mj_transcript/
turn_context.rs

1//! Process-local classification evidence using the shared transcript summary.
2use agent_client_protocol::schema::v1::{ContentBlock, SessionUpdate};
3use mj_core::activity::ActivityFacts;
4use mj_core::activity::verdict::*;
5use mj_core::config::HarnessKind;
6use std::sync::{Arc, Mutex};
7
8#[derive(Debug, Default)]
9struct TurnContextState {
10    generation: u64,
11    parent_activity: Option<std::time::Instant>,
12    decision_log: Option<mj_core::jev::DecisionLog>,
13    user_prompt_tail: String,
14    message_id: Option<String>,
15    assistant_text_tail: String,
16    last_completed_message: String,
17    summary: crate::summary::TranscriptSummary,
18    background_commands: usize,
19    queued_commands: usize,
20    background_inventory: Vec<mj_core::relay::BackgroundCommand>,
21    native_agent_ids: Vec<String>,
22    session_id: String,
23}
24
25/// The prompt that became visible to the harness at this durable observation.
26/// A queued or unconfirmed steering prompt has not been delivered yet.
27pub fn delivered_prompt_command_id(observation: &mj_core::relay::RelayObservation) -> Option<&str> {
28    use mj_core::relay::{RelayCommandOutcome, RelayObservation};
29    match observation {
30        RelayObservation::CommandStarted { command_id, .. } => Some(command_id),
31        RelayObservation::CommandCompleted {
32            outcome: RelayCommandOutcome::Steered { queued_command_id },
33            ..
34        } => Some(queued_command_id),
35        _ => None,
36    }
37}
38
39/// The relay and runtime share this small process-local evidence accumulator.
40#[derive(Debug, Default, Clone)]
41pub struct TurnContext(Arc<Mutex<TurnContextState>>);
42
43impl TurnContext {
44    /// Process-local parent clock; child traffic and inventory updates never mark it.
45    pub fn mark_parent_activity(&self) {
46        self.0
47            .lock()
48            .expect("turn context lock poisoned")
49            .parent_activity = Some(std::time::Instant::now());
50    }
51
52    pub fn parent_activity(&self) -> Option<std::time::Instant> {
53        self.0
54            .lock()
55            .expect("turn context lock poisoned")
56            .parent_activity
57    }
58
59    pub fn set_decision_log(&self, log: mj_core::jev::DecisionLog) {
60        self.0
61            .lock()
62            .expect("turn context lock poisoned")
63            .decision_log = Some(log);
64    }
65    pub fn decision_log(&self) -> Option<mj_core::jev::DecisionLog> {
66        self.0
67            .lock()
68            .expect("turn context lock poisoned")
69            .decision_log
70            .clone()
71    }
72    pub fn reset(&self, prompt: &str) {
73        let mut state = self.0.lock().expect("turn context lock poisoned");
74        let generation = state.generation.wrapping_add(1);
75        *state = TurnContextState {
76            generation,
77            parent_activity: Some(std::time::Instant::now()),
78            decision_log: state.decision_log.clone(),
79            user_prompt_tail: tail(prompt, USER_PROMPT_BYTES),
80            background_commands: state.background_commands,
81            queued_commands: state.queued_commands,
82            background_inventory: std::mem::take(&mut state.background_inventory),
83            native_agent_ids: std::mem::take(&mut state.native_agent_ids),
84            session_id: std::mem::take(&mut state.session_id),
85            summary: std::mem::take(&mut state.summary),
86            ..Default::default()
87        };
88    }
89
90    /// Record durable transcript observations through the same path during live work and replay.
91    pub fn observe_relay(
92        &self,
93        observation: &mj_core::relay::RelayObservation,
94        prompt: Option<&str>,
95    ) {
96        use mj_core::relay::{RelayCommandOutcome, RelayObservation};
97        match observation {
98            RelayObservation::CommandStarted { .. }
99            | RelayObservation::CommandCompleted {
100                outcome: RelayCommandOutcome::Steered { .. },
101                ..
