Record verdicts, promote lessons and choose round lessons in the TUI (#174)

CHANGELOG.md

@@ -37,7 +37,16 @@ All notable changes to SyncMill are documented here.
   checkpoint and before the task, with the excerpts and memory context shrinking
   to make room. A resumed conversation keeps what earlier rounds sent it, so a
   round that would resume one in a room with earlier lessons adds a note about
-  `--fresh-session`. TUI-launched rounds do not send lessons yet.
+  `--fresh-session`.
+- The TUI Rooms tab records verdicts and promotes lessons (#174): `j` records
+  accepted/rejected on an applied room, bound to the patch digest the dialog
+  showed; `p` promotes a lesson from a judged room and refuses if the verdict
+  changed while the dialog was open (`RoomCoordinator.promote_lesson` gains
+  `expected_judgment_id`). The room pane shows the outcome. TUI rounds now send
+  the project's approved lessons: the round dialog lists them with checkboxes,
+  applies the CLI's skip and budget rules live, and refuses to start when the
+  prompt leaves no room for the checked lessons; the history note shows there,
+  before you confirm, for the lineage you pick.
   The round conversation record is now `schema_version` 2: it adds the
   selected lesson ids and each member's delivery, shown by `room compare`,
   `room export` and the TUI (`rounds[].conversation.lessons`,
@@ -70,6 +79,16 @@ All notable changes to SyncMill are documented here.
 
 ### Fixed
 
+- The TUI no longer exits when the board's decisions or the selected item
+  cannot be read during a Board-tab refresh; it shows `DEGRADED` and keeps
+  the last view, as it already did for the item list.
+- A TUI round, verdict or promotion acts on the room its dialog was opened
+  for. A round dialog used to launch on whichever room was selected when you
+  confirmed, which a refresh could have changed meanwhile.
+- A verdict recorded from any surface now marks an open TUI Compare or
+  Details snapshot as changed; a verdict changes neither the room's revision
+  nor its candidates, so the snapshot used to keep showing the old outcome
+  without saying so.
 - `syncmill start` suggests verification commands only from files whose
   name matches exactly, so each suggestion names a file that is on disk. On a
   case-insensitive filesystem (the macOS default) a repository with only a

