feat(feedback): Phase 2 — HITL thumbs + correction capture with prompt-conditioning loop

Closes the human-in-the-loop gap from docs/brief/feature-set.md and
Phase 2 of the 2026-04-23 feature-complete roadmap.

Storage (kon-storage):
- Migration v10 adds the `feedback` table: (target_type, target_id,
  rating, original_text, corrected_text, context_json, profile_id,
  created_at) with CHECK constraints on target_type and rating, plus
  indexes on (target_type, rating, created_at DESC) for prompt-time
  retrieval and (profile_id, target_type, created_at DESC) for
  per-profile scoping.
- New public API: `FeedbackTargetType`, `RecordFeedbackParams`,
  `FeedbackRow`, `record_feedback`, `list_feedback_examples`.
- Tests updated — the RB-02 rollback regression now discovers the
  real max version at runtime instead of hard-coding v10 for its
  poison migration.

LLM (kon-llm):
- `prompts::FeedbackExample` — local shape for few-shot exemplars so
  kon-llm stays independent of kon-storage.
- `prompts::build_conditioned_system_prompt` — appends a "here is
  the style this user prefers" block to the base system prompt
  when examples are available; returns the base prompt unchanged
  when empty, so new users and early sessions see generic output.
- `LlmEngine::decompose_task_with_feedback` and
  `LlmEngine::extract_tasks_with_feedback` thread examples through
  to the builder. The old one-arg variants are preserved and now
  call through with an empty slice.
- 4 unit tests covering empty, empty-input-skip, correction-wins,
  and thumbs-up-only fallback.

Tauri (src-tauri):
- New commands::feedback module: `record_feedback`,
  `list_feedback_examples_cmd`.
- `decompose_and_store` and `extract_tasks_from_transcript_cmd`
  now fetch the last 5 positive/neutral feedback rows for their
  target type and pass them through to the LLM, wiring the
  learning loop end-to-end.
- Shared `to_llm_examples` helper parses the `context_json.input`
  field (where the recorder stashes the parent task text / transcript
  chunk) back into the exemplar shape.

Frontend (MicroSteps.svelte):
- Thumbs-up and thumbs-down buttons on every micro-step row.
  Hover-revealed; the vote recolours the icon; clicking again
  clears the local highlight (the row itself stays in the audit
  trail).
- Pencil icon + double-click to edit step text. Save flows through
  update_task_cmd for persistence and records a correction feedback
  row with (original_text, corrected_text) — the highest-value
  training signal.
- Parent task text is captured in context_json.input at record time
  so the prompt builder can reconstruct the (input, preferred-output)
  pair on subsequent decompositions.
- Feedback capture is best-effort — a record_feedback failure never
  interrupts the primary action.

What's deferred to a later phase:
- Thumbs + corrections on extracted tasks (same pipeline, different
  surface — probably TasksPage after the AI-extraction path)
- Thumbs on transcript cleanup output
- Semantic retrieval over the feedback corpus (once there is enough
  data to justify embedding infrastructure; the storage shape is
  already ready for it)

crates/llm/src/lib.rs

@@ -240,11 +240,30 @@ impl LlmEngine {
     }
 
     pub fn decompose_task(&self, task_text: &str) -> Result<Vec<String>, EngineError> {
+        self.decompose_task_with_feedback(task_text, &[])
+    }
+
+    /// Same as `decompose_task` but allows callers to pass recent HITL
+    /// feedback rows so the system prompt gets conditioned on the
+    /// user's preferred decomposition style. The `examples` vec is
+    /// rendered into a few-shot block appended to the base system
+    /// prompt by `prompts::build_conditioned_system_prompt`.
+    ///
+    /// Callers should pass most-recent-first; older examples still
+    /// participate but weigh less because of their position in the
+    /// prompt. Empty slice keeps behaviour identical to `decompose_task`.
+    pub fn decompose_task_with_feedback(
+        &self,
+        task_text: &str,
+        examples: &[prompts::FeedbackExample],
+    ) -> Result<Vec<String>, EngineError> {
         let model = self.loaded_model_arc()?;
+        let system =
+            prompts::build_conditioned_system_prompt(prompts::DECOMPOSE_TASK_SYSTEM, examples);
         let prompt = render_chat_prompt(
             &model,
             &[
-                ("system", prompts::DECOMPOSE_TASK_SYSTEM),
+                ("system", system.as_str()),
                 ("user", &format!("Task: {task_text}")),
             ],
         )?;
@@ -261,15 +280,27 @@ impl LlmEngine {
     }
 
     pub fn extract_tasks(&self, transcript: &str) -> Result<Vec<String>, EngineError> {
+        self.extract_tasks_with_feedback(transcript, &[])
+    }
+
+    /// Feedback-conditioned variant of `extract_tasks`. See
+    /// `decompose_task_with_feedback` for the `examples` semantics.
+    pub fn extract_tasks_with_feedback(
+        &self,
+        transcript: &str,
+        examples: &[prompts::FeedbackExample],
+    ) -> Result<Vec<String>, EngineError> {
         if transcript.trim().is_empty() {
             return Ok(Vec::new());
         }
 
