Aleksandr Nogikh

pkg/aflow: fix context compression failure on new Gemini models On new Gemini models, context compression caused subsequent API calls to fail with HTTP 400 errors (mostly affecting the Vertex AI API) because cryptographic thought signatures on preserved function calls no longer matched the truncated history. Additionally, mutating the multi-turn history for compression confused summarizer models into executing tool calls instead of summarizing, and unescaped XML framing tags in untrusted history opened prompt injection risks. Refactor context compression to invoke the summarizer with a single prompt using XML-tagged history, auto-fill bypass thought signatures in the Gemini adapter, and disarm reserved XML tags in untrusted payloads.

syz-cluster: evaluate unreproduced diff fuzzing crashes with AI triage Many kernel crashes discovered during patch diff fuzzing lack standalone reproducers or stem from pre-existing base kernel bugs triggered during execution. Relying strictly on reproducer verification misses legitimate regressions introduced by a patch series, while reporting all unverified crashes creates excessive noise. Introduce AI finding triage at the end of the patch fuzzing workflow to evaluate unreproduced crash candidates against the patch series and kernel source. Automatically confirm and report findings that the AI verifies as introduced by the patch series, while categorizing rejected candidates as base kernel bugs.

dashboard/app: use helper methods for AI job and report polling Reduce boilerplate in dashboard AI test cases by consolidating agent and report polling into helper methods. Manually constructing poll requests and confirmation payloads across test functions makes the test suite verbose and harder to maintain. Replace repetitive agent and report polling checks with pollAIJob, pollAIWorkflow, pollAndConfirmReport, and assertNoPendingReport.

pkg/aflow: support Suggested-by tags in patch iteration workflow Credit reviewers who provide idea suggestions during automated patch review. Support extracting and preserving Suggested-by tags across patch iterations and include them when rendering patch commit descriptions.

dashboard/app: allow disabling AI comment replies per stage Automatic comment replying is still an experimental feature. We want to avoid sending unmoderated text replies to public mailing lists, and for now it is cleaner to only allow conversational replies during the moderation stage. This also simplifies the overall implementation of #7671. Add a ReplyToComments configuration option to AIPatchStageConfig. When disabled for a stage, pass the flag to the patch iteration workflow to skip generating text replies and suppress creating external report entries if no new patch version was produced.

dashboard/app: add AI jobs statistics graphs page Provide monthly visualization charts for AI token consumption per model, finished job counts, and job type distributions to track LLM usage and agent workload trends. Aggregate token spans and jobs by month in Spanner queries, expose charts to admins via Google Charts, and add unit tests for graph builders and endpoint access control.

pkg/aflow: support TargetVMArch across actions and workflows For targets where userspace differs from the VM architecture (e.g. 386 on amd64), actions previously failed by attempting to build and run a 32-bit kernel instead of the host VM kernel. Propagate TargetVMArch across workflow inputs and actions. Build kernels and configure runners using TargetVMArch, format RawTarget accordingly, and enable CONFIG_IA32_EMULATION when building amd64 kernels for 386.

pkg/report: improve title and guilty file extraction for iomap crashes Warnings in fs/iomap (such as ifs_free) are triggered when a filesystem violates folio state invariants. Currently, not skipping ifs_free and truncation helpers merges unrelated bugs across different filesystems into a single "WARNING in ifs_free" bucket, and guilty file extraction fails to prefer fs/<fs>/* over sibling helper directory fs/iomap. Skip ifs_free and folio truncation/invalidation helpers in stack traces, and treat fs/iomap and fs/netfs as having "fs" directory depth during guilty file extraction so calling filesystems take precedence.

pkg/aflow/flow/patching: add patch refinement loop for formatting Ensure generated kernel patches comply with kernel coding standards and checkpatch.pl before completion or resubmission. Add checkpatch and clang-format tools. Remove mandatory clang-format-diff.py from crash.TestPatch in favor of tool-based formatting. Introduce patchRefinementLoop in formatting.go to run checkpatch.pl after patch generation and delegate style fixes to a patch-formatter agent. Update verdictAgent to separate StyleItems from CodeItems. Export sandboxed patchdiff.Diff.

pkg/cover/backend, pkg/aflow: resolve source lines to coverage PCs Workflows like seed generation need to determine which KCOV coverage callbacks cover a given file and line. Previously, seedgen parsed DWARF line tables independently, duplicating logic already available in pkg/cover/backend. Add ResolveLineToPCs to backend.Impl to map a file and line number to its covering basic block entry PC. Switch seedgen to this method and remove its duplicate DWARF parsing.