CLAUDE.md

@@ -31,7 +31,7 @@ uv run syncmill room export <room-id> --include patch:<cand> [--include output:|
 # room rounds record their effective starting tree + gate settings (v2 adds the gate's env_allowlist: null = inherited, [] distinct) in rounds.provenance_json (rooms.round_provenance_status, writer and readers share it, readers accept v1+v2; NULL = legacy); compare/export/TUI show them as starting tree / gate settings beside the report-derived gate profile, whose JSON fields are unchanged; a v1 round's environment reads not recorded and never groups as same settings; `room apply --verify` still runs the current `[gate]` and prints a `verify:` notice (`rooms.verify_settings_notice`) unless the round's v2 record equals it
 uv run syncmill room select <room-id> <cand> [--expected-revision N --expected-digest SHA256 --rationale TEXT]   # verifies the artifact, records a bound selection (rooms.selection_json); apply refuses an inconsistent record
 uv run syncmill room judge <room-id> accepted|rejected [--expected-digest SHA256 --rationale TEXT]   # #174: append-only human verdict on the APPLIED candidate (room_judgments, bound to candidate + patch digest + applied run; never inferred from select/apply, never touches memory, no state/revision change); rooms.room_outcome_status is the one read model (not_selected / selected_not_applied / applied_unreviewed / accepted / rejected / unknown — a damaged latest record is unknown, never an older verdict) shown by compare/export as room.outcome
-uv run syncmill room promote <room-id> --lesson TEXT [--scope GOAL --supersedes ID]   # #174 PR 2: needs the board; only a judged room (accepted/rejected); a board decision bound to project_id with provenance_json (room/judgment/verdict/candidate/digest/run) and a unique source_key (retry returns the existing decision, never reactivates a superseded one); `run` shows then injects Board.project_lessons (project-wide scope only; board.select_lessons budget; --skip-lesson/--no-lessons, forwarded in _inner_argv) via Supervisor.run(lessons=) AFTER routing classified the raw prompt; board items get unbound + own-project decisions; room rounds: `room round`/`start` show them (cli._show_round_lessons, rooms.round_lessons_fit) and pass RoundLesson pairs to run_round(lessons=) (detached: a second anonymous fd, --lessons-fd); the coordinator never opens the board; lessons sit after the checkpoint, before the task, excerpts budgeted around them; TUI rounds send none (note only)
+uv run syncmill room promote <room-id> --lesson TEXT [--scope GOAL --supersedes ID]   # #174 PR 2: needs the board; only a judged room (accepted/rejected); a board decision bound to project_id with provenance_json (room/judgment/verdict/candidate/digest/run) and a unique source_key (retry returns the existing decision, never reactivates a superseded one); `run` shows then injects Board.project_lessons (project-wide scope only; board.select_lessons budget; --skip-lesson/--no-lessons, forwarded in _inner_argv) via Supervisor.run(lessons=) AFTER routing classified the raw prompt; board items get unbound + own-project decisions; room rounds: `room round`/`start` show them (cli._show_round_lessons, rooms.round_lessons_fit) and pass RoundLesson pairs to run_round(lessons=) (detached: a second anonymous fd, --lessons-fd); the coordinator never opens the board; lessons sit after the checkpoint, before the task, excerpts budgeted around them; TUI (Rooms `j` verdict / `p` promote, `tui.py` `_confirm_room_judge`/`_confirm_lesson_promote`): the round dialog reads `project_lessons` once on the Textual thread (`RoundLessonsView`), recomputes `select_lessons(skip=unchecked)` live, refuses (never drops) when `round_lessons_fit` fails on `round_prompt`'s fallback rule, and launches with exactly those pairs; promote passes `expected_judgment_id` (checked on the same outcome read as the provenance); the round/verdict/promote dialog callbacks act on the captured room id + store root
 uv run syncmill conversations list
 uv run syncmill conversations show <run-id>
 uv run syncmill conversations close <run-id>

docs/architecture/tui.md

@@ -86,6 +86,29 @@ the store root changes or the store becomes unavailable. Both
 views read rows from the snapshot (no artifact read on the event loop), so their
 gate/eligibility/usage/blocker wording is the CLI's; the Output view (`o`)
 interprets none of it.
+Rooms review actions (#174): `j` records a verdict (`RoomCoordinator.judge`,
+`surface="tui"`, fenced on the patch digest the dialog showed), `p` promotes a
+lesson (`promote_lesson(expected_judgment_id=)`, refused when the latest
+judgment is no longer the one the dialog showed), and the `x` round dialog
+shows the project's approved lessons (`RoundLessonsView`, read once when it
+opens) and recomputes the CLI's `select_lessons(skip=)` on every checkbox,
+prompt or lineage change. It refuses instead of dropping lessons when the
+prompt leaves no room (`round_prompt` gives the fit check and the launch one
+fallback rule), and the worker gets exactly the shown pairs. The round, verdict
+and promote dialogs capture `(room id, store root)` when they open; its callback acts on those and
+refuses when the store was rebound, never reading the ambient selection. Judge
+and promote run on the Textual thread on the existing connections (SQLite is
+thread-affine; the board connection must not enter a worker), accepting up to
+the stores' 5 s busy timeout of blocking under write contention. The evidence
+snapshot also records the room's judgment count, read before the build: a
+verdict changes neither revision nor candidates, so `changed` compares the
+count too; a failed count read, at the build or now, degrades to a "cannot
+tell" line instead of aborting the staged room update. Checkbox labels built from stored text go
+through `plain(...)` (Textual parses a `str` label as markup); the
+`markup_sinks` guard flags any other non-literal label. A lesson's checkbox
+label is only its id: a `Checkbox` label never wraps, so a long lesson would be
+cut off at the dialog edge while every word of it is sent. The full text is a
+wrapping `PlainStatic` under it.
 EVERY poll path is degrade-guarded (`_ROOM_POLL_ERRORS`/`_SESSION_POLL_ERRORS`/
 `_JOB_POLL_ERRORS`) — an unguarded read escapes the interval callback and kills
 the app, so a store failure must show DEGRADED and keep the last good view.