         let model = self.loaded_model_arc()?;
+        let system =
+            prompts::build_conditioned_system_prompt(prompts::EXTRACT_TASKS_SYSTEM, examples);
         let prompt = render_chat_prompt(
             &model,
             &[
-                ("system", prompts::EXTRACT_TASKS_SYSTEM),
+                ("system", system.as_str()),
                 ("user", &format!("Transcript:\n{transcript}")),
             ],
         )?;

crates/llm/src/prompts.rs

@@ -10,3 +10,113 @@ output a JSON array of action items the speaker committed to. Each item must \
 be a short imperative sentence. Omit observations, wishes, and background \
 context that are not explicit commitments. Output an empty array if there are \
 no action items.";
+
+/// Compact representation of a human-in-the-loop feedback example used
+/// for few-shot prompt conditioning. Built by kon-storage and fed to the
+/// prompt builder below; we keep this struct local to the LLM crate so
+/// kon-llm does not depend on kon-storage.
+#[derive(Debug, Clone)]
+pub struct FeedbackExample {
+    /// What the AI was given as input (e.g. the parent task text, or
+    /// the transcript chunk). Kept verbatim.
+    pub input: String,
+    /// What the AI produced originally. `None` if the user only
+    /// gave a thumbs-up without a prior edit (positive signal
+    /// without a paired correction).
+    pub original_output: Option<String>,
+    /// What the user changed it to. `None` for thumbs-only rows.
+    /// This is the highest-value signal — when present, inject it
+    /// as the "good" output in the few-shot example.
+    pub corrected_output: Option<String>,
+}
+
+/// Render a feedback example into the exemplar block used in prompt
+/// conditioning. Returns `None` for rows that carry no usable pairing
+/// (e.g. a thumbs-up with no input context).
+fn render_feedback_exemplar(ex: &FeedbackExample) -> Option<String> {
+    if ex.input.trim().is_empty() {
+        return None;
+    }
+    let good = ex
+        .corrected_output
+        .as_deref()
+        .or(ex.original_output.as_deref())?;
+    let good = good.trim();
+    if good.is_empty() {
+        return None;
+    }
+    Some(format!("Input: {}\nGood output: {}", ex.input.trim(), good))
+}
+
+/// Build a system prompt that combines the base task system prompt
+/// with a few-shot block assembled from recent HITL examples. If no
+/// usable examples are available, returns the base prompt unchanged
+/// so early users see the generic behaviour and the LLM is not
+/// confused by an empty exemplar section.
+///
+/// The exemplars are ordered most-recent-first (caller's order is
+/// preserved) so the LLM weights the user's current style over
+/// earlier noise, mirroring what a human reviewer would do.
+pub fn build_conditioned_system_prompt(base: &str, examples: &[FeedbackExample]) -> String {
+    let rendered: Vec<String> = examples
+        .iter()
+        .filter_map(render_feedback_exemplar)
+        .collect();
+    if rendered.is_empty() {
+        return base.to_string();
+    }
+    let block = rendered
+        .iter()
+        .map(|s| format!("- {s}"))
+        .collect::<Vec<_>>()
+        .join("\n");
+    format!(
+        "{base}\n\nHere are examples of the style this user prefers, in the \
+         user's own words. Match this style closely when producing your output:\n{block}"
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn builds_plain_prompt_when_no_examples() {
+        let out = build_conditioned_system_prompt(DECOMPOSE_TASK_SYSTEM, &[]);
+        assert_eq!(out, DECOMPOSE_TASK_SYSTEM);
+    }
+
+    #[test]
+    fn skips_empty_input_examples() {
+        let examples = vec![FeedbackExample {
+            input: String::new(),
+            original_output: None,
+            corrected_output: Some("ignored".into()),
+        }];
+        let out = build_conditioned_system_prompt(DECOMPOSE_TASK_SYSTEM, &examples);
+        assert_eq!(out, DECOMPOSE_TASK_SYSTEM);
+    }
+
+    #[test]
+    fn prefers_corrected_over_original() {
+        let examples = vec![FeedbackExample {
+            input: "Clean room".into(),
+            original_output: Some("Organise your bedroom".into()),
+            corrected_output: Some("Pick up one shirt from the floor".into()),
+        }];
+        let out = build_conditioned_system_prompt(DECOMPOSE_TASK_SYSTEM, &examples);
+        assert!(out.contains("Pick up one shirt from the floor"));
+        assert!(!out.contains("Organise your bedroom"));
+    }
+
+    #[test]
+    fn falls_back_to_original_when_no_correction() {
+        let examples = vec![FeedbackExample {
+            input: "Write report".into(),
+            original_output: Some("Open a blank document".into()),
+            corrected_output: None,
+        }];
+        let out = build_conditioned_system_prompt(DECOMPOSE_TASK_SYSTEM, &examples);
+        assert!(out.contains("Open a blank document"));
+    }
+}

crates/storage/src/database.rs

@@ -889,6 +889,151 @@ pub async fn list_recent_errors(pool: &SqlitePool, limit: i64) -> Result<Vec<Err
         .collect())
 }
 
