#20675: fix(minimax): enable tool-use/result repair and drop malformed assistant turns

## Problem MiniMax has the same strict `tool_call`/`tool_result` pairing requirement as Anthropic and Google, but `repairToolUseResultPairing` and `allowSyntheticToolResults` were never enabled for it in `resolveTranscriptPolicy`. This caused sessions to fail with: > **Jinja template rendering failed: Message has tool role, but there was no previous assistant message with a tool call!** Two root causes compound the issue: 1. **Missing repair flag** — orphaned/displaced `toolResult` messages are never moved or synthetically filled, so MiniMax sees a tool-result turn with no preceding tool-call turn. 2. **MiniMax API bug** — MiniMax sometimes returns `finish_reason: "tool_calls"` (`stopReason=toolUse`) with no actual function-call blocks in the content. These text-only turns end up immediately before a real tool-calling assistant turn, creating consecutive assistant messages. MiniMax's Jinja2 template then misattributes the subsequent tool results to the wrong turn. The comment in `session-transcript-repair.ts` already named MiniMax explicitly as a strict provider — the flag just wasn't set. ## Changes **`src/agents/transcript-policy.ts`** - Add `MINIMAX_MODEL_HINTS` constant and `isMiniMaxModel()` helper (mirrors existing `isMistralModel()` pattern) - Enable `repairToolUseResultPairing` for MiniMax - Enable `allowSyntheticToolResults` for MiniMax **`src/agents/session-transcript-repair.ts`** (in `repairToolCallInputs`) - Drop assistant messages with `content.length === 0` — these are always error artifacts from failed/interrupted API calls and create consecutive assistant turns that strict providers reject - Drop assistant messages where `stopReason === "toolUse"` but no tool-call blocks are present — handles the MiniMax API bug where it signals tool intent but emits no function calls ## Testing - [x] Verified locally against a corrupted MiniMax session that was producing the Jinja error — MiniMax replies successfully after the fix - [x] Linter clean (`pnpm check`) - [ ] Unit tests (the repair logic is covered by existing tests; no new test cases added — happy to add if maintainers prefer) ## AI-assisted Built with Claude (Sonnet 4.6). I understand what the code does and verified it locally. <!-- greptile_comment --> <h3>Greptile Summary</h3> Enables transcript repair for MiniMax provider to handle strict tool-call/tool-result pairing requirements. After MiniMax switched to `anthropic-messages` API (PR #15297), it inherited Anthropic's strict turn ordering constraints but repair mechanisms were never enabled. The fix adds: - MiniMax detection helper in `transcript-policy.ts` (mirrors existing `isMistralModel()` pattern) - Enables `repairToolUseResultPairing` and `allowSyntheticToolResults` for MiniMax - Drops malformed assistant messages (empty content arrays or `stopReason=toolUse` with no actual tool calls) to prevent consecutive assistant turns that trigger Jinja validation errors Changes follow established patterns for Anthropic/Google providers and address documented MiniMax API bugs where `finish_reason: "tool_calls"` is returned without function-call content. <h3>Confidence Score: 4/5</h3> - Safe to merge with minimal risk - addresses a documented bug with defensive repair logic - The implementation follows established patterns for Anthropic/Google strict providers and addresses a real failure mode (Jinja template errors). The defensive repair logic is sound but lacks test coverage for MiniMax-specific scenarios. The changes are scoped and well-documented, though comprehensive unit tests would strengthen confidence. - No files require special attention - both changes are straightforward and follow existing patterns _{Last reviewed commit: 3b58ee6} <!-- greptile_other_comments_section --> _{(3/5) Reply to the agent's comments like "Can you suggest a fix for this @greptileai?" or ask follow-up questions!} <!-- /greptile_comment -->