docs/user-guide.ko.md

@@ -888,8 +888,11 @@ Room 라운드에도 같은 프로젝트 전체 교훈이 들어갑니다. `room
 내용이 그대로 남습니다. 교훈을 빼면 다시 보내지 않을 뿐, 이미 보낸 것을 거둘 수는 없습니다.
 이전 라운드에 교훈이 있던 방에서 대화를 이어 가는 라운드라면 목록에 `--fresh-session`을
 안내하는 문장이 붙습니다. 이 안내는 목록을 보여 준 시점의 방 상태를 기준으로 하므로,
-detached worker가 시작하기 전에 끝난 라운드는 반영하지 않습니다. TUI에서 시작한 라운드는 아직 교훈을 보내지 않고 같은 안내를
-보여 줍니다. `room compare`는 라운드가 고른 교훈과 후보마다 교훈이 들어갔는지를 보여 줍니다.
+detached worker가 시작하기 전에 끝난 라운드는 반영하지 않습니다. TUI에서 시작한 라운드도
+교훈을 보냅니다. 라운드 대화상자가 교훈마다 체크박스를 보여 주고(체크 해제 = `--skip-lesson`),
+예산 때문에 빠진 교훈을 표시하며, 프롬프트가 체크한 교훈을 담을 자리를 남기지 않으면 교훈을
+빼고 진행하는 대신 시작을 거부합니다. 이전 교훈 안내는 고른 lineage에 맞춰 대화상자 안에
+나옵니다. `room compare`는 라운드가 고른 교훈과 후보마다 교훈이 들어갔는지를 보여 줍니다.
 
 `room apply --verify`는 다른 `--verify`와 마찬가지로 라운드에 기록된 설정이 아니라 현재
 `[gate]`로 검증합니다. 기록을 다시 쓰면 지금 설정한 것보다 넓은 환경이 verify에 넘어갈 수
@@ -1065,7 +1068,13 @@ Rooms의 `a`는 prompt와 member만 기록하며 provider를 호출하지 않습
 잘림 상태). 완료된 후보는 Diff로, 실패한 후보는 Output으로 열려요. `R`은
 `room recover`로 **프로젝트 전체**의 죽은 room 워커/중단된 apply를 복구하고,
 `k`는 선택한 room을
-닫습니다(후보와 영수증은 보존).
+닫습니다(후보와 영수증은 보존). room 화면에는 결과(outcome)가 나옵니다. `j`는 적용된 room에
+대한 판정을 기록합니다(`room judge`: accepted 또는 rejected를 고르고 이유를 덧붙일 수 있으며,
+판정은 대화상자에 보인 patch digest에 묶입니다). `p`는 판정된 room에서 교훈을 승격합니다
+(`room promote`, board 필요). 대화상자는 교훈에 기록될 출처를 보여 주고, 열려 있는 동안 판정이
+바뀌었으면 거부합니다. `x`의 라운드 대화상자는 프로젝트의 승인된 교훈을 보여 줍니다(Rooms 참고).
+이 대화상자들은 그사이 선택이 바뀌어도 처음 연 room에만 적용됩니다. 어디서든 판정이 기록되면
+열려 있는 Compare/Details 스냅샷에 바뀌었다는 표시가 붙습니다.
 
