After answering OpenCode question, composer looks idle while tools keep running

[satyalyadav]

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/web

Steps to reproduce

1. Start a thread with OpenCode as harness.
2. Let the agent ask a question with the OpenCode question tool.
3. Answer it. In my run the question was which baseline to use and I picked "Same model, all providers (Recommended)".
4. Watch the timeline and composer right after the answer.

Expected behavior

After I answer, the thread keeps showing working until the turn ends. New tool activity appears in the timeline and the composer stays in a working state.

Actual behavior

After I answer, the timeline shows "User input submitted Same model, all providers (Recommended)" and a collapsed "Used 1 tool and ran 3 commands" row, then the composer says "Ask anything" and looks enabled. No assistant text streams. It looks dead or already replied. The backend is still busy and keeps starting tools.

Impact

Major degradation or frequent failure. I cannot tell if work continues without checking logs, so I may send a duplicate message or kill a healthy run.

Version or commit

t3 v0.0.42 (t3 --version, WSL runtime server on port 3773)
OpenCode 1.18.33 (opencode --version)
Node v24.18.0, Bun 1.3.13

Environment

OS: Ubuntu 24.04.4 LTS
Workspace: /home/satyal, project global
Thread title: Benchmark NVIDIA GLM 5.3 Flash
Thread ID: 5a32c675-900a-417f-bac9-085a7121725b
OpenCode session: ses_f153ceb3bffezezZTSN2UGO4gv
Model selection: opencode/muse-spark-1.3-contributor-free, variant xhigh, agent build
Runtime mode: full-access, interaction mode default

Local DB state at time of check (~/.t3/userdata/state.sqlite):
projection_thread_sessions for this thread: status running, active_turn_id opencode-turn-9a9ee513-0339-4b14-89b4-42d4dd64798a, updated_at 2026-09-29T01:47:20.389Z, provider_instance_id opencode.
projection_threads: pending_approval_count 0, pending_user_input_count 0.

Logs, stack traces, or screenshots

Provider log ~/.t3/userdata/logs/provider/events.5a32c675-900a-417f-bac9-085a7121725b.log:

• 2026-09-29T01:25:25Z question.asked, tool question, state running, plus user-input.requested
• 2026-09-29T01:25:47Z item.completed for the question with answer "Same model, all providers (Recommended)"
• 2026-09-29T01:25:51Z onward new item.started for bash, todowrite, write
• session.status stayed busy through at least 2026-09-29T01:29:23Z, no turn.completed in the log window

Screenshot from the app shows the initial benchmark prompt, then "User input submitted Same model, all providers (Recommended)", then collapsed tool rows, then an enabled "Ask anything" composer. I can attach the screenshot and a sanitized log slice on request. Full logs contain local paths so I did not paste them here.

Workaround

Check provider logs or state.sqlite to confirm the session is still running, instead of trusting the composer state. Refresh was not needed here since the run was healthy, only the status display was wrong.

Notes

OpenCode resumed correctly in this run, so this is not anomalyco/opencode #36382 where the backend hangs after Confirm. Related T3 Code items that look distinct: PR #12900 fixes the "User input submitted" label hiding the question, #9318 covers Submit being swallowed around pending input, #1048 and #2886 cover stale turn state where the UI says working after the backend is done. This report is the reverse, idle looking composer while busy.

[juliusmarminge]

Triage

This matches current main as well as your 0.0.42 build, and it is not one of the threads you listed.

The timeline gap is in the work log. OpenCode maps a tool part that is still pending to item.started, and only emits item.updated once that part is running (apps/server/src/provider/Layers/OpenCodeAdapter.ts, message.part.updated). The web work log then drops every tool.started activity and only renders tool.updated / tool.completed (apps/web/src/session-logic.ts, deriveWorkLogEntries). So the item.started rows in the provider log for bash, todowrite, and write do not create timeline rows. Until those parts move to running or completed, the thread stays on the pre-answer collapsed group ("Used 1 tool and ran 3 commands") plus the folded "User input submitted …" row. That label is current behavior. #12900 only changes it to show the question text. It does not keep the turn looking live.

