# Potential Cleanup to do
> Note: Those are just internal bookkeeping notes. They will eventually be done if needed, and should not be picked up unless they are part of another fix or enhancement.
## 2026-10-14 - PR 332 related - openai_resp done arguments
- Refresh `call_id` and `name` from `output_item.done` when the entry was created from an `output_item.added` that carried empty values.
- Emit a final `ToolCallChunk` when `output_item.done` overrides the accumulated arguments, so streaming consumers see the complete arguments.
- Finalize `in_progress_tool_calls` on the `ResponseIncomplete` and EOF paths. Today `captured_tool_calls` only reads `captured_data.tool_calls`, which is populated in `ResponseCompleted`, so those accumulated calls are dropped.
- Convert `tests/openai_resp_done_arguments.rs` into a yakbak protocol-fixture cassette under `tests/data/yakbak/openai_resp/`, add a `test_yakbak_openai_resp_done_arguments_override` case in `tests/tests_yakbak_openai_resp.rs`, and delete the raw TCP test.
- In the new test, use `TestResult<()>` with `?`, and check the length before indexing `calls[0]`, or use `.first().ok_or(...)`.
- In `tests/tests_yakbak_openai_resp.rs`, replace the byte-index string slices (`&reasoning[..reasoning.len().min(60)]` and the `content` tail slice) with char-boundary-safe truncation, since the content can be multi-byte (Japanese).
- In `Cargo.toml`, extend the `native-tls` comment to mention re-adding `system-proxy`.
## 2026-10-15 - PR 334 related - v0.7.0 migration guide note
- Context: PR 334 changes behavior for `default-features = false` builds. They no longer read macOS or Windows OS proxy settings unless `system-proxy` is enabled.
- Why: the `CHANGELOG.md` entry is already added under API Breaking Changes. The v0.7.0 migration guide is maintainer-owned and was not updated.
- Suggested change: add a short migration note to the v0.7.0 migration guide that shows `features = ["system-proxy"]` as the way to keep the old behavior. Keep the wording aligned with the README note.
## 2026-10-15 - PR 334 related - proxy setters overwrite each other
- Context: in `src/client/web_config.rs`, `with_proxy_url`, `with_https_proxy_url`, and `with_all_proxy_url` each overwrite the same stored proxy, so only one proxy is kept.
- Why: this is not introduced by PR 334, but it can surprise callers who set an HTTP proxy and then an HTTPS proxy and expect both to apply.
- Suggested change: decide the intended semantics first. Either document that the last setter wins, or store the proxies separately per scheme and apply them in `apply_to_builder`. Add a unit test for the chosen behavior.
## 2026-10-15 - PR 335 related - confirm Vertex live tool tests
- Context: `VertexAdapter::to_gemini_web_request_data` calls `GeminiAdapter::build_gemini_request_payload`, so PR 335 also sends `functionCall.id` and `functionResponse.id` to Vertex.
- Why: the PR author ran only the Gemini live tests. Vertex defines the same `id` fields and is expected to accept them, but this is not verified.
- Suggested change: run the live Vertex tool tests in `tests/tests_p_vertex.rs`, especially `test_tool_gemini_full_flow_ok` and `test_tool_gemini_simple_ok`, before or just after merging. If one fails with 400, apply the same name and id rule that the Gemini path uses.
## 2026-10-15 - PR 335 related - infallible gemini part helpers
- Context: `gemini_function_call_part` and `gemini_function_response_part` return `Result<Value>` only because `x_insert` is fallible.
- Why: the reviewer noted this is acceptable and matches the file's style. Building the object as a `serde_json::Map` would make the helpers infallible and simpler to call.
- Suggested change: optional. Build the `functionCall` and `functionResponse` objects as `serde_json::Map` values, then return `Value` directly, if this makes the callers simpler. Keep it only if it reads clearly.