 실행은 Textual이 아니라 분리된 백그라운드 매니저가 맡아요. `q`는 화면만 닫고 실행은
 계속합니다. 다시 `syncmill tui`를 열면 실행 상태와 로컬 rolling 로그에 재연결돼요.

docs/user-guide.md

@@ -964,8 +964,11 @@ conversation keeps whatever earlier rounds sent it: leaving a lesson out stops i
 being sent again, it cannot take it back. When a round would resume a conversation
 in a room whose earlier rounds had lessons, the list adds a note pointing to
 `--fresh-session`. The note reflects the room as of the listing: a round that
-finishes before a detached worker starts is not taken into account. Rounds started from the TUI do not send lessons yet and show
-the same note. `room compare` shows which lessons a round selected and, per
+finishes before a detached worker starts is not taken into account. Rounds started from
+the TUI send lessons too: the round dialog lists them with a checkbox each (unchecking is
+`--skip-lesson`), names any left out by the budget, and refuses to start, instead of
+dropping them, when the prompt leaves no room for the checked ones; the history note
+appears in the dialog for the lineage you pick. `room compare` shows which lessons a round selected and, per
 candidate, whether they were included.
 
 `room apply --verify` runs the current `[gate]`, like every `--verify`, not the
@@ -1175,7 +1178,15 @@ code, per-check gate results, files, digests, and truncation state). A
 completed candidate opens on Diff; a failed one opens on Output. `R` recovers
 dead room workers/interrupted applies for the whole project via `room recover`,
 and `k` closes the
-selected room (candidates and receipts are preserved).
+selected room (candidates and receipts are preserved). The room pane shows its
+outcome. `j` records your verdict on an applied room (`room judge`: you pick
+accepted or rejected, add an optional rationale, and the verdict is bound to the
+patch digest the dialog showed). `p` promotes a lesson from a judged room
+(`room promote`, needs the board); the dialog shows the source the lesson will
+record and refuses if the verdict changed while it was open. `x` lists the
+project's approved lessons in the round dialog (see Rooms). These dialogs act on
+the room they were opened for, even if the selection moves meanwhile. A verdict
+recorded anywhere marks an open Compare or Details snapshot as changed.
 
 Runs are owned by a detached manager, not Textual. Only one TUI-managed worker per
 board is allowed, but ordinary external `board run` processes remain legal. Pressing

src/syncmill/rooms.py

@@ -2959,6 +2959,15 @@ def list_judgments(self, room_id: str) -> list[dict[str, Any]]:
         ).fetchall()
         return [dict(row) for row in rows]
 
+    def judgment_count(self, room_id: str) -> int:
+        """How many judgment rows the room holds: a cheap freshness signal for
+        a view built from an earlier read (a verdict changes neither the
+        room's revision nor its candidates)."""
+        row = self._conn.execute(
+            "SELECT COUNT(*) FROM room_judgments WHERE room_id=?", (room_id,)
+        ).fetchone()
+        return int(row[0])
+
     def outcome(self, room_id: str) -> OutcomeCheck:
         """The room's human outcome, read in one snapshot."""
         with self._read_snapshot() as conn:
@@ -4242,13 +4251,18 @@ def promote_lesson(
         *,
         scope: str = "",
         supersedes: str | None = None,
+        expected_judgment_id: str | None = None,
     ) -> tuple[Any, bool]:
         """Promote a short, human-written lesson from a judged room into the
         board's decisions, bound to this project (#174).
 
         Only a room whose latest judgment binds (accepted or rejected) can be
         promoted: the lesson's provenance names that judgment, and a retry of
-        the same promotion returns the decision it already created."""
+        the same promotion returns the decision it already created.
+        ``expected_judgment_id`` refuses the promotion unless that judgment is
+        still the latest, checked on the same read the provenance comes from,
+        so a verdict recorded after the person reviewed it is never promoted
+        in its name."""
         text = normalize_lesson(lesson)
         if not isinstance(scope, str):
             raise RoomError("lesson scope must be text")
@@ -4263,6 +4277,14 @@ def promote_lesson(
                 f"room outcome is {outcome.status}{detail}; record a verdict with "
                 "`syncmill room judge` before promoting a lesson"
             )
+        if (
+            expected_judgment_id is not None
+            and judgment["judgment_id"] != expected_judgment_id
+        ):
+            raise RoomError(
+                "the room's verdict changed since you reviewed it; review the "
+                "promotion again"
+            )
         provenance = {
             "kind": "room-judgment",
             "room_id": room_id,