use read::ReadTool;
969// pub use search_cache;
970
971pub use crate::mcp::McpTool;
972pub use ask::{AskBridge, AskTool};
973pub use ast_edit::AstEditTool;
974pub use ast_grep::AstGrepTool;
975pub use commit::CommitTool;
976pub use context7::{Context7QueryDocsTool, Context7ResolveLibraryIdTool};
977pub use memory_edit::MemoryEditTool;
978pub use memory_recall::MemoryRecallTool;
979pub use memory_reflect::MemoryReflectTool;
980pub use memory_retain::MemoryRetainTool;
981pub use subagent::SubagentTool;
982pub use write::WriteTool;
983
984/// Tool registry for managing available tools
985#[derive(Clone)]
986pub struct ToolRegistry {
987    tools: Arc<parking_lot::RwLock<std::collections::HashMap<String, Arc<dyn AgentTool>>>>,
988    /// Optional MCP manager, set by `with_builtins_cwd()` so the TUI and
989    /// other consumers can reach the live MCP state (Phase 2+).
990    mcp_manager: Arc<parking_lot::RwLock<Option<Arc<crate::mcp::McpManager>>>>,
991}
992
993impl Default for ToolRegistry {
994    fn default() -> Self {
995        Self::new()
996    }
997}
998
999impl ToolRegistry {
1000    /// Creates an empty tool registry.
1001    pub fn new() -> Self {
1002        Self {
1003            tools: Arc::new(parking_lot::RwLock::new(std::collections::HashMap::new())),
1004            mcp_manager: Arc::new(parking_lot::RwLock::new(None)),
1005        }
1006    }
1007
1008    /// Attach an `McpManager` to this registry. Replaces any previous one.
1009    pub fn set_mcp_manager(&self, mgr: Arc<crate::mcp::McpManager>) {
1010        *self.mcp_manager.write() = Some(mgr);
1011    }
1012
1013    /// Get the attached `McpManager`, if any.
1014    pub fn mcp_manager(&self) -> Option<Arc<crate::mcp::McpManager>> {
1015        self.mcp_manager.read().clone()
1016    }
1017
1018    /// Register a tool
1019    pub fn register(&self, tool: impl AgentTool + 'static) {
1020        let name = tool.name().to_string();
1021        self.tools.write().insert(name, Arc::new(tool));
1022    }
1023
1024    /// Register a tool that is already wrapped in an `Arc`.
1025    /// This is the primary path for extensions that produce `Arc<dyn AgentTool>`.
1026    pub fn register_arc(&self, tool: Arc<dyn AgentTool>) {
1027        let name = tool.name().to_string();
1028        self.tools.write().insert(name, tool);
1029    }
1030
1031    /// Get a tool by name
1032    pub fn get(&self, name: &str) -> Option<Arc<dyn AgentTool>> {
1033        self.tools.read().get(name).cloned()
1034    }
1035
1036    /// Unregister a tool by name.
1037    /// Returns `true` if the tool was present and removed.
1038    pub fn unregister(&self, name: &str) -> bool {
1039        self.tools.write().remove(name).is_some()
1040    }
1041
1042    /// List all registered tool names
1043    pub fn names(&self) -> Vec<String> {
1044        self.tools.read().keys().cloned().collect()
1045    }
1046
1047    /// Get all tool definitions
1048    pub fn definitions(&self) -> Vec<ToolDefinition> {
1049        self.tools
1050            .read()
1051            .values()
1052            .map(|t| t.to_definition())
1053            .collect()
1054    }
1055
1056    /// Get all tools as a slice
1057    pub fn get_tools(&self) -> Vec<Arc<dyn AgentTool>> {
1058        self.tools.read().values().cloned().collect()
1059    }
1060
1061    /// Check whether all tools in `required` are registered.
1062    ///
1063    /// Useful for validating program/module dependencies before execution.
