From 19badfe8b7cf03439a7e0287cd8cf7600d804716 Mon Sep 17 00:00:00 2001 From: Hare Date: Mon, 11 May 2026 22:22:36 +0900 Subject: [PATCH 1/3] fix: bundle anthropic assistant bursts --- .../llm_client/scheme/anthropic/request.rs | 175 ++++++++++++++---- 1 file changed, 142 insertions(+), 33 deletions(-) diff --git a/crates/llm-worker/src/llm_client/scheme/anthropic/request.rs b/crates/llm-worker/src/llm_client/scheme/anthropic/request.rs index 835840f7..62fac8d1 100644 --- a/crates/llm-worker/src/llm_client/scheme/anthropic/request.rs +++ b/crates/llm-worker/src/llm_client/scheme/anthropic/request.rs @@ -242,10 +242,13 @@ impl AnthropicScheme { /// - Tool calls are content parts within assistant messages /// - Tool results are content parts within user messages /// - /// Each non-`Message` item produces exactly one content part, so - /// "last part for the item" is always well-defined. For breakpoint - /// `Message` items the output is forced into the array form so a - /// marker has a part to attach to. + /// Assistant-side items are accumulated until a user/system message or + /// tool result boundary so one logical assistant burst becomes one + /// Anthropic assistant message content array. Pending parts carry their + /// origin item index; when flushed, the final part for each item records + /// the `(msg_idx, part_idx)` used by breakpoint attachment. User/system + /// `Message` items keep the single-text shorthand unless a breakpoint + /// needs a concrete part to live on. fn convert_items_to_messages( &self, items: &[Item], @@ -261,19 +264,6 @@ impl AnthropicScheme { for (i, item) in items.iter().enumerate() { match item { Item::Message { role, content, .. } => { - flush_pending( - &mut messages, - &mut pending_assistant, - "assistant", - &mut locations, - ); - flush_pending(&mut messages, &mut pending_user, "user", &mut locations); - - let anthropic_role = match role { - Role::User | Role::System => "user", - Role::Assistant => "assistant", - }; - let parts: Vec = content .iter() .map(|p| match p { @@ -284,27 +274,43 @@ impl AnthropicScheme { }) .collect(); - let force_parts = breakpoints.contains(&i); - let msg_idx = messages.len(); + match role { + Role::Assistant => { + flush_pending(&mut messages, &mut pending_user, "user", &mut locations); + pending_assistant.extend(parts.into_iter().map(|part| (i, part))); + } + Role::User | Role::System => { + flush_pending( + &mut messages, + &mut pending_assistant, + "assistant", + &mut locations, + ); + flush_pending(&mut messages, &mut pending_user, "user", &mut locations); - // Preserve the single-text shorthand unless a - // breakpoint needs a concrete part to live on. - if parts.len() == 1 && !force_parts { - if let AnthropicContentPart::Text { text, .. } = &parts[0] { + let force_parts = breakpoints.contains(&i); + let msg_idx = messages.len(); + + // Preserve the single-text shorthand unless a + // breakpoint needs a concrete part to live on. + if parts.len() == 1 && !force_parts { + if let AnthropicContentPart::Text { text, .. } = &parts[0] { + messages.push(AnthropicMessage { + role: "user".to_string(), + content: AnthropicContent::Text(text.clone()), + }); + continue; + } + } + + let last_part_idx = parts.len().saturating_sub(1); messages.push(AnthropicMessage { - role: anthropic_role.to_string(), - content: AnthropicContent::Text(text.clone()), + role: "user".to_string(), + content: AnthropicContent::Parts(parts), }); - continue; + locations[i] = Some((msg_idx, last_part_idx)); } } - - let last_part_idx = parts.len().saturating_sub(1); - messages.push(AnthropicMessage { - role: anthropic_role.to_string(), - content: AnthropicContent::Parts(parts), - }); - locations[i] = Some((msg_idx, last_part_idx)); } Item::ToolCall { @@ -626,6 +632,109 @@ mod tests { out } + #[test] + fn assistant_burst_bundles_reasoning_text_and_tool_call() { + let scheme = AnthropicScheme::new(); + let request = Request::new() + .user("question?") + .item(Item::reasoning("thinking").with_signature("SIG-A")) + .item(Item::assistant_message("answer")) + .item(Item::tool_call("c1", "tool_a", r#"{"x":1}"#)); + + let req = scheme.build_request("claude-sonnet-4-20250514", &request, &cap_explicit()); + + assert_eq!(req.messages.len(), 2, "messages: {:?}", req.messages); + assert_eq!(req.messages[0].role, "user"); + assert_eq!(req.messages[1].role, "assistant"); + let AnthropicContent::Parts(parts) = &req.messages[1].content else { + panic!("assistant burst must be emitted as content parts"); + }; + assert_eq!(parts.len(), 3, "parts: {:?}", parts); + assert!(matches!(parts[0], AnthropicContentPart::Thinking { .. })); + assert!(matches!(parts[1], AnthropicContentPart::Text { .. })); + assert!(matches!(parts[2], AnthropicContentPart::ToolUse { .. })); + } + + #[test] + fn tool_result_and_user_messages_bound_assistant_bursts() { + let scheme = AnthropicScheme::new(); + let request = Request::new() + .user("question?") + .item(Item::reasoning("thinking").with_signature("SIG-A")) + .item(Item::assistant_message("answer")) + .item(Item::tool_call("c1", "tool_a", "{}")) + .item(Item::tool_result("c1", "result")) + .item(Item::assistant_message("final")) + .user("follow up"); + + let req = scheme.build_request("claude-sonnet-4-20250514", &request, &cap_explicit()); + + let roles: Vec<&str> = req.messages.iter().map(|msg| msg.role.as_str()).collect(); + assert_eq!( + roles, + vec!["user", "assistant", "user", "assistant", "user"] + ); + + let AnthropicContent::Parts(first_assistant) = &req.messages[1].content else { + panic!("first assistant burst must be content parts"); + }; + assert_eq!(first_assistant.len(), 3); + assert!(matches!( + first_assistant[0], + AnthropicContentPart::Thinking { .. } + )); + assert!(matches!( + first_assistant[1], + AnthropicContentPart::Text { .. } + )); + assert!(matches!( + first_assistant[2], + AnthropicContentPart::ToolUse { .. } + )); + + let AnthropicContent::Parts(tool_result) = &req.messages[2].content else { + panic!("tool result must be content parts"); + }; + assert_eq!(tool_result.len(), 1); + assert!(matches!( + tool_result[0], + AnthropicContentPart::ToolResult { .. } + )); + + let AnthropicContent::Parts(second_assistant) = &req.messages[3].content else { + panic!("second assistant burst must be content parts"); + }; + assert_eq!(second_assistant.len(), 1); + assert!(matches!( + second_assistant[0], + AnthropicContentPart::Text { .. } + )); + } + + #[test] + fn assistant_message_breakpoint_maps_to_text_part_inside_burst() { + let scheme = AnthropicScheme::new(); + let mut request = Request::new().items(vec![ + Item::user_message("question?"), + Item::reasoning("thinking").with_signature("SIG-A"), + Item::assistant_message("answer"), + Item::tool_call("c1", "tool_a", "{}"), + Item::user_message("next"), + ]); + request.cache_anchor = Some(2); + + let req = scheme.build_request("claude-sonnet-4-20250514", &request, &cap_explicit()); + let AnthropicContent::Parts(parts) = &req.messages[1].content else { + panic!