llm/
context.rs

1use serde::{Deserialize, Serialize};
2
3use crate::catalog::LlmModel;
4use crate::chat_message::AssistantReasoning;
5use crate::model_settings::ModelSettings;
6use crate::reasoning::ReasoningEffort;
7use crate::types::IsoString;
8
9use super::{ChatMessage, ToolDefinition};
10
11#[doc = include_str!("docs/context.md")]
12#[derive(Debug, Clone, Serialize, Deserialize)]
13pub struct Context {
14    messages: Vec<ChatMessage>,
15    tools: Vec<ToolDefinition>,
16    #[serde(skip)]
17    reasoning_effort: Option<ReasoningEffort>,
18    #[serde(skip)]
19    model_settings: ModelSettings,
20    #[serde(skip)]
21    prompt_cache_key: Option<String>,
22}
23
24impl Context {
25    pub fn new(messages: Vec<ChatMessage>, tools: Vec<ToolDefinition>) -> Self {
26        Self {
27            messages,
28            tools,
29            reasoning_effort: None,
30            model_settings: ModelSettings::default(),
31            prompt_cache_key: None,
32        }
33    }
34
35    pub fn prompt_cache_key(&self) -> Option<&str> {
36        self.prompt_cache_key.as_deref()
37    }
38
39    pub fn set_prompt_cache_key(&mut self, key: Option<String>) {
40        self.prompt_cache_key = key;
41    }
42
43    pub fn reasoning_effort(&self) -> Option<ReasoningEffort> {
44        self.reasoning_effort
45    }
46
47    pub fn set_reasoning_effort(&mut self, effort: Option<ReasoningEffort>) {
48        self.reasoning_effort = effort;
49    }
50
51    pub fn model_settings(&self) -> &ModelSettings {
52        &self.model_settings
53    }
54
55    pub fn set_model_settings(&mut self, settings: ModelSettings) {
56        self.model_settings = settings;
57    }
58
59    pub fn add_message(&mut self, message: ChatMessage) {
60        self.messages.push(message);
61    }
62
63    pub fn set_tools(&mut self, tools: Vec<ToolDefinition>) {
64        self.tools = tools;
65    }
66
67    pub fn set_system_content(&mut self, content: String) {
68        if let Some(ChatMessage::System { content: existing, .. }) = self.messages.first_mut() {
69            *existing = content;
70        } else {
71            self.messages.insert(0, ChatMessage::System { content, timestamp: IsoString::now() });
72        }
73    }
74
75    pub fn messages(&self) -> &Vec<ChatMessage> {
76        &self.messages
77    }
78
79    pub fn tools(&self) -> &Vec<ToolDefinition> {
80        &self.tools
81    }
82
83    /// Returns the number of messages in the context
84    pub fn message_count(&self) -> usize {
85        self.messages.len()
86    }
87
88    /// Estimate total token count using the ~4 bytes/token heuristic.
89    /// Includes messages and tool definitions. Used for pre-flight overflow detection.
90    pub fn estimated_token_count(&self) -> u32 {
91        let message_bytes: usize = self.messages.iter().map(ChatMessage::estimated_bytes).sum();
92        let tool_bytes: usize = self
93            .tools
94            .iter()
95            .map(|tool| tool.name.len() + tool.description.len() + tool.parameters.to_string().len())
96            .sum();
97        let total_bytes = message_bytes + tool_bytes;
98        u32::try_from(total_bytes / 4).unwrap_or(u32::MAX)
99    }
100
101    /// Build an assistant turn and its tool call results and append them to messages.
102    pub fn push_assistant_turn(
103        &mut self,
104        content: &str,
105        reasoning: AssistantReasoning,
106        completed_tools: Vec<Result<super::ToolCallResult, super::ToolCallError>>,
107    ) {
108        let tool_requests: Vec<_> = completed_tools
109            .iter()
110            .map(|result| match result {
111                Ok(r) => {
112                    super::ToolCallRequest { id: r.id.clone(), name: r.name.clone(), arguments: r.arguments.clone() }
113                }
114                Err(e) => super::ToolCallRequest {
115                    id: e.id.clone(),
116                    name: e.name.clone(),
117                    arguments: e.arguments.clone().unwrap_or_default(),
118                },
119            })
120            .collect();
121
122        self.messages.push(ChatMessage::Assistant {
123            content: content.to_string(),
124            reasoning,
125            timestamp: IsoString::now(),
126            tool_calls: tool_requests,
127        });
128
129        for result in completed_tools {
130            self.messages.push(ChatMessage::ToolCallResult(result));
131        }
132    }
133
134    /// Return a copy with encrypted reasoning projected onto `model`.
