opencode/big-pickle: chain-of-thought leaks into ext part in long sessions (no separate easoning part) #56342

Summary

With opencode/big-pickle, assistant messages in long sessions store the model's chain-of-thought inside the text part instead of a separate reasoning part, so the internal planning becomes part of the visible answer (and of the stored transcript / title generation).

Measured over 21,181 assistant messages on a local opencode.db: 2,549 (12%) begin with planning text, and 100% of those cases are messages that have no reasoning part at all.

Environment

• opencode version: 2.0.32
• OS: Windows 11 Pro
• Terminal: VS Code integrated terminal
• Shell: PowerShell
• Install/channel: release

Reproduction

1. Set "model": "opencode/big-pickle" as default and work in a session with a large context (long conversation, large system prompt).
2. Ask for any task that requires planning.
3. Inspect the assistant message (TUI, API, or the stored parts): its content array contains a single type: "text" part whose text starts with the chain-of-thought in English (The user says ... Let me ... I need to ... I should ...), followed by the actual answer. There is no type: "reasoning" part for that message.
4. Same model, small prompt (2 messages): the reasoning is delivered separately — observed reasoning(84) + text(2).

Minimal deterministic repro not isolated yet; reproduced from stored sessions at scale, plus the small-prompt control.

Expected Behavior

The chain-of-thought should be delivered/persisted as a separate type: "reasoning" part (as happens with small prompts), keeping the text part limited to the visible answer.

Actual Behavior

In long sessions the reasoning is written into the text part. Consequences:

• the visible answer starts with internal planning;
• auto-generated session titles end up containing the reasoning (some titles literally begin with the model's planning sentence);
• transcript consumers that only look at text parts (e.g. tooling) see the model "talking to itself" about the rules.

Measurements (local DB)

Schema used (opencode v2 SQLite): message, part (data JSON with type), session_message (type=assistant, data.content[]), session_v2 (model).

Part types in part for all sessions: tool 28106 · step-start 26728 · step-finish 26078 · text 18894 · reasoning 14335 · patch 735 · compaction 248.

For opencode/big-pickle (238 sessions):

Correlation that isolates the symptom:

hard leak with reasoning = 0 on every day: when the reasoning part exists, the text part is clean; when it does not, planning shows up in text. On 2026-10-15 the ratio inverts (8 with vs 634 without).

Additional Context

• cli.json already has session.thinking: "hide", and opencode models reports opencode/big-pickle with variants: [] — there is no toggle to flip.
• The model name/cost entry: opencode/big-pickle, variants: [], cost 0 in/out.
• Workarounds tried locally: (1) a hard prompt rule "never emit reasoning in the visible body" — reduces but does not eliminate; (2) a plugin that detects planning markers in the text part and injects a warning into the system prompt — works as a post-hoc guard, but the current v2 hook set has no way to rewrite the visible text (experimental.text.complete no longer exists), so the offending turn is already sent.
• Question: is splitting expected from the provider adapter (e.g. reasoning_content / reasoning items) or from core when persisting parts? If the opencode provider needs an explicit option for this case (reasoningSummary, include, textVerbosity, …), please point to it — happy to test.

Suggested labels

bug, windows