Add task context to shadow skill selection (#39008)

## Why

Short continuation prompts such as `continue` do not contain enough context for
skill selection on their own.

## What changed

- Add the `task_context_fusion_v1` shadow selector, combining the current request
  with up to two prior substantive requests and recently relevant skills.
- Record explicit skill intent and successful skill invocations for future turns,
  while excluding same-turn observations from predictions.
- Bound retained requests, augmented queries, and skill history, including safe
  truncation at UTF-8 character boundaries.

## Testing

Add unit and extension tests for continuation prompts, explicit intent, turn
isolation, cold thread state, bounded history, and unchanged control selectors.

GitOrigin-RevId: 72eca3f0d64620a0d24e95d5675f126ac982f8c7

codex-rs/ext/skills/src/extension.rs

@@ -395,11 +395,12 @@ where
                 let shadow_selected_entries =
                     collect_explicit_skill_mentions(&input.user_input, &shadow_catalog);
                 Some(self.shadow_selection.run(
-                    &input.user_input,
+                    &input,
                     &shadow_catalog,
                     &shadow_selected_entries,
                     host_snapshot.as_deref(),
                     Arc::clone(&thread_state.recent_skill_invocations),
+                    Arc::clone(&thread_state.shadow_task_context),
                 ))
             } else {
                 None

codex-rs/ext/skills/src/shadow_selection_experiment/mod.rs

@@ -1,5 +1,9 @@
 // This shadow-selection experiment is temporary and should be removed after evaluation.
 
+mod task_context;
+
+pub(crate) use task_context::ShadowTaskContext;
+
 use std::collections::HashMap;
 use std::collections::HashSet;
 use std::collections::VecDeque;
@@ -10,6 +14,7 @@ use std::time::Duration;
 use std::time::Instant;
 
 use crate::HostSkillsSnapshot;
+use codex_extension_api::TurnInputContext;
 use codex_otel::MetricsClient;
 use codex_protocol::user_input::UserInput;
 
@@ -64,14 +69,16 @@ impl ShadowSelectionExperiment {
 
     pub(crate) fn run(
         &self,
-        inputs: &[UserInput],
+        input: &TurnInputContext,
         catalog: &SkillCatalog,
         explicitly_selected: &[SkillCatalogEntry],
         host_snapshot: Option<&HostSkillsSnapshot>,
         recent_skill_invocations: Arc<RecentSkillInvocations>,
+        task_context: Arc<ShadowTaskContext>,
     ) -> ShadowSelectionTurnState {
-        let query = build_shadow_query(inputs);
+        let query = build_shadow_query(&input.user_input);
         let query_script = query_script_tag(&query.text);
+        let task_snapshot = task_context.begin_turn(&input.turn_id, &query, &input.user_input);
         let explicitly_selected_skill_resources = explicitly_selected
             .iter()
             .map(|entry| normalize_skill_resource(entry.main_prompt.as_str()))
@@ -132,9 +139,19 @@ impl ShadowSelectionExperiment {
             lru_selector.clone(),
             routing_selector.clone(),
         );
-        let mut ranked_selections = Vec::with_capacity(self.selectors.len() + 5);
+        let task_selector = LruPlusLexicalCharacterRoutingSkillSelector::new(
+            LruSkillSelector::new(
+                task_snapshot
+                    .recent_skills
+                    .iter()
+                    .filter_map(|resource| eligible_skill_ids_by_resource.get(resource).copied())
+                    .collect(),
+            ),
+            routing_selector.clone(),
+        );
+        let mut ranked_selections = Vec::with_capacity(self.selectors.len() + 6);
 
-        for selector in self
+        for (method, selector, query) in self
             .selectors
             .iter()
             .map(std::convert::AsRef::as_ref)
@@ -145,14 +162,21 @@ impl ShadowSelectionExperiment {
                 &lru_plus_character_selector as &dyn CheapSkillSelector,
                 &lru_plus_lexical_character_selector as &dyn CheapSkillSelector,
             ])
+            .map(|selector| (selector.method(), selector, &query))
+            .chain([(
+                "task_context_fusion_v1",
+                &task_selector as &dyn CheapSkillSelector,
+                &task_snapshot.query,
+            )])
         {
+            let query_script = query_script_tag(&query.text);
             let start = Instant::now();
             let selection =
                 selector.select(&query.text, &documents, /*limit*/ MAX_SHADOW_RESULTS);
             let duration = start.elapsed();
             let selected_ids = sanitize_selected_ids(&selection, &eligible_ids);
             self.record_metrics(ShadowSelectionObservation {
-                method: selector.method(),
+                method,
                 selection: &selection,
                 query_truncated_before_selection: query.truncated,
                 query_script,
@@ -161,14 +185,14 @@ impl ShadowSelectionExperiment {
                 duration,
             });
             ranked_selections.push(RankedSelection {
-                method: selector.method(),
+                method,
                 skill_resources: selected_ids
                     .iter()
                     .map(|id| normalize_skill_resource(catalog.entries[*id].main_prompt.as_str()))
                     .collect(),
             });
             tracing::debug!(
-                method = selector.method(),
+                method,
                 catalog_entries = documents.len(),
                 selected_entries = selected_ids.len(),
                 query_terms = selection.query_term_count,
@@ -179,12 +203,29 @@ impl ShadowSelectionExperiment {
             );
         }
 