102            } => {
103                if let Some(prompt) = prompt {
104                    self.reset(prompt);
105                    self.0
106                        .lock()
107                        .expect("turn context lock poisoned")
108                        .summary
109                        .push_user(prompt);
110                }
111            }
112            RelayObservation::SessionUpdate { update } => self.observe(update),
113            RelayObservation::TerminalOutput {
114                terminal_id,
115                output,
116                truncated,
117                exit_code,
118                signal,
119            } => {
120                self.observe_terminal(&mj_core::transcript::TerminalOutputRecord {
121                    terminal_id: terminal_id.clone(),
122                    output: output.clone(),
123                    truncated: *truncated,
124                    exit_code: *exit_code,
125                    signal: signal.clone(),
126                });
127            }
128            RelayObservation::CommandCompleted {
129                outcome: RelayCommandOutcome::ContextCleared { .. },
130                ..
131            } => self.clear_history(),
132            _ => {}
133        }
134    }
135
136    /// Invalidate a pending verdict at a lifecycle boundary without discarding evidence.
137    pub fn invalidate(&self) {
138        let mut state = self.0.lock().expect("turn context lock poisoned");
139        state.generation = state.generation.wrapping_add(1);
140    }
141
142    pub fn generation(&self) -> u64 {
143        self.0
144            .lock()
145            .expect("turn context lock poisoned")
146            .generation
147    }
148
149    pub fn counts(&self) -> (usize, usize) {
150        let state = self.0.lock().expect("turn context lock poisoned");
151        (state.background_commands, state.queued_commands)
152    }
153
154    pub fn set_counts(&self, background_commands: usize, queued_commands: usize) {
155        let mut state = self.0.lock().expect("turn context lock poisoned");
156        if (state.background_commands, state.queued_commands)
157            != (background_commands, queued_commands)
158        {
159            state.generation = state.generation.wrapping_add(1);
160            state.background_commands = background_commands;
161            state.queued_commands = queued_commands;
162        }
163    }
164
165    pub fn set_session_id(&self, session_id: &str) {
166        self.0
167            .lock()
168            .expect("turn context lock poisoned")
169            .session_id = session_id.into();
170    }
171
172    pub fn session_id(&self) -> String {
173        self.0
174            .lock()
175            .expect("turn context lock poisoned")
176            .session_id
177            .clone()
178    }
179
180    /// Compare identities as well as counts; identical level reports are not activity.
181    pub fn set_background_inventory(
182        &self,
183        mut commands: Vec<mj_core::relay::BackgroundCommand>,
184        mut native_agent_ids: Vec<String>,
185    ) {
186        commands.sort_by(|a, b| a.id.cmp(&b.id));
187        native_agent_ids.sort();
188        let mut state = self.0.lock().expect("turn context lock poisoned");
189        if state.background_inventory != commands || state.native_agent_ids != native_agent_ids {
190            state.background_inventory = commands;
191            state.native_agent_ids = native_agent_ids;
192            state.generation = state.generation.wrapping_add(1);
193        }
194    }
195
196    pub fn observe(&self, update: &SessionUpdate) {
197        let mut state = self.0.lock().expect("turn context lock poisoned");
198        state.parent_activity = Some(std::time::Instant::now());
199        if matches!(
200            update,
201            SessionUpdate::AgentMessageChunk(_)
202                | SessionUpdate::AgentThoughtChunk(_)
203                | SessionUpdate::ToolCall(_)
204                | SessionUpdate::ToolCallUpdate(_)
205        ) {
206            state.generation = state.generation.wrapping_add(1);
207        }
208        state.summary.observe(update);
209        match update {
210            SessionUpdate::AgentMessageChunk(chunk) => {
211                let id = chunk.message_id.as_ref().map(ToString::to_string);
212                if id != state.message_id {
213                    complete_message(&mut state);
214                    state.message_id = id;
215                }
216                if let ContentBlock::Text(content) = &chunk.content {
217                    // Trim each chunk before appending, so a huge chunk never grows retained state.
