fix(session): preserve Anthropic thinking for errored same-model replay (fixes #38620)

[optamus-ai]

Issue for this PR

Closes #38620

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Fixes V2 replay failure where errored assistant messages lose Anthropic thinking metadata while retaining tool_use, producing invalid follow-up requests when extended thinking is enabled.

Root cause: packages/core/src/session/runner/to-llm-message.ts:73 required message.error === undefined for reuseProviderMetadata. An aborted/failed step therefore dropped all reasoning provider state (anthropic.signature / anthropic.redactedData) while still replaying tool_use parts. Anthropic requires tool_use to be preceded by its thinking/redacted_thinking block when thinking is enabled → 400.

Fix (minimal, preserves replay safety): Relax guard only when same-model and message contains Anthropic thinking. For sameModel && (error === undefined || hasAnthropicThinking), reuse provider metadata so signed/redacted thinking is retained alongside tool_use. Preserves deliberate guard for other providers (existing session-runner-message.test.ts expects OpenAI drop) and does not silently discard valid provider state. Alternative of dropping tool calls would discard valid state.

Preserved: sameModel check, ToolOutput handling, SessionCompaction token estimation, security/permission unchanged.

How did you verify your code works?

• Reproduced with deterministic test: errored Anthropic thinking sig_test + tool_use with same model → before fix, reasoning lost signature; after fix, both preserved and reasoningIdx < toolIdx.
• Verified existing session-runner-message.test.ts still passes (6 tests), including drops provider-native continuation metadata from failed assistant turns for OpenAI (still drops).
• Added session-replay-38620.test.ts with 3 cases: errored Anthropic same-model preserved, errored OpenAI still drops, different model drops even for Anthropic — all pass via bun test.
• git diff — 1 file to-llm-message.ts 12+1, 1 new test file 131 lines, Buffer.byteLength not needed here.
• Real git clone --depth 1, checkout -b, commit 4bd9bbc, push to optamus-ai/opencode, gh pr create — history df35e84 preserved.

Screenshots / recordings

N/A — non-UI replay invariant.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Based on my search, I found several related PRs that address similar issues around Anthropic thinking preservation and session replay:

Potential related PRs:

1. PR fix(core): keep reasoning provider metadata on errored turns for valid Anthropic replay #44054 - fix(core): keep reasoning provider metadata on errored turns for valid Anthropic replay

   • Directly addresses keeping reasoning metadata on errored turns for Anthropic replay — this appears to be the closest match to PR fix(session): preserve Anthropic thinking for errored same-model replay (fixes #38620) #45951

2. PR fix(core): drop tool parts when a failed turn loses its reasoning block #40157 - fix(core): drop tool parts when a failed turn loses its reasoning block

   • Relates to handling tool parts when reasoning blocks are lost during failed turns

3. PR fix(core): replay thinking safely and drop unsettled tool calls from errored messages #42735 - fix(core): replay thinking safely and drop unsettled tool calls from errored messages

   • Addresses replaying thinking and managing tool calls in errored messages

4. PR fix: preserve thinking/redacted_thinking blocks in Anthropic message transforms #23755 - fix: preserve thinking/redacted_thinking blocks in Anthropic message transforms

   • Related to preserving thinking/redacted_thinking blocks in Anthropic messages

5. PR fix(opencode): drop truncated reasoning from replayed history #40148 - fix(opencode): drop truncated reasoning from replayed history

   • Addresses reasoning handling in replay scenarios

Why they may be related: These PRs all tackle similar problems in the session replay pipeline involving Anthropic's extended thinking feature, provider metadata preservation, and handling of tool_use parts alongside reasoning blocks during failed/errored turns.

PR #44054 appears to be the most directly related, as it shares the exact same goal of preserving reasoning provider metadata on errored turns for valid Anthropic replay scenarios.