코드 탭 컨텍스트 영속화와 Auto budget 복구 적용

- AxAgentExecutionEngine에서 시스템 프롬프트 중복을 제거하고 structured tool_use/tool_result 전사본을 conversation.Messages로 동기화해 다음 턴과 저장 이력에서도 코드 작업 컨텍스트가 유지되도록 수정 - AgentQueryContextBuilder와 ContextCondenser에 post-compact tool snippet 복원, recent window 확대, tool result 보존 강화 로직을 추가해 장기 코드 실행 중 빌드/파일 근거 손실을 줄임 - MaxContextTokens=0 Auto 모드를 AppSettings, SettingsService 마이그레이션, 설정 UI, 오버레이 UI, 컨텍스트 사용량 표시, LLM 요청 본문에 연결하고 Auto 모드에서는 provider output cap 강제 주입을 제거 - 관련 회귀 테스트와 문서 README/DEVELOPMENT/CODE_CONTEXT_RELIABILITY_PLAN을 갱신하고 깨진 진단 문자열 기대값을 영어 기준으로 정리 검증: - dotnet build src/AxCopilot/AxCopilot.csproj -c Release -v minimal -p:OutputPath=bin\\verify_context_reliability_followup\\ -p:IntermediateOutputPath=obj\\verify_context_reliability_followup\\ - dotnet test src/AxCopilot.Tests/AxCopilot.Tests.csproj -c Release -v minimal --filter "AxAgentExecutionEngineTests|AgentQueryContextBuilderTests|ContextCondenserTests|SettingsServiceTests|AgentLoopDiagnosticsFormatterTests" -p:OutputPath=bin\\verify_context_reliability_followup_tests\\ -p:IntermediateOutputPath=obj\\verify_context_reliability_followup_tests\\ - dotnet test src/AxCopilot.Tests/AxCopilot.Tests.csproj -c Release -v minimal --filter "AgentLoopQueryAssemblyServiceTests|AgentLoopPreLlmStageServiceTests|AgentLoopLlmRequestPreparationServiceTests|AgentMessageInvariantHelperTests|CodeTaskWorkingSetServiceTests|AgentLoopE2ETests" -p:OutputPath=bin\\verify_context_reliability_followup_tests2\\ -p:IntermediateOutputPath=obj\\verify_context_reliability_followup_tests2\\ - dotnet build src/AxCopilot/AxCopilot.csproj -c Release -v minimal -p:OutputPath=bin\\verify_context_reliability_final\\ -p:IntermediateOutputPath=obj\\verify_context_reliability_final\\
2026-04-16 07:45:36 +09:00
parent c6e5abfa50
commit e07b6dbed0
23 changed files with 851 additions and 268 deletions
--- a/README.md
+++ b/README.md
@@ -1,63 +1,15 @@
 # AX Commander

- Update: 2026-04-16 06:41 (KST)
- Continued the Code-tab structural refactor by extracting the dispatch/stream stage into [AgentLoopLlmDispatchStageService.cs](</E:/AX Copilot - Codex/src/AxCopilot/Services/Agent/AgentLoopLlmDispatchStageService.cs>).
- The new dispatch stage now owns:
-  - workflow `query_context` / `llm_request` logging
-  - stream preview and tool-ready thinking updates
-  - read-only tool prefetch handoff
-  - the recovery-capable `SendWithToolsWithRecoveryAsync(...)` dispatch itself
- [AgentLoopService.cs](</E:/AX Copilot - Codex/src/AxCopilot/Services/Agent/AgentLoopService.cs>) now treats LLM dispatch as a staged service call instead of building the stream callback inline, which keeps the loop closer to the staged `claw-code` shape.
- [StreamingToolExecutionCoordinator.cs](</E:/AX Copilot - Codex/src/AxCopilot/Services/Agent/StreamingToolExecutionCoordinator.cs>) was also rewritten with English-only status strings so the active stream/wait path no longer carries mojibake text, and [AgentLoopPreLlmStageService.cs](</E:/AX Copilot - Codex/src/AxCopilot/Services/Agent/AgentLoopPreLlmStageService.cs>) now returns clean English missing-tool failures.
