server: remember a tools-aware visible KV checkpoint after tool-call turns

[Flor1an-B]

Fixes the root cause reported in #691 (and the same-shaped tool-continuation gap discussed in #609/#611, for the OpenAI-compatible chat/completions path once a conversation actually has tools).

Root cause

chat_history_uses_tool_context() makes render_deepseek_chat_prompt_text() always keep the opening <think> tag for a historical assistant turn once any tool schema is present — <think>{reasoning}</think> — rendering {reasoning} from whatever reasoning_content the client echoes back for that turn. Most OpenAI-compatible chat/completions clients never replay reasoning_content, so that renders as a literal <think></think>, while the live KV session actually contains the real sampled reasoning bytes for that turn.

The existing visible-checkpoint mechanism already handles exactly this mismatch for the toolless case (remember_thinking_checkpoint() / should_remember_thinking_checkpoint()), but both are explicitly gated off whenever request.has_tools is true, and the tool-call-finish path (canonicalize_tool_checkpoint(), or the raw-DSML-replay fallback when should_canonicalize_tool_checkpoint() is false) only ever calls thinking_live_clear() afterwards. So a tools-enabled conversation never gets a live-continuation match once a non-replaying client's rendered history diverges from the live tokens — which happens on essentially every turn, since thinking_live_visible_prefix_prompt() (the matching side) is already engine-agnostic and untouched by has_tools; only the remembering side was gated off.

Fix

Mirrors the toolless mechanism instead of introducing a new one:

• should_remember_thinking_checkpoint() only excludes has_tools for GLM, whose history renderer keeps the opening <think> tag either way (append_glm_assistant_message_prefix()) and isn't affected by this bug.
• New build_tool_context_thinking_visible_text() mirrors build_toolless_thinking_visible_text() but keeps the opening <think> tag, matching what chat_history_uses_tool_context() will actually render for this turn once it becomes history (the toolless helper strips it, which is correct only when tool context is absent).
• New remember_tool_thinking_checkpoint() remembers prompt_text + build_tool_checkpoint_suffix(..., reasoning=NULL, ...) as the visible key after a tool-call finish, instead of unconditionally clearing it. A client that does replay real reasoning still gets an ordinary token-prefix hit first (before this fallback is ever consulted), so this is purely additive — it doesn't change behavior for clients that already work today.

Both new builders/functions are DeepSeek/DSML-syntax only (gated on model_syntax != SERVER_MODEL_SYNTAX_GLM); GLM keeps its current toolless-only behavior unchanged, since I haven't verified this mechanism against GLM's history-rendering shape and didn't want to guess at a model family I can't test.

Testing

• Full existing ds4_test suite passes unchanged.
• Two new regression tests: test_tool_context_thinking_visible_text_keeps_think_tag (verifies the two builders produce different shapes for the identical request, which is the whole point of not reusing the toolless one) and test_tool_checkpoint_suffix_omits_reasoning (verifies the remembered suffix omits reasoning while still reproducing the exact tool-call DSML).
• Updated test_thinking_checkpoint_remember_gate to cover the new GLM/non-GLM split explicitly.

Happy to add a live A/B log-based repro (thinking live continuation match=visible-prefix vs token-mismatch) if useful, similar to what was posted on #611 — didn't include one here since the existing regression tests already pin down the exact byte-shape logic directly.

🤖 Generated with Claude Code