memory: implement flat extract staging

.yoi/tickets/00001KXRM6G0G/item.md

@@ -1,14 +1,16 @@
 ---
 title: 'Flat candidate staging record schemaを実装する'
-state: 'ready'
+state: 'closed'
 created_at: '2026-07-17T18:07:40Z'
-updated_at: '2026-07-17T18:10:00Z'
+updated_at: '2026-07-17T19:56:08Z'
 assignee: null
+queued_by: 'yoi ticket'
+queued_at: '2026-07-17T19:48:08Z'
 ---
 
 ## 背景
 
-旧 staging は extract run ごとの batch payload (`decisions` / `discussions` / `attempts` / `requests`) と entry-level source refs を前提にしていた。しかし、consolidation から見ると batch は session に紐いた不必要な集団になり、record ごとの discard / merge / promote / defer が曖昧になる。
+旧 staging は extract run ごとの batch payload (`decisions` / `discussions` / `attempts` / `requests`) と entry-level source refs を前提にしていた。しかし、consolidation から見ると batch は session に紐づいた不必要な集団になり、record ごとの discard / merge / promote / defer が曖昧になる。
 
 新方針では、extract は記憶化を検討すべき candidate を kind taxonomy に基づいて切り出し、candidate ごとに flat staging record を作る。
 

.yoi/tickets/00001KXRM6G0G/resolution.md

@@ -0,0 +1,9 @@
+Implemented flat candidate staging record schema.
+
+The memory extract staging path now treats staging as flat candidate records: one extract run can produce zero or more staging files, and each staging file is one consolidation decision unit. The transitional `write_extracted` path now accepts `candidates[]`, and the host expands each candidate into a flat `StagingRecord` with schema version, ids, source, kind, claim, usefulness, optional staleness, bounded evidence, and source refs.
+
+Validation passed:
+- `cargo fmt --check`
+- `cargo test -p memory`
+- `cargo test -p worker`
+- `nix build .#yoi`

.yoi/tickets/00001KXRM6G0G/thread.md

@@ -13,4 +13,80 @@ LocalTicketBackend によって作成されました。
 Created to replace the previous batch staging direction. This Ticket owns the new flat staging record schema: one candidate per staging record, with kinds `preference`, `working_assumption`, `constraint`, `decision`, `open_question`, and `lesson`.
 
 
+---
+
+<!-- event: state_changed author: "yoi ticket" at: 2026-07-17T19:48:08Z from: ready to: queued reason: queued field: state -->
+
+## State changed
+
+Ticket を `yoi ticket` が queued にしました。
+
+
+---
+
+<!-- event: state_changed author: "yoi ticket" at: 2026-07-17T19:48:08Z from: queued to: inprogress reason: cli_state field: state -->
+
+## State changed
+
+State changed to `inprogress`.
+
+
+---
+
+<!-- event: implementation_report author: hare at: 2026-07-17T19:56:08Z -->
+
+## Implementation report
+
+Implemented flat candidate staging record schema.
+
+Changes:
+- Replaced extract payload with `candidates[]` based on narrow candidate kinds: `preference`, `working_assumption`, `constraint`, `decision`, `open_question`, `lesson`.
+- Added flat `StagingRecord` schema with `schema_version`, `id`, `extract_run_id`, record-level `source`, `kind`, `claim`, `why_useful`, optional `staleness`, bounded `evidence[]`, and `source_refs[]`.
+- Changed staging writer to write one JSON file per candidate. Empty payload writes no files.
+- Updated transitional `write_extracted` tool and extract prompt to candidate extraction instead of activity-log extraction.
+- Updated consolidation staging/input paths and tests to read/render flat records.
+- Updated worker extract path to handle multiple staged candidate records per extract run.
+
+Validation:
+- `cargo fmt --check`
+- `cargo test -p memory`
+- `cargo test -p worker`
+- `nix build .#yoi`
+
+
+---
+
+<!-- event: state_changed author: "yoi ticket" at: 2026-07-17T19:56:08Z from: inprogress to: done reason: cli_state field: state -->
+
+## State changed
+
+State changed to `done`.
+
+
+---
+
+<!-- event: state_changed author: hare at: 2026-07-17T19:56:08Z from: done to: closed reason: closed field: state -->
+
+## State changed
+
+Ticket を closed にしました。
+
+
+---
+
+<!-- event: close author: hare at: 2026-07-17T19:56:08Z status: closed -->
+
+## 完了
+
+Implemented flat candidate staging record schema.
+
+The memory extract staging path now treats staging as flat candidate records: one extract run can produce zero or more staging files, and each staging file is one consolidation decision unit. The transitional `write_extracted` path now accepts `candidates[]`, and the host expands each candidate into a flat `StagingRecord` with schema version, ids, source, kind, claim, usefulness, optional staleness, bounded evidence, and source refs.
+
+Validation passed:
+- `cargo fmt --check`
+- `cargo test -p memory`
+- `cargo test -p worker`
+- `nix build .#yoi`
+
+
 ---

crates/memory/src/consolidate/input.rs

@@ -206,7 +206,10 @@ pub fn render_tidy_hints(tidy: &TidyHints) -> String {
 mod tests {
     use super::*;
     use crate::consolidate::tidy::{SimilarSlugCluster, SourcesOverflow};
-    use crate::extract::{DecisionEntry, ExtractedPayload, write_staging};
+    use crate::extract::{
+        CandidateKind, ExtractedCandidate, ExtractedPayload, STAGING_SCHEMA_VERSION,
+        StagingEvidence, StagingRecord, write_staging,
+    };
     use crate::schema::{EvidenceKind, SourceEvidenceRef, SourceRef};
     use chrono::Utc;
     use std::path::Path;
@@ -222,6 +225,18 @@ mod tests {
         std::fs::write(p, content).unwrap();
     }
 
