fix(runtime,web): route inline think tags into reasoning lane #109

Merged

salmaelsoly merged 2 commits from integration_hide_inline_think_tags into integration

2026-06-11 12:53:21 +00:00

salmaelsoly commented

2026-06-11 12:16:22 +00:00

Member

Summary

Models that inline their reasoning as literal <think>...</think> tags in the assistant content were having those tags rendered as plain text in the chat. This change routes that content into the existing reasoning lane so it appears in the collapsible ThinkingPane, both live during streaming and after reload.

Closes #104

Changes

New think_tags module in hero_shrimp_runtime: incremental ThinkTagFilter for streaming (handles markers split across SSE chunks, held-back partial markers, unclosed trailing <think>) and split_think_tags for whole responses; 13 unit tests
Streaming path in llm/completion/mod.rs routes every content delta through the filter, emitting llm:delta with kind: "content" or kind: "reasoning"; non-streaming OpenAI-compatible path splits content and pushes extracted segments as ReasoningBlock entries; Anthropic branch and trajectory export untouched
Web UI fallback in store.ts (extractThinkBlocks applied in stream flush, turn:end, and history load) so messages already persisted with raw tags render clean
ChatThread.tsx: settled assistant messages with reasoning render a collapsed ThinkingPane
Regenerated embedded Vite bundle

Test Results

cargo test --workspace: 2740 passed, 0 failed, 18 ignored
cargo clippy --workspace -- -D warnings: clean

## Summary Models that inline their reasoning as literal `<think>...</think>` tags in the assistant content were having those tags rendered as plain text in the chat. This change routes that content into the existing reasoning lane so it appears in the collapsible ThinkingPane, both live during streaming and after reload. ## Related Issue Closes https://forge.ourworld.tf/lhumina_code/hero_shrimp/issues/104 ## Changes - New `think_tags` module in `hero_shrimp_runtime`: incremental `ThinkTagFilter` for streaming (handles markers split across SSE chunks, held-back partial markers, unclosed trailing `<think>`) and `split_think_tags` for whole responses; 13 unit tests - Streaming path in `llm/completion/mod.rs` routes every content delta through the filter, emitting `llm:delta` with `kind: "content"` or `kind: "reasoning"`; non-streaming OpenAI-compatible path splits content and pushes extracted segments as `ReasoningBlock` entries; Anthropic branch and trajectory export untouched - Web UI fallback in `store.ts` (`extractThinkBlocks` applied in stream flush, `turn:end`, and history load) so messages already persisted with raw tags render clean - `ChatThread.tsx`: settled assistant messages with reasoning render a collapsed ThinkingPane - Regenerated embedded Vite bundle ## Test Results - cargo test --workspace: 2740 passed, 0 failed, 18 ignored - cargo clippy --workspace -- -D warnings: clean

salmaelsoly added 1 commit

2026-06-11 12:16:23 +00:00