1064    ///
1065    /// # Example
1066    ///
1067    /// ```
1068    /// use oxi_agent::ToolRegistry;
1069    /// let registry = ToolRegistry::new();
1070    /// assert!(!registry.has_all(&["read", "write"]));
1071    /// ```
1072    pub fn has_all(&self, required: &[&str]) -> bool {
1073        let tools = self.tools.read();
1074        required.iter().all(|name| tools.contains_key(*name))
1075    }
1076
1077    /// Return the subset of `required` tool names that are **not** registered.
1078    ///
1079    /// # Example
1080    ///
1081    /// ```
1082    /// use oxi_agent::ToolRegistry;
1083    /// let registry = ToolRegistry::new();
1084    /// let missing = registry.missing(&["read", "exec", "nonexistent"]);
1085    /// assert_eq!(missing, vec!["read", "exec", "nonexistent"]);
1086    /// ```
1087    pub fn missing<'a>(&self, required: &[&'a str]) -> Vec<&'a str> {
1088        let tools = self.tools.read();
1089        required
1090            .iter()
1091            .filter(|name| !tools.contains_key(**name))
1092            .copied()
1093            .collect()
1094    }
1095
1096    /// Create a registry with all built-in tools
1097    ///
1098    /// # Examples
1099    ///
1100    /// ```
1101    /// use oxi_agent::ToolRegistry;
1102    /// let registry = ToolRegistry::with_builtins();
1103    /// let tools = registry.names();
1104    /// assert!(tools.contains(&"read".to_string()));
1105    /// assert!(tools.contains(&"write".to_string()));
1106    /// assert!(tools.contains(&"bash".to_string()));
1107    /// ```
1108    pub fn with_builtins() -> Self {
1109        Self::with_builtins_cwd(PathBuf::from("."), &[])
1110    }
1111
1112    /// Create a registry with all built-in tools, using the given cwd.
1113    ///
1114    /// Pass `disabled_tools` to selectively disable built-in tools
1115    /// (e.g. `["web_search", "github_search"]` for a minimal setup).
1116    pub fn with_builtins_cwd(cwd: PathBuf, disabled_tools: &[String]) -> Self {
1117        let registry = Self::new();
1118        let disabled: std::collections::HashSet<&str> =
1119            disabled_tools.iter().map(|s| s.as_str()).collect();
1120
1121        // Helper to create shared cache on demand
1122        let cache_once: std::cell::OnceCell<Arc<search_cache::SearchCache>> =
1123            std::cell::OnceCell::new();
1124
1125        // MCP: use OnceCell to avoid re-creating McpManager on repeated calls
1126        let mcp_once: std::cell::OnceCell<Arc<crate::mcp::McpManager>> = std::cell::OnceCell::new();
1127        let mcp_manager = mcp_once.get_or_init(crate::mcp::McpManager::spawn).clone();
1128
1129        // Register all builtin tools — essential ones ignore disabled list
1130        let mut all_tools: Vec<Box<dyn AgentTool>> = vec![
1131            Box::new(ReadTool::with_cwd(cwd.clone())),
1132            Box::new(WriteTool::with_cwd(cwd.clone())),
1133            Box::new(AstGrepTool::with_cwd(cwd.clone())),
1134            Box::new(BashTool::with_cwd(cwd.clone())),
1135            Box::new(EditTool::with_cwd(cwd.clone())),
1136            Box::new(GrepTool::with_cwd(cwd.clone())),
1137            Box::new(FindTool::with_cwd(cwd.clone())),
1138            Box::new(LsTool::with_cwd(cwd.clone())),
1139            Box::new(web_search::WebSearchTool::new(
1140                cache_once
1141                    .get_or_init(|| Arc::new(search_cache::SearchCache::new()))
1142                    .clone(),
1143            )),
1144            Box::new(search_cache::GetSearchResultsTool::new(
1145                cache_once
1146                    .get_or_init(|| Arc::new(search_cache::SearchCache::new()))
1147                    .clone(),
1148            )),
1149            Box::new(github::GitHubTool::new(
1150                cache_once
1151                    .get_or_init(|| Arc::new(search_cache::SearchCache::new()))
1152                    .clone(),
1153            )),
1154            Box::new(SubagentTool::with_cwd(cwd.clone())),
1155            Box::new(todo::TodoTool),
1156            Box::new(memory_recall::MemoryRecallTool),
1157            Box::new(memory_reflect::MemoryReflectTool),
1158            Box::new(memory_retain::MemoryRetainTool),
1159            Box::new(memory_edit::MemoryEditTool),
1160        ];
1161
1162        all_tools.push(Box::new(crate::mcp::McpTool::new(mcp_manager.clone())));
1163
1164        // Phase 3: register direct MCP tools from the metadata cache.