+        // Explicit intent is a relevance signal even if the subsequent prompt read fails.
+        // Keep it out of the implicit-only controls and freeze predictions before recording it.
+        for entry in explicitly_selected.iter().filter(|entry| {
+            entry.is_model_visible()
+                && matches!(
+                    &entry.authority.kind,
+                    SkillSourceKind::Host | SkillSourceKind::Orchestrator
+                )
+        }) {
+            task_context.record(
+                &input.turn_id,
+                normalize_skill_resource(entry.main_prompt.as_str()),
+            );
+        }
+
         ShadowSelectionTurnState {
             ranked_selections,
+            turn_id: input.turn_id.clone(),
             query_script,
             eligible_skill_resources,
             seen_skill_resources: Mutex::new(HashSet::new()),
             recent_skill_invocations,
+            task_context,
         }
     }
 
@@ -204,6 +245,9 @@ impl ShadowSelectionExperiment {
         state
             .recent_skill_invocations
             .record(skill_resource.clone());
+        state
+            .task_context
+            .record(&state.turn_id, skill_resource.clone());
         let Some(metrics_client) = self.metrics_client.as_ref() else {
             return;
         };
@@ -274,10 +318,12 @@ impl ShadowSelectionExperiment {
 
 pub(crate) struct ShadowSelectionTurnState {
     ranked_selections: Vec<RankedSelection>,
+    turn_id: String,
     query_script: &'static str,
     eligible_skill_resources: HashSet<String>,
     seen_skill_resources: Mutex<HashSet<String>>,
     recent_skill_invocations: Arc<RecentSkillInvocations>,
+    task_context: Arc<ShadowTaskContext>,
 }
 
 #[derive(Default)]
@@ -435,6 +481,7 @@ fn is_cjk(character: char) -> bool {
     )
 }
 
+#[derive(Clone, Debug, PartialEq, Eq)]
 struct ShadowQuery {
     text: String,
     truncated: bool,
@@ -479,5 +526,5 @@ fn push_bounded(destination: &mut String, value: &str) -> bool {
 }
 
 #[cfg(test)]
-#[path = "shadow_selection_experiment_tests.rs"]
+#[path = "experiment_tests.rs"]
 mod tests;