+    fn candidate_payload(kind: CandidateKind, claim: &str) -> ExtractedPayload {
+        ExtractedPayload {
+            candidates: vec![ExtractedCandidate {
+                kind,
+                claim: claim.into(),
+                why_useful: "useful for consolidation".into(),
+                staleness: None,
+                evidence_ids: Vec::new(),
+            }],
+        }
+    }
+
     #[test]
     fn build_includes_all_sections_when_populated() {
         let dir = tempfile::TempDir::new().unwrap();
@@ -238,13 +253,13 @@ mod tests {
                 n = now()
             ),
         );
-        let (_id, _) = write_staging(
+        let _written = write_staging(
             &layout,
             SourceRef {
                 segment_id: "s".into(),
                 range: [0, 1],
             },
-            ExtractedPayload::default(),
+            candidate_payload(CandidateKind::Preference, "Prefer concise tickets"),
         )
         .unwrap();
         let staging = crate::consolidate::staging::list_staging_entries(&layout);
@@ -281,17 +296,27 @@ mod tests {
     fn staging_render_preserves_entry_source_refs() {
         let dir = tempfile::TempDir::new().unwrap();
         let layout = WorkspaceLayout::new(dir.path().to_path_buf());
-        let (_id, _) = write_staging(
+        std::fs::create_dir_all(layout.staging_dir()).unwrap();
-            &layout,
+        let id = uuid::Uuid::now_v7();
-            SourceRef {
+        let record = StagingRecord {
+            schema_version: STAGING_SCHEMA_VERSION,
+            id: id.to_string(),
+            extract_run_id: "run-1".into(),
+            source: SourceRef {
                 segment_id: "segment-record".into(),
                 range: [0, 10],
             },
-            ExtractedPayload {
+            kind: CandidateKind::Decision,
-                decisions: vec![DecisionEntry {
+            claim: "Keep flat staging records".into(),
-                    options: vec!["preserve".into()],
+            why_useful: "consolidation input is lossless JSON".into(),
-                    chosen: "preserve".into(),
+            staleness: None,
-                    rationale: "consolidation input is lossless JSON".into(),
+            evidence: vec![StagingEvidence {
+                id: "ev-1".into(),
+                kind: EvidenceKind::new(EvidenceKind::MESSAGE),
+                entry_range: Some([3, 4]),
+                excerpt: Some("bounded excerpt".into()),
+                summary: Some("bounded summary".into()),
+            }],
             source_refs: vec![SourceEvidenceRef {
                 session_id: Some("session-1".into()),
                 segment_id: Some("segment-1".into()),
@@ -301,9 +326,10 @@ mod tests {
                 label: Some("user message".into()),
                 summary: Some("bounded summary".into()),
             }],
-                }],
+        };
-                ..Default::default()
+        std::fs::write(
-            },
+            layout.staging_dir().join(format!("{id}.json")),
+            serde_json::to_string_pretty(&record).unwrap(),
         )
         .unwrap();
         let staging = crate::consolidate::staging::list_staging_entries(&layout);

crates/memory/src/consolidate/lock.rs

@@ -200,7 +200,7 @@ fn pid_is_alive(_pid: u32) -> bool {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::extract::{ExtractedPayload, write_staging};
+    use crate::extract::{CandidateKind, ExtractedCandidate, ExtractedPayload, write_staging};
     use crate::schema::SourceRef;
 
     fn make_layout() -> (tempfile::TempDir, WorkspaceLayout) {
@@ -210,6 +210,18 @@ mod tests {
         (dir, layout)
     }
 
+    fn candidate_payload(claim: &str) -> ExtractedPayload {
+        ExtractedPayload {
+            candidates: vec![ExtractedCandidate {
+                kind: CandidateKind::Lesson,
+                claim: claim.into(),
+                why_useful: "useful for test".into(),
+                staleness: None,
+                evidence_ids: Vec::new(),
+            }],
+        }
+    }
+
     #[test]
     fn acquire_writes_lock_file() {
         let (_dir, layout) = make_layout();
@@ -257,24 +269,28 @@ mod tests {
     #[test]
     fn release_drops_consumed_entries_and_unlinks_lock() {
         let (_dir, layout) = make_layout();
-        let (id_a, _) = write_staging(
+        let id_a = write_staging(
             &layout,
             SourceRef {
                 segment_id: "s".into(),
                 range: [0, 0],
             },
-            ExtractedPayload::default(),
+            candidate_payload("a"),
         )
-        .unwrap();
+        .unwrap()
-        let (id_b, _) = write_staging(
+        .remove(0)
+        .id;
+        let id_b = write_staging(
             &layout,
             SourceRef {
                 segment_id: "s".into(),
                 range: [1, 1],
             },
-            ExtractedPayload::default(),
+            candidate_payload("b"),
         )
-        .unwrap();
+        .unwrap()
+        .remove(0)
+        .id;
 
         let lock = StagingLock::acquire(&layout, std::process::id(), "worker", vec![id_a]).unwrap();
         let lock_path = lock.path().to_path_buf();

crates/memory/src/consolidate/staging.rs

@@ -116,11 +116,19 @@ pub fn list_staging_entries_snapshot(layout: &WorkspaceLayout) -> StagingEntries
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::extract::{ExtractedPayload, write_staging};
+    use crate::extract::{CandidateKind, ExtractedCandidate, ExtractedPayload, write_staging};
     use crate::schema::SourceRef;
 
-    fn empty_payload() -> ExtractedPayload {
+    fn candidate_payload(claim: &str) -> ExtractedPayload {
-        ExtractedPayload::default()
+        ExtractedPayload {
+            candidates: vec![ExtractedCandidate {
+                kind: CandidateKind::Lesson,
+                claim: claim.into(),
+                why_useful: "useful for later consolidation".into(),
+                staleness: None,
+                evidence_ids: Vec::new(),
+            }],
+        }
     }
 
     fn source(segment_id: &str, range: [u64; 2]) -> SourceRef {
@@ -135,9 +143,18 @@ mod tests {
         let tmp = tempfile::TempDir::new().unwrap();
         let layout = WorkspaceLayout::new(tmp.path().to_path_buf());
 
-        let (id1, _) = write_staging(&layout, source("s", [0, 1]), empty_payload()).unwrap();
+        let id1 = write_staging(&layout, source("s", [0, 1]), candidate_payload("one"))
-        let (id2, _) = write_staging(&layout, source("s", [2, 3]), empty_payload()).unwrap();
+            .unwrap()
-        let (id3, _) = write_staging(&layout, source("s", [4, 5]), empty_payload()).unwrap();
+            .remove(0)
+            .id;
+        let id2 = write_staging(&layout, source("s", [2, 3]), candidate_payload("two"))
+            .unwrap()
+            .remove(0)
+            .id;
+        let id3 = write_staging(&layout, source("s", [4, 5]), candidate_payload("three"))
+            .unwrap()
+            .remove(0)
+            .id;
 
         let entries = list_staging_entries(&layout);
         let ids: Vec<Uuid> = entries.iter().map(|e| e.id).collect();
@@ -148,13 +165,15 @@ mod tests {
     fn skips_lock_file_and_counts_invalid_json() {
         let tmp = tempfile::TempDir::new().unwrap();
         let layout = WorkspaceLayout::new(tmp.path().to_path_buf());
-        let (_id, _) = write_staging(&layout, source("s", [0, 1]), empty_payload()).unwrap();
+        let _id = write_staging(&layout, source("s", [0, 1]), candidate_payload("kept"))
+            .unwrap()
+            .remove(0)
+            .id;
 