## 2026-10-15 - PR 335 related - align gemini_ix id policy
- Context: `gemini_ix` in `into_ix_input_steps` sends `call_id` as it is. This includes empty strings and synthetic `call#{name}#{n}` ids, which the `gemini` adapter now filters with `gemini_call_id`.
- Why: this is pre-existing and does not affect matching within `gemini_ix`. It only matters when histories from the `gemini` adapter are replayed through `gemini_ix`, and then the two adapters follow different id rules.
- Suggested change: optionally apply the same empty and synthetic id filter in `gemini_ix`, or share the helper. Check the Interactions API tests before changing it, since the current behavior is accepted by the existing tests.
## 2026-10-15 - PR 336 related - pastey adapter identifiers
- Context: `src/adapter/macros/dispatcher_macros.rs` uses `paste!` only to concatenate fixed tokens such as `[<OpenAI Adapter>]`.
- Why: the concatenation is cosmetic, because the tokens are fixed and known. Writing the identifiers directly would remove the direct `pastey` dependency from `genai`. `pastey` would still be pulled in transitively by `macro_rules_attribute`, so this is a cleanup, not a dependency fix.
- Suggested change: replace the `[<X Adapter>]` forms with literal identifiers such as `OpenAIAdapter`, then remove the direct `pastey` dependency from `Cargo.toml` and the `paste` import. Confirm the dispatch match still expands the same way.
## 2026-10-15 - PR 336 related - pre-merge dependency checks
- Context: PR 336 replaces `paste` with `pastey` and states that `cargo build`, `cargo clippy --all-targets`, `cargo test --lib`, and the yakbak replay tests pass.
- Why: a few checks were not reported. The `bedrock-sigv4` cfg arm of `dispatch_adapter!` was not explicitly built, `Cargo.lock` consistency was not mentioned, and `paste` leaving the tree was not confirmed.
- Suggested change: run `cargo tree -i paste` (expect nothing), `cargo build --features bedrock-sigv4`, and confirm `Cargo.lock` is updated if it is tracked. Record the results in the PR.
## 2026-10-19 - PR 339 related - openai_resp terminal paths still drop captured state
- Context: PR 339 (`openai_resp - keep a stream's tool calls when the response is incomplete`) moved the tool-call capture out of `RespStreamEvent::ResponseCompleted` into `OpenAIRespStreamer::capture_tool_calls(&mut self, output: &[Value])` in `src/adapter/adapters/openai_resp/streamer.rs`, and now calls it from both `ResponseCompleted` and `ResponseIncomplete`.
- Resolved by PR 339: `response.incomplete` no longer loses the calls accumulated in `in_progress_tool_calls`. That closes the `ResponseIncomplete` half of the third bullet in `## 2026-10-14 - PR 332 related - openai_resp done arguments`. That bullet is kept verbatim for history, and only its EOF half is still open.
- Potential bug 1, EOF path: the `Poll::Ready(None)` branch (the `None => { if !self.done { ... } }` arm of `poll_next`) builds its `InterStreamEnd` from `captured_data` only and never calls `capture_tool_calls`, so `in_progress_tool_calls`, which is filled by `response.function_call_arguments.delta` and `response.custom_tool_call_input.delta`, is dropped.
- It also leaves `captured_stop_reason` as `None`, because `captured_data.stop_reason` is only set by `response.completed`, `response.incomplete`, and `response.failed`.
- This is the same failure mode PR 339 fixed, on the path taken when the event source ends without a terminal event: a dropped connection, a proxy closing the stream, or a backend that streams deltas only.
- Suggested change: call `self.capture_tool_calls(&[])` before building the end in that branch (empty `output`, since there is no terminal response to fall back to), and add a test whose stream stops right after the argument deltas.
- Potential bug 2, `response.incomplete` drops reasoning signatures: `OutputItemDone` pushes a `type: "reasoning"` item's `encrypted_content` into `captured_data.thought_signatures`, and `ResponseCompleted` moves them into `InterStreamEnd::captured_thought_signatures`. The `ResponseIncomplete` branch hard-codes `captured_thought_signatures: None`, and it also skips the `response.output` scan that `ResponseCompleted` does, so blocks finalized before the cap are lost.