The composer is a separate signal. The placeholder stays "Ask anything…" for both an idle thread and a running one (apps/web/src/components/chat/ChatComposer.tsx). Working is the stop control, and derivePhase shows it only when session.status === "running" (apps/web/src/session-logic.ts). With no live tool row, a running session should also append a Thinking row (apps/web/src/components/chat/MessagesTimeline.logic.ts). projection_thread_sessions.status = running with an active turn, and no turn.completed, is exactly that state: the session is set to running on turn.started and back to ready only on turn.completed / turn.aborted. If the screenshot has neither Stop nor Thinking, the client phase was not running even though SQLite was, and that second gap is still open. A typed send while Stop is showing queues behind the current turn. It does not start a second one.

Not the same as #9318 (Submit swallowed while a question is still pending; here pending_user_input_count is 0), #1048 / #2886 (UI stuck working after the provider is done), or anomalyco/opencode#36382 (you already confirmed OpenCode kept running).

Please drop the screenshot and a sanitized slice of events.5a32c675-900a-417f-bac9-085a7121725b.log from 01:25:25Z through the first post-answer item.updated or item.completed, if any. The useful check is whether Stop and Thinking were actually absent, and whether those bash/todowrite/write parts ever left pending.

[satyalyadav]

Sanitized slice from events.5a32c675-900a-417f-bac9-085a7121725b.log, 01:25:25Z through first post-answer item.completed. IDs shortened, bash output truncated. Checked for tokens, keys, and auth material. None present. Local home path left as ~ in one command.

Timeline:

• 01:25:25.079Z question.asked, id que_...FnJK: "Which OpenCode baseline do you want for the NVIDIA GLM 5.3 Flash speed test?", header "Baseline models", 3 options, multiple false, tool call call_...eb2f
• 01:25:25.081Z user-input.requested, requestId que_...FnJK
• 01:25:46.947Z user-input.resolved, answers: question-0-baseline-models = "Same model, all providers (Recommended)"
• 01:25:47.002Z question.replied, answers [["Same model, all providers (Recommended)"]]
• 01:25:47.002Z message.part.updated, tool question, state status completed, title "Asked 1 question", output "User has answered your questions..."
• 01:25:46.940Z item.completed, itemType dynamic_tool_call, status completed, title "Asked 1 question", tool question
• 01:25:47.126Z session.status busy
• 01:25:51.943Z item.started, itemType command_execution, title bash, tool bash, state status pending, input {}
• 01:25:59.521Z item.completed, tool bash, state status completed. Input command was a read only probe, truncated here: ls -la ~/.local/share/opencode/log/, grep for opencode URLs in config, opencode --help | head -n 80. Output truncated to first file listing lines plus public opencode.ai and npm registry URLs. No secrets.
• 01:25:59.646Z and later session.status still busy. No turn.completed in this window.

Answers to triage checks:

• Yes, those post-answer bash parts start as pending with empty input, then move to completed at 01:25:59.521Z.
• Screenshot from the report shows no Stop control and no Thinking row, only "User input submitted Same model, all providers (Recommended)", a collapsed "Used 1 tool and ran 3 commands" group, and an enabled "Ask anything" composer.
• pending_user_input_count was 0 in projection_threads at check time, projection_thread_sessions.status was running with active turn opencode-turn-9a9ee513-0339-4b14-89b4-42d4dd64798a.

[kvnloo]

I think there are two independent UX contracts worth pinning here:

1. tool.started is already durable evidence that work exists. The work log should render a live/pending row for it instead of waiting for tool.updated / tool.completed.
2. Answering a question is not a terminal turn transition. If the canonical session/turn is still running, the composer + timeline must continue to project working state even when no assistant text is currently visible.

That gives a useful negative control: user-input.resolved → item.started(pending) → session.status busy must never render like ready.

I would keep the fixes at those owners rather than adding an OpenCode-specific spinner: lifecycle events own activity visibility; canonical session/turn state owns whether the thread is still working.

[kvnloo]

Two independent gaps from the trace now have draft fixes:

• fix(web): show tool.started rows while work is pending #14225 — tool.started is visible immediately as inProgress and collapses in place into later lifecycle updates. This removes the post-answer blank timeline while OpenCode tools are still only pending.
• fix(web): keep running phase while a turn is still active #14226 — composer/queue phase remains running when canonical turn ownership still exists (session.activeTurnId or latestTurn.state === running) even if one session-status projection temporarily reads ready.

Both are provider-neutral. No OpenCode-specific spinner/state was added: lifecycle events own work visibility; session/turn ownership owns whether the thread is still working.

#12900 remains complementary: it improves the question/answer label but does not own liveness.