-        // Drop a non-UUID json file, an unparsable UUID-named json file, a
+        // Drop a non-UUID json file, an unparsable UUID-named json file, an
-        // legacy source UUID-named json file, and a bare lock file alongside.
+        // old batch-schema UUID-named json file, and a bare lock file alongside.
         // Lock files are not `.json`; invalid `.json` files are surfaced
         // separately instead of being mistaken for an empty staging directory.
-        // Legacy `source.session_id` staging remains readable for compatibility.
         std::fs::write(layout.staging_dir().join("not-a-uuid.json"), "{}").unwrap();
         let bad_id = Uuid::now_v7();
         std::fs::write(layout.staging_dir().join(format!("{bad_id}.json")), "{").unwrap();
@@ -174,17 +193,12 @@ mod tests {
         std::fs::write(layout.staging_dir().join(".consolidation.lock"), "{}").unwrap();
 
         let entries = list_staging_entries(&layout);
-        assert_eq!(entries.len(), 2);
+        assert_eq!(entries.len(), 1);
 
         let snapshot = list_staging_entries_snapshot(&layout);
-        assert_eq!(snapshot.entries.len(), 2);
+        assert_eq!(snapshot.entries.len(), 1);
-        assert_eq!(snapshot.invalid_count, 2);
+        assert_eq!(snapshot.invalid_count, 3);
-        assert!(
+        assert_eq!(snapshot.entries[0].record.claim, "kept");
-            snapshot
-                .entries
-                .iter()
-                .any(|entry| entry.record.source.segment_id == "legacy-session")
-        );
     }
 
     #[test]

crates/memory/src/extract/input.rs

@@ -11,8 +11,8 @@ use llm_engine::Item;
 pub fn build_extract_input(items: &[Item]) -> String {
     let mut out = String::new();
     out.push_str(
-        "Extract activity logs from the conversation slice below. \
+        "Extract memory candidates from the conversation slice below. \
-         Follow the system prompt's schema strictly and call `write_extracted` once.\n\n",
+         Follow the system prompt's candidate schema strictly and call `write_extracted` once.\n\n",
     );
     out.push_str("## Conversation slice\n");
     out.push_str(&render_items(items));

crates/memory/src/extract/mod.rs

@@ -1,4 +1,4 @@
-//! extract: 活動抽出。
+//! extract: memory candidate extraction.
 //!
 //! 通常 Worker の post-run hook で発火する disposable Engine と、その
 //! 出力を `<workspace>/.yoi/memory/_staging/<id>.json` に書き出す
@@ -24,10 +24,11 @@ mod tool;
 
 pub use input::build_extract_input;
 pub use payload::{
-    AttemptEntry, DecisionEntry, DiscussionEntry, ExtractedPayload, RequestEntry, StagingRecord,
+    CandidateKind, ExtractedCandidate, ExtractedPayload, STAGING_SCHEMA_VERSION, StagingEvidence,
+    StagingRecord,
 };
 pub use pointer::{ExtractPointerPayload, fold_pointer};
-pub use staging::{StagingError, write_staging};
+pub use staging::{StagingWriteResult, write_staging};
 pub use tool::{ExtractWorkerContext, write_extracted_tool};
 
 /// session-store `LogEntry::Extension` で使う domain 名。

crates/memory/src/extract/payload.rs

@@ -1,254 +1,210 @@
-//! extract 抽出の出力 schema。
+//! extract staging schema.
 //!
-//! LLM は [`ExtractedPayload`] そのもの（record-level source 抜き）を返し、Worker 側
+//! Extract produces memory-candidate records, not activity-log batches. During
-//! ラッパーが [`StagingRecord`] に組み立てて staging へ書き出す。
+//! the transitional `write_extracted` path the model submits an
-//! source は機械付与する契約 (`docs/plan/memory.md` §Extract)。
+//! [`ExtractedPayload`] containing `candidates[]`; the host expands each
+//! candidate into one [`StagingRecord`] and attaches record ids, extract-run ids,
+//! record-level source, evidence snippets, and source anchors mechanically.
 
 use schemars::JsonSchema;
 use serde::{Deserialize, Serialize};
 
-use crate::schema::{SourceEvidenceRef, SourceRef};
+use crate::schema::{EvidenceKind, SourceEvidenceRef, SourceRef};
 
-/// LLM が返す活動ログ候補の集合。すべて optional（空配列は許容）。
+/// Current flat staging schema version.
+pub const STAGING_SCHEMA_VERSION: u32 = 2;
+
+/// Candidate kinds that extract is allowed to stage.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum CandidateKind {
+    Preference,
+    WorkingAssumption,
+    Constraint,
+    Decision,
+    OpenQuestion,
+    Lesson,
+}
+
+impl CandidateKind {
+    pub fn as_str(&self) -> &'static str {
+        match self {
+            CandidateKind::Preference => "preference",
+            CandidateKind::WorkingAssumption => "working_assumption",
+            CandidateKind::Constraint => "constraint",
+            CandidateKind::Decision => "decision",
+            CandidateKind::OpenQuestion => "open_question",
+            CandidateKind::Lesson => "lesson",
+        }
+    }
+}
+
+/// Model-submitted candidate before host-side staging metadata is attached.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+pub struct ExtractedCandidate {
+    /// Kind of candidate. This is intentionally a narrow taxonomy, not an
+    /// activity-log category.
+    pub kind: CandidateKind,
+    /// Concise candidate claim.
+    pub claim: String,
+    /// Why this may be useful for future work or consolidation.
+    pub why_useful: String,
+    /// Optional invalidation / staleness hint.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub staleness: Option<String>,
+    /// Evidence ids selected by the extract worker. The transitional
+    /// `write_extracted` path may leave this empty until `session-explore`
+    /// provides host-issued evidence ids.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub evidence_ids: Vec<String>,
+}
+
+/// Transitional extract output: zero or more flat candidates.
 #[derive(Debug, Clone, Default, Serialize, Deserialize, JsonSchema)]
 pub struct ExtractedPayload {
     #[serde(default)]
-    pub decisions: Vec<DecisionEntry>,
+    pub candidates: Vec<ExtractedCandidate>,
-    #[serde(default)]
-    pub discussions: Vec<DiscussionEntry>,
-    #[serde(default)]
-    pub attempts: Vec<AttemptEntry>,
-    #[serde(default)]
-    pub requests: Vec<RequestEntry>,
 }
 
 impl ExtractedPayload {
-    /// すべての配列が空であれば true。空ペイロードは
-    /// "Nothing to save" 扱いで staging への書き込みを省いてよい。
     pub fn is_empty(&self) -> bool {
-        self.decisions.is_empty()
+        self.candidates.is_empty()
-            && self.discussions.is_empty()
-            && self.attempts.is_empty()
-            && self.requests.is_empty()
     }
 }
 
