Dmitry Vyukov

This commit **enhances error handling** within the **`aflow` package's LLM agent** by recognizing an additional API error code. It **fixes** an issue where the system did not properly handle API error code 499 (cancelled) during LLM error parsing. Specifically, the `parseLLMErrorImpl` function in `pkg/aflow/llm_agent.go` now includes this error in the list that triggers a retry mechanism. This ensures that **cancelled LLM API requests** are correctly identified and lead to a retry, improving the overall robustness of LLM interactions. A new test case has been added to verify this enhanced retry logic.

This commit introduces a **bug fix** within the **`pkg/aflow`** module, specifically addressing an issue with **MCP tool naming conventions**. It **modifies the `registerMCP` function** in `mcp.go` to **replace hyphens with underscores** in **MCP tool names** during registration. This change is critical to prevent the generation of **invalid Python method names** in downstream **MCP clients**, ensuring proper client-side integration and functionality. A new test in `pkg/aflow/flow/flows_test.go` has also been added to validate that MCP tool names adhere to this corrected format, preventing future regressions.

This commit introduces a **new capability** to the **`pkg/aflow`** package by adding a `session-initializer` **MCP tool**, which is crucial for populating session state and substituting workflow arguments in the Multi-Component Pipeline system. It significantly **enhances the MCP system's data flow** by modifying `registerMCPAction` to return structured action results, making critical outputs like `ReproducedCrashReport` directly accessible for subsequent tools. Additionally, the commit includes a **refactoring** of schema generation within `pkg/aflow/schema.go` and updates `docs/mcp.md` with manual testing instructions for the **MCP server**, effectively **fixing #6929**.

This commit **refactors** the **`aflow` package** by **renaming output fields** within the **`action/crash` module**. Specifically, the `BugTitle` and `CrashReport` fields in the `reproduceResult` struct are now `ReproducedBugTitle` and `ReproducedCrashReport` to prevent naming conflicts with original bug data. This change ensures that workflows consuming these outputs, such as those in the **`flow/patching` module**, can clearly distinguish between original and reproduced crash information. The update prevents potential data clashes and enhances the clarity of crash reproduction results within the **`aflow` system**.

This commit **disables specific tests** for the **`grepper` tool** within `pkg/aflow/tool` on non-Linux operating systems. It addresses a **test failure** on platforms like OpenBSD, where the error message for "bad expression" differs from the expected output. By adding a `//go:build linux` directive to `grepper_test.go`, the **test suite** will now exclusively run on **Linux**, preventing CI instability on other OSes. This is a **maintenance fix** to ensure test reliability across different environments.

This commit **enhances error reporting** by systematically displaying verbose output from failed binary executions across the system. It introduces a new utility function, `osutil.VerboseMessage`, to extract detailed error information, which is then integrated into the **`syz-agent`** and the **`syz-aflow`** tool. This change **refactors** existing custom error handling in `pkg/aflow` modules, replacing ad-hoc solutions with a unified approach. The primary impact is providing **more useful error messages** for operations like `git` or `make` failures, significantly improving debugging and diagnostic capabilities.

This commit **repairs** a **bug** in the **AI job access check** within the `dashboard/app` module, specifically affecting the `filterJobsAccess` function. Previously, the system would attempt to select more than 1000 bugs, causing a `datastore_v3: BAD_REQUEST` API error. The fix **limits the number of bugs processed** to avoid this API failure, ensuring the **AI job access check** operates reliably and correctly filters access levels for AI jobs.

This commit **enhances the `pkg/aflow` module** by preventing the registration of **MCP tools with duplicate names**, which previously led to unpredictable tool selection. It introduces validation logic in `pkg/aflow/mcp.go` to ensure unique tool identifiers, while also ignoring specific tools like `set-results` and those created in tests. As part of this **bug fix**, the conflicting `crash-reproducer` tool in `pkg/aflow/tool/syzlang` is **renamed to `reproduce-crash`** to resolve an existing naming conflict. This **refactoring** and **validation addition** significantly improves the **robustness and predictability** of the `aflow` system's tool management.

This commit **enhances error handling** within the **`pkg/aflow`** module's **LLM agent** by expanding the `parseLLMErrorImpl` function. It now correctly identifies and categorizes **gateway timeouts** and a specific **'quota exceeded' message** from Generative AI services. This **bug fix** improves the robustness of the system by enabling more precise retry logic and better operational resilience when interacting with LLMs. New test cases have been added in `llm_agent_test.go` to validate these expanded error parsing capabilities, ensuring the system can more effectively manage transient and resource-related issues.

This commit introduces a **new capability** within the **`aflow` package** by replacing the `Provide` action with a dedicated `Flow.Consts` field for defining fixed initial inputs. This **refactoring** clarifies that these inputs are flow properties rather than actions, integrating them directly into the flow registration and execution context (`pkg/aflow/execute.go`). The change simplifies syntax by using a map for constants and required **maintenance** updates to existing test helpers and numerous test cases, along with new tests for `Flow.Consts` functionality. This ensures a more semantically accurate and less verbose way to configure initial flow states.