+// --- Feedback (HITL) -------------------------------------------------------
+//
+// Phase 2 of the feature-complete roadmap: capture thumbs + corrections on
+// AI-generated output so the prompt builder can inject recent examples as
+// few-shot exemplars. Storage-only here; the prompt-conditioning logic lives
+// in kon-llm. Retrieval returns the most recent rows, narrowed to the
+// active profile when provided so feedback does not cross profiles.
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum FeedbackTargetType {
+    MicroStep,
+    TaskExtraction,
+    Cleanup,
+}
+
+impl FeedbackTargetType {
+    pub fn as_str(self) -> &'static str {
+        match self {
+            FeedbackTargetType::MicroStep => "microstep",
+            FeedbackTargetType::TaskExtraction => "task_extraction",
+            FeedbackTargetType::Cleanup => "cleanup",
+        }
+    }
+
+    /// Parse the database `target_type` string back into the enum.
+    /// Named `parse` rather than `from_str` so it does not collide with
+    /// the `std::str::FromStr` trait — the trait is overkill here
+    /// because callers never want a `FromStr::Err` and already know the
+    /// set of valid values at the call site.
+    pub fn parse(s: &str) -> Option<Self> {
+        match s {
+            "microstep" => Some(FeedbackTargetType::MicroStep),
+            "task_extraction" => Some(FeedbackTargetType::TaskExtraction),
+            "cleanup" => Some(FeedbackTargetType::Cleanup),
+            _ => None,
+        }
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct RecordFeedbackParams {
+    pub target_type: FeedbackTargetType,
+    pub target_id: Option<String>,
+    /// -1 = thumbs down, 0 = correction (neutral), +1 = thumbs up.
+    pub rating: i8,
+    pub original_text: Option<String>,
+    pub corrected_text: Option<String>,
+    pub context_json: Option<String>,
+    pub profile_id: Option<String>,
+}
+
+#[derive(Debug, Clone)]
+pub struct FeedbackRow {
+    pub id: i64,
+    pub target_type: String,
+    pub target_id: Option<String>,
+    pub rating: i64,
+    pub original_text: Option<String>,
+    pub corrected_text: Option<String>,
+    pub context_json: Option<String>,
+    pub profile_id: String,
+    pub created_at: String,
+}
+
+pub async fn record_feedback(pool: &SqlitePool, params: RecordFeedbackParams) -> Result<i64> {
+    if !matches!(params.rating, -1..=1) {
+        return Err(KonError::StorageError(format!(
+            "invalid feedback rating {} (must be -1, 0, or 1)",
+            params.rating
+        )));
+    }
+    let profile_id = params
+        .profile_id
+        .unwrap_or_else(|| crate::DEFAULT_PROFILE_ID.to_string());
+    let row = sqlx::query(
+        "INSERT INTO feedback (
+            target_type, target_id, rating,
+            original_text, corrected_text, context_json, profile_id
+         ) VALUES (?, ?, ?, ?, ?, ?, ?)
+         RETURNING id",
+    )
+    .bind(params.target_type.as_str())
+    .bind(params.target_id)
+    .bind(params.rating as i64)
+    .bind(params.original_text)
+    .bind(params.corrected_text)
+    .bind(params.context_json)
+    .bind(profile_id)
+    .fetch_one(pool)
+    .await
+    .map_err(|e| KonError::StorageError(format!("record_feedback failed: {e}")))?;
+    Ok(row.get::<i64, _>("id"))
+}
+
+/// Fetch the most recent feedback rows for a given target type, scoped to
+/// the active profile. Used by the prompt builder to gather few-shot
+/// exemplars. Orders by `created_at DESC` so the most recent corrections
+/// outweigh older ones — the user's style drifts, and we want the LLM
+/// to track the current preference.
+///
+/// `min_rating` filters out thumbs-down examples when the caller only
+/// wants positive reinforcement; pass `-1` to include everything.
+pub async fn list_feedback_examples(
+    pool: &SqlitePool,
+    target_type: FeedbackTargetType,
+    limit: i64,
+    min_rating: i8,
+    profile_id: Option<&str>,
+) -> Result<Vec<FeedbackRow>> {
+    let pid = profile_id.unwrap_or(crate::DEFAULT_PROFILE_ID);
+    let rows = sqlx::query(
+        "SELECT id, target_type, target_id, rating,
+                original_text, corrected_text, context_json,
+                profile_id, created_at
+         FROM feedback
+         WHERE target_type = ?
+           AND profile_id = ?
+           AND rating >= ?
+         ORDER BY created_at DESC
+         LIMIT ?",
+    )
+    .bind(target_type.as_str())
+    .bind(pid)
+    .bind(min_rating as i64)
+    .bind(limit)
+    .fetch_all(pool)
+    .await
+    .map_err(|e| KonError::StorageError(format!("list_feedback_examples failed: {e}")))?;
+
+    Ok(rows
+        .into_iter()
+        .map(|r| FeedbackRow {
+            id: r.get("id"),
+            target_type: r.get("target_type"),
+            target_id: r.get("target_id"),
+            rating: r.get("rating"),
+            original_text: r.get("original_text"),
+            corrected_text: r.get("corrected_text"),
+            context_json: r.get("context_json"),
+            profile_id: r.get("profile_id"),
+            created_at: r.get("created_at"),
+        })
+        .collect())
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