-/// 判断したこと（選択肢 + 選んだ + 根拠）。
+/// Bounded evidence snippet copied into a flat staging record.
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
-pub struct DecisionEntry {
+pub struct StagingEvidence {
-    /// 検討された選択肢の列挙。
+    pub id: String,
-    pub options: Vec<String>,
+    pub kind: EvidenceKind,
-    /// 採用された選択肢。
-    pub chosen: String,
-    /// 採用理由 / 根拠。
-    pub rationale: String,
-    /// Host-resolved anchors backing this individual claim.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    #[schemars(skip)]
-    pub source_refs: Vec<SourceEvidenceRef>,
-}
-
-/// 議論したこと（トピック + 論点）。結論が出ていなくてもよい。
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-pub struct DiscussionEntry {
-    /// 議論の主題。
-    pub topic: String,
-    /// 主題の中で挙がった論点 / 観点。
-    pub points: Vec<String>,
-    /// Host-resolved anchors backing this individual claim.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    #[schemars(skip)]
-    pub source_refs: Vec<SourceEvidenceRef>,
-}
-
-/// 試したこと（試行 + 結果 + 成否）。
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-pub struct AttemptEntry {
-    /// 何を試したか。
-    pub action: String,
-    /// 試した結果。
-    pub result: String,
-    /// 試行が目的に対して成功したか。失敗 / 部分成功も含めて bool で表現する。
-    pub succeeded: bool,
-    /// Host-resolved anchors backing this individual claim.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    #[schemars(skip)]
-    pub source_refs: Vec<SourceEvidenceRef>,
-}
-
-/// ユーザー submit の構造化要約。
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-pub struct RequestEntry {
-    /// ユーザーの意図 / ゴール。
-    pub intent: String,
-    /// 対象ファイル / モジュール / 機能（任意）。
     #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub target: Option<String>,
+    pub entry_range: Option<[u64; 2]>,
-    /// 一文サマリ。
+    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub summary: String,
+    pub excerpt: Option<String>,
-    /// Host-resolved anchors backing this individual claim.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub summary: Option<String>,
-    #[schemars(skip)]
-    pub source_refs: Vec<SourceEvidenceRef>,
 }
 
-/// staging に書き出される 1 ファイル分のレコード。
+/// One flat staging record. One record is one consolidation decision unit.
-///
-/// `source` は Worker 側ラッパーが segment_id と log entry range を
-/// 機械付与する。LLM はこのフィールドを見ない / 推論しない。
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct StagingRecord {
+    pub schema_version: u32,
+    pub id: String,
+    pub extract_run_id: String,
     pub source: SourceRef,
-    #[serde(flatten)]
+    pub kind: CandidateKind,
-    pub payload: ExtractedPayload,
+    pub claim: String,
+    pub why_useful: String,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub staleness: Option<String>,
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub evidence: Vec<StagingEvidence>,
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub source_refs: Vec<SourceEvidenceRef>,
+}
+
+impl StagingRecord {
+    pub fn from_candidate(
+        id: impl Into<String>,
+        extract_run_id: impl Into<String>,
+        source: SourceRef,
+        candidate: ExtractedCandidate,
+        evidence: Vec<StagingEvidence>,
+        source_refs: Vec<SourceEvidenceRef>,
+    ) -> Self {
+        Self {
+            schema_version: STAGING_SCHEMA_VERSION,
+            id: id.into(),
+            extract_run_id: extract_run_id.into(),
+            source,
+            kind: candidate.kind,
+            claim: candidate.claim,
+            why_useful: candidate.why_useful,
+            staleness: candidate.staleness,
+            evidence,
+            source_refs,
+        }
+    }
 }
 
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::schema::{EvidenceKind, SourceEvidenceRef};
 
-    fn record_source() -> serde_json::Value {
+    fn source() -> SourceRef {
-        serde_json::json!({
+        SourceRef {
-            "source": {
-                "segment_id": "seg-old",
-                "range": [1, 5]
-            }
-        })
-    }
-
-    #[test]
-    fn old_staging_json_without_entry_source_refs_deserializes() {
-        let mut raw = record_source();
-        raw["decisions"] = serde_json::json!([
-            {
-                "options": ["keep", "drop"],
-                "chosen": "keep",
-                "rationale": "compatible"
-            }
-        ]);
-        raw["discussions"] = serde_json::json!([
-            {
-                "topic": "compatibility",
-                "points": ["missing source_refs should default empty"]
-            }
-        ]);
-        raw["attempts"] = serde_json::json!([
-            {
-                "action": "deserialize old staging",
-                "result": "ok",
-                "succeeded": true
-            }
-        ]);
-        raw["requests"] = serde_json::json!([
-            {
-                "intent": "preserve old JSON",
-                "summary": "old payload has no entry anchors"
-            }
-        ]);
-
-        let record: StagingRecord = serde_json::from_value(raw).unwrap();
-
-        assert_eq!(record.source.segment_id, "seg-old");
-        assert!(record.payload.decisions[0].source_refs.is_empty());
-        assert!(record.payload.discussions[0].source_refs.is_empty());
-        assert!(record.payload.attempts[0].source_refs.is_empty());
-        assert!(record.payload.requests[0].source_refs.is_empty());
-        let serialized = serde_json::to_string(&record).unwrap();
-        assert!(!serialized.contains("\"source_refs\""));
-    }
-
-    #[test]
-    fn new_staging_json_roundtrips_entry_source_refs() {
-        let evidence = SourceEvidenceRef {
-            session_id: Some("session-1".into()),
-            segment_id: Some("segment-1".into()),
-            entry_range: Some([10, 12]),
-            evidence_id: Some("ev-1".into()),
-            evidence_kind: Some(EvidenceKind::new(EvidenceKind::TOOL_RESULT)),
-            label: Some("cargo test result".into()),
-            summary: Some("bounded host summary".into()),
-        };
-        let record = StagingRecord {
-            source: SourceRef {
             segment_id: "segment-record".into(),
             range: [0, 20],
-            },
+        }
-            payload: ExtractedPayload {
+    }
-                decisions: vec![DecisionEntry {
+
-                    options: vec!["a".into(), "b".into()],
+    #[test]
-                    chosen: "a".into(),
+    fn extracted_payload_empty_when_no_candidates() {
-                    rationale: "evidence-backed".into(),
+        assert!(ExtractedPayload::default().is_empty());
-                    source_refs: vec![evidence.clone()],
+        let payload = ExtractedPayload {
+            candidates: vec![ExtractedCandidate {
+                kind: CandidateKind::Decision,
+                claim: "Use flat staging records".into(),
+                why_useful: "Consolidation can resolve candidates independently".into(),
+                staleness: None,
+                evidence_ids: Vec::new(),
             }],
-                discussions: vec![DiscussionEntry {
-                    topic: "anchor shape".into(),
-                    points: vec!["entry refs roundtrip".into()],
-                    source_refs: vec![SourceEvidenceRef {
-                        evidence_kind: Some(EvidenceKind::new(EvidenceKind::MESSAGE)),
-                        ..Default::default()
-                    }],
-                }],
-                attempts: vec![AttemptEntry {
-                    action: "serialize".into(),
-                    result: "contains source refs".into(),
-                    succeeded: true,
-                    source_refs: vec![SourceEvidenceRef {
-                        evidence_kind: Some(EvidenceKind::new(EvidenceKind::FILE_REF)),
-                        evidence_id: Some("file:crates/memory/src/extract/payload.rs".into()),
-                        ..Default::default()
-                    }],
-                }],
-                requests: vec![RequestEntry {
-                    intent: "keep anchors".into(),
-                    target: Some("memory staging".into()),
-                    summary: "entry-level anchors survive".into(),
-                    source_refs: vec![SourceEvidenceRef {
-                        evidence_kind: Some(EvidenceKind::new(EvidenceKind::TICKET_REF)),
-                        evidence_id: Some("00001KXNYXNM6".into()),
-                        ..Default::default()
-                    }],
-                }],
-            },
         };
+        assert!(!payload.is_empty());
+    }
+
+    #[test]
+    fn flat_staging_record_roundtrips() {
+        let evidence = StagingEvidence {
+            id: "E001".into(),
+            kind: EvidenceKind::new(EvidenceKind::MESSAGE),
+            entry_range: Some([10, 12]),
+            excerpt: Some("extract candidate taxonomy".into()),
+            summary: Some("User and assistant discussed staging kinds".into()),
+        };
+        let source_ref = SourceEvidenceRef {
+            segment_id: Some("segment-1".into()),
+            entry_range: Some([10, 12]),
+            evidence_id: Some("E001".into()),
+            evidence_kind: Some(EvidenceKind::new(EvidenceKind::MESSAGE)),
+            label: Some("design discussion".into()),
+            summary: Some("bounded host summary".into()),
+            ..Default::default()
+        };
+        let candidate = ExtractedCandidate {
+            kind: CandidateKind::Constraint,
+            claim: "Extract stages candidates but does not write Memory".into(),
+            why_useful: "Preserves extract/consolidation boundary".into(),
+            staleness: Some("Revisit if the execution boundary changes".into()),
+            evidence_ids: vec!["E001".into()],
+        };
+        let record = StagingRecord::from_candidate(
+            "stg-1",
+            "run-1",
+            source(),
+            candidate,
+            vec![evidence],
+            vec![source_ref],
+        );
 