codex-rs/ext/skills/src/shadow_selection_experiment/task_context.rs

@@ -0,0 +1,155 @@
+use std::collections::VecDeque;
+use std::sync::Mutex;
+use std::sync::PoisonError;
+
+use codex_protocol::user_input::UserInput;
+use codex_utils_string::take_bytes_at_char_boundary;
+
+use super::ShadowQuery;
+
+const MAX_PRIOR_REQUESTS: usize = 2;
+const MAX_REQUEST_BYTES: usize = 2 * 1024;
+const MAX_QUERY_BYTES: usize = 4 * 1024;
+const MAX_RECENT_SKILLS: usize = 50;
+
+/// Shadow-only relevance history. New and reconstructed thread runtimes start cold.
+#[derive(Default)]
+pub(crate) struct ShadowTaskContext(Mutex<TaskContextState>);
+
+#[derive(Default)]
+struct TaskContextState {
+    prior_requests: VecDeque<ShadowQuery>,
+    recent_skills: VecDeque<String>,
+    pending: Option<PendingTurn>,
+}
+
+struct PendingTurn {
+    id: String,
+    request: Option<ShadowQuery>,
+    recent_skills: VecDeque<String>,
+}
+
+#[derive(Debug, PartialEq, Eq)]
+pub(super) struct TaskContextSnapshot {
+    pub(super) query: ShadowQuery,
+    pub(super) recent_skills: Vec<String>,
+}
+
+impl ShadowTaskContext {
+    pub(super) fn begin_turn(
+        &self,
+        turn_id: &str,
+        current: &ShadowQuery,
+        inputs: &[UserInput],
+    ) -> TaskContextSnapshot {
+        let mut state = self.0.lock().unwrap_or_else(PoisonError::into_inner);
+        if state.pending.as_ref().is_none_or(|turn| turn.id != turn_id) {
+            if let Some(previous) = state.pending.take() {
+                if let Some(request) = previous.request {
+                    state
+                        .prior_requests
+                        .retain(|prior| prior.text != request.text);
+                    state.prior_requests.push_front(request);
+                    state.prior_requests.truncate(MAX_PRIOR_REQUESTS);
+                }
+                for resource in previous.recent_skills.into_iter().rev() {
+                    remember_skill(&mut state.recent_skills, resource);
+                }
+            }
+            state.pending = Some(PendingTurn {
+                id: turn_id.to_string(),
+                request: None,
+                recent_skills: VecDeque::new(),
+            });
+        }
+
+        let retained = take_bytes_at_char_boundary(&current.text, MAX_REQUEST_BYTES);
+        let text = take_bytes_at_char_boundary(&current.text, MAX_QUERY_BYTES);
+        let mut query = ShadowQuery {
+            text: text.to_string(),
+            truncated: current.truncated || text.len() < current.text.len(),
+        };
+        for prior in &state.prior_requests {
+            if prior.text == retained {
+                continue;
+            }
+            if !query.text.is_empty() && query.text.len() < MAX_QUERY_BYTES {
+                query.text.push('\n');
+            }
+            let part = take_bytes_at_char_boundary(
+                &prior.text,
+                MAX_QUERY_BYTES.saturating_sub(query.text.len()),
+            );
+            query.text.push_str(part);
+            query.truncated |= prior.truncated || part.len() < prior.text.len();
+            if part.len() < prior.text.len() {
+                break;
+            }
+        }
+        let recent_skills = state.recent_skills.iter().cloned().collect();
+        if is_substantive(&current.text, inputs)
+            && let Some(pending) = state.pending.as_mut()
+        {
+            pending.request = Some(ShadowQuery {
+                text: retained.to_string(),
+                truncated: current.truncated || retained.len() < current.text.len(),
+            });
+        }
+        TaskContextSnapshot {
+            query,
+            recent_skills,
+        }
+    }
+
+    /// Records relevance evidence for future turns, never the active turn's predictions.
+    pub(super) fn record(&self, turn_id: &str, resource: String) {
+        let mut state = self.0.lock().unwrap_or_else(PoisonError::into_inner);
+        if let Some(pending) = state.pending.as_mut().filter(|turn| turn.id == turn_id) {
+            remember_skill(&mut pending.recent_skills, resource);
+        }
+    }
+}
+
+fn remember_skill(skills: &mut VecDeque<String>, resource: String) {
+    skills.retain(|previous| previous != &resource);
+    skills.push_front(resource);
+    skills.truncate(MAX_RECENT_SKILLS);
+}
+
+fn is_substantive(text: &str, inputs: &[UserInput]) -> bool {
+    if inputs.iter().any(|input| {
+        matches!(input, UserInput::Skill { name, .. } | UserInput::Mention { name, .. } if !name.trim().is_empty())
+    }) {
+        return true;
+    }
+    let normalized = text
+        .split(|character: char| !character.is_alphanumeric())
+        .filter(|part| !part.is_empty())
+        .map(str::to_lowercase)
+        .collect::<Vec<_>>()
+        .join(" ");
+    !matches!(
+        normalized.as_str(),
+        "" | "yes"
+            | "yep"
+            | "yeah"
+            | "ok"
+            | "okay"
+            | "sure"
+            | "go"
+            | "go ahead"
+            | "continue"
+            | "please continue"
+            | "proceed"
+            | "do it"
+            | "do that"
+            | "try again"
+            | "retry"
+            | "thanks"
+            | "thank you"
+    )
+}
+
+#[cfg(test)]
+#[path = "task_context_tests.rs"]
+mod tests;