("assistant burst must be content parts"); + }; + + assert!(matches!(parts[0], AnthropicContentPart::Thinking { .. })); + assert!(matches!(parts[1], AnthropicContentPart::Text { .. })); + assert!(matches!(parts[2], AnthropicContentPart::ToolUse { .. })); + assert_eq!(part_cache_control(&parts[1]), Some(CacheControl::Ephemeral)); + assert_eq!(part_cache_control(&parts[2]), Some(CacheControl::Ephemeral)); + } + /// Convenience: a turn that ends with one assistant text, one tool /// call/result pair, and a final assistant text. Produced at /// `history[head..]` indices shown alongside, so tests can reason From d18f536945ea3d42d287c024636faeaa527d4342 Mon Sep 17 00:00:00 2001 From: Hare Date: Mon, 11 May 2026 22:23:38 +0900 Subject: [PATCH 2/3] review: anthropic assistant burst bundling --- tickets/anthropic-assistant-burst-bundling.md | 5 +++ ...thropic-assistant-burst-bundling.review.md | 36 +++++++++++++++++++ 2 files changed, 41 insertions(+) create mode 100644 tickets/anthropic-assistant-burst-bundling.review.md diff --git a/tickets/anthropic-assistant-burst-bundling.md b/tickets/anthropic-assistant-burst-bundling.md index 4cf2c81a..77f4bbb8 100644 --- a/tickets/anthropic-assistant-burst-bundling.md +++ b/tickets/anthropic-assistant-burst-bundling.md @@ -29,3 +29,8 @@ Anthropic Messages API は user/assistant の交互を要求し、同一論理 t - モデル世代別の thinking keep/strip デフォルト分岐（reasoning-history-persist のフォローアップ候補と同じ扱い） - `clear_thinking_20251015` context-edit - prune.rs の reasoning aware 化 + +## レビュー状態 + +- `19badfe fix: bundle anthropic assistant bursts` を review 済み。結果は `tickets/anthropic-assistant-burst-bundling.review.md`。 +- 判断: approve / merge 可。 diff --git a/tickets/anthropic-assistant-burst-bundling.review.md b/tickets/anthropic-assistant-burst-bundling.review.md new file mode 100644 index 00000000..88e0e847 --- /dev/null +++ b/tickets/anthropic-assistant-burst-bundling.review.md @@ -0,0 +1,36 @@ +# Review: Anthropic assistant burst bundling + +## 対象 + +- Ticket: `tickets/anthropic-assistant-burst-bundling.md` +- Branch: `anthropic-assistant-burst-bundling` +- Reviewed commit: `19badfe fix: bundle anthropic assistant bursts` + +## 確認内容 + +- `Item::Reasoning` / `Item::Message(Role::Assistant)` / `Item::ToolCall` が pending assistant parts に積まれ、user/system message または tool result まで 1 つの Anthropic `assistant` message として flush される。 +- parts の順序は history arrival 順を維持している。 +- `Item::ToolResult` は pending assistant を先に flush してから user message の `tool_result` part として扱われ、assistant burst の境界になっている。 +- user/system message も pending assistant / pending user を flush してから user message として出力され、assistant burst の境界になっている。 +- breakpoint / `cache_control` の item index → `(msg_idx, part_idx)` mapping は pending parts に origin item index を持たせる既存構造で維持されている。assistant text item も pending 経由になったため、assistant burst 内の text part に marker を付けられる。 +- user/system の single-text shorthand は breakpoint が無い場合に維持されている。assistant 側は burst flush によって `Parts` に統一されるが、ticket の許容範囲内。 +- protocol / scope / history persistence / prompt context processing には触れていない。 + +## 追加テスト + +- `assistant_burst_bundles_reasoning_text_and_tool_call` +- `tool_result_and_user_messages_bound_assistant_bursts` +- `assistant_message_breakpoint_maps_to_text_part_inside_burst` + +## 検証 + +- `cargo test -p llm-worker` passed +- `cargo fmt --check` failed due existing unrelated rustfmt diffs outside this ticket's changed file. +- `cargo fmt -p llm-worker --check` failed due existing unrelated rustfmt diffs in other llm-worker files; `crates/llm-worker/src/llm_client/scheme/anthropic/request.rs` was not reported. +- `git diff --check` passed. + +## 判断 + +Approve. + +Ticket の要件通り、Anthropic wire 上の隣接 assistant message 分割を解消し、thinking / text / tool_use を 1 assistant message の content parts に束ねている。変更は projection とそのテストに限定されており、既存設計を歪める追加抽象や範囲外の reasoning policy 変更は入っていない。 From f444b387bec09963c3c64c5141e044ddede36fdd Mon Sep 17 00:00:00 2001 From: Hare Date: Mon, 11 May 2026 22:23:53 +0900 Subject: [PATCH 3/3] docs(tickets): complete anthropic assistant burst bundling --- TODO.md | 1 - tickets/anthropic-assistant-burst-bundling.md | 36 ------------------- ...thropic-assistant-burst-bundling.review.md | 36 ------------------- 3 files changed, 73 deletions(-) delete mode 100644 tickets/anthropic-assistant-burst-bundling.md delete mode 100644 tickets/anthropic-assistant-burst-bundling.review.md diff --git a/TODO.md b/TODO.md index 677a25fc..f49cb45e 100644 --- a/TODO.md +++ b/TODO.md @@ -13,7 +13,6 @@ - Exchange / Turn / Call セマンティクス整理 → [tickets/exchange-turn-call-semantics.md](tickets/exchange-turn-call-semantics.md) - llm-worker のエラー耐性 - ストリーム途中失敗時の継続 → [tickets/llm-worker-stream-continuation.md](tickets/llm-worker-stream-continuation.md) -- llm-worker: Anthropic projection で assistant ターン内ブロックを 1 message に束ねる → [tickets/anthropic-assistant-burst-bundling.md](tickets/anthropic-assistant-burst-bundling.md) - ネイティブ GUI クライアント MVP → [tickets/native-gui-mvp.md](tickets/native-gui-mvp.md) - E2E テストハーネス（`tests/e2e/`、opt-in） → [tickets/e2e-harness.md](tickets/e2e-harness.md) - TUI 拡充 diff --git a/tickets/anthropic-assistant-burst-bundling.md b/tickets/anthropic-assistant-burst-bundling.md deleted file mode 100644 index 77f4bbb8..00000000 --- a/tickets/anthropic-assistant-burst-bundling.md +++ /dev/null @@ -1,36 +0,0 @@ -# Anthropic projection: assistant ターン内ブロックを 1 message に束ねる - -## 背景 - -`crates/llm-worker/src/llm_client/scheme/anthropic/request.rs` の `convert_items_to_messages` は、Worker が 1 ターンで生成する `[Reasoning, assistant_message, ToolCall]` の連列を、Anthropic wire 上で **複数の隣接した assistant message** に分割している。 - -具体的には: -- `Item::Reasoning` を `pending_assistant` に push -- 次の `Item::Message { Role::Assistant }` が到来すると `pending_assistant` を flush し、自分自身は別 message として messages に直 push -- 続く `Item::ToolCall` は再び `pending_assistant` に積まれ、turn 末で flush され 3 つ目の assistant message に - -結果として 1 turn が `assistant[Thinking] / assistant[text] / assistant[tool_use]` の 3 message に展開される。 - -Anthropic Messages API は user/assistant の交互を要求し、同一論理 turn 内の thinking/text/tool_use は **1 つの assistant message の `content` 配列** に並べる仕様。新世代 Claude (Opus 4.5+/Sonnet 4.6+) で thinking signature を round-trip する際、隣接 assistant message に分かれていると signature の文脈が崩れて 400 になる懸念がある（reasoning-history-persist のレビュー指摘）。 - -なお、本バグは reasoning-history-persist で導入されたものではなく、`assistant_message` + `tool_call` の組合せで以前から存在していた pre-existing な分割。