         let json = serde_json::to_string_pretty(&record).unwrap();
+        assert!(json.contains("\"schema_version\": 2"));
+        assert!(json.contains("\"kind\": \"constraint\""));
         assert!(json.contains("source_refs"));
-        assert!(json.contains("tool_result"));
+        assert!(json.contains("evidence"));
-        assert!(json.contains("entry_range"));
         let parsed: StagingRecord = serde_json::from_str(&json).unwrap();
 
-        let source_ref = &parsed.payload.decisions[0].source_refs[0];
+        assert_eq!(parsed.schema_version, STAGING_SCHEMA_VERSION);
-        assert_eq!(source_ref.session_id.as_deref(), Some("session-1"));
+        assert_eq!(parsed.kind, CandidateKind::Constraint);
-        assert_eq!(source_ref.segment_id.as_deref(), Some("segment-1"));
+        assert_eq!(parsed.evidence[0].id, "E001");
-        assert_eq!(source_ref.entry_range, Some([10, 12]));
+        assert_eq!(parsed.source_refs[0].evidence_id.as_deref(), Some("E001"));
-        assert_eq!(source_ref.evidence_id.as_deref(), Some("ev-1"));
+    }
-        assert_eq!(
+
-            source_ref.evidence_kind.as_ref().map(EvidenceKind::as_str),
+    #[test]
-            Some(EvidenceKind::TOOL_RESULT)
+    fn candidate_kind_serializes_as_snake_case() {
-        );
+        let json = serde_json::to_string(&CandidateKind::WorkingAssumption).unwrap();
-        assert_eq!(source_ref.label.as_deref(), Some("cargo test result"));
+        assert_eq!(json, "\"working_assumption\"");
-        assert_eq!(source_ref.summary.as_deref(), Some("bounded host summary"));
+        let parsed: CandidateKind = serde_json::from_str("\"open_question\"").unwrap();
-        assert_eq!(
+        assert_eq!(parsed, CandidateKind::OpenQuestion);
-            parsed.payload.discussions[0].source_refs[0]
-                .evidence_kind
-                .as_ref()
-                .map(EvidenceKind::as_str),
-            Some(EvidenceKind::MESSAGE)
-        );
-        assert_eq!(
-            parsed.payload.attempts[0].source_refs[0]
-                .evidence_kind
-                .as_ref()
-                .map(EvidenceKind::as_str),
-            Some(EvidenceKind::FILE_REF)
-        );
-        assert_eq!(
-            parsed.payload.requests[0].source_refs[0]
-                .evidence_kind
-                .as_ref()
-                .map(EvidenceKind::as_str),
-            Some(EvidenceKind::TICKET_REF)
-        );
     }
 }

crates/memory/src/extract/staging.rs

@@ -1,9 +1,12 @@
-//! `<workspace>/.yoi/memory/_staging/<id>.json` への書き出しヘルパー。
+//! extract staging writer.
 //!
-//! 1 件 1 ファイル、UUIDv7 命名（短命なので衝突回避と順序を兼ねる）。
+//! Staging is flat: one file is one candidate and one consolidation decision
-//! `source` を機械付与した [`StagingRecord`] 形式で保存する。
+//! unit. The transitional extract tool still submits `ExtractedPayload` with
+//! `candidates[]`; this writer expands it into one [`StagingRecord`] per
+//! candidate.
 
 use std::fs;
+use std::io;
 use std::path::PathBuf;
 
 use uuid::Uuid;
@@ -12,104 +15,108 @@ use crate::extract::payload::{ExtractedPayload, StagingRecord};
 use crate::schema::SourceRef;
 use crate::workspace::WorkspaceLayout;
 
-/// staging 書き出し時のエラー。
+/// Filesystem result for a single staged candidate.
-#[derive(Debug, thiserror::Error)]
+#[derive(Debug, Clone)]
-pub enum StagingError {
+pub struct StagingWriteResult {
-    #[error("failed to create staging dir {}: {source}", .path.display())]
+    pub id: Uuid,
-    CreateDir {
+    pub path: PathBuf,
-        path: PathBuf,
-        #[source]
-        source: std::io::Error,
-    },
-    #[error("failed to write staging file {}: {source}", .path.display())]
-    Write {
-        path: PathBuf,
-        #[source]
-        source: std::io::Error,
-    },
-    #[error("failed to serialize staging record: {0}")]
-    Serialize(#[from] serde_json::Error),
 }
 
