You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Codex goal resume fails with EmptyInput while native work continues #20232
Sending /goal resume in a blocked Codex goal thread shows “The provider could not start this turn.” Two attempts failed. Provider events continued after T3 marked the run failed.
Diagnosis
The logs show a native turn starting during thread/resume. T3 then sends another turn/start request with input: []. Codex rejects it with failed to submit turn input: EmptyInput.
The affected release’s runGoalCommand implementation passes an empty input array for resume. That branch also exists in inspected main commit e3b35d1b2440b104f0bc2e357377c348121234dd.
Steps to reproduce
Observed twice in the affected thread. A clean minimal reproduction has not been run.
Open a Codex thread with an existing blocked goal.
Send /goal resume.
Observe the startup error and inspect the provider log.
Expected: the goal resumes with a tracked run and consistent UI state.
Actual: the empty start request fails, while native work can continue behind a failed UI state.
Version
Observed on 0.0.46-nightly.20261015.3158, then again after updating to 0.1.1-nightly.20261016.3190. Provider initialization logs confirm both versions.
Environment
Windows x64, OS build 10.0.26200; T3 Code desktop with local server; Codex app-server 0.162.1; model gpt-6.1-sol.
Evidence
Selected fields from the provider log on 2026-10-16, UTC. Paths, identifiers, and unrelated payloads omitted:
Later provider events include command execution and active goal updates, while T3’s stored run state is failed.
Related issues
#18833: resumed native turns interfere with compaction tracking; different operation and error. PR #17589: an open fix for EmptyInput during automatic restart continuation. This report concerns explicit /goal resume. Issue and PR searches found no exact issue. No shipped fix was verified.
Fix applied or workaround
No fix applied. A normal continuation message is a proposed workaround, but it has not been verified. No local T3 source or database edits were made.
Filed by
Prepared by GPT-6.1-Sol through the Codex harness in T3 Code, following the official triage playbook in the existing session. The t3 triage launcher was not run.
In runGoalCommand (apps/server/src/orchestration-v2/Adapters/CodexAdapterV2.ts), every non-set command that reaches the activation step, including /goal resume, calls startNativeTurn(turnInput, []). The only flip to status: "active" comes after that turn/start. Codex rejects the empty turn/start with EmptyInput, so the run is marked failed before the goal is ever reactivated. Your log also shows Codex had already started a native turn during thread/resume, which then keeps running with no T3 run attached to it.
This is separate from #17589, which only retries EmptyInput on the restart-continuation path, and from #18833.
Likely fix direction: for resume, set the goal to active (or send a non-empty continuation) instead of starting a turn with empty input, and attach the run to any turn Codex has already started.
This still reproduces on 0.1.3-nightly.20261023.3579 (Windows desktop, local server) with Codex app-server 0.164.0. Main at 22dc1da still passes [] for non-set goal commands (CodexAdapterV2.ts L6515-6523).
Below are selected protocol log fields after a server restart, with two threads that both had an active goal. Only method, request id, input length and goal status are shown.
20:11:30.314 out thread/resume thread B
20:11:30.632 in thread/goal/updated status=active thread B
20:11:30.660 in turn/started status=inProgress thread B (no T3 request; Codex started it)
20:12:18.016 out thread/inject_items thread B
20:12:19.209 out thread/goal/get thread B
20:12:19.211 out turn/start input=[] id=16 thread B
20:12:19.214 in error -32603 "failed to submit turn input: EmptyInput" id=16
Thread A had the same /goal resume sequence and ended in EmptyInput at 20:12:56.850. Both threads went through T3's CodexAdapterV2 on the same restart.
The Codex 0.164.0 source shows why /goal resume fails every time while that turn is running:
turn/start tries to steer a running turn first. EmptyInput is returned only on that steer path (session/turn_input.rs#L860-L861). An empty turn/start on an idle thread starts a turn normally.
The native work is not lost. In this log, turn B ran to completion at 20:14:34 while T3 showed the run as failed. Adopting the already-running turn when the goal is active, as #20247 does, looks like the right fix for these logs. The restart-continuation side of the same race is covered in #21610.
Activity
juliusmarminge commentedon Oct 16, 2026
Note
Grok responding on behalf of Julius.
This is a bug, confirmed on main.
In
runGoalCommand(apps/server/src/orchestration-v2/Adapters/CodexAdapterV2.ts), every non-setcommand that reaches the activation step, including/goal resume, callsstartNativeTurn(turnInput, []). The only flip tostatus: "active"comes after thatturn/start. Codex rejects the emptyturn/startwithEmptyInput, so the run is marked failed before the goal is ever reactivated. Your log also shows Codex had already started a native turn duringthread/resume, which then keeps running with no T3 run attached to it.This is separate from #17589, which only retries
EmptyInputon the restart-continuation path, and from #18833.Likely fix direction: for resume, set the goal to
active(or send a non-empty continuation) instead of starting a turn with empty input, and attach the run to any turn Codex has already started.Thanks for the precise logs.
amanthanvi commentedon Oct 24, 2026
This still reproduces on
0.1.3-nightly.20261023.3579(Windows desktop, local server) with Codex app-server0.164.0. Main at 22dc1da still passes[]for non-setgoal commands (CodexAdapterV2.tsL6515-6523).Below are selected protocol log fields after a server restart, with two threads that both had an active goal. Only method, request id, input length and goal status are shown.
Thread A had the same
/goal resumesequence and ended inEmptyInputat 20:12:56.850. Both threads went through T3'sCodexAdapterV2on the same restart.The Codex 0.164.0 source shows why
/goal resumefails every time while that turn is running:thread/resume, anactivegoal is restored and its continuation turn starts on an idle thread. This isrestore_after_resume/continue_if_idle.turn/starttries to steer a running turn first.EmptyInputis returned only on that steer path (session/turn_input.rs#L860-L861). An emptyturn/starton an idle thread starts a turn normally.The native work is not lost. In this log, turn B ran to completion at 20:14:34 while T3 showed the run as failed. Adopting the already-running turn when the goal is active, as #20247 does, looks like the right fix for these logs. The restart-continuation side of the same race is covered in #21610.