135    ///
136    /// Encrypted content is opaque to every model but the one that produced it,
137    /// so it survives only when its source model matches. Passing `None` — the
138    /// model is not in the catalog, so nothing can be claimed to match — drops
139    /// all of it.
140    pub fn filter_encrypted_reasoning(&self, model: Option<&LlmModel>) -> Self {
141        let messages = self
142            .messages
143            .iter()
144            .map(|message| match message {
145                ChatMessage::Assistant { content, reasoning, timestamp, tool_calls } => ChatMessage::Assistant {
146                    content: content.clone(),
147                    reasoning: AssistantReasoning {
148                        summary_text: reasoning.summary_text.clone(),
149                        encrypted_content: reasoning
150                            .encrypted_content
151                            .as_ref()
152                            .filter(|encrypted| model.is_some_and(|model| &encrypted.model == model))
153                            .cloned(),
154                    },
155                    timestamp: timestamp.clone(),
156                    tool_calls: tool_calls.clone(),
157                },
158                other => other.clone(),
159            })
160            .collect();
161        Self { messages, ..self.clone() }
162    }
163
164    /// Clear all non-system messages, retaining only system prompts.
165    pub fn clear_conversation(&mut self) {
166        self.messages.retain(super::chat_message::ChatMessage::is_system);
167    }
168
169    /// Replace all non-system messages while preserving the system prompt and runtime state.
170    pub fn replace_conversation(&mut self, messages: Vec<ChatMessage>) {
171        self.messages = self
172            .messages
173            .drain(..)
174            .filter(ChatMessage::is_system)
175            .chain(messages.into_iter().filter(|m| !m.is_system()))
176            .collect();
177    }
178
179    /// Get all non-system messages for summarization
180    pub fn messages_for_summary(&self) -> Vec<&ChatMessage> {
181        self.messages.iter().filter(|msg| !msg.is_system()).collect()
182    }
183
184    /// Create a new context with all messages replaced by a summary.
185    /// Preserves the system prompt and tools.
186    pub fn with_compacted_summary(&self, summary: &str) -> Context {
187        let system_messages: Vec<_> = self.messages.iter().filter(|msg| msg.is_system()).cloned().collect();
188
189        let non_system_count = self.messages.len() - system_messages.len();
190
191        let mut messages = system_messages;
192        if non_system_count > 0 {
193            messages.push(ChatMessage::Summary {
194                content: summary.to_string(),
195                timestamp: IsoString::now(),
196                messages_compacted: non_system_count,
197            });
198        }
199
200        Context { messages, ..self.clone() }
201    }
202}
203
204#[cfg(test)]
205mod tests {
206    use super::*;
207
208    use crate::ToolCallResult;
209    use crate::catalog::LlmModel;
210
211    fn create_test_context() -> Context {
212        let messages = vec![
213            ChatMessage::system("You are a helpful assistant."),
214            ChatMessage::user("Hello"),
215            ChatMessage::Assistant {
216                content: "Hi there!".to_string(),
217                reasoning: AssistantReasoning::default(),
218                timestamp: IsoString::now(),
219                tool_calls: vec![],
220            },
221            ChatMessage::ToolCallResult(Ok(ToolCallResult {
222                id: "1".to_string(),
223                name: "tool1".to_string(),
224                arguments: "{}".to_string(),
225                result: "Result 1".to_string(),
226            })),
227            ChatMessage::ToolCallResult(Ok(ToolCallResult {
228                id: "2".to_string(),
229                name: "tool2".to_string(),
230                arguments: "{}".to_string(),
231                result: "Result 2".to_string(),
232            })),
233            ChatMessage::ToolCallResult(Ok(ToolCallResult {
234                id: "3".to_string(),
235                name: "tool3".to_string(),
236                arguments: "{}".to_string(),
237                result: "Result 3".to_string(),
238            })),
239        ];
240        Context::new(messages, vec![])
241    }
242
243    #[test]
244    fn replace_conversation_preserves_system_message() {
245        let mut ctx = create_test_context();
246        ctx.replace_conversation(vec![ChatMessage::user("new")]);
247
248        assert_eq!(ctx.message_count(), 2);
249        assert!(ctx.messages()[0].is_system());
250        assert!(matches!(ctx.messages()[1], ChatMessage::User { .. }));
251    }
252
253    #[test]
254    fn replace_conversation_replaces_old_non_system_messages() {
255        let mut ctx = create_test_context();
256        ctx.replace_conversation(vec![ChatMessage::Assistant {
257            content: "replacement".to_string(),
258            reasoning: AssistantReasoning::default(),
259            timestamp: IsoString::now(),
260            tool_calls: vec![],
261        }]);
262
263        assert_eq!(ctx.message_count(), 2);
264        assert!(
265            ctx.messages()
266                .iter()
267                .all(|message| { !matches!(message, ChatMessage::User { .. } | ChatMessage::ToolCallResult(_)) })
268        );
269        assert!(matches!(ctx.messages()[1], ChatMessage::Assistant { ref content, .. } if content == "replacement"));
270    }
271
272    #[test]
273    fn replace_conversation_filters_incoming_system_messages() {
274        let mut ctx = create_test_context();
275        ctx.replace_conversation(vec![ChatMessage::system("wrong system"), ChatMessage::user("kept")]);
276
277        assert_eq!(ctx.message_count(), 2);
278        assert!(
279            matches!(ctx.messages()[0], ChatMessage::System { ref content, .. } if content == "You are a helpful assistant.")