- Impact: an incomplete answer keeps its text and its tool calls but not the continuation blocks OpenAI expects on the next turn, so replaying that assistant message is inconsistent with a completed answer.
- Suggested change: `take()` `self.captured_data.thought_signatures` there as well, or fold the terminal paths into one helper that builds `InterStreamEnd` for completed, incomplete, and EOF.
- Potential bug 3, `response.incomplete` ignores `capture_usage`: `ResponseCompleted` stores usage only under `if self.options.capture_usage`, while `ResponseIncomplete` uses `response.usage.map(Into::into)` directly, so usage is reported on an incomplete answer even when the option is off.
- Suggested change: gate it like the completed branch, ideally through the same shared helper.
- Test suite: `tests/openai_resp_incomplete_tool_calls.rs` spins up a raw `TcpListener`, the same pattern the 2026-10-14 item already wants converted. Fold it into that work: a yakbak cassette under `tests/data/yakbak/openai_resp/` plus a case in `tests/tests_yakbak_openai_resp.rs`, then delete the raw TCP server.
## 2026-10-19 - PR 340 related - anthropic max_tokens tool-call test and streamer follow-ups
- Context: PR 340 captures a `tool_use` block cut off by `max_tokens` at `message_stop`, through `AnthropicStreamer::capture_tool_call(id, name, input, cut_off)` in `src/adapter/adapters/anthropic/streamer.rs`, called from both `content_block_stop` and `message_stop`.
- Test file: `tests/anthropic_max_tokens_tool_calls.rs` spins up a raw `TcpListener`, and its name breaks the `tests/tests_*.rs` convention.
- Suggested change: fold it into `tests/tests_yakbak_anthropic.rs` with a cassette under `tests/data/yakbak/anthropic/`, then delete the raw TCP server, the same conversion the 2026-10-14 item wants for the OpenAI Responses tests.
- Nit, helper placement: `capture_tool_call` sits in the top inherent `impl` block, while the file keeps its private helpers (`capture_usage`, `parse_message_data`) in the bottom `// Support` impl.
- Suggested change: move it there, and consider replacing the `cut_off: bool` argument with two named wrappers or a small enum, since `.., false)` and `.., true)` do not read at the call sites.
- Missing test: no coverage for `capture_tool_calls == false`, which is exactly the guard that moved out of the `content_block_stop` match arm and into the method.
- Follow-up fix, not a regression: `StreamEnd::into_assistant_message_for_tool_use()` now returns `Some(..)` for a stream whose call was truncated mid-input, so replaying that message serializes `tool_use.input` as a JSON string, and Anthropic requires an object. The same exposure already exists for malformed non-truncated arguments.
- Suggested change: in `into_anthropic_request_parts`, coerce a non-object `ToolCall.fn_arguments` to `{}` with a warn, so a captured truncated call cannot produce a request Anthropic rejects.
- Out of scope, unchanged: a `thinking` block cut off by `max_tokens` is still dropped, because only a block that reaches `content_block_stop` is turned into an `InterStreamThoughtBlock`.
## 2026-10-00 - Fix: finalize OpenAI Responses tool calls on incomplete and EOF
Behavior change, needs yakbak coverage. Commit prefix: `- openai_resp - Finalize tool calls on incomplete and EOF`
- Refresh `call_id` and `name` from `output_item.done` when the entry came from an `output_item.added` with empty values.
- Emit a final `ToolCallChunk` when `output_item.done` overrides the accumulated arguments.
- Finalize `in_progress_tool_calls` on the `ResponseIncomplete` and EOF (`None`) paths so accumulated calls are not dropped.
- Add a yakbak cassette test and remove the raw TCP test. Blocked until the test files are provided.