- Added [AgentLoopLlmDispatchStageServiceTests.cs](</E:/AX Copilot - Codex/src/AxCopilot.Tests/Services/AgentLoopLlmDispatchStageServiceTests.cs>) to lock:
-  - staged dispatch argument handoff
-  - run-state retry reset after a successful dispatch
-  - streaming preview / tool-ready / retry-reset event emission
-  - read-only prefetch delegation through the coordinator
- Validation:
-  - `dotnet build src/AxCopilot/AxCopilot.csproj -c Release -v minimal -p:OutputPath=bin\\verify_llm_dispatch_stage_escalated2\\ -p:IntermediateOutputPath=obj\\verify_llm_dispatch_stage_escalated2\\` warnings 0 / errors 0
-  - `dotnet test src/AxCopilot.Tests/AxCopilot.Tests.csproj -c Release -v minimal --filter "AgentLoopLlmDispatchStageServiceTests|AgentLoopPreLlmStageServiceTests|AgentLoopQueryAssemblyServiceTests|AgentLoopIterationPreparationServiceTests|AgentLoopLlmRequestPreparationServiceTests|AgentMessageInvariantHelperTests|CodeTaskWorkingSetServiceTests|AgentLoopE2ETests" -p:OutputPath=bin\\verify_llm_dispatch_stage_tests_escalated2\\ -p:IntermediateOutputPath=obj\\verify_llm_dispatch_stage_tests_escalated2\\` passed 43
-
- Update: 2026-04-16 02:13 (KST)
- Continued the Code-tab structural refactor by extracting the pre-LLM stage into [AgentLoopPreLlmStageService.cs](</E:/AX Copilot - Codex/src/AxCopilot/Services/Agent/AgentLoopPreLlmStageService.cs>).
- The new stage service now owns:
-  - iteration thinking-summary selection
-  - Gemini free-tier delay planning
-  - user-prompt submit hook planning and payload generation
-  - missing-tool guard construction
-  - final request assembly handoff through the staged query/history pipeline
- [AgentLoopService.cs](</E:/AX Copilot - Codex/src/AxCopilot/Services/Agent/AgentLoopService.cs>) now reads more like an orchestrator: history/query assembly, pre-LLM stage planning, LLM dispatch, tool execution, and recovery.
- Added [AgentLoopPreLlmStageServiceTests.cs](</E:/AX Copilot - Codex/src/AxCopilot.Tests/Services/AgentLoopPreLlmStageServiceTests.cs>) to lock:
-  - free-tier delay planning
-  - prompt-submit hook dedup by fingerprint
-  - runtime-policy missing-tool failure shaping
-  - working-set-backed request assembly
- Validation:
-  - `dotnet build src/AxCopilot/AxCopilot.csproj -c Release -v minimal -p:OutputPath=bin\\verify_pre_llm_stage_structure\\ -p:IntermediateOutputPath=obj\\verify_pre_llm_stage_structure\\` warnings 0 / errors 0
-  - `dotnet test src/AxCopilot.Tests/AxCopilot.Tests.csproj -c Release -v minimal --filter "AgentLoopPreLlmStageServiceTests|AgentLoopQueryAssemblyServiceTests|AgentLoopIterationPreparationServiceTests|AgentLoopLlmRequestPreparationServiceTests|AgentMessageInvariantHelperTests|SessionLearningCollectorTests|CodeTaskWorkingSetServiceTests|AgentLoopE2ETests" -p:OutputPath=bin\\verify_pre_llm_stage_structure_tests\\ -p:IntermediateOutputPath=obj\\verify_pre_llm_stage_structure_tests\\` passed 60
-
- Update: 2026-04-16 02:05 (KST)
- Continued the Code-tab context reliability work with a structural refactor that moves query/history assembly closer to the staged shape used by `claw-code`.
- Added `src/AxCopilot/Services/Agent/AgentLoopQueryAssemblyService.cs` so `AgentLoopService` no longer directly owns all of the following responsibilities in one place:
-  - refreshing session learnings for non-Code tabs
-  - projecting queued commands into the message list
-  - preparing the query window
-  - appending Code working-set supplemental context before request dispatch
- `src/AxCopilot/Services/Agent/AgentLoopService.cs` now delegates staged assembly to `AgentLoopQueryAssemblyService.PrepareHistory(...)` and `PrepareRequest(...)`, leaving the loop focused more narrowly on orchestration, hooks, tool execution, and retry flow.
- Added `src/AxCopilot.Tests/Services/AgentLoopQueryAssemblyServiceTests.cs` to lock:
-  - non-Code session-learning injection
-  - Code-tab skip behavior for session learnings
-  - Code working-set supplemental message injection before the tool reminder
- Rewrote `src/AxCopilot.Tests/Services/SessionLearningCollectorTests.cs` in English-only comments and assertions so the test file itself also follows the new encoding/comment rule.