280        );
281        assert!(matches!(ctx.messages()[1], ChatMessage::User { .. }));
282    }
283
284    #[test]
285    fn replace_conversation_does_not_change_tools() {
286        let tool = ToolDefinition::new("read_file", "Reads a file", serde_json::json!({}));
287        let mut ctx = Context::new(vec![ChatMessage::system("system")], vec![tool.clone()]);
288        ctx.replace_conversation(vec![ChatMessage::user("new")]);
289
290        assert_eq!(ctx.tools(), &vec![tool]);
291    }
292
293    #[test]
294    fn test_message_count() {
295        let ctx = create_test_context();
296        assert_eq!(ctx.message_count(), 6);
297    }
298
299    #[test]
300    fn test_with_compacted_summary_preserves_system_prompt() {
301        let ctx = create_test_context();
302        let compacted = ctx.with_compacted_summary("This is a summary of previous conversation.");
303
304        assert_eq!(compacted.message_count(), 2);
305        assert!(compacted.messages()[0].is_system());
306        assert!(compacted.messages()[1].is_summary());
307    }
308
309    #[test]
310    fn test_with_compacted_summary_empty_context() {
311        let ctx = Context::new(vec![ChatMessage::system("System")], vec![]);
312        let compacted = ctx.with_compacted_summary("Summary");
313
314        assert_eq!(compacted.message_count(), 1);
315    }
316
317    #[test]
318    fn test_messages_for_summary() {
319        let ctx = create_test_context();
320        let msgs = ctx.messages_for_summary();
321
322        assert_eq!(msgs.len(), 5);
323        assert!(msgs.iter().all(|m| !m.is_system()));
324    }
325
326    #[test]
327    fn test_prompt_cache_key_default_is_none() {
328        let ctx = create_test_context();
329        assert_eq!(ctx.prompt_cache_key(), None);
330    }
331
332    #[test]
333    fn test_prompt_cache_key_set_and_get() {
334        let mut ctx = create_test_context();
335        ctx.set_prompt_cache_key(Some("session-123".to_string()));
336        assert_eq!(ctx.prompt_cache_key(), Some("session-123"));
337
338        ctx.set_prompt_cache_key(None);
339        assert_eq!(ctx.prompt_cache_key(), None);
340    }
341
342    #[test]
343    fn test_prompt_cache_key_preserved_through_compaction() {
344        let mut ctx = create_test_context();
345        ctx.set_prompt_cache_key(Some("session-abc".to_string()));
346        let compacted = ctx.with_compacted_summary("Summary");
347        assert_eq!(compacted.prompt_cache_key(), Some("session-abc"));
348    }
349
350    #[test]
351    fn test_prompt_cache_key_preserved_through_projection() {
352        let model: LlmModel = "anthropic:claude-opus-4-6".parse().unwrap();
353        let mut ctx = Context::new(vec![ChatMessage::user("Hello")], vec![]);
354        ctx.set_prompt_cache_key(Some("session-xyz".to_string()));
355        let projected = ctx.filter_encrypted_reasoning(Some(&model));
356        assert_eq!(projected.prompt_cache_key(), Some("session-xyz"));
357    }
358
359    #[test]
360    fn test_reasoning_effort_default_is_none() {
361        let ctx = create_test_context();
362        assert_eq!(ctx.reasoning_effort(), None);
363    }
364
365    #[test]
366    fn test_reasoning_effort_set_and_get() {
367        let mut ctx = create_test_context();
368        ctx.set_reasoning_effort(Some(crate::ReasoningEffort::High));
369        assert_eq!(ctx.reasoning_effort(), Some(crate::ReasoningEffort::High));
370
371        ctx.set_reasoning_effort(None);
372        assert_eq!(ctx.reasoning_effort(), None);
373    }
374
375    #[test]
376    fn test_reasoning_effort_preserved_through_compaction() {
377        let mut ctx = create_test_context();
378        ctx.set_reasoning_effort(Some(crate::ReasoningEffort::Medium));
379        let compacted = ctx.with_compacted_summary("Summary");
380        assert_eq!(compacted.reasoning_effort(), Some(crate::ReasoningEffort::Medium));
381    }
382
383    #[test]
384    fn test_estimated_token_count() {
385        use crate::ToolDefinition;
386
387        // "You are a helpful assistant." = 28 bytes
388        // "Hello" = 5 bytes
389        // "Hi there!" = 9 bytes (assistant, no reasoning, no tool calls)
390        // 3 tool results: "Result 1" (8) + "tool1" (5) + "{}" (2) = 15 each = 45 total
391        // Total message bytes = 28 + 5 + 9 + 45 = 87
392        let ctx = create_test_context();
393        let base_estimate = ctx.estimated_token_count();