crates/storage/src/lib.rs

@@ -10,10 +10,11 @@ pub use database::{
     add_profile_term, complete_subtask_and_check_parent, complete_task, count_transcripts,
     create_profile, delete_profile, delete_profile_term, delete_task, delete_transcript,
     get_profile, get_setting, get_task_by_id, get_transcript, init, insert_subtask, insert_task,
-    insert_transcript, list_profile_terms, list_profiles, list_recent_errors, list_subtasks,
-    list_tasks, list_transcripts, list_transcripts_paged, log_error, search_transcripts,
-    set_setting, uncomplete_task, update_profile, update_task, update_transcript,
-    update_transcript_meta, ErrorLogRow, InsertTranscriptParams, ProfileRow, ProfileTermRow,
+    insert_transcript, list_feedback_examples, list_profile_terms, list_profiles,
+    list_recent_errors, list_subtasks, list_tasks, list_transcripts, list_transcripts_paged,
+    log_error, record_feedback, search_transcripts, set_setting, uncomplete_task, update_profile,
+    update_task, update_transcript, update_transcript_meta, ErrorLogRow, FeedbackRow,
+    FeedbackTargetType, InsertTranscriptParams, ProfileRow, ProfileTermRow, RecordFeedbackParams,
     TaskRow, TranscriptRow,
 };
 pub use file_storage::{app_data_dir, crashes_dir, database_path, logs_dir, recordings_dir};

crates/storage/src/migrations.rs

@@ -334,6 +334,49 @@ const MIGRATIONS: &[(i64, &str, &str)] = &[
         FROM transcripts;
         "#,
     ),
+    (
+        10,
+        "feedback: HITL thumbs + correction capture",
+        r#"
+        -- Feedback rows capture human-in-the-loop signal on AI-generated
+        -- output. Two flavours bundled into one table:
+        --   - thumbs (rating = -1 | +1, original_text optional, corrected_text NULL)
+        --   - correction (rating defaults to +1, original_text + corrected_text present)
+        --
+        -- `target_type` names the producing surface:
+        --   'microstep'  — subtask decomposition from DECOMPOSE_TASK_SYSTEM
+        --   'task_extraction' — tasks lifted from a transcript (EXTRACT_TASKS_SYSTEM)
+        --   'cleanup'    — transcript cleanup output
+        --
+        -- `target_id` is the surface-specific identifier where one exists
+        -- (subtask id, task id, transcript id). NULL is allowed because
+        -- not every feedback event has a stable target id yet.
+        --
+        -- `context_json` carries the input the AI was conditioned on
+        -- (parent task text, transcript chunk, etc.) so future prompt
+        -- builders can reconstruct the original I/O pair for few-shot
+        -- injection or semantic retrieval.
+        CREATE TABLE feedback (
+            id INTEGER PRIMARY KEY AUTOINCREMENT,
+            target_type TEXT NOT NULL
+                CHECK (target_type IN ('microstep', 'task_extraction', 'cleanup')),
+            target_id TEXT,
+            rating INTEGER NOT NULL
+                CHECK (rating IN (-1, 0, 1)),
+            original_text TEXT,
+            corrected_text TEXT,
+            context_json TEXT,
+            profile_id TEXT NOT NULL DEFAULT '00000000-0000-0000-0000-000000000001'
+                REFERENCES profiles(id) ON DELETE RESTRICT,
+            created_at TEXT NOT NULL DEFAULT (datetime('now'))
+        );
+
+        CREATE INDEX idx_feedback_target_type_rating
+            ON feedback(target_type, rating, created_at DESC);
+        CREATE INDEX idx_feedback_profile
+            ON feedback(profile_id, target_type, created_at DESC);
+        "#,
+    ),
 ];
 
 /// Split SQL into individual statements, respecting BEGIN...END trigger blocks.
@@ -483,7 +526,7 @@ mod tests {
             .fetch_one(&pool)
             .await
             .unwrap();
-        assert_eq!(count, 9);
+        assert_eq!(count, 10);
 
         sqlx::query("INSERT INTO settings (key, value) VALUES ('test', 'value')")
             .execute(&pool)
@@ -502,7 +545,7 @@ mod tests {
             .fetch_one(&pool)
             .await
             .unwrap();
-        assert_eq!(count, 9);
+        assert_eq!(count, 10);
     }
 
     #[tokio::test]
@@ -859,8 +902,11 @@ mod tests {
     // The poisoned migration below first creates `poison_marker`
     // (syntactically valid, would succeed against any SQLite) and then
     // runs a guaranteed-invalid function call. Under the new atomic
-    // implementation, neither `poison_marker` nor the v9 row should
+    // implementation, neither `poison_marker` nor the poison row should
     // survive the failed call.
+    //
+    // Version number must sit above the real MIGRATIONS max so the
+    // baseline migrate cleanly finishes first.
     #[tokio::test]
     async fn multi_statement_migration_rolls_back_on_failure() {
         let pool = SqlitePoolOptions::new()
@@ -871,8 +917,18 @@ mod tests {
 
         run_migrations(&pool).await.expect("baseline migrate");
 
-        const POISON: &[(i64, &str, &str)] = &[(
-            10,
+        // Discover the real max version so the poison migration is
+        // always exactly one past the end of MIGRATIONS, regardless of
+        // how many real migrations we add in future.
+        let real_max: i64 =
+            sqlx::query_scalar("SELECT COALESCE(MAX(version), 0) FROM schema_version")
+                .fetch_one(&pool)
+                .await
+                .expect("read schema_version");
+        let poison_version = real_max + 1;
+
+        let poison: &[(i64, &str, &str)] = &[(
+            poison_version,
             "rb-02 atomicity poison",
             r#"
             CREATE TABLE poison_marker (id INTEGER PRIMARY KEY);
@@ -880,7 +936,7 @@ mod tests {
             "#,
         )];
 
