Taras Madan

pkg/spanner: unify Spanner client abstraction and switch to emulator-based testing - Refactored pkg/coveragedb, pkg/cover/heatmap, syz-cluster, and dashboard/app to use concrete *spanner.Client instead of mockable interface proxy. - Moved Spanner Emulator control code, ParseURI, and database setup to a shared library utility in pkg/spanner/emulator.go. - Replaced mock-based testing with Spanner Emulator-based testing across all test cases. - Removed deprecated interfaces from pkg/spanner and deleted the spannerclient/ proxy layer and its mocks.

dashboard/app: hide AI job errors for non-admins Only expose error messages and error summaries on AI job pages, trajectory logs, and job review history to users with admin privileges. This prevents leaking potentially sensitive details to non-admin users. Add a test case to verify error visibility restrictions.

pkg/aflow: explicitly pass model to generateContent in llm_agent Previously, generateContent implicitly used the main agent model (a.Model) for all requests. This caused compressContext to execute on the main agent's model config instead of GoodBalancedModel. We know the original intention was to use GoodBalancedModel because: 1. In compressContext, the trajectory span is explicitly initialized with GoodBalancedModel: Model: ctx.modelName(GoodBalancedModel) 2. The comment above the config explicitly says: "Lightweight config targeting the Flash model." Fix this by explicitly passing the model to generateContent and generateContentCached, and passing GoodBalancedModel from the compressor.

coveragedb: add sessions table to track active sessions for GC * Introduce 'sessions' table to track active and incomplete sessions. * SaveMergeResult leaves completed sessions in the 'sessions' table. * DeleteGarbage deletes active sessions that are older than 1 week and not in merge_history. * Include auto-migration logic for legacy abandoned sessions in DeleteGarbage.

dashboard/app: migrate header navigation to bootstrap Migrate the custom header and sub-navigation bar to standard Bootstrap components (.navbar, .nav-pills, .dropdown). This simplifies the codebase and introduces several layout improvements: - Use Bootstrap's flex alignment (ms-auto) to push the "Send us feedback" button to the far right. - Integrate Bootstrap's .dropdown-toggle to show standard caret symbols on expandable menus. - Standardize layout rules with flex stretching to ensure all navigation buttons render at the exact same height regardless of content or state. - Align topbar metadata links vertically with the brand logo.

pkg/aflow, dashboard: support model pools with fallbacks and abstract tiers Introduces support for configuring fallback model pools in workflows. LLMAgent and LLMTool Model fields can now be assigned abstract tier constants (e.g. BestExpensiveModels, GoodBalancedModels) or comma separated concrete fallback pools. The workflow runner will attempt to execute each concrete model from the pool in sequence, falling back to the next candidate in the list if the current one fails (e.g. on quota errors). Start spans log an empty Model field to avoid sending fallback pool noise to the dashboard, and are updated to show only the successful concrete model in the final span once the task successfully executes.

all: migrate golang.org/x/exp/maps to standard maps Migrated all usages of golang.org/x/exp/maps to the standard library maps package (and slices package where iterators are collected). Removed golang.org/x/exp from direct dependencies in go.mod.

pkg/coveragedb: clean up orphaned functions

dashboard/app: support flexible coverage export with filtering Add `with-covered` and `with-uncovered` query parameters to the public JSON coverage API to allow filtering of covered and uncovered line coordinates. Add `filepath` query parameter to allow path prefix filtering, which is passed down to Spanner query to speed up retrieval.

pkg/aflow: convert model configurations to ModelType enum Introduces the ModelType enum to represent model tier configurations (BestExpensiveModel and GoodBalancedModel) instead of raw strings. This provides typing for LLMAgent and LLMTool model configurations. Unrecognized/concrete model names (e.g. from tests or overrides) are resolved via the default case in modelName.