394
395        // With no tools, estimate = message_bytes / 4
396        assert_eq!(base_estimate, 87 / 4);
397
398        let tool = ToolDefinition::new("read_file", "Reads a file", serde_json::json!({}));
399        let ctx_with_tools = Context::new(ctx.messages().clone(), vec![tool]);
400        let with_tools_estimate = ctx_with_tools.estimated_token_count();
401        assert_eq!(with_tools_estimate, (87 + 9 + 12 + 2) / 4);
402        assert!(with_tools_estimate > base_estimate);
403    }
404
405    #[test]
406    fn compaction_drops_encrypted_reasoning() {
407        let model: LlmModel = "anthropic:claude-opus-4-6".parse().unwrap();
408        let ctx = Context::new(
409            vec![
410                ChatMessage::user("Hello"),
411                ChatMessage::Assistant {
412                    content: "I see.".to_string(),
413                    reasoning: AssistantReasoning {
414                        summary_text: Some("thinking".to_string()),
415                        encrypted_content: Some(crate::EncryptedReasoningContent {
416                            id: "r_test".to_string(),
417                            model,
418                            content: "blob".to_string(),
419                        }),
420                    },
421                    timestamp: IsoString::now(),
422                    tool_calls: vec![],
423                },
424            ],
425            vec![],
426        );
427        let compacted = ctx.with_compacted_summary("Summary of conversation");
428
429        for msg in compacted.messages() {
430            if let ChatMessage::Assistant { reasoning, .. } = msg {
431                assert!(reasoning.encrypted_content.is_none(), "compaction should drop encrypted reasoning");
432            }
433        }
434    }
435
436    #[test]
437    fn projected_for_keeps_matching_model() {
438        let model: LlmModel = "anthropic:claude-opus-4-6".parse().unwrap();
439        let ctx = Context::new(
440            vec![ChatMessage::Assistant {
441                content: "reply".to_string(),
442                reasoning: AssistantReasoning {
443                    summary_text: Some("think".to_string()),
444                    encrypted_content: Some(crate::EncryptedReasoningContent {
445                        id: "r_test".to_string(),
446                        model: model.clone(),
447                        content: "blob".to_string(),
448                    }),
449                },
450                timestamp: IsoString::now(),
451                tool_calls: vec![],
452            }],
453            vec![],
454        );
455        let projected = ctx.filter_encrypted_reasoning(Some(&model));
456        if let ChatMessage::Assistant { reasoning, .. } = &projected.messages()[0] {
457            assert!(reasoning.encrypted_content.is_some());
458            assert_eq!(reasoning.summary_text.as_deref(), Some("think"));
459        } else {
460            panic!("expected assistant message");
461        }
462    }
463
464    #[test]
465    fn projected_for_strips_non_matching_model() {
466        let model_a: LlmModel = "anthropic:claude-opus-4-6".parse().unwrap();
467        let model_b: LlmModel = "anthropic:claude-sonnet-4-5-20250929".parse().unwrap();
468        let ctx = Context::new(
469            vec![ChatMessage::Assistant {
470                content: "reply".to_string(),
471                reasoning: AssistantReasoning {
472                    summary_text: Some("think".to_string()),
473                    encrypted_content: Some(crate::EncryptedReasoningContent {
474                        id: "r_test".to_string(),
475                        model: model_a,
476                        content: "blob".to_string(),
477                    }),
478                },
479                timestamp: IsoString::now(),
480                tool_calls: vec![],
481            }],
482            vec![],
483        );
484        let projected = ctx.filter_encrypted_reasoning(Some(&model_b));
485        if let ChatMessage::Assistant { reasoning, .. } = &projected.messages()[0] {
486            assert!(reasoning.encrypted_content.is_none());
487            assert_eq!(reasoning.summary_text.as_deref(), Some("think"));
488        } else {
489            panic!("expected assistant message");
490        }
491    }
492}
llm/context.rs

llm/
context.rs