-        let result = run_migrations_slice(&pool, POISON).await;
+        let result = run_migrations_slice(&pool, poison).await;
         assert!(
             result.is_err(),
             "poisoned migration must return Err, got: {result:?}"
@@ -896,14 +952,14 @@ mod tests {
             "poison_marker must not exist; got: {marker:?}"
         );
 
-        // `schema_version` must not include v10 — version insert is part
-        // of the same transaction that rolled back.
+        // `schema_version` must not include the poison version — version
+        // insert is part of the same transaction that rolled back.
         let max: i64 = sqlx::query_scalar("SELECT COALESCE(MAX(version), 0) FROM schema_version")
             .fetch_one(&pool)
             .await
             .expect("read schema_version");
         assert_eq!(
-            max, 9,
+            max, real_max,
             "schema_version must not advance past the failed migration"
         );
     }

src-tauri/src/commands/feedback.rs

@@ -0,0 +1,110 @@
+// Tauri commands for human-in-the-loop feedback capture and retrieval.
+// Phase 2 of the feature-complete roadmap: thumbs + correction capture
+// on AI-generated output feeds a few-shot loop that conditions future
+// prompts on the user's preferred style.
+
+use serde::{Deserialize, Serialize};
+
+use kon_storage::{
+    list_feedback_examples as db_list_feedback_examples, record_feedback as db_record_feedback,
+    FeedbackRow, FeedbackTargetType, RecordFeedbackParams,
+};
+
+use crate::AppState;
+
+#[derive(Debug, Clone, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub struct RecordFeedbackInput {
+    /// One of "microstep", "task_extraction", "cleanup".
+    pub target_type: String,
+    /// Optional surface-specific id (subtask id, task id, transcript id).
+    #[serde(default)]
+    pub target_id: Option<String>,
+    /// -1 = thumbs down, 0 = correction (neutral), +1 = thumbs up.
+    pub rating: i8,
+    #[serde(default)]
+    pub original_text: Option<String>,
+    #[serde(default)]
+    pub corrected_text: Option<String>,
+    /// Freeform JSON context: e.g. the parent task text, the transcript
+    /// chunk the AI was given, etc. Used later by the prompt builder
+    /// to reconstruct the (input, preferred-output) pair.
+    #[serde(default)]
+    pub context_json: Option<String>,
+    #[serde(default)]
+    pub profile_id: Option<String>,
+}
+
+#[derive(Debug, Clone, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub struct FeedbackDto {
+    pub id: i64,
+    pub target_type: String,
+    pub target_id: Option<String>,
+    pub rating: i64,
+    pub original_text: Option<String>,
+    pub corrected_text: Option<String>,
+    pub context_json: Option<String>,
+    pub profile_id: String,
+    pub created_at: String,
+}
+
+impl From<FeedbackRow> for FeedbackDto {
+    fn from(r: FeedbackRow) -> Self {
+        Self {
+            id: r.id,
+            target_type: r.target_type,
+            target_id: r.target_id,
+            rating: r.rating,
+            original_text: r.original_text,
+            corrected_text: r.corrected_text,
+            context_json: r.context_json,
+            profile_id: r.profile_id,
+            created_at: r.created_at,
+        }
+    }
+}
+
+fn parse_target_type(raw: &str) -> Result<FeedbackTargetType, String> {
+    FeedbackTargetType::parse(raw).ok_or_else(|| format!("unknown feedback target_type: {raw}"))
+}
+
+#[tauri::command]
+pub async fn record_feedback(
+    state: tauri::State<'_, AppState>,
+    input: RecordFeedbackInput,
+) -> Result<i64, String> {
+    let target_type = parse_target_type(&input.target_type)?;
+    db_record_feedback(
+        &state.db,
+        RecordFeedbackParams {
+            target_type,
+            target_id: input.target_id,
+            rating: input.rating,
+            original_text: input.original_text,
+            corrected_text: input.corrected_text,
+            context_json: input.context_json,
+            profile_id: input.profile_id,
+        },
+    )
+    .await
+    .map_err(|e| e.to_string())
+}
+
+#[tauri::command]
+pub async fn list_feedback_examples_cmd(
+    state: tauri::State<'_, AppState>,
+    target_type: String,
+    limit: Option<i64>,
+    min_rating: Option<i8>,
+    profile_id: Option<String>,
+) -> Result<Vec<FeedbackDto>, String> {
+    let target = parse_target_type(&target_type)?;
+    let limit = limit.unwrap_or(8).clamp(1, 64);
+    let min_rating = min_rating.unwrap_or(0).clamp(-1, 1);
+    let rows =
+        db_list_feedback_examples(&state.db, target, limit, min_rating, profile_id.as_deref())
+            .await
+            .map_err(|e| e.to_string())?;
+    Ok(rows.into_iter().map(FeedbackDto::from).collect())
+}

src-tauri/src/commands/mod.rs

@@ -1,6 +1,7 @@
 pub mod audio;
 pub mod clipboard;
 pub mod diagnostics;
+pub mod feedback;
 pub mod hardware;
 pub mod hotkey;
 pub mod live;

src-tauri/src/commands/tasks.rs

@@ -6,12 +6,14 @@ use serde::{Deserialize, Serialize};
 
 use uuid::Uuid;
 