-/// `payload` を `source` で wrap して staging に書き出す。
+/// Write one flat staging JSON file per extracted candidate.
 ///
-/// 戻り値は割り当てられた staging file の (id, path)。`payload` が
+/// Returns an empty vector when `payload` has no candidates.
-/// 完全に空の場合は呼び出し側が事前に `is_empty()` で skip 推奨だが、
-/// この関数は空でも正規に書き出す（仕様 §Extract で空配列許容と
-/// 明記されており、書く / 書かないの判断は呼び出し側に委ねる）。
 pub fn write_staging(
     layout: &WorkspaceLayout,
     source: SourceRef,
     payload: ExtractedPayload,
-) -> Result<(Uuid, PathBuf), StagingError> {
+) -> io::Result<Vec<StagingWriteResult>> {
-    let staging_dir = layout.staging_dir();
+    if payload.candidates.is_empty() {
-    fs::create_dir_all(&staging_dir).map_err(|source| StagingError::CreateDir {
+        return Ok(Vec::new());
-        path: staging_dir.clone(),
+    }
-        source,
-    })?;
 
+    let dir = layout.staging_dir();
+    fs::create_dir_all(&dir)?;
+    let extract_run_id = Uuid::now_v7().to_string();
+    let mut written = Vec::with_capacity(payload.candidates.len());
+
+    for candidate in payload.candidates {
         let id = Uuid::now_v7();
-    let path = staging_dir.join(format!("{id}.json"));
+        let record = StagingRecord::from_candidate(
-    let record = StagingRecord { source, payload };
+            id.to_string(),
-    let json = serde_json::to_string_pretty(&record)?;
+            extract_run_id.clone(),
-    fs::write(&path, json).map_err(|source| StagingError::Write {
+            source.clone(),
-        path: path.clone(),
+            candidate,
-        source,
+            Vec::new(),
-    })?;
+            Vec::new(),
-    Ok((id, path))
+        );
+        let path = dir.join(format!("{}.json", id));
+        let bytes = serde_json::to_vec_pretty(&record).map_err(io::Error::other)?;
+        fs::write(&path, bytes)?;
+        written.push(StagingWriteResult { id, path });
+    }
+
+    Ok(written)
 }
 
 #[cfg(test)]
 mod tests {
     use super::*;
 
-    use crate::extract::payload::{DecisionEntry, ExtractedPayload};
+    use crate::extract::payload::{CandidateKind, ExtractedCandidate};
+
+    fn layout() -> WorkspaceLayout {
+        let dir = tempfile::tempdir().unwrap();
+        // leak tempdir for the duration of the test process; sufficient for unit tests
+        let path = dir.keep();
+        WorkspaceLayout::new(path)
+    }
+
+    fn source() -> SourceRef {
+        SourceRef {
+            segment_id: "segment-1".into(),
+            range: [1, 3],
+        }
+    }
+
+    fn candidate(kind: CandidateKind, claim: &str) -> ExtractedCandidate {
+        ExtractedCandidate {
+            kind,
+            claim: claim.into(),
+            why_useful: "useful for future consolidation".into(),
+            staleness: None,
+            evidence_ids: Vec::new(),
+        }
+    }
 
     #[test]
-    fn writes_record_with_machine_attached_source() {
+    fn writes_one_file_per_candidate() {
-        let tmp = tempfile::TempDir::new().unwrap();
+        let layout = layout();
-        let layout = WorkspaceLayout::new(tmp.path().to_path_buf());
-
-        let source = SourceRef {
-            segment_id: "sess-1".into(),
-            range: [3, 7],
-        };
         let payload = ExtractedPayload {
-            decisions: vec![DecisionEntry {
+            candidates: vec![
-                options: vec!["a".into(), "b".into()],
+                candidate(CandidateKind::Preference, "Prefer implementation tickets"),
-                chosen: "a".into(),
+                candidate(CandidateKind::Decision, "Use flat staging records"),
-                rationale: "shorter".into(),
+            ],
-                source_refs: Vec::new(),
-            }],
-            ..Default::default()
         };
-        let (id, path) = write_staging(&layout, source.clone(), payload).unwrap();
+        let results = write_staging(&layout, source(), payload).unwrap();
-        assert_eq!(path.parent().unwrap(), layout.staging_dir());
+        assert_eq!(results.len(), 2);
-        assert!(
+        assert_ne!(results[0].id, results[1].id);
-            path.file_name()
-                .unwrap()
-                .to_string_lossy()
-                .contains(&id.to_string())
-        );
 
-        let written: StagingRecord =
+        let first = fs::read_to_string(&results[0].path).unwrap();
-            serde_json::from_str(&fs::read_to_string(&path).unwrap()).unwrap();
+        let second = fs::read_to_string(&results[1].path).unwrap();
-        assert_eq!(written.source.segment_id, "sess-1");
+        let first_record: StagingRecord = serde_json::from_str(&first).unwrap();
-        assert_eq!(written.source.range, [3, 7]);
+        let second_record: StagingRecord = serde_json::from_str(&second).unwrap();
-        assert_eq!(written.payload.decisions.len(), 1);
+
+        assert_eq!(first_record.kind, CandidateKind::Preference);
+        assert_eq!(second_record.kind, CandidateKind::Decision);
+        assert_eq!(first_record.extract_run_id, second_record.extract_run_id);
+        assert_eq!(first_record.source.segment_id, "segment-1");
     }
 
     #[test]
-    fn empty_payload_is_written_verbatim() {
+    fn empty_payload_writes_nothing() {
-        let tmp = tempfile::TempDir::new().unwrap();
+        let layout = layout();
-        let layout = WorkspaceLayout::new(tmp.path().to_path_buf());
+        let results = write_staging(&layout, source(), ExtractedPayload::default()).unwrap();
-        let source = SourceRef {
+        assert!(results.is_empty());
-            segment_id: "sess".into(),
+        assert!(!layout.staging_dir().exists());
-            range: [0, 0],
-        };
-        let (_, path) = write_staging(&layout, source, ExtractedPayload::default()).unwrap();
-        let written: StagingRecord =
-            serde_json::from_str(&fs::read_to_string(&path).unwrap()).unwrap();
-        assert!(written.payload.is_empty());
     }
 }

crates/memory/src/extract/tool.rs

@@ -12,10 +12,10 @@ use llm_engine::tool::{Tool, ToolDefinition, ToolError, ToolMeta, ToolOutput};
 
 use crate::extract::payload::ExtractedPayload;
 
