Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: restore Deep Research plan approval (#10134)

Wasim Yousef Said71b7b30672b4
diff --git a/studio/backend/core/research_runs.py b/studio/backend/core/research_runs.py
index 6a5fcff4d80..060c8e619b3 100644
--- a/studio/backend/core/research_runs.py
+++ b/studio/backend/core/research_runs.py
@@ -1916,8 +1916,6 @@ class ResearchSupervisor:
preview_labels = True,
)
plan = _parse_and_validate_plan(response, planning_reasoning, max_steps)
- # Arming research in the composer is the approval, so the plan is queued as it is
- # stored rather than parked for a second confirmation.
try:
result = await asyncio.to_thread(
db.set_plan,
@@ -1925,7 +1923,6 @@ class ResearchSupervisor:
plan,
None,
self.worker_id,
- auto_approve = True,
)
except db.ResearchConflictError:
if await asyncio.to_thread(db.is_cancel_requested, run["id"]):
diff --git a/studio/backend/storage/research_runs_db.py b/studio/backend/storage/research_runs_db.py
index 34b6a05ebf5..44f7779d475 100644
--- a/studio/backend/storage/research_runs_db.py
+++ b/studio/backend/storage/research_runs_db.py
@@ -652,18 +652,9 @@ def set_plan(
plan: dict,
expected_revision: int | None = None,
worker_id: str | None = None,
- auto_approve: bool = False,
) -> dict:
- """Store the plan and either park the run for review or queue it in the same write.
-
- Arming research in the composer is the approval, so the worker queues its own plan with
- ``auto_approve``. Done in one transaction: a separate approve call left a window where
- the run sat at awaiting_approval, flashing the review card and letting a concurrent plan
- edit fail the run. The endpoint still parks a hand-edited plan for the user to confirm.
- """
raw, digest = canonical_plan(plan)
steps = plan.get("steps") or []
- status = "queued" if auto_approve else "awaiting_approval"
conn = get_connection()
try:
conn.execute("BEGIN IMMEDIATE")
@@ -690,9 +681,9 @@ def set_plan(
revision += 1
conn.execute(
"UPDATE research_runs SET plan_json = ?, plan_revision = ?, plan_hash = ?, "
- "status = ?, error_message = NULL, lease_owner = NULL, "
+ "status = 'awaiting_approval', error_message = NULL, lease_owner = NULL, "
"lease_expires_at = NULL, updated_at = ? WHERE id = ?",
- (raw, revision, digest, status, now_ms(), run_id),
+ (raw, revision, digest, now_ms(), run_id),
)
conn.execute("DELETE FROM research_plan_steps WHERE run_id = ?", (run_id,))
conn.executemany(
@@ -707,14 +698,12 @@ def set_plan(
run_id,
"plan.ready",
{
- "status": status,
+ "status": "awaiting_approval",
"plan": plan,
"planRevision": revision,
"planHash": digest,
},
)
- if auto_approve:
- _event_locked(conn, run_id, "run.approved", {"status": status})
_commit_event(conn)
return {"plan": plan, "planRevision": revision, "planHash": digest}
except Exception:
diff --git a/studio/backend/tests/test_deep_research_handoff_simulation.py b/studio/backend/tests/test_deep_research_handoff_simulation.py
index 8c06c3fee30..6ec858b0097 100644
--- a/studio/backend/tests/test_deep_research_handoff_simulation.py
+++ b/studio/backend/tests/test_deep_research_handoff_simulation.py
@@ -339,7 +339,7 @@ def test_the_handed_off_question_is_what_actually_gets_researched(research_home,
assert RAW_MESSAGE not in prompts[0].split("Latest research request:")[-1]
-def test_planning_no_longer_stops_to_ask_for_approval(research_home, monkeypatch):
+def test_planning_waits_for_plan_approval_before_research(research_home, monkeypatch):
from core import research_runs as worker
research_db.create_run(
@@ -371,7 +371,11 @@ def test_planning_no_longer_stops_to_ask_for_approval(research_home, monkeypatch
claimed = research_db.claim_next(supervisor.worker_id)
asyncio.run(supervisor._plan(claimed))
- assert research_db.get_run("run-1")["status"] == "queued"
+ planned = research_db.get_run("run-1")
+ assert planned["status"] == "awaiting_approval"
+ assert research_db.claim_next("worker-2") is None
+ assert research_db.approve("run-1", planned["planRevision"], planned["planHash"]) == "queued"
+ assert research_db.claim_next("worker-2") is not None
# ── What the change must not break ────────────────────────────────
diff --git a/studio/backend/tests/test_research_runs_storage.py b/studio/backend/tests/test_research_runs_storage.py
index 1b00bdc774f..764fa883d78 100644
--- a/studio/backend/tests/test_research_runs_storage.py
+++ b/studio/backend/tests/test_research_runs_storage.py
@@ -1780,14 +1780,14 @@ def test_supervisor_planning_and_research_are_durable_with_mocked_io(
planning = research_db.claim_next(supervisor.worker_id)
asyncio.run(supervisor._process(planning))
planned = research_db.get_run("run-1")
- # Planning approves itself: arming research in the composer is the approval, so there is
- # no second confirmation between the plan and the evidence gathering.
