V2 TUI silently succeeds on empty length turns until manual compaction #52015

Activity

1. github-actions

   assigned

   nexxeln

   on Sep 29, 2026
2. github-actions commented on Sep 29, 2026

   

   github-actionsBot

   on Sep 29, 2026 – with GitHub Actions

   Contributor

   More actions

   ---

   This issue might be a duplicate of existing issues. Please check:

   • Empty response with finish_reason "length" and zero tokens is treated as success #50949: Same finish_reason: "length" empty/zero-output response is treated as a successful turn with no visible error.
   • bug: recover context-limited length completions before durable output #44609: Context-limited length completion before assistant output, with compaction and retry proposed as recovery.
   • anthropic: adaptive-thinking turn silently ends with max_tokens and zero output tokens when reasoning exhausts the 32k budget #51517: In v2.0.18, reasoning exhausts the token budget and produces a length finish with zero visible output while the session succeeds.
   • Turns truncated at the output limit are recorded as normal completions #40146: Broader report of truncated length turns being recorded as normal completions, especially when no usable output was produced.
   • v2: empty reasoning-only response is recorded as successful completion #37372: Related V2 report where an assistant response with no usable output is recorded as successful.
3. Veld101 commented on Sep 29, 2026

   

   Veld101

   on Sep 29, 2026

   More actions

   This is #50949 (already flagged as the duplicate). A fix is open in #51741.

   The change: a normalized length finish with no streamed output now fails the Step with a visible error instead of settling as a success, so an empty turn stops looking like a healthy one. Partial output is unaffected - a length finish that streamed any content still settles normally with its visible text - and a length finish with tool calls still continues.

   Your detail that manual compaction recovers the session is consistent with the same code path: nothing marks the empty turn as failed, so there is no reason to surface an error or recover automatically. #51741 makes the failure visible; it does not add automatic recovery, which is the separate ask tracked in #44609.
4. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   More actions

   Thanks for the detailed report. I reproduced this with a simulated model whose responses always finish with reason length and contain no text, reasoning or tool calls.

   • v2.0.18: each prompt produces one assistant message with finish: "length", empty content and 0 output tokens. The session goes idle, and the TUI shows no answer and no error. No automatic compaction is attempted.
   • Latest dev (86ce19d96): the output-limit continuation added in feat(core): continue responses after output token limits #53876 runs, so each prompt now produces three empty length turns. The TUI still shows no error, and compaction still isn't attempted.

   The recording below shows the v2.0.18 run, and the screenshot shows the empty turns on dev. The script below reproduces it.

   Related work: open PRs #56673 and #51741 both handle turns that finish with no output, and #44609 proposed compacting and retrying in this case. Until then, a manual compaction like the one you used is the workaround.

   recording.mp4

   

   Reproduction script (drive.ts)

   // Reproduces #52015: a provider step that finishes with finish_reason "length"
   // and no content is settled as a successful, silent turn.
   import { Effect, Stream } from "effect"
   import { Llm, OpenCodeDriver } from "opencode-drive"
   
   export default OpenCodeDriver.use(
     {
       ...(process.env.OC_DEV ? { opencode: { dev: process.env.OC_DEV } } : {}),
       tui: { recording: true },
     },
     ({ ui, llm, opencode }) =>
       Effect.gen(function* () {
         let calls = 0
         // Every model request: empty stream, finish reason "length", no text/reasoning/tool calls.
         yield* llm.serve(() => {
           calls++
           return Stream.make(Llm.finish("length"))
         })
         yield* ui.screenshot("01-home")
         yield* ui.submit("Summarize the project status")
         yield* Effect.sleep("6 seconds")
         yield* ui.screenshot("02-after-first-prompt")
         yield* ui.submit("hello?")
         yield* Effect.sleep("6 seconds")
         yield* ui.screenshot("03-after-second-prompt")
         yield* Effect.log(`model requests served: ${calls}`)
         // Dump the stored assistant messages: finish, tokens and any error.
         const sessions: any = yield* opencode.session.list()
         for (const session of (sessions.items ?? sessions.data ?? sessions) as any[]) {
           const messages: any = yield* opencode.message.list({ sessionID: session.id } as any)
           for (const m of (messages.items ?? messages.data ?? messages) as any[])
             if (m.type === "assistant")
               yield* Effect.log(JSON.stringify({ finish: m.finish, tokens: m.tokens, error: m.error, content: m.content }))
         }
       }),
   )

   Run it with opencode-drive: opencode-drive run ./drive.ts. It drives opencode from your PATH; set OC_DEV to the path of an OpenCode checkout to run that from source instead.
5. opencodejr

   added

   reproduced

   on Oct 28, 2026