-const WRITE_EXTRACTED_DESCRIPTION: &str = "Submit the final activity-log JSON for this slice. \
+const WRITE_EXTRACTED_DESCRIPTION: &str = "Submit extracted memory-candidate JSON for this slice. \
-Pass an object with `decisions`, `discussions`, `attempts`, and `requests` arrays (any may be empty). \
+Pass an object with a `candidates` array. Each candidate must have `kind`, `claim`, and `why_useful`; \
-Call this exactly once and end the turn. Do not include `source`, session metadata, or free-form prose — \
+`staleness` and `evidence_ids` are optional. Call this exactly once and end the turn. Do not include \
-the wrapper attaches provenance mechanically.";
+record ids, source anchors, session metadata, or free-form prose — the wrapper attaches staging metadata mechanically.";
 
 /// extract sub-Engine の出力受け口。`ExtractedPayload` 1 件をホストする。
 #[derive(Debug, Default)]
@@ -63,11 +63,8 @@ impl Tool for WriteExtractedTool {
             ToolError::InvalidArgument(format!("invalid write_extracted input: {e}"))
         })?;
         let summary = format!(
-            "Recorded activity log: decisions={} discussions={} attempts={} requests={}",
+            "Recorded memory candidates: candidates={}",
-            payload.decisions.len(),
+            payload.candidates.len(),
-            payload.discussions.len(),
-            payload.attempts.len(),
-            payload.requests.len(),
         );
         {
             let mut guard = self
@@ -116,20 +113,17 @@ mod tests {
         let ctx = Arc::new(ExtractWorkerContext::new());
         let tool: Arc<dyn Tool> = Arc::new(WriteExtractedTool { ctx: ctx.clone() });
         let input = serde_json::json!({
-            "decisions": [{
+            "candidates": [{
-                "options": ["a", "b"],
+                "kind": "decision",
-                "chosen": "a",
+                "claim": "Use flat staging",
-                "rationale": "test"
+                "why_useful": "Consolidation can resolve candidates independently"
-            }],
+            }]
-            "discussions": [],
-            "attempts": [],
-            "requests": []
         })
         .to_string();
         let out = tool.execute(&input, Default::default()).await.unwrap();
-        assert!(out.summary.contains("decisions=1"));
+        assert!(out.summary.contains("candidates=1"));
         let payload = ctx.take_payload().unwrap();
-        assert_eq!(payload.decisions.len(), 1);
+        assert_eq!(payload.candidates.len(), 1);
         assert_eq!(ctx.call_count(), 1);
     }
 
@@ -138,22 +132,21 @@ mod tests {
         let ctx = Arc::new(ExtractWorkerContext::new());
         let tool: Arc<dyn Tool> = Arc::new(WriteExtractedTool { ctx: ctx.clone() });
 
-        let first =
+        let first = serde_json::json!({"candidates": []}).to_string();
-            serde_json::json!({"decisions": [], "discussions": [], "attempts": [], "requests": []})
-                .to_string();
         tool.execute(&first, Default::default()).await.unwrap();
 
         let second = serde_json::json!({
-            "decisions": [],
+            "candidates": [{
-            "discussions": [],
+                "kind": "lesson",
-            "attempts": [{"action": "x", "result": "ok", "succeeded": true}],
+                "claim": "Validation should use Nix build",
-            "requests": []
+                "why_useful": "Packaging can fail independently"
+            }]
         })
         .to_string();
         tool.execute(&second, Default::default()).await.unwrap();
 
         let payload = ctx.take_payload().unwrap();
-        assert_eq!(payload.attempts.len(), 1);
+        assert_eq!(payload.candidates.len(), 1);
         assert_eq!(ctx.call_count(), 2);
     }
 

crates/worker/src/worker.rs

@@ -3172,15 +3172,15 @@ impl<C: LlmClient, St: Store> Worker<C, St> {
         });
 
         let source_segment_id = self.segment_state.segment_id();
-        let staging_id = if payload.is_empty() {
+        let staging_results = if payload.is_empty() {
-            String::new()
+            Vec::new()
         } else {
             let source = memory::schema::SourceRef {
                 segment_id: source_segment_id.to_string(),
                 range: [start_entry as u64, end_entry as u64],
             };
-            let (id, _) = match extract::write_staging(&layout, source, payload) {
+            match extract::write_staging(&layout, source, payload) {
-                Ok(result) => result,
+                Ok(results) => results,
                 Err(err) => {
                     let usage = usage_capture
                         .lock()
@@ -3197,9 +3197,12 @@ impl<C: LlmClient, St: Store> Worker<C, St> {
                     );
                     return Err(WorkerError::ExtractStaging(err));
                 }
+            }
         };
-            id.to_string()
+        let staging_id = staging_results
-        };
+            .first()
+            .map(|result| result.id.to_string())
+            .unwrap_or_default();
 
         let pointer_payload = extract::ExtractPointerPayload {
             processed_through_entry: end_entry,
@@ -3220,16 +3223,12 @@ impl<C: LlmClient, St: Store> Worker<C, St> {
             .expect("extract_pointer poisoned") = Some(pointer_payload);
 
         let mut extract_audit = extract_audit_base;
-        if !staging_id.is_empty() {
+        extract_audit.staging_count = staging_results.len();
-            extract_audit.staging_count = 1;
+        for result in &staging_results {
-            extract_audit.staging_ids.push(staging_id.clone());
+            extract_audit.staging_ids.push(result.id.to_string());
-            extract_audit.staging_paths.push(
+            extract_audit
-                layout
+                .staging_paths
-                    .staging_dir()
+                .push(result.path.display().to_string());
-                    .join(format!("{staging_id}.json"))
-                    .display()
-                    .to_string(),
-            );
         }
         let usage = usage_capture
             .lock()
@@ -4819,7 +4818,7 @@ pub enum WorkerError {
     Skill(#[from] SkillClientError),
 
     #[error("memory extract staging write failed: {0}")]
-    ExtractStaging(#[source] memory::extract::StagingError),
+    ExtractStaging(#[source] std::io::Error),
 
     #[error("memory consolidation lock acquisition failed: {0}")]
     ConsolidationLock(#[source] memory::consolidate::LockError),

crates/worker/tests/consolidation_test.rs

@@ -24,7 +24,7 @@ use llm_engine::Engine;
 use llm_engine::llm_client::event::{Event as LlmEvent, ResponseStatus, StatusEvent};
 use llm_engine::llm_client::{ClientError, LlmClient, Request};
 use memory::WorkspaceLayout;
-use memory::extract::{ExtractedPayload, write_staging};
+use memory::extract::{CandidateKind, ExtractedCandidate, ExtractedPayload, write_staging};
 use memory::schema::SourceRef;
 use session_store::FsStore;
 use session_store::{CombinedStore, FsWorkerStore};
@@ -182,18 +182,32 @@ async fn make_worker_with(
     .unwrap()
 }
 
