llama.cpp b11538, InvalidRequestException (400 instead of 500), agent follow-ups closed

[bernardladenthin]

Summary

• llama.cpp b11534 → b11538 (one step, 45 KiB): #30229 (CUDA round fix), #30160 (graph reorder for get_rows), #30228 (UI Models Manager: GET /models gains context_length; common_get_decision_type replaced by common_get_gguf_info(), which no project C++ called), #28331 (llama-bench -fitc). All 11 patches apply unchanged, the drop-checks are unchanged, the two history rows are written. RouterModel.getContextLength() exposes the new router field (0 from an older server or a GGUF without the key); the value.* PIT gate at 100 % is covered by RouterModelTest / RouterModelsResponseParserTest.
• C2: InvalidRequestException (a LlamaException) from the JNI layer, and OpenAiCompatServer answers 400 instead of 500 for a request the native layer rejects. Upstream's server has answered these with 400 since b11337 (#29060); the Java transport reported server_error, because the JNI layer threw the same LlamaException for a rejected request and a failed inference. Native side: c_llama_invalid_request next to c_llama_error, chosen by three helpers with upstream's own rules — throw_invalid_request for request parsing (also behind parse_json_params and require_json_field), throw_by_exception_type (std::invalid_argument / common_json_error → 400, the rule of upstream's ex_wrapper; tokenization, format_prompt_infill, the /v1/systemone handler), and error_class_for for a failed task result via the new json_helpers get_result_error_code (400 for ERROR_TYPE_INVALID_REQUEST and ERROR_TYPE_EXCEED_CONTEXT_SIZE). The message of a rejected request is the bare reason now, no longer the JSON object the old throw_invalid_request wrapped it in (CHANGELOG "Changed"). Java side: the server catches the typed exception together with the IllegalArgumentException of its own validation at the ten request sites, on every surface (OpenAI, Ollama, Anthropic, Responses; an invalid_request_error SSE event once a stream has started); a plain LlamaException stays 500. The Server layer may now see the Exception layer (LlamaArchitectureTest).
• D1/D2 (agent follow-ups) were already done; the TODO entries were stale. The approval gate exists (ApprovalMode, ConsoleApprovalStrategy, AgentRunner.approval with Atmosphere's ToolApprovalPolicy, --auto; pinned by ApprovalWireTest) — entry removed. Tool rounds across REPL turns ride as AgentSession.toolNote in front of the next user message, and the asked-for form (full ChatMessage list incl. tool_calls) is impossible on atmosphere-ai 4.0.72: AbstractAgentRuntime.assembleMessages rebuilds every history entry as new ChatMessage(role, content) (checked in the bytecode) — entry rewritten with that reason and the reopen condition. docs/handover/local-agent-pr-a.md stays with a status note, because deliverable 6 carries the local agent's B2 measurements and the four OPEN items the Windows CPU-variants job still needs; TODO.md "CPU variants" points at it.

Test plan

• Affected unit / integration tests pass locally — C++ 607/607 (ctest, +4 for get_result_error_code); Java suite model-free 1845 run / 0 failures / 285 skipped (-Dnet.ladenthin.llama.test.ngl=0; +6 model-free HTTP tests: typed rejection is 400 for chat, embeddings, completions, Ollama and Anthropic, an SSE error event when streaming, a plain LlamaException 500); ErrorHandlingTest tightened to the typed exception plus an empty input array and a malformed body; model-backed 400 cases in OpenAiServerEmbeddingsIntegrationTest ("input":[], unknown encoding_format) and OpenAiServerCompletionIntegrationTest (no prompt) — these need CI's models; spotless:check, compile spotbugs:check (0), javadoc:jar (0 errors, 0 warnings), check-natives.py, check-patches.py, check-release-gate.py, check-run-scripts.py, check-shared-files.py, the 113 buildcheck unit tests, reuse lint all green; clang-format 23.1.3 applied.
• CI is green on this branch
• Docs / CHANGELOG updated where applicable — CHANGELOG (Added / Changed / Fixed), CLAUDE.md (pin, "The JNI exception boundary", the json_helpers function list, the test table, the FindClass list), TODO.md, docs/history/llama-cpp-breaking-changes.md, Javadoc on every LlamaModel entry point that can throw the new exception.

Related issues / PRs

Follows #496, #497 and #498. Closes the TODO.md entries "400 vs. 500 for an invalid request body", "Approval for destructive tools" and (rewritten) "Tool rounds are not carried across REPL turns".

Checklist

• I have read CONTRIBUTING.md and CODE_OF_CONDUCT.md
• My commits follow Conventional Commits
• No security-sensitive changes (if there are, I have notified the maintainer privately per SECURITY.md)

🤖 Generated with Claude Code

https://claude.ai/code/session_01AytmJF9faEiQEVt6eetQS2

---

Generated by Claude Code

[sonarqubecloud]

Quality Gate failed

Failed conditions
61.9% Coverage on New Code (required ≥ 80%)

See analysis details on SonarQube Cloud