- Validation:
-  - `dotnet build src/AxCopilot/AxCopilot.csproj -c Release -v minimal -p:OutputPath=bin\\verify_query_assembly_structure\\ -p:IntermediateOutputPath=obj\\verify_query_assembly_structure\\` warnings 0 / errors 0
-  - `dotnet test src/AxCopilot.Tests/AxCopilot.Tests.csproj -c Release -v minimal --filter "AgentLoopQueryAssemblyServiceTests|AgentLoopIterationPreparationServiceTests|AgentLoopLlmRequestPreparationServiceTests|AgentMessageInvariantHelperTests|SessionLearningCollectorTests|CodeTaskWorkingSetServiceTests|AgentLoopE2ETests" -p:OutputPath=bin\\verify_query_assembly_structure_tests\\ -p:IntermediateOutputPath=obj\\verify_query_assembly_structure_tests\\` passed 56
-
- Update: 2026-04-16 01:28 (KST)
- Added an encoding rule to `AGENTS.md`: comments inside code files must be written in English only, and any broken mojibake strings found in touched code files should be rewritten in English before commit.
- Reviewed the recent Code tab runs again. On `2026-04-16`, the request message count still grew during long runs (`messages=7 -> 125`, and another run `118 -> 139`), so the main issue is not raw context length. The problem is context fidelity: build/file evidence is being compacted too aggressively while tool-trace repair noise is repeatedly inserted.
- Captured the remediation roadmap in `docs/CODE_CONTEXT_RELIABILITY_PLAN.md`. The plan covers workspace-context bootstrap repair, a dedicated Code working-set memory layer, task-aware pruning, tool-trace invariant hardening, and prompt/encoding cleanup.
- The plan references `claw-code/.../src/query.ts`, `history.ts`, and `memory-context.md`, and it also incorporates external research from Anthropic Claude Code memory docs, the OpenAI practical guide to building agents, and the SWE-Pruner paper.
+- Update: 2026-04-16 07:40 (KST)
+- Code tab context reliability was hardened to keep durable tool transcript history, auto-resolve the context budget when `MaxContextTokens = 0`, restore recent compacted tool snippets after boundary compaction, and deduplicate repeated system prompts before each loop request.
+- `src/AxCopilot/Services/Agent/AxAgentExecutionEngine.cs` now preserves full `ChatMessage` metadata during turn cloning and syncs structured tool-use and tool-result transcript messages back into `conversation.Messages`, so the next turn and persisted history keep the same repair context.
+- `src/AxCopilot/Services/Agent/AgentQueryContextBuilder.cs` and `src/AxCopilot/Services/Agent/ContextCondenser.cs` now protect a larger recent window, keep more tool-result detail, and inject compacted tool snippets into the post-compact context block instead of only count-based summaries.
+- `src/AxCopilot/Models/AppSettings.cs`, `src/AxCopilot/Services/SettingsService.cs`, `src/AxCopilot/ViewModels/SettingsViewModel.cs`, `src/AxCopilot/Views/SettingsWindow.xaml`, `src/AxCopilot/Views/AgentSettingsWindow.xaml.cs`, `src/AxCopilot/Views/ChatWindow.OverlaySettingsPresentation.cs`, and `src/AxCopilot/Views/ChatWindow.ContextUsagePresentation.cs` now treat `MaxContextTokens = 0` as Auto and expose that choice consistently in settings and live UI.
+- `src/AxCopilot/Services/LlmService.cs` and `src/AxCopilot/Services/LlmService.ToolUse.cs` now omit provider output token caps when the shared context budget is set to Auto, which avoids forcing the old 32K cap into OpenAI-compatible, Gemini, IBM, and tool-calling request payloads.