1165        for def in mcp_manager.direct_tools_from_cache() {
1166            all_tools.push(Box::new(crate::mcp::McpDirectTool::new(
1167                mcp_manager.clone(),
1168                def,
1169            )));
1170        }
1171
1172        // Remember the manager on the registry so the TUI can reach it.
1173        registry.set_mcp_manager(mcp_manager);
1174
1175        all_tools.push(Box::new(context7::Context7ResolveLibraryIdTool::new()));
1176        all_tools.push(Box::new(context7::Context7QueryDocsTool::new()));
1177        all_tools.push(Box::new(generate_image::GenerateImageTool::new()));
1178        all_tools.push(Box::new(commit::CommitTool::unconfigured()));
1179        all_tools.push(Box::new(ast_edit::AstEditTool::new()));
1180        all_tools.push(Box::new(lsp::LspTool));
1181        all_tools.push(Box::new(eval_tool::EvalTool));
1182        all_tools.push(Box::new(checkpoint_tool::CheckpointTool));
1183        all_tools.push(Box::new(checkpoint_tool::RewindTool));
1184        all_tools.push(Box::new(hub_tool::HubTool));
1185        all_tools.push(Box::new(yield_tool::YieldTool));
1186        all_tools.push(Box::new(goal_tool::GoalTool));
1187        all_tools.push(Box::new(review_tool::ReviewTool));
1188        all_tools.push(Box::new(learn_tool::LearnTool));
1189        all_tools.push(Box::new(manage_skill_tool::ManageSkillTool));
1190        all_tools.push(Box::new(inspect_image_tool::InspectImageTool));
1191        all_tools.push(Box::new(computer_tool::ComputerTool));
1192        all_tools.push(Box::new(tts_tool::TtsTool));
1193        all_tools.push(Box::new(vibe_tool::VibeTool));
1194        // debug_tool — DAP-backed debugger integration.
1195        // Most actions are validated scaffolds (route through xd://debug);
1196        // real launch/attach/breakpoint control is wired via the harness device.
1197        all_tools.push(Box::new(debug_tool::DebugTool));
1198
1199        for tool in all_tools {
1200            if tool.essential() || !disabled.contains(tool.name()) {
1201                // web_search ↔ get_search_results coupling
1202                if tool.name() == "get_search_results" && disabled.contains("web_search") {
1203                    continue;
1204                }
1205                registry.register_arc(Arc::from(tool));
1206            }
1207        }
1208
1209        registry
1210    }
1211
1212    /// Extend this registry with all tools from another registry.
1213    ///
1214    /// Useful for composing tool sets from multiple sources
1215    /// (e.g., coding tools + kernel tools + browser tools).
1216    ///
1217    /// # Example
1218    ///
1219    /// ```ignore
1220    /// let base = ToolRegistry::new();
1221    /// base.extend_from(&other_registry);
1222    /// ```
1223    pub fn extend_from(&self, other: &ToolRegistry) {
1224        for name in other.names() {
1225            if let Some(tool) = other.get(&name) {
1226                self.register_arc(tool);
1227            }
1228        }
1229    }
1230
1231    /// Create registry with selected builtins only.
1232    pub fn with_selected_tools(cwd: PathBuf, names: &[&str]) -> Self {
1233        let full = Self::with_builtins_cwd(cwd, &[]);
1234        let registry = Self::new();
1235        let set: std::collections::HashSet<&str> = names.iter().copied().collect();
1236        for name in full.names() {
1237            if set.contains(name.as_str())
1238                && let Some(tool) = full.get(&name)
1239            {
1240                registry.register_arc(tool);
1241            }
1242        }
1243        registry
1244    }
1245}