Repository navigation
llm codec tests: F25 REHOME rows 107 and 109 at 1.5.5 (stacked on #857) - #884
Merged
Merged
Conversation
…ni stop scan, candidates count, tool ids, transcription usage, Anthropic response_format) - Responses: a translated stream stamped created_at 0 (1970); it stamps the host's wall clock again, as 1.5.5 stamped now_unix_secs() (v1.5.5 proto/openai_responses/writer.rs:479). - Gemini/Cohere: the relay's stop-reason scan is structural (member at a container depth), so a model-written tool `args` object holding "finishReason":"MALFORMED_FUNCTION_CALL" no longer faults a shared lane. - Gemini: candidatesTokenCount is the whole output again, with no thoughtsTokenCount (v1.5.5 proto/gemini/writer.rs:962-964); streamed tool ids after text count the text block as 1.5.5 did (v1.5.5 proto/gemini/reader.rs:895), matching the buffered ids; a duration-billed transcription writes no invented usageMetadata.audioDurationSeconds (v1.5.5 handlers/gemini.rs:259-265). - Anthropic: a lane without native structured outputs drops response_format again (v1.5.5 proto/anthropic/writer.rs:343-347) instead of a synthetic forced tool that overwrote the caller's tool_choice and streamed back as a tool call. Unit goldens resp_r2g_reasoning.json returns to its pre-971e2b6508 bytes.
…rted to their codec homes Each legacy busbar-llm engine test is ported at equal strength (its literal figures) onto the served codec path its F25 v2 row names: Bedrock tool_choice none forwards and reports the drop (108); a native-structured-output Anthropic lane rides output_config.format and reports a schema-less json_object (110); an Anthropic tool id reaches an OpenAI client remapped and unleaked (350); the router shim keys strip per egress and the model survives strip-then-rewrite, with the same-dialect relay sending the caller bytes (504, 505); the Cohere stop-sequence cap clamps six to five and forwards five whole (613, 614).
…d on #857) Row 107 (egress_dropped_controls_audit_tests.rs:26) pinned the post-1.5.5 synthetic busbar_response_format forced tool; per ARCHITECT it is ported at 1.5.5: an OpenAI json_object onto a default Anthropic lane forwards with no response_format, tool_choice or tools and reports the dropped control response_format (v1.5.5 proto/anthropic/writer.rs:343-347). Row 109 (egress_dropped_controls_audit_tests.rs:146) ports the per-dialect dropped-control lists with the Anthropic default lane at 1.5.5 (response_format), Bedrock response_format + tool_choice=none, OpenAI none, and a native lane response_format.
…h sides appended tests at the same file ends; kept both
Collaborator
Author
|
Rule 20 note from #1031 (row 155, draft): disjoint hunks in the gemini/responses writers (a per-stream translate_cap replaces the contract reader), proto_codec.rs (ProtocolWriter::set_translate_cap) and a few tests. A test merge is clean. |
…onse_format drop in busbar-llm's audit tests, the tool-args stop test) - gemini/handler.rs write_transcription_response: the single-arm `match` is an `if let` (clippy::single_match). It writes the same bytes. - busbar-llm engine/tests/egress_dropped_controls_audit_tests.rs: the two tests that pinned the post-1.5.5 `busbar_response_format` forced tool now hold 1.5.5 (v1.5.5 crates/busbar/src/proto/anthropic/writer.rs:343-347; ARCHITECT ruling 2026-10-13, F25 rows 107/109, as #884 ports them). - openai_to_anthropic_response_format_forwards_dropped_and_audited: the default Anthropic lane forwards with no response_format, tool_choice or tools, and records the `response_format on anthropic` degraded audit event. - egress_dropped_controls_reports_the_right_controls_per_dialect: the default Anthropic lane reports ["response_format"]. - gemini tests a_finish_reason_inside_tool_args_is_not_the_stop_reason: the raw scan reads the candidate's own `STOP` (EndTurn), never the args' MALFORMED_FUNCTION_CALL (Error). The test had also compared it to read_response's stop reason, which promotes a tool-call STOP to ToolUse from the content; a token-only scan cannot see that, and the fault check does not need it. The test now asserts each side for what it is.
…llm-codec-rehome-rf
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #857. This branch contains #857's commit, 9d86513, which restores 1.5.5's
response_formatdrop on a default Anthropic lane. Its own change is the one test commit on top. Once #857 lands, the diff shrinks to that commit.What. These two F25 coverage-map v2 REHOME rows depend on #857:
crates/busbar-llm/src/engine/tests/egress_dropped_controls_audit_tests.rs:26. The legacy test pinned the syntheticbusbar_response_formatforced tool, which was post-1.5.5 drift. The ARCHITECT ruling of 2026-10-13 is that 1.5.5 wins, so it is ported at 1.5.5. Testanthropic/tests/ir_mapping_q57_tests.rs::openai_json_object_onto_a_default_anthropic_lane_is_dropped_and_reported: the body forwards non-empty, with noresponse_format,tool_choiceortools, andresponse_formatappears in the lane's dropped controls. 1.5.5 source:crates/busbar/src/proto/anthropic/writer.rs:343-347. The map row is marked COVERED-at-1.5.5.egress_dropped_controls_audit_tests.rs:146. Testcodec/tests/drops_tests.rs::egress_dropped_controls_reports_the_right_controls_per_dialect, through the chat handle:["response_format"]. That is 1.5.5; the legacy test expected[], pinning the forced tool.["response_format", "tool_choice=none"].[].["response_format"].Tests only; not
$.