+- Verification:
+  - `dotnet build src/AxCopilot/AxCopilot.csproj -c Release -v minimal -p:OutputPath=bin\\verify_context_reliability_followup\\ -p:IntermediateOutputPath=obj\\verify_context_reliability_followup\\` warning 0 / error 0
+  - `dotnet test src/AxCopilot.Tests/AxCopilot.Tests.csproj -c Release -v minimal --filter "AxAgentExecutionEngineTests|AgentQueryContextBuilderTests|ContextCondenserTests|SettingsServiceTests|AgentLoopDiagnosticsFormatterTests" -p:OutputPath=bin\\verify_context_reliability_followup_tests\\ -p:IntermediateOutputPath=obj\\verify_context_reliability_followup_tests\\` passed 50
+  - `dotnet test src/AxCopilot.Tests/AxCopilot.Tests.csproj -c Release -v minimal --filter "AgentLoopQueryAssemblyServiceTests|AgentLoopPreLlmStageServiceTests|AgentLoopLlmRequestPreparationServiceTests|AgentMessageInvariantHelperTests|CodeTaskWorkingSetServiceTests|AgentLoopE2ETests" -p:OutputPath=bin\\verify_context_reliability_followup_tests2\\ -p:IntermediateOutputPath=obj\\verify_context_reliability_followup_tests2\\` passed 38

 - 업데이트: 2026-04-16 00:57 (KST)
 - AX Agent 앱 생성 메시지의 가로 폭과 정렬을 다시 다듬었습니다. `src/AxCopilot/Views/ChatWindow.ResponsePresentation.cs`에 `GetAgentEventMaxWidth()`를 추가해 앱이 그리는 진행/도구/완료 카드 폭만 별도로 줄였고, `src/AxCopilot/Views/ChatWindow.AgentEventRendering.cs`, `src/AxCopilot/Views/ChatWindow.V2LiveProgressPresentation.cs`, `src/AxCopilot/Views/ChatWindow.V2AgentEventPresentation.cs`는 해당 폭을 사용하면서 중앙 정렬 대신 좌측 기준으로 붙도록 맞췄습니다.
@@ -2433,37 +2385,3 @@ MIT License
 - 검증:
  - `dotnet build src/AxCopilot/AxCopilot.csproj -c Release -v minimal -p:OutputPath=bin\\verify_project_scaffold_layout\\ -p:IntermediateOutputPath=obj\\verify_project_scaffold_layout\\` 경고 0 / 오류 0
  - `dotnet test src/AxCopilot.Tests/AxCopilot.Tests.csproj -c Release -v minimal --filter "IntentGateServiceTests|ProjectScaffoldProfileCatalogTests|SkillServiceRuntimePolicyTests|AgentLoopCodeQualityTests" -p:OutputPath=bin\\verify_project_scaffold_layout_tests\\ -p:IntermediateOutputPath=obj\\verify_project_scaffold_layout_tests\\` 통과 183
-업데이트: 2026-04-16 01:41 (KST)
- Code 탭 컨텍스트 신뢰성 보강 1차 구현을 반영했습니다. `src/AxCopilot/Services/Agent/CodeTaskWorkingSetService.cs`를 추가해 최근 생성 디렉터리, 최근 읽기/쓰기 파일, 최신 build/test 진단, 다음 복구 초점을 하나의 working set 블록으로 유지하고, 각 반복의 실제 LLM 요청에 보조 system context로 주입합니다.
- `src/AxCopilot/Services/Agent/AgentQueryContextBuilder.cs`, `AgentToolResultBudget.cs`, `AgentLoopIterationPreparationService.cs`, `AgentLoopLlmRequestPreparationService.cs`를 함께 조정해 Code 탭에서는 더 넓은 protected recent window와 더 큰 tool_result budget을 사용하도록 바꿨습니다. `build_run`, `test_loop`, `file_read`, `multi_read`, `lsp_code_intel`, `git_tool` 같은 고가치 evidence는 기본 탭보다 덜 잘리도록 보호합니다.
- `src/AxCopilot/Views/ChatWindow.UtilityPresentation.cs`는 `.ax-context.md`가 아직 없더라도 첫 요청 시점에 workspace context 생성을 바로 시작하고, 생성이 완료되기 전에는 언어 워크플로우 힌트를 bootstrap context로 넣도록 바꿨습니다. 덕분에 완전한 빈 작업 폴더에서도 첫 루프부터 최소한의 프로젝트 힌트가 LLM에 전달됩니다.
- `src/AxCopilot/Services/Agent/AgentLoopService.cs`는 Code 탭에서 generic `session_learnings` 주입을 줄이고, 대신 working set과 query-context diagnostics를 반복마다 `WorkflowLogService.LogTransition(..., \"query_context\", ...)`로 남깁니다. 이제 로그에서 query-view 범위, protected recent 값, supplemental context 수, estimated send token, working-set 요약을 함께 확인할 수 있습니다.
