PR-Agent flagged that the leaf coalesceKey omitted `stage`, so two
adjacent no-requestId logs differing only by stage (e.g. one logging
`route` and the next `dispatch`) would merge into a single `×2`
row, hiding a distinct entry.
Add `stage` to the coalesce tuple. New unit test asserts that two
leaves identical except for stage stay as separate rows.
Refs #1138, #1151
PR-Agent flagged that the trace-child coalesce was hiding legitimate
repeated stage emissions: a request that retries an upstream call or
emits the same stage twice for any reason would show only the first
occurrence behind a `×N` badge, with subsequent attempts no longer
selectable or inspectable.
The original noise problem motivating intra-trace coalesce was the
149-stage `web-server:http` self-polling trace, but that's already
hidden by default via the `hideDashboardInternals` filter. Users who
opt in to see internals are deliberately debugging the dashboard and
SHOULD see all 149 stages — collapsing them defeats the opt-in.
Drop `coalesceChildren` entirely. Trace children render uncoalesced
so retries, duplicated stages, and multi-attempt traces stay
individually inspectable. Standalone-leaf coalesce stays — that's
the actual user-facing dedup the PR was about.
Tests: drop the 7 `coalesceChildren` cases; keep 16 covering
`deriveStageHint` and `deriveTraceGroups`.
Refs #1138, #1151
Three component tests under jsdom: skeleton render, search filter wiring,
row-click -> detail-panel-header flow. react-virtuoso is mocked at module
level (vi.mock) so jsdom can render rows without layout measurements.
Production behavior unchanged.
Refs #1142, #1138
Mirror the server-side CodexQuotaWindow metadata extension on the UI
side: api-client.ts gains category/cadence/featureLabel; utils.ts
breakdown returns additionalWindows so Spark quota does not pollute
core 5h/weekly badges; quota-tooltip-content renders Spark inline with
prettified labels (GPT-5.3-Codex-Spark -> Codex Spark). Adds breakdown
unit tests and i18n strings for English/Chinese/Vietnamese/Japanese.
- fall back to stale CLIProxy version caches instead of surfacing 500s
- remove duplicate dashboard update-check fetches and disable retry/focus refetch
- add backend and UI regressions for the degraded-path behavior
- keep raw JSON model mappings aligned with requested=upstream semantics
- block malformed mapping lines instead of coercing them on save
- render saved mappings as requested to upstream and add regression coverage
Refs #941
- add the huggingface preset and hf alias to the shared preset catalog
- default the preset to Factory Droid because the documented path is generic chat completions
- cover the CLI and dialog behavior with preset metadata and UI tests
- keep requested=upstream semantics stable in the AI Providers editors
- preview the client-visible ANTHROPIC_MODEL instead of the upstream target
- add UI coverage for alias parsing and formatting helpers
Refs #941