codex-rs/ext/skills/src/shadow_selection_experiment/task_context_tests.rs

@@ -0,0 +1,112 @@
+use pretty_assertions::assert_eq;
+
+use super::*;
+
+fn query(text: &str) -> ShadowQuery {
+    ShadowQuery {
+        text: text.to_string(),
+        truncated: false,
+    }
+}
+
+fn begin(context: &ShadowTaskContext, turn_id: &str, text: &str) -> TaskContextSnapshot {
+    context.begin_turn(turn_id, &query(text), &[])
+}
+
+#[test]
+fn prior_requests_survive_continuations_and_keep_short_substantive_input() {
+    let context = ShadowTaskContext::default();
+    begin(&context, "1", "Fix lint errors.");
+    assert_eq!(
+        query("continue\nFix lint errors."),
+        begin(&context, "2", "continue").query
+    );
+    begin(&context, "3", "OK!!!");
+    begin(&context, "4", "deploy");
+    begin(&context, "5", "修复测试");
+    assert_eq!(
+        query("continue\n修复测试\ndeploy"),
+        begin(&context, "6", "continue").query
+    );
+    assert_eq!(
+        query("deploy\n修复测试"),
+        begin(&context, "7", "deploy").query
+    );
+    assert!(is_substantive(
+        "go",
+        &[UserInput::Mention {
+            name: "go".to_string(),
+            path: "skill://go".to_string(),
+        }]
+    ));
+}
+
+#[test]
+fn current_turn_observations_are_pending_and_late_observations_are_ignored() {
+    let context = ShadowTaskContext::default();
+    begin(&context, "1", "continue");
+    context.record("1", "a".to_string());
+    assert_eq!(
+        Vec::<String>::new(),
+        begin(&context, "1", "continue").recent_skills
+    );
+    assert_eq!(vec!["a"], begin(&context, "2", "continue").recent_skills);
+    context.record("1", "late".to_string());
+    context.record("2", "a".to_string());
+    context.record("2", "b".to_string());
+    assert_eq!(vec!["a"], begin(&context, "2", "continue").recent_skills);
+    assert_eq!(
+        TaskContextSnapshot {
+            query: query("continue"),
+            recent_skills: vec!["b".to_string(), "a".to_string()],
+        },
+        begin(&context, "3", "continue")
+    );
+    assert_eq!(
+        TaskContextSnapshot {
+            query: query("continue"),
+            recent_skills: Vec::new(),
+        },
+        begin(&ShadowTaskContext::default(), "3", "continue")
+    );
+}
+
+#[test]
+fn history_and_augmented_query_are_bounded_at_utf8_boundaries() {
+    let context = ShadowTaskContext::default();
+    let first = "界".repeat(MAX_REQUEST_BYTES);
+    let initial = begin(&context, "1", &first);
+    assert_eq!(
+        ShadowQuery {
+            text: take_bytes_at_char_boundary(&first, MAX_QUERY_BYTES).to_string(),
+            truncated: true,
+        },
+        initial.query
+    );
+    for index in 0..=MAX_RECENT_SKILLS {
+        context.record("1", format!("skill-{index}"));
+    }
+    begin(&context, "2", &"é".repeat(MAX_REQUEST_BYTES));
+    context.record("2", "skill-1".to_string());
+    let expected = format!(
+        "continue\n{}\n{}",
+        "é".repeat(MAX_REQUEST_BYTES / 2),
+        take_bytes_at_char_boundary(&first, MAX_REQUEST_BYTES)
+    );
+    assert_eq!(
+        TaskContextSnapshot {
+            query: ShadowQuery {
+                text: take_bytes_at_char_boundary(&expected, MAX_QUERY_BYTES).to_string(),
+                truncated: true,
+            },
+            recent_skills: std::iter::once("skill-1".to_string())
+                .chain(
+                    (2..=MAX_RECENT_SKILLS)
+                        .rev()
+                        .map(|index| format!("skill-{index}"))
+                )
+                .collect(),
+        },
+        begin(&context, "3", "continue")
+    );
+}

codex-rs/ext/skills/src/state.rs

@@ -23,6 +23,7 @@ use crate::provider::SkillListQuery;
 use crate::provider::SkillReadRequest;
 use crate::shadow_selection_experiment::RecentSkillInvocations;
 use crate::shadow_selection_experiment::ShadowSelectionTurnState;
+use crate::shadow_selection_experiment::ShadowTaskContext;
 use crate::sources::SkillProviders;
 
 const MAX_CACHED_ORCHESTRATOR_RESOURCES: usize = 100;
@@ -41,6 +42,7 @@ pub(crate) struct SkillsThreadState {
     orchestrator_cache: Mutex<Option<Arc<OrchestratorGenerationCache>>>,
     shadow_selection_turn: Mutex<Option<ShadowSelectionTurn>>,
     pub(crate) recent_skill_invocations: Arc<RecentSkillInvocations>,
+    pub(crate) shadow_task_context: Arc<ShadowTaskContext>,
 }
 
 impl SkillsThreadState {
@@ -53,6 +55,7 @@ impl SkillsThreadState {
             orchestrator_cache: Mutex::new(None),
             shadow_selection_turn: Mutex::new(None),
             recent_skill_invocations: Arc::new(RecentSkillInvocations::default()),
+            shadow_task_context: Arc::new(ShadowTaskContext::default()),
         }
     }