ai: gpt-5.6-sol still dies with "reasoning start before end: rs_...:1" on 2.0.14 after #45789

[vvo]

Summary

gpt-5.6-sol turns still die with reasoning start before end: rs_...:1 on 2.0.14, which includes the fix from #45789 (for #36892 / #43312). The turn fails as an untyped Failed to drain Session defect, idle_outcome is failed, and the TUI shows an empty assistant message with a cursor and no error. 7 crashes across 4 sessions in 5 days, all on openai/gpt-5.6-sol#medium through the Vercel AI Gateway provider with reasoningSummary: "auto". Removing reasoningSummary is the only workaround found.

Environment

• opencode version: 2.0.14 (also reproduced on 0.0.0-beta-19086, beta channel)
• OS: Darwin 25.6.0 (darwin arm64)
• Terminal: iTerm.app, TERM=xterm-256color, COLORTERM=truecolor
• Shell: /bin/zsh
• Install/channel: Homebrew anomalyco/tap/opencode-v2, latest channel
• Active plugins: two local TUI plugins (opencode-cost-details, opencode-prs, both from github.com/vvo/opencode-plugins), -opencode.sidebar.context disabled. Crashes happen in the server, plugins are TUI-only.

Reproduction

Not deterministic. Roughly 1 crash per 20-30 sol turns in long sessions (100+ messages, 200k+ context).

1. Global config:

   "providers": {
     "vercel": {
       "settings": { "gateway": { "caching": "auto" } },
       "models": {
         "openai/gpt-5.6-sol": {
           "settings": { "reasoningEffort": "medium", "reasoningSummary": "auto" }
         }
       }
     }
   }

2. Run a long session on vercel/openai/gpt-5.6-sol#medium.
3. Keep prompting. Eventually a turn ends immediately after the reasoning starts.

Expected Behavior

Either the overlapping summary part is normalized (close the open fragment, keep going) or the turn fails with a typed provider error that the TUI renders, as #45789 intended.

Actual Behavior

Server log (2.0.14):

level=ERROR message="Failed to drain Session"
cause="Error: reasoning start before end: rs_00db537c08807b5a016ab28c9ed7e887d1a46353013482c653:1
    at <anonymous> (/$bunfs/root/chunk-bam104a1.js:350:1615)
    at runLoop (/$bunfs/root/chunk-cp4m0mhv.js:24:2279)
    at evaluate (/$bunfs/root/chunk-cp4m0mhv.js:24:1518)
    at <anonymous> (/$bunfs/root/chunk-cp4m0mhv.js:24:6004)
    at processTicksAndRejections (native:7:39)
    at SessionRunner.publishLLMEvent (/$bunfs/root/chunk-bam104a1.js:350:11556)
    at SessionRunner.publishLLMEvent (definition) (/$bunfs/root/chunk-bam104a1.js:350:9184)
    at SessionStep.attempt (...)"
sessionID=ses_f376c02cdffe9hOZjBNXdDlU4D

Stored assistant message: two reasoning parts, the first with empty text and only a generationId, the second with text and time.completed. No time.completed on the message, no error, no finish. Next stored record is {"type":"idle","outcome":"failed"}.

"content": [
  { "type": "reasoning", "text": "", "state": { "generationId": "gen_..." },
    "time": { "created": 1790086302993, "completed": 1790086303014 } },
  { "type": "reasoning", "text": "**Updating service integration plan**\n\n...",
    "time": { "created": 1790086303027, "completed": 1790086303157 } }
]

The TUI shows the user prompt, one "Thought:" line, then an empty assistant block with a cursor. No error, no retry.

Additional Context

All occurrences from ~/.local/share/opencode/log/opencode.log, every one on openai/gpt-5.6-sol variant medium, every fragment id ends in :1:

time (UTC)	version	session	fragment
2026-09-17 21:34	beta-19086	ses_f4eb62340ffe...	rs_...:1
2026-09-18 08:05	beta-19086	ses_f4c7d5da5ffe...	rs_...:1
2026-09-18 08:15	beta-19086	ses_f4c7d5da5ffe...	rs_...:1
2026-09-18 09:31	beta-19086	ses_f4c637183ffe...	rs_...:1
2026-09-22 11:35	beta-19086	ses_f376c02cdffe...	rs_...:1
2026-09-22 11:52	beta-19086	ses_f376c02cdffe...	rs_...:1
2026-09-22 14:11	2.0.14	ses_f376c02cdffe...	rs_...:1

Same sessions also ran claude-fable-5.1, gpt-6-astra and glm-5.3-flash turns with zero crashes. So it is sol summary parts, not context size alone.

Relation to #36892 and #43312: same error string and same publish-llm-event die path. #45789 landed in 2.0.14 (merge 074413a), so the missing-boundary normalization it added does not cover this interleaving. The empty first reasoning part with only a generationId looks like output_item.added opened :0, then reasoning_summary_part.added for index 1 arrived while :0 was still active. The provider stream is not retained so I cannot say whether the Vercel AI Gateway reorders the events or OpenAI emits them that way.

Workaround: removing reasoningSummary: "auto" from the sol model settings. No crash since, at the cost of not seeing reasoning summaries.

Happy to run a debug build or add stream capture if that helps.

[rekram1-node]

vercel ai gateway still uses the legacy aisdk route, should be fixed in next v2 release