218                    state
219                        .assistant_text_tail
220                        .push_str(&tail(&content.text, ASSISTANT_TEXT_BYTES));
221                    state.assistant_text_tail =
222                        tail(&state.assistant_text_tail, ASSISTANT_TEXT_BYTES);
223                }
224            }
225            SessionUpdate::ToolCall(_) => {
226                complete_message(&mut state);
227            }
228            SessionUpdate::ToolCallUpdate(_) | SessionUpdate::AgentThoughtChunk(_) => {
229                complete_message(&mut state)
230            }
231            _ => {}
232        }
233    }
234
235    pub fn observe_terminal(&self, record: &mj_core::transcript::TerminalOutputRecord) {
236        let mut state = self.0.lock().expect("turn context lock poisoned");
237        state.summary.observe_terminal(record);
238        state.generation = state.generation.wrapping_add(1);
239    }
240
241    pub fn mark_earlier_history_omitted(&self) {
242        self.0
243            .lock()
244            .expect("turn context lock poisoned")
245            .summary
246            .mark_earlier_history_omitted();
247    }
248
249    pub fn clear_history(&self) {
250        let mut state = self.0.lock().expect("turn context lock poisoned");
251        state.summary = Default::default();
252        state.user_prompt_tail.clear();
253        state.assistant_text_tail.clear();
254        state.last_completed_message.clear();
255        state.message_id = None;
256        state.generation = state.generation.wrapping_add(1);
257    }
258
259    pub fn evidence(
260        &self,
261        harness: HarnessKind,
262        phase: TurnPhase,
263        facts: &ActivityFacts,
264        now_ms: i64,
265    ) -> TurnEvidence {
266        let state = self.0.lock().expect("turn context lock poisoned");
267        let summary = state.summary.latest_user_messages();
268        let mut evidence = TurnEvidence {
269            harness,
270            phase,
271            silent_for_s: facts
272                .last_acp_activity_at_ms
273                .map_or(0, |last| now_ms.saturating_sub(last).max(0) as u64 / 1000),
274            tools_in_flight: facts
275                .tools_in_flight
276                .iter()
277                .take(IN_FLIGHT_TOOLS)
278                .map(|tool| ToolEvidence {
279                    title: tail(
280                        &state
281                            .summary
282                            .entries
283                            .iter()
284                            .find(|e| e.id == tool.tool_call_id)
285                            .map(|e| e.text.clone())
286                            .unwrap_or_else(|| "unknown tool".into()),
287                        TOOL_TITLE_BYTES,
288                    ),
289                    running_s: now_ms.saturating_sub(tool.started_at_ms).max(0) as u64 / 1000,
290                })
291                .collect(),
292            transcript_summary: summary.render(48 * 1024),
293            background_commands: facts.background_commands,
294            queued_commands: facts.queued_commands,
295            user_prompt_tail: state.user_prompt_tail.clone(),
296            assistant_text_tail: if state.assistant_text_tail.is_empty() {
297                state.last_completed_message.clone()
298            } else {
299                state.assistant_text_tail.clone()
300            },
301        };
302        let mut limit = 48 * 1024;
303        while serde_json::to_vec(&evidence)
304            .expect("serialize evidence")
305            .len()
306            > 60 * 1024
307        {
308            limit /= 2;
309            evidence.transcript_summary = summary.render(limit);
310        }
311        evidence
312    }
313}
314
315fn complete_message(state: &mut TurnContextState) {
316    if !state.assistant_text_tail.is_empty() {
317        state.last_completed_message = std::mem::take(&mut state.assistant_text_tail);
318    }
319    state.message_id = None;
320}
321
322fn tail(text: &str, maximum_bytes: usize) -> String {
323    let start = text.floor_char_boundary(text.len().saturating_sub(maximum_bytes));
324    let mut value = text[start..].to_owned();