## 2026-10-00 - Refactor: literal adapter identifiers in dispatcher macros
High-risk macro refactor. Commit prefix: `> macros - Use literal adapter identifiers`
- `src/adapter/macros/dispatcher_macros.rs`: replace `[<X Adapter>]` forms with literal identifiers such as `OpenAIAdapter`, and remove the `paste` import.
- `Cargo.toml`: remove the direct `pastey` dependency if nothing else uses it.
- Verify with `cargo tree -i pastey`, `cargo build`, and `cargo build --features bedrock-sigv4`.
## 2026-10-27 - OpenAI-compatible streaming and shared tool tests
- Context: PR 347 changed `src/adapter/adapters/openai/streamer.rs` so one SSE message can emit reasoning, content, and every tool call through a `pending_events` queue. This affects all adapters using `OpenAIStreamer`, not only OpenAI.
- Shared live tool tests: `tests/support/common_tests.rs` is used by multiple provider suites. The review identified intermittent failures when a model returned Fahrenheit instead of Celsius, or multiple tool calls while `common_test_tool_full_flow_ok` answered only one. Check the current file before changing it: later workbench notes report that the tests were already hardened.
- If still needed, answer every assistant tool call with a matching `ToolResponse`, including the function name for providers such as Gemini. Alternatively, send only the one assistant call being answered.
- Keep streaming and non-streaming argument checks aligned. Account for valid model variance in units, capitalization, location formatting, and parallel calls without losing the check that captured arguments are parsed JSON objects.
- Avoid requiring final prose to repeat facts absent from the supplied tool responses. Run the affected OpenAI live tests and at least one other provider suite after changes; live tests require credentials.
- Synthetic streaming tests: `tests/openai_stream_shared_chunks.rs` was reported to use a local raw TCP SSE server. Consider moving its support into shared test infrastructure or using yakbak, consistent with the existing OpenAI Responses and Anthropic cassette follow-ups above. Cover multiple tool calls, reasoning and content together, content and tool calls together, finish-reason frames, and disabled capture options.
- Streamer polish, optional: in `src/adapter/adapters/openai/streamer.rs`, consider a warning for malformed tool-call entries instead of silently skipping them, and document the invariant that queued events are drained before another SSE message is read. Verify existing behavior before changing it.
- Tool-call event semantics: `capture_tool_call` returns accumulated arguments when capture is enabled, rather than necessarily returning only the latest fragment. Consumers that concatenate successive `ToolCallChunk` arguments can over-accumulate. Clarify the intended contract and test both enabled and disabled capture before changing behavior; this is not specific to PR 347.
## 2026-10-27 - Cohere request coverage and embedding errors
- Context: PR 351 changes Cohere v1 `chat_history` entries from `content` to `message` in `src/adapter/adapters/cohere/adapter_impl.rs`. Its unit tests are in `src/adapter/adapters/cohere/adapter_impl_tests.rs`.
- Add a multi-turn live test in `tests/tests_p_cohere.rs` to confirm that the provider accepts and uses history. The existing unit tests validate the outbound payload without a network call. Also consider a unit test for system-message `preamble` handling.
- In `src/adapter/adapters/cohere/embed.rs`, replace the fabricated serde error created with `serde_json::from_str::<()>(...).unwrap_err()` for missing embedding data with an appropriate existing or dedicated error. The current input always fails to parse, but the `unwrap_err()` becomes panic-capable if that assumption changes. Preserve useful error context and check compatibility of any public error change.
- Optional test cleanup: align setup, execution, and check markers in `adapter_impl_tests.rs`, and consider deriving test endpoint and authentication fixtures from adapter defaults if that makes the tests clearer.
## 2026-10-27 - Vertex model listing and Claude streaming
- Context: PR 352 selects `:streamRawPredict` for Vertex Claude streaming and retains `:rawPredict` for non-streaming requests. PR 353 updates the hardcoded Vertex model list and moves URL tests to `src/adapter/adapters/vertex/adapter_impl_tests.rs`.