+use kon_llm::prompts::FeedbackExample as LlmFeedbackExample;
 use kon_storage::{
     complete_subtask_and_check_parent as db_complete_subtask, complete_task as db_complete_task,
     delete_task as db_delete_task, get_task_by_id as db_get_task,
     insert_subtask as db_insert_subtask, insert_task as db_insert_task,
-    list_subtasks as db_list_subtasks, list_tasks as db_list_tasks,
-    uncomplete_task as db_uncomplete_task, update_task as db_update_task, TaskRow,
+    list_feedback_examples as db_list_feedback_examples, list_subtasks as db_list_subtasks,
+    list_tasks as db_list_tasks, uncomplete_task as db_uncomplete_task,
+    update_task as db_update_task, FeedbackRow, FeedbackTargetType, TaskRow,
 };
 
 use crate::AppState;
@@ -166,6 +168,34 @@ pub async fn uncomplete_task_cmd(
         .map_err(|e| e.to_string())
 }
 
+/// Convert HITL feedback rows fetched from storage into the few-shot
+/// exemplar shape the LLM crate consumes. We reconstruct the `input`
+/// (parent task text, transcript chunk) from `context_json` where the
+/// recorder has stored it. Rows without usable input are dropped —
+/// the prompt builder filters them too, but doing it here keeps the
+/// exemplar list tight and the prompt budget predictable.
+fn to_llm_examples(rows: Vec<FeedbackRow>) -> Vec<LlmFeedbackExample> {
+    rows.into_iter()
+        .filter_map(|r| {
+            let ctx: serde_json::Value =
+                serde_json::from_str(r.context_json.as_deref().unwrap_or("{}")).ok()?;
+            let input = ctx
+                .get("input")
+                .and_then(|v| v.as_str())
+                .map(str::to_string)
+                .unwrap_or_default();
+            if input.trim().is_empty() {
+                return None;
+            }
+            Some(LlmFeedbackExample {
+                input,
+                original_output: r.original_text,
+                corrected_output: r.corrected_text,
+            })
+        })
+        .collect()
+}
+
 #[tauri::command]
 pub async fn decompose_and_store(
     state: tauri::State<'_, AppState>,
@@ -176,12 +206,23 @@ pub async fn decompose_and_store(
         .map_err(|e| e.to_string())?
         .ok_or_else(|| format!("Task {parent_task_id} not found"))?;
 
+    // Pull recent micro-step feedback so the system prompt gets
+    // conditioned on the user's preferred decomposition style. We
+    // cap at 5 examples to keep the prompt under budget regardless
+    // of how much feedback has been captured.
+    let examples = db_list_feedback_examples(&state.db, FeedbackTargetType::MicroStep, 5, 0, None)
+        .await
+        .map(to_llm_examples)
+        .unwrap_or_default();
+
     let engine = state.llm_engine.clone();
     let parent_text = parent.text.clone();
-    let steps = tokio::task::spawn_blocking(move || engine.decompose_task(&parent_text))
-        .await
-        .map_err(|e| e.to_string())?
-        .map_err(|e| e.to_string())?;
+    let steps = tokio::task::spawn_blocking(move || {
+        engine.decompose_task_with_feedback(&parent_text, &examples)
+    })
+    .await
+    .map_err(|e| e.to_string())?
+    .map_err(|e| e.to_string())?;
 
     let mut created = Vec::new();
     for text in steps {
@@ -205,8 +246,14 @@ pub async fn extract_tasks_from_transcript_cmd(
     state: tauri::State<'_, AppState>,
     transcript: String,
 ) -> Result<Vec<String>, String> {
+    let examples =
+        db_list_feedback_examples(&state.db, FeedbackTargetType::TaskExtraction, 5, 0, None)
+            .await
+            .map(to_llm_examples)
+            .unwrap_or_default();
+
     let engine = state.llm_engine.clone();
-    tokio::task::spawn_blocking(move || engine.extract_tasks(&transcript))
+    tokio::task::spawn_blocking(move || engine.extract_tasks_with_feedback(&transcript, &examples))
         .await
         .map_err(|e| e.to_string())?
         .map_err(|e| e.to_string())

src-tauri/src/lib.rs

@@ -288,6 +288,9 @@ pub fn run() {
             commands::tasks::extract_tasks_from_transcript_cmd,
             commands::tasks::list_subtasks_cmd,
             commands::tasks::complete_subtask_cmd,
+            // HITL feedback (Phase 2 roadmap)
+            commands::feedback::record_feedback,
+            commands::feedback::list_feedback_examples_cmd,
             // Profiles + profile terms (canonical SQLite-backed profile CRUD) — Task 12
             commands::profiles::list_profiles_cmd,
             commands::profiles::get_profile_cmd,

src/lib/components/FocusTimer.svelte

@@ -15,8 +15,26 @@
   // follows the sensory-zone theme switcher in Settings.
 