This commit **enhances the `pkg/aflow/flow/repro` module** by providing **syzkaller program and syscall syntax documentation** directly to the LLM agent. A new file, `docs/docs.go`, was added to embed this crucial information, which is then integrated into the agent's prompt within `pkg/aflow/flow/repro/repro.go`. This **new capability** aims to improve the LLM agent's understanding and accuracy when generating or analyzing syzkaller programs, ensuring it has all necessary syntax details. The change directly impacts the **AI agent's ability to process syzkaller-related tasks** more effectively.

This commit **enhances the crash reproducer agent** within the `pkg/aflow/flow/repro` module by **integrating codesearch and grepper tools**. This **new feature** allows the agent to access and analyze kernel source code, which is critical for effectively creating bug reproducers. Additionally, the commit includes a minor **refinement to the agent's prompt instructions**. This improvement significantly aids in the debugging process by providing the agent with better context and search capabilities, thereby streamlining the generation of accurate bug reproducers.

This commit introduces the **new capability to collect execution coverage** during syzkaller program testing and crash reproduction. It involves **refactoring** the `pkg/aflow/action/crash` module to integrate coverage collection into its reproduction logic and extending the `pkg/instance` package's program execution environment to handle and parse coverage data. Core components like `pkg/bisect`, `syz-ci/jobs`, and `syz-cluster/pkg/retest` are updated to propagate the new `collectCoverage` parameter. This **foundational work** provides the necessary plumbing for future features that will utilize this coverage information, such as improved crash analysis or seed generation, addressing requirements like those in #6878. While not yet fully wired into existing workflows, it establishes the infrastructure for advanced coverage-driven development.

This commit introduces a new `Tools` utility function within the **`pkg/aflow`** module to safely combine multiple tool slices. This **new capability** prevents potential aliasing issues and subtle bugs that can arise when managing tool sets for agents. Existing tool combination logic in `pkg/aflow/flow/assessment` and `pkg/aflow/flow/patching` sub-modules has been **refactored** to leverage this new `aflow.Tools` function. This change improves code robustness and simplifies tool management across various agent flows by providing a dedicated, safe mechanism for tool aggregation.

This commit **refactors** the **`syz-agent`**'s update logic by simplifying its handling of `SyzkallerCommit` data, as this information is now consistently provided by the dashboard. Specifically, the `CheckLatest` function within `pkg/updater/updater.go` has been **renamed to `checkLatest`**, making it an unexported, internal utility. This **maintenance** work ensures that the agent's codebase remains lean and avoids redundant logic, improving the overall clarity and efficiency of the **`pkg/updater`** module. The change reflects that the `syz-agent` no longer needs to independently determine `SyzkallerCommit` information, as the dashboard consistently supplies it with job crash data.

This commit **refactors** the handling of `syzkaller` commit information within the **`pkg/aflow/flow/repro`** package. It streamlines the `ReproInputs` struct by removing the `SyzkallerCommit` field, simplifying its definition. Instead, the `syzkaller` git revision is now provided directly during the flow's initialization, ensuring this critical detail is managed more explicitly at the point of use. This **refactoring** improves the clarity and directness of how `syzkaller` versioning is integrated into the `repro` process, affecting the internal structure and initialization logic of `repro` inputs.

This commit introduces a **new capability** to programmatically retrieve a list of system call description files, primarily affecting the `pkg/aflow/tool/syzlang` package. It embeds `linux/*.txt` files into the `sys/syz-sysgen` binary using `go:embed` to make them accessible via the new `DescriptionFiles` function. This functionality is then integrated into the **reproducibility flow** within `pkg/aflow/flow/repro`, ensuring that bug templates now include the relevant description files for improved debugging. A generic `Provide` function was also added to `pkg/aflow/action.go` to support this integration, enhancing the overall `aflow` ecosystem.

This commit **fixes a data race** within the **`dashboard/app` module** concerning the `localUIConfig`. The **bug fix** addresses an issue where concurrent access to the UI configuration could lead to unpredictable behavior or test failures. It achieves this by modifying `dashboard/app/app_test.go` to ensure `localUIConfig` is properly initialized and checked via `checkConfig` within the test setup's `init` function. This change significantly improves the **robustness of the dashboard's UI configuration handling** and enhances the reliability of its test suite by preventing race conditions during configuration setup.

This commit introduces a **new capability** to automatically apply an **'actionable' label** to bugs following **AI moderation**, enabling their auto-upstreaming. The `aiBugLabel` function within the **dashboard's AI component** (`dashboard/app/ai.go`) now leverages a newly defined `ModerationOutputs` struct from `pkg/aflow/ai/ai.go` to determine actionability. This change standardizes the output of the **Aflow moderation flow** and integrates the new `ActionableLabel` constant into the system. The primary impact is streamlining the bug triaging process by automating the identification and upstreaming of actionable issues.

This commit introduces a **new capability** to the **dashboard application**, allowing users to efficiently filter and view only **AI jobs that require moderation**. The backend logic in `dashboard/app/ai.go` was updated within the `handleAIJobsPage` function to incorporate this new 'Needs Moderation' filter and enhance the overall job filtering mechanism. Concurrently, the `dashboard/app/templates/ai_jobs.html` template was modified to integrate this filter into the user interface and improve usability by automatically submitting the filter form when a workflow selection changes, removing the need for a manual submit button. This enhancement significantly **streamlines the moderation workflow** for AI jobs, making it easier for administrators to identify and act on tasks requiring their attention.