Reasoning が同じ flush 経路を継承した形。 - -## 要件 - -- 同一論理ターンに属する `Item::Reasoning` / `Item::Message(Assistant)` / `Item::ToolCall` を、Anthropic wire 上の **1 つの assistant message の `content` 配列** に束ねる -- 順序は arrival 順 (= history 順)。Anthropic 仕様の典型は thinking → text → tool_use -- user / system role の `Item::Message` や `Item::ToolResult` を境界として assistant burst を区切る -- 既存の breakpoint (cache_control) 計算が壊れないこと: 各 item のオリジン index → (msg_idx, part_idx) マッピングは flush_pending 経由で記録されているので、Item::Message(Assistant) も pending を経由するように揃えれば自然に追従する -- Single-text 専用の `AnthropicContent::Text` shorthand は assistant burst 内 1 part のみのときに限定して維持するか、簡潔さのために常に `Parts` 形式に統一するかは実装時に判断 -- 既存テスト群（`completed_turn`, `single_text_message_uses_text_shorthand_without_breakpoint`, `breakpoint_on_tool_result_head` 等）の意図を逸脱しないよう更新 - -## スコープ外 - -- モデル世代別の thinking keep/strip デフォルト分岐（reasoning-history-persist のフォローアップ候補と同じ扱い） -- `clear_thinking_20251015` context-edit -- prune.rs の reasoning aware 化 - -## レビュー状態 - -- `19badfe fix: bundle anthropic assistant bursts` を review 済み。結果は `tickets/anthropic-assistant-burst-bundling.review.md`。 -- 判断: approve / merge 可。 diff --git a/tickets/anthropic-assistant-burst-bundling.review.md b/tickets/anthropic-assistant-burst-bundling.review.md deleted file mode 100644 index 88e0e847..00000000 --- a/tickets/anthropic-assistant-burst-bundling.review.md +++ /dev/null @@ -1,36 +0,0 @@ -# Review: Anthropic assistant burst bundling - -## 対象 - -- Ticket: `tickets/anthropic-assistant-burst-bundling.md` -- Branch: `anthropic-assistant-burst-bundling` -- Reviewed commit: `19badfe fix: bundle anthropic assistant bursts` - -## 確認内容 - -- `Item::Reasoning` / `Item::Message(Role::Assistant)` / `Item::ToolCall` が pending assistant parts に積まれ、user/system message または tool result まで 1 つの Anthropic `assistant` message として flush される。 -- parts の順序は history arrival 順を維持している。 -- `Item::ToolResult` は pending assistant を先に flush してから user message の `tool_result` part として扱われ、assistant burst の境界になっている。 -- user/system message も pending assistant / pending user を flush してから user message として出力され、assistant burst の境界になっている。 -- breakpoint / `cache_control` の item index → `(msg_idx, part_idx)` mapping は pending parts に origin item index を持たせる既存構造で維持されている。assistant text item も pending 経由になったため、assistant burst 内の text part に marker を付けられる。 -- user/system の single-text shorthand は breakpoint が無い場合に維持されている。assistant 側は burst flush によって `Parts` に統一されるが、ticket の許容範囲内。 -- protocol / scope / history persistence / prompt context processing には触れていない。 - -## 追加テスト - -- `assistant_burst_bundles_reasoning_text_and_tool_call` -- `tool_result_and_user_messages_bound_assistant_bursts` -- `assistant_message_breakpoint_maps_to_text_part_inside_burst` - -## 検証 - -- `cargo test -p llm-worker` passed -- `cargo fmt --check` failed due existing unrelated rustfmt diffs outside this ticket's changed file. -- `cargo fmt -p llm-worker --check` failed due existing unrelated rustfmt diffs in other llm-worker files; `crates/llm-worker/src/llm_client/scheme/anthropic/request.rs` was not reported. -- `git diff --check` passed. - -## 判断 - -Approve. - -Ticket の要件通り、Anthropic wire 上の隣接 assistant message 分割を解消し、thinking / text / tool_use を 1 assistant message の content parts に束ねている。変更は projection とそのテストに限定されており、既存設計を歪める追加抽象や範囲外の reasoning policy 変更は入っていない。