   import { onMount, onDestroy } from "svelte";
-  import { X, Plus } from "lucide-svelte";
+  import { X, Plus, ExternalLink } from "lucide-svelte";
   import { focusTimer } from "$lib/stores/focusTimer.svelte.js";
+  import { hasTauriRuntime } from "$lib/utils/runtime.js";
+
+  // Hide the "pop out" button inside the float window itself — opening
+  // a second float from a float would be silly and would re-mount the
+  // same component. Detect via URL rather than a prop so we do not
+  // have to thread context through every mount site.
+  let isSecondaryWindow = $state(false);
+  if (typeof window !== "undefined") {
+    isSecondaryWindow = window.location.pathname.startsWith("/float")
+      || window.location.pathname.startsWith("/viewer");
+  }
+
+  function handlePopOut() {
+    // Mirror the button in TasksPage.svelte — opens the always-on-top
+    // Now list + pinned timer in one floating window.
+    if (!hasTauriRuntime()) return;
+    window.open("/float", "_blank", "width=380,height=520");
+  }
 
   const RING_SIZE = 64;
   const RING_STROKE = 5;
@@ -148,6 +166,16 @@
         >
           <Plus size={14} aria-hidden="true" />
         </button>
+        {#if !isSecondaryWindow}
+          <button
+            class="icon-btn"
+            onclick={handlePopOut}
+            aria-label="Pop out timer + Now list into floating window"
+            title="Pop out (keeps timer + tasks on top)"
+          >
+            <ExternalLink size={14} aria-hidden="true" />
+          </button>
+        {/if}
         <button
           class="icon-btn"
           onclick={handleCancel}

src/lib/components/MicroSteps.svelte

@@ -1,8 +1,8 @@
 <script lang="ts">
   import { invoke } from '@tauri-apps/api/core';
-  import { ListTree, Check, Timer, Loader2 } from 'lucide-svelte';
+  import { ListTree, Check, Timer, Loader2, ThumbsUp, ThumbsDown, Pencil } from 'lucide-svelte';
 
-  let { parentTaskId, reduceMotion = false } = $props();
+  let { parentTaskId, parentTaskText = '', reduceMotion = false } = $props();
 
   interface Subtask {
     id: string;
@@ -15,6 +15,15 @@
   let error = $state('');
   let decomposing = $state(false);
 
+  // Per-step UI state. Keyed by subtask id so we never lose state when
+  // the list reorders. Values:
+  //   rating[id] — 1 | -1 — the thumbs vote the user gave this session
+  //   editing[id] — true while the user is editing the step text
+  //   draft[id]   — the in-flight edit value before save
+  let rating = $state<Record<string, 1 | -1 | undefined>>({});
+  let editing = $state<Record<string, boolean>>({});
+  let draft = $state<Record<string, string>>({});
+
   async function loadSubtasks() {
     loading = true;
     error = '';
@@ -53,6 +62,88 @@
     }));
   }
 