325    mj_core::transcript::truncate_string_start(&mut value, maximum_bytes);
326    value
327}
328
329#[cfg(test)]
330mod tests {
331    use super::*;
332    use mj_core::activity::InFlightToolCall;
333    use serde_json::json;
334    fn message(id: &str, text: &str) -> SessionUpdate {
335        serde_json::from_value(json!({"sessionUpdate":"agent_message_chunk","messageId":id,"content":{"type":"text","text":text}})).unwrap()
336    }
337
338    fn tool(title: &str) -> SessionUpdate {
339        serde_json::from_value(json!({"sessionUpdate":"tool_call","toolCallId":title,"title":title,"status":"in_progress"})).unwrap()
340    }
341
342    #[test]
343    fn diagnostic_logging_does_not_change_evidence_or_generation_and_survives_reset() {
344        let context = TurnContext::default();
345        context.reset("Implement the parser");
346        let generation = context.generation();
347        let before = serde_json::to_value(context.evidence(
348            HarnessKind::Codex,
349            TurnPhase::Running,
350            &Default::default(),
351            0,
352        ))
353        .unwrap();
354        let dir = tempfile::tempdir().unwrap();
355        let log = mj_core::jev::DecisionLog::open(dir.path().into()).unwrap();
356        context.set_decision_log(log.clone());
357        let attempt = log.start("s", "activity", "Who acts?", "Current request");
358        attempt.finish("unchanged", "Kept runtime facts");
359        assert_eq!(context.generation(), generation);
360        assert_eq!(
361            serde_json::to_value(context.evidence(
362                HarnessKind::Codex,
363                TurnPhase::Running,
364                &Default::default(),
365                0
366            ))
367            .unwrap(),
368            before
369        );
370        context.reset("New work");
371        assert!(context.decision_log().is_some());
372    }
373
374    #[test]
375    fn verdict_uses_latest_delivered_user_without_tool_history_but_keeps_live_facts() {
376        use mj_core::relay::RelayObservation;
377        let context = TurnContext::default();
378        let start = |id: &str| RelayObservation::CommandStarted {
379            command_id: id.into(),
380            started_at_ms: 0,
381        };
382        context.observe_relay(&start("old"), Some("OLD REQUEST"));
383        context.observe(&message("old", "OLD ANSWER"));
384        context.observe(&tool("active-build"));
385        context.observe_relay(&start("new"), Some("CURRENT REQUEST"));
386        let noisy: SessionUpdate = serde_json::from_value(json!({
387            "sessionUpdate":"tool_call", "toolCallId":"noisy", "title":"noisy", "status":"completed",
388            "rawInput":{"command":"TOOL_BODY".repeat(20_000)}
389        })).unwrap();
390        context.observe(&noisy);
391        context.observe(&message("new", "CURRENT ANSWER"));
392        let facts = ActivityFacts {
393            background_commands: 2,
394            queued_commands: 3,
395            tools_in_flight: vec![InFlightToolCall {
396                tool_call_id: "active-build".into(),
397                title: Some("active-build".into()),
398                status: agent_client_protocol::schema::v1::ToolCallStatus::InProgress,
399                started_at_ms: 1_000,
400            }],
401            ..Default::default()
402        };
403        let evidence = context.evidence(HarnessKind::Codex, TurnPhase::Running, &facts, 5_000);
404        assert!(evidence.transcript_summary.contains("CURRENT REQUEST"));
405        assert!(evidence.transcript_summary.contains("CURRENT ANSWER"));
406        for excluded in [
407            "OLD REQUEST",
408            "OLD ANSWER",
409            "TOOL_BODY",
410            "<tool",
411            "bytes omitted",
412        ] {
413            assert!(
414                !evidence.transcript_summary.contains(excluded),
415                "{excluded}"
416            );
417        }
418        assert_eq!(evidence.user_prompt_tail, "CURRENT REQUEST");
419        assert_eq!(evidence.assistant_text_tail, "CURRENT ANSWER");
420        assert_eq!(evidence.background_commands, 2);
421        assert_eq!(evidence.queued_commands, 3);
422        assert_eq!(evidence.tools_in_flight.len(), 1);