- assert planned["status"] == "queued"
+ assert planned["status"] == "awaiting_approval"
assert planned["planRevision"] == 1
assert planned["assistantMessageId"] is None
+ assert research_db.claim_next(supervisor.worker_id) is None
+ assert research_db.approve("run-1", planned["planRevision"], planned["planHash"]) == "queued"
running = research_db.claim_next(supervisor.worker_id)
- assert running is not None # planning released its lease; the queued run starts immediately
+ assert running is not None
asyncio.run(supervisor._process(running))
completed = research_db.get_run("run-1")
diff --git a/studio/frontend/src/features/chat/components/research-activity-panel.tsx b/studio/frontend/src/features/chat/components/research-activity-panel.tsx
index c33e95d844b..82e21edda6c 100644
--- a/studio/frontend/src/features/chat/components/research-activity-panel.tsx
+++ b/studio/frontend/src/features/chat/components/research-activity-panel.tsx
@@ -600,7 +600,7 @@ function PlanReview({ runId }: { runId: string }): ReactElement | null {
open={open}
onOpenChange={(nextOpen) => setOpen(runId, nextOpen)}
>
- <DialogContent className="max-h-[min(680px,calc(100dvh-6rem))] grid-rows-[auto_minmax(0,1fr)_auto] gap-0 overflow-hidden p-0 sm:max-w-3xl [&>[data-slot=dialog-close]]:right-6 [&>[data-slot=dialog-close]]:top-6">
+ <DialogContent className="max-h-[min(680px,calc(100dvh-6rem))] grid-rows-[auto_minmax(0,1fr)_auto] gap-0 overflow-hidden p-0 data-open:animate-none sm:max-w-3xl [&>[data-slot=dialog-close]]:right-6 [&>[data-slot=dialog-close]]:top-6">
<DialogHeader className="border-b border-border/70 px-7 pb-4 pt-6 pr-16">
<DialogTitle>Review the research plan</DialogTitle>
<DialogDescription className="max-w-2xl leading-relaxed">
diff --git a/studio/frontend/tests/research-activity-progress.test.ts b/studio/frontend/tests/research-activity-progress.test.ts
index 48d9453be6a..99a52672238 100644
--- a/studio/frontend/tests/research-activity-progress.test.ts
+++ b/studio/frontend/tests/research-activity-progress.test.ts
@@ -368,9 +368,12 @@ test("a stalled watcher does not stop events reaching the store", async () => {
const iterator = watchResearchRun(RUN_ID)[Symbol.asyncIterator]();
await iterator.next();
+ const plan = { title: "Plan", steps: [{ title: "One", query: "q" }] };
const planned = run({
status: "awaiting_approval",
- plan: { title: "Plan", steps: [{ title: "One", query: "q" }] },
+ plan,
+ planRevision: 1,
+ planHash: "plan-hash",
lastEventSeq: 1,
});
ingest(planned, event(1, "plan.ready", { planRevision: 1 }, planned));
@@ -379,6 +382,13 @@ test("a stalled watcher does not stop events reaching the store", async () => {
useResearchRunStore.getState().sessions[RUN_ID].run.status,
"awaiting_approval",
);
+
+ assert.deepEqual(useResearchRunStore.getState().planReviewByRunId[RUN_ID], {
+ revision: 1,
+ open: true,
+ editing: false,
+ draft: plan,
+ });
const next = await iterator.next();
assert.equal(next.value?.status, "awaiting_approval");
await iterator.return?.(undefined);