- `src/AxCopilot/Services/LlmService.ToolUse.cs`는 historical tool-call sanitization 결과를 assistant flatten / orphan conversion 건수로 요약 로그에 남깁니다. 사후 보정은 유지하되, 보정 빈도를 추적할 수 있게 만들어 이후 invariant hardening 후속 작업의 기준선을 확보했습니다.
- 검증:
-  - `dotnet build src/AxCopilot/AxCopilot.csproj -c Release -v minimal -p:OutputPath=bin\\verify_context_reliability_full\\ -p:IntermediateOutputPath=obj\\verify_context_reliability_full\\` 경고 0 / 오류 0
-  - `dotnet test src/AxCopilot.Tests/AxCopilot.Tests.csproj -c Release -v minimal --filter "AgentQueryContextBuilderTests|AgentToolResultBudgetTests|AgentLoopIterationPreparationServiceTests|AgentLoopLlmRequestPreparationServiceTests|CodeTaskWorkingSetServiceTests|AgentLoopCodeQualityTests" -p:OutputPath=bin\\verify_context_reliability_full_tests\\ -p:IntermediateOutputPath=obj\\verify_context_reliability_full_tests\\` 통과 150
- 업데이트: 2026-04-16 01:57 (KST)
- Code 탭 컨텍스트 신뢰성 보강 2차 작업을 반영했습니다.
-  - `src/AxCopilot/Services/Agent/AgentMessageInvariantHelper.cs`
-    - 전송 직전 `tool_use/tool_result` 이력을 다시 검사하는 `NormalizeHistoricalToolTrace(...)`를 추가했습니다.
-    - 결과가 없는 structured assistant tool-call은 plain assistant transcript로 평탄화하고, 대응 assistant가 없는 `tool_result`는 plain user transcript로 바꿔 request payload 자체가 더 일관된 상태로 나가도록 조정했습니다.
-  - `src/AxCopilot/Services/Agent/AgentLoopLlmRequestPreparationService.cs`
-    - query view를 그대로 참조하지 않고 deep clone한 뒤 tool-trace normalization을 적용하도록 바꿨습니다.
-    - supplemental message도 clone해서 원본 대화 이력을 직접 변형하지 않도록 했습니다.
-    - normalization 결과(`flattened structured assistant`, `converted orphan tool_result`)를 preparation result에 함께 남깁니다.
-  - `src/AxCopilot/Services/Agent/AgentLoopContextReliability.cs`
-    - `query_context` workflow log에 `tool_trace_repair=assistants:X/orphan_results:Y`를 추가해, 전송 직전 어떤 보정이 들어갔는지 실행 단위로 추적 가능하게 했습니다.
-  - `src/AxCopilot/Services/Agent/AgentLoopDiagnosticsFormatter.cs`
-    - active Code 경로에서 보이던 깨진 컨텍스트 압축 완료 문구를 영어 기준으로 정리했습니다.
-  - `src/AxCopilot/Services/Agent/SessionLearningCollector.cs`
-    - 주석과 주입 메시지, 학습 추출 요약 문자열을 영어 기준으로 전면 정리해 인코딩 손상 시에도 깨진 문자열이 다시 노출되지 않게 했습니다.
- 테스트:
-  - `src/AxCopilot.Tests/Services/AgentMessageInvariantHelperTests.cs`
-    - missing tool-result가 있는 structured assistant flatten
-    - orphan `tool_result` plain transcript 변환
-    - valid structured tool pair 보존
-  - `src/AxCopilot.Tests/Services/AgentLoopLlmRequestPreparationServiceTests.cs`
-    - request preparation 단계가 broken tool trace를 normalize하면서도 원본 query messages는 보존하는지 검증
- 검증:
-  - `dotnet build src/AxCopilot/AxCopilot.csproj -c Release -v minimal -p:OutputPath=bin\\verify_tool_trace_hardening\\ -p:IntermediateOutputPath=obj\\verify_tool_trace_hardening\\` 경고 0 / 오류 0
-  - `dotnet test src/AxCopilot.Tests/AxCopilot.Tests.csproj -c Release -v minimal --filter "AgentMessageInvariantHelperTests|AgentLoopLlmRequestPreparationServiceTests|AgentQueryContextBuilderTests|CodeTaskWorkingSetServiceTests|AgentLoopE2ETests" -p:OutputPath=bin\\verify_tool_trace_hardening_tests\\ -p:IntermediateOutputPath=obj\\verify_tool_trace_hardening_tests\\` 통과 34