+fn staging_payload(claim: String) -> ExtractedPayload {
+    ExtractedPayload {
+        candidates: vec![ExtractedCandidate {
+            kind: CandidateKind::Lesson,
+            claim,
+            why_useful: "useful for consolidation trigger tests".into(),
+            staleness: None,
+            evidence_ids: Vec::new(),
+        }],
+    }
+}
+
 fn write_n_staging(layout: &WorkspaceLayout, n: usize) -> Vec<uuid::Uuid> {
     let mut ids = Vec::new();
     for i in 0..n {
-        let (id, _) = write_staging(
+        let id = write_staging(
             layout,
             SourceRef {
                 segment_id: format!("s-{i}"),
                 range: [i as u64, i as u64],
             },
-            ExtractedPayload::default(),
+            staging_payload(format!("candidate-{i}")),
         )
-        .unwrap();
+        .unwrap()
+        .remove(0)
+        .id;
         ids.push(id);
     }
     ids

resources/prompts/internal/memory_extract_system.md

@@ -1,40 +1,63 @@
-You are the activity extractor for a Yoi memory subsystem.
+You are a Yoi memory extract worker.
 
-Your single job: read the supplied conversation slice and emit a structured JSON record of "what happened" via the `write_extracted` tool. You are not consolidating, summarising, or generating memory — that is the consolidation worker's job.
+Your job is to read the supplied conversation slice and extract only memory candidates that may be worth later consolidation. Do not produce activity logs.
 
-# Memory language
+## Language
 
-- `language`: `{{ language }}`.
+- `language`: `{{language}}`
-- Write extracted fact strings (`rationale`, `topic`, `points`, `action`, `result`, `intent`, `summary`, etc.) in this language.
+- Write candidate claims, usefulness, and staleness text in this language.
-- Preserve code identifiers, paths, command names, quoted user text, logs, and external proper nouns when translation would reduce fidelity.
+- Preserve literal identifiers, paths, commands, branch names, issue IDs, tool names, model names, and quoted user/system text as-is.
+- If the configured language is unclear, use English.
 
-# Hard rules
+Call `write_extracted` exactly once with an object of this shape:
 
-- Call `write_extracted` exactly once. Do not narrate, ask questions, or send any other tool output.
+```json
-- The argument is an object with four arrays: `decisions`, `discussions`, `attempts`, `requests`. Any of them may be empty. If nothing in the slice is worth recording, call `write_extracted({"decisions": [], "discussions": [], "attempts": [], "requests": []})` and stop.
+{
-- Do NOT include `source`, `session_id`, entry indices, timestamps, or any provenance metadata. The wrapper attaches them mechanically.
+  "candidates": [
-- Do NOT add free-form commentary, summaries, or explanatory prose outside the schema fields.
+    {
+      "kind": "preference",
+      "claim": "...",
+      "why_useful": "...",
+      "staleness": "...",
+      "evidence_ids": []
+    }
+  ]
+}
+```
 
-# Extraction guidance
+Allowed candidate kinds:
 
-- `decisions`: judgements made during the slice. Each entry needs `options` (the alternatives considered), `chosen` (what was picked), and `rationale` (why).
+- `preference`: durable user/workspace preference or working style, not a one-off instruction.
-- `discussions`: topics that were debated. `topic` plus `points` (the considerations raised). Open / unresolved discussions are valid.
+- `working_assumption`: provisional assumption that affects future design/implementation and may later change.
-- `attempts`: things that were tried. `action`, `result`, and a `succeeded` boolean. Partial success is `false` with the result text describing the partial outcome.
+- `constraint`: boundary, invariant, or prohibition that future work/review should respect.
-- `requests`: structured summaries of user submissions. `intent` (what the user wants), optional `target` (file / module / feature), and a one-line `summary`.
+- `decision`: choice with alternatives/chosen/rationale; not a mere fact or progress note.
+- `open_question`: unresolved question that affects follow-up work and has a concrete next action.
+- `lesson`: reusable learning from validation/failure/attempts that can improve future work.
 
-# Quality bar
+Required fields per candidate:
 
-- Drop one-off chit-chat, shallow questions, and turn-by-turn progress noise. Keep entries with long-term reference value.
+- `kind`: one of the allowed candidate kinds.
-- Do not duplicate content already captured by static project docs (AGENTS.md, plan documents) — those are not "what happened in this slice".
+- `claim`: concise statement of the candidate.
-- Prefer concise, fact-shaped strings. Do not pad rationale or summary fields.
+- `why_useful`: why this candidate may be useful for future consolidation.
 
-# Anti-noise rules
+Optional fields:
 
-Authoritative project records (issue trackers, task boards, planning documents, changelogs, version-control history, generated reports) are the source of truth for their exact contents. Memory must not mirror those records verbatim or maintain a parallel state ledger, but it may capture durable project-management facts, workflow constraints, recurring patterns, and abstractions when they will help future work.
+- `staleness`: when this candidate should be revisited or invalidated.
+- `evidence_ids`: leave empty in this transitional path unless host-issued evidence ids are present.
 
-- `attempts`: skip actions whose only substance is maintaining an authoritative record or moving an item through an external lifecycle. Keep attempts for outcomes that are not captured by that record itself: build / test outcomes, external API responses, observed bug reproductions, design experiments, and process lessons that inform later judgement.
+Do not extract:
-- `discussions`: skip transient triage that goes stale within the day — immediate scheduling, checklist-style state reads, or short-lived sequencing choices. Keep discussions whose points outlive the session (architectural trade-offs, durable process constraints, recurring workflow questions).
-- `decisions`: the rationale must be a design / policy / process / approach reason, not "we did X in this session". Recording that an item was filed, completed, or moved through a lifecycle is NOT a decision; recording the durable policy or abstraction behind that workflow can be.
-- Avoid copying titles, bodies, checklists, raw statees, or short-lived identifiers from authoritative project records. If a record is only meaningful as an exact state mirror or with a transient identifier, the record itself is probably session-local and should be skipped.
 
-When you have produced the JSON, call `write_extracted` and end the turn. No follow-up text.
+- tool-call chronology;
+- file read/write history;
+- generic progress updates;
+- current-focus updates;
+- one-off chit-chat;
+- resolved local confusion;
+- assistant self-corrections without durable consequence;
+- authoritative Ticket/docs/git facts copied verbatim;
+- validation results unless they imply a reusable lesson, active blocker, or authority evidence;
+- implementation details that belong only in commit diff.
+
+Prefer no candidates over noisy candidates. If nothing is worth staging, call `write_extracted` with `{"candidates": []}`.
+
+Do not include record ids, source anchors, session metadata, free-form prose, or raw tool output content. The host attaches staging metadata mechanically.