423        assert!(evidence.tools_in_flight[0].title.contains("active-build"));
424        assert_eq!(evidence.tools_in_flight[0].running_s, 4);
425        let full = context.0.lock().unwrap().summary.render(256 * 1024);
426        assert!(full.contains("OLD REQUEST") && full.contains("TOOL_BODY"));
427    }
428
429    #[test]
430    fn evidence_caps_utf8_text_tools_and_durations() {
431        let context = TurnContext::default();
432        context.reset(&"é".repeat(20_000));
433        context.observe(&message("first", &"🦀".repeat(20_000)));
434        let first = context.evidence(
435            HarnessKind::Claude,
436            TurnPhase::Running,
437            &ActivityFacts::default(),
438            0,
439        );
440        assert_eq!(first.user_prompt_tail.len(), USER_PROMPT_BYTES);
441        assert_eq!(first.assistant_text_tail.len(), ASSISTANT_TEXT_BYTES);
442        for n in 0..20 {
443            context.observe(&tool(&format!("{n}{}", "é".repeat(300))));
444        }
445        let facts = ActivityFacts {
446            last_acp_activity_at_ms: Some(1_000),
447            background_commands: 3,
448            queued_commands: 2,
449            tools_in_flight: (0..100)
450                .map(|n| InFlightToolCall {
451                    tool_call_id: n.to_string(),
452                    title: Some("🦀".repeat(200)),
453                    status: agent_client_protocol::schema::v1::ToolCallStatus::InProgress,
454                    started_at_ms: 2_000,
455                })
456                .collect(),
457            ..Default::default()
458        };
459        let evidence = context.evidence(HarnessKind::Claude, TurnPhase::Running, &facts, 95_000);
460        assert_eq!(evidence.tools_in_flight.len(), IN_FLIGHT_TOOLS);
461        assert!(
462            evidence
463                .tools_in_flight
464                .iter()
465                .all(|tool| tool.title.len() <= TOOL_TITLE_BYTES && tool.running_s == 93)
466        );
467        assert!(evidence.transcript_summary.len() <= 48 * 1024);
468        assert_eq!(evidence.assistant_text_tail, first.assistant_text_tail);
469        assert_eq!(evidence.silent_for_s, 94);
470        let json = serde_json::to_value(evidence).unwrap();
471        assert_eq!(json["phase"], "running");
472        assert_eq!(json["harness"], "claude");
473        assert_eq!(json["background_commands"], 3);
474        assert_eq!(json["queued_commands"], 2);
475        assert!(serde_json::to_vec(&json).unwrap().len() < 64 * 1024);
476    }
477
478    #[test]
479    fn message_boundaries_and_prompt_reset_do_not_mix_answers() {
480        let context = TurnContext::default();
481        context.reset("First prompt");
482        context.observe(&message("one", "Old answer"));
483        context.observe(&message("two", "New "));
484        context.observe(&message("two", "answer?"));
485        let evidence = context.evidence(
486            HarnessKind::Claude,
487            TurnPhase::Replied,
488            &ActivityFacts::default(),
489            0,
490        );
491        assert_eq!(evidence.assistant_text_tail, "New answer?");
492        let generation = context.generation();
493        context.set_counts(1, 2);
494        assert_ne!(context.generation(), generation);
495        let generation = context.generation();
496        context.set_counts(1, 2);
497        assert_eq!(context.generation(), generation);
498        assert_eq!(context.counts(), (1, 2));
499        context.invalidate();
500        assert_ne!(context.generation(), generation);
501        assert_eq!(context.counts(), (1, 2));
502        let generation = context.generation();
503        context.reset("Second prompt");
504        assert_eq!(context.counts(), (1, 2));
505        assert_ne!(context.generation(), generation);
506        let evidence = context.evidence(
507            HarnessKind::Claude,
508            TurnPhase::Running,
509            &ActivityFacts::default(),
510            0,
511        );
512        assert_eq!(evidence.user_prompt_tail, "Second prompt");
513        assert!(evidence.assistant_text_tail.is_empty());
514        assert!(!evidence.transcript_summary.contains("<user>"));
515    }
516}
mj_transcript/turn_context.rs

mj_transcript/
turn_context.rs