+  // --- HITL feedback --------------------------------------------------------
+  //
+  // All three paths (thumbs up, thumbs down, correction-via-edit) route
+  // into the same `record_feedback` command. The parent task text is the
+  // "input" the AI was given, so it travels in context_json so the prompt
+  // builder can reconstruct the (input, good-output) pair.
+
+  function feedbackContextJson() {
+    return JSON.stringify({ input: parentTaskText ?? '' });
+  }
+
+  async function recordThumb(step: Subtask, ratingValue: 1 | -1) {
+    // Toggle: if the user already voted the same way, clear it (record
+    // rating 0 means correction, not a thumb-off — we just skip the
+    // re-record and drop the local highlight). Unvoting isn't stored;
+    // the audit trail stays immutable.
+    if (rating[step.id] === ratingValue) {
+      const next = { ...rating };
+      delete next[step.id];
+      rating = next;
+      return;
+    }
+    rating = { ...rating, [step.id]: ratingValue };
+    try {
+      await invoke('record_feedback', {
+        input: {
+          targetType: 'microstep',
+          targetId: step.id,
+          rating: ratingValue,
+          originalText: step.text,
+          correctedText: null,
+          contextJson: feedbackContextJson(),
+        },
+      });
+    } catch (_) { /* feedback capture is best-effort, never fatal */ }
+  }
+
+  function startEdit(step: Subtask) {
+    editing = { ...editing, [step.id]: true };
+    draft = { ...draft, [step.id]: step.text };
+  }
+
+  function cancelEdit(stepId: string) {
+    const nextE = { ...editing }; delete nextE[stepId]; editing = nextE;
+    const nextD = { ...draft }; delete nextD[stepId]; draft = nextD;
+  }
+
+  async function saveEdit(step: Subtask) {
+    const next = (draft[step.id] ?? '').trim();
+    cancelEdit(step.id);
+    if (!next || next === step.text) return;
+    const original = step.text;
+    // Update in-memory first so the UI is snappy; roll back if the
+    // persistence call fails so we never show stale-but-different text.
+    const idx = subtasks.findIndex(s => s.id === step.id);
+    if (idx >= 0) subtasks[idx] = { ...subtasks[idx], text: next };
+    try {
+      await invoke('update_task_cmd', {
+        id: step.id,
+        patch: { text: next },
+      });
+      // Record correction as the highest-value feedback signal.
+      await invoke('record_feedback', {
+        input: {
+          targetType: 'microstep',
+          targetId: step.id,
+          rating: 0,
+          originalText: original,
+          correctedText: next,
+          contextJson: feedbackContextJson(),
+        },
+      }).catch(() => {});
+    } catch (_) {
+      if (idx >= 0) subtasks[idx] = { ...subtasks[idx], text: original };
+    }
+  }
+
+  function handleEditKeydown(evt: KeyboardEvent, step: Subtask) {
+    if (evt.key === 'Enter') { evt.preventDefault(); saveEdit(step); }
+    else if (evt.key === 'Escape') { evt.preventDefault(); cancelEdit(step.id); }
+  }
+
   $effect(() => {
     if (parentTaskId) loadSubtasks();
   });
@@ -97,10 +188,64 @@
             <Check size={9} aria-hidden="true" />
           {/if}
         </button>
-        <span class="text-[12px] flex-1 min-w-0 {step.done ? 'line-through text-text-tertiary' : 'text-text-secondary'} truncate">
-          {step.text}
-        </span>
-        {#if !step.done}
+
+        {#if editing[step.id]}
+          <!-- svelte-ignore a11y_autofocus — deliberate: inline edit
+               is user-initiated and focus must land on the input to
+               match the UX pattern users expect from any task app. -->
+          <input
+            type="text"
+            bind:value={draft[step.id]}
+            onkeydown={(e) => handleEditKeydown(e, step)}
+            onblur={() => saveEdit(step)}
+            class="text-[12px] flex-1 min-w-0 bg-bg-input border border-accent rounded px-1.5 py-0.5 text-text focus:outline-none"
+            autofocus
+            data-no-transition
+          />
+        {:else}
+          <button
+            type="button"
+            class="text-[12px] flex-1 min-w-0 {step.done ? 'line-through text-text-tertiary' : 'text-text-secondary'} truncate text-left cursor-text bg-transparent border-0 p-0"
+            ondblclick={() => !step.done && startEdit(step)}
+            disabled={step.done}
+            aria-label="Double-click to edit this step"
+            title="Double-click to edit"
+          >{step.text}</button>
+        {/if}
+
+        {#if !step.done && !editing[step.id]}
+          <!-- HITL feedback: thumbs vote + pencil edit. All three
+               route into record_feedback and feed the prompt-conditioning
+               loop. See docs/roadmap/2026-04-23-... Phase 2. -->
+          <button
+            class="opacity-0 group-hover:opacity-100 p-0.5 text-text-tertiary hover:text-success
+              {rating[step.id] === 1 ? '!opacity-100 text-success' : ''}"
+            onclick={() => recordThumb(step, 1)}
+            aria-label={rating[step.id] === 1 ? 'Remove thumbs up' : 'Thumbs up — this is a good step'}
+            title="Thumbs up — train the model on this style"
+            style={reduceMotion ? '' : 'transition: opacity var(--duration-ui), color var(--duration-ui)'}
+          >
+            <ThumbsUp size={10} aria-hidden="true" />
+          </button>
+          <button
+            class="opacity-0 group-hover:opacity-100 p-0.5 text-text-tertiary hover:text-danger
+              {rating[step.id] === -1 ? '!opacity-100 text-danger' : ''}"
+            onclick={() => recordThumb(step, -1)}
+            aria-label={rating[step.id] === -1 ? 'Remove thumbs down' : 'Thumbs down — this misses the mark'}
+            title="Thumbs down — avoid this style"
+            style={reduceMotion ? '' : 'transition: opacity var(--duration-ui), color var(--duration-ui)'}
+          >
+            <ThumbsDown size={10} aria-hidden="true" />
+          </button>
+          <button
+            class="opacity-0 group-hover:opacity-100 p-0.5 text-text-tertiary hover:text-accent"
+            onclick={() => startEdit(step)}
+            aria-label="Edit this step (the correction trains future suggestions)"
+            title="Edit — this is the strongest training signal"
+            style={reduceMotion ? '' : 'transition: opacity var(--duration-ui)'}
+          >
+            <Pencil size={10} aria-hidden="true" />
+          </button>
           <button
             class="opacity-0 group-hover:opacity-100 flex items-center gap-1 text-[10px] text-text-tertiary hover:text-accent"
             onclick={() => startTimer(step.id)}

src/lib/components/WipTaskList.svelte

@@ -107,7 +107,7 @@
           </div>
           <!-- Micro-steps panel (expanded) -->
           {#if expandedTaskIds.has(task.id)}
-            <MicroSteps parentTaskId={task.id} />
+            <MicroSteps parentTaskId={task.id} parentTaskText={task.text} />
           {/if}
         </div>
       {/each}

src/routes/float/+layout@.svelte

@@ -12,6 +12,7 @@
     PREFERENCES_CHANGED_EVENT,
   } from "$lib/stores/preferences.svelte.js";
   import Titlebar from "$lib/components/Titlebar.svelte";
+  import FocusTimer from "$lib/components/FocusTimer.svelte";
   import { loadOsInfo, isLinux } from "$lib/utils/osInfo.js";
 
   let { children } = $props();
@@ -90,3 +91,9 @@
     {@render children()}
   </div>
 </div>
+
+<!-- Focus timer also visible in the always-on-top float window so a
+     running countdown stays with the Now list. The component is a
+     global overlay (position: fixed) so it pins to the top-right of
+     this window independent of the Tasks content below. -->
+<FocusTimer />