- Run the Claude streaming live tests in `tests/tests_p_vertex.rs` against a configured Vertex project, including simple and capture-all cases if available. This verifies that the Vertex stream's SSE event names are handled by `src/adapter/adapters/anthropic/streamer.rs`; the URL unit tests alone cannot verify the live response format.
- In `src/adapter/adapters/vertex/adapter_impl_tests.rs`, consider checking that Gemini streaming URL selection remains unchanged, and replacing test-only `.expect("service url")` calls with propagated errors where practical.
- Verify the current Vertex model names and retirement status against Google documentation before revising `all_model_names` in `src/adapter/adapters/vertex/adapter_impl.rs`. The PR 353 review specifically called for checking the removal of Gemini 2.5 entries and the proposed `gemini-3.8-flash` replacement. Do not assume those names or retirement dates remain correct. Update related examples, tests, and documentation only when supported by current provider information.
- Review the unknown-publisher fallback in `VertexAdapter::get_service_url`: it warns and defaults to Google, while other paths validate the publisher. Consider propagating an error if the trait and callers allow it, and test the behavior rather than assuming every caller validates first.
- Separate optional safety cleanup: `gemini_tool_config` in `src/adapter/adapters/gemini/adapter_impl.rs` uses mutable `serde_json::Value` indexing for `allowedFunctionNames`. Its current object shape makes that assignment safe, but mutable indexing can panic if the shape changes. Consider a fallible insertion consistent with the file's style.
## 2026-10-27 - Stop-reason mapping and documentation
- Context: PR 354 updates `StopReason::from(String)` in `src/chat/chat_response.rs` so additional refusal, content-filter, and model-length strings map to existing variants rather than `Other`. Raw provider strings remain available, but the selected enum variant and its serialized representation change for those strings.
- Check that the `ContentFilter` variant and `is_max_tokens` documentation describe the complete current mapping, including refusals and context-window exhaustion. Keep the documentation table aligned and run `cargo fmt --check` after any code changes.
- Review remaining provider termination reasons deliberately rather than assuming they belong to an existing variant. Examples raised in review include OpenAI Responses `max_output_tokens` and Gemini `MALFORMED_FUNCTION_CALL`, `TOO_MANY_TOOL_CALLS`, and `UNEXPECTED_TOOL_CALL`. Verify the actual provider field carrying each value and the adapter's extraction path before adding mappings and tests.
- Any further mapping changes are observable to callers matching `Other` and can change serialized variants. Preserve raw values and document the compatibility impact in the appropriate release notes.
## 2026-10-27 - OpenAI-compatible provider follow-ups
- Context: PR 350 adds Runcrate and PR 355 adds CloudGPU using the shared OpenAI-compatible adapter machinery. Both route through explicit namespaces and add public `AdapterKind` variants, which can affect downstream exhaustive matches.
- In `src/adapter/adapters/all_adapters_tests.rs`, retain or extend namespace, endpoint, authentication, model-name preservation, and embedding support tests when changing either provider. CloudGPU was reported to reject embeddings, while Runcrate was added with embedding support. Recheck provider documentation and behavior before changing those capabilities.
- Review the shared OpenAI model-name handling for names ending in reasoning qualifiers such as `-low` or `-high`. The PR 350 review noted that this existing behavior may alter an explicitly namespaced Runcrate model ID. Check `src/adapter/adapters/openai/adapter_shared.rs` and the related model parsing code before deciding whether a provider-specific exception is warranted; add regression coverage for any change.
- Optional documentation consistency: compare provider ordering and link formatting in `README.md`, `docs/for-llm/api-reference-for-llm.md`, and `src/adapter/adapter_kind.rs`. Runcrate and CloudGPU documentation was reported to use a mix of plain and tracking-parameter URLs; avoid changing links without checking their intended destinations.
- Check `src/adapter/adapters/custom/adapter_impl.rs` for the reported stale `// TODO FOR LLM` concerning endpoint and authentication methods. Remove or revise it only after confirming the current implementation and the project's comment-preservation rules.