Fighter90/career-ops-ui v1.258.0

on GitHub

[1.258.0] — 2026-10-25

Six fixes from the agent fleet, one train: the #/scan company counter stops mixing universes, deep research demands sources, form errors light up the field, loading states get skeletons and an instant h1, and evaluations cache their 37k-token prompt prefix.

Fixed

• #/scan «Active companies» counter (CAR-100): the numerator counted distinct companies across ALL result rows (1554 on prod — hh.ru alone contributes hundreds) while the denominator counted API-backed tracked companies (~90) — two universes in one fraction. Pairing (a): the numerator now counts companies among rows the API-backed tracked scan itself fetched (matched by _entry ?? company; hh.ru/telegram employer rows drop out); the pre-scan fallback moved into the same universe (it could read «104/90»).
• Deep research fabrication (CAR-110): the deep prompt now carries a hard contract — every factual claim ends with [source: domain.com] (cannot name it → do not state it; inventing a marker is worse than writing "unknown"), unknown facts are written unknown, and the report MUST end with a ## Data freshness section. Freshness anchors: today's UTC date + an optional company url on POST /api/deep (bounded, never fetched server-side).
• Form errors highlight the field (CAR-130): a failed POST marks the first form field aria-invalid + .is-error and focuses it (WCAG 3.3.1); cleared on edit. The network-error hint is deployment-aware — loopback keeps the dev-command hint, a deployed origin gets «server is not responding» instead of bash web-ui/bin/start.md advice (the connection banner had the same hardcode).
• Loading states (CAR-112): #/map and #/tracker render their h1 instantly (the router focus target existed only after data arrived) and a shared skeleton helper (Skeleton.rows, aria-hidden shimmer rows inside a role="status" region reusing common.loading) replaces bare «Loading…» text on tracker, scan results and the map shell; prefers-reduced-motion stops the pulse.
• Evaluate prompt caching (CAR-111): buildEvaluationPromptParts splits the ~37k-token prompt into a byte-stable prefix (CV + profile + mode files + instructions) and the JD tail; the Anthropic path sends [prefix + cache_control: ephemeral] + [JD] — the prefix bills at the cache-read rate (≈10× cheaper) on every call after the first. Missing/mismatched prefix fails open to the byte-identical old request; the OpenAI-compatible cascade is untouched (server-side caching there).

Notes

• All six fixes came from the agent fleet with red→green evidence (5 + 7 + 7 + 7 + 8 new cases) and review rounds; the first CAR-111 CodeQL alerts (#579/#580) were fixed in code — the test stub parses the URL host instead of substring matching, and a dead full-prompt string was dropped.
• The score backfill ran on prod (v1.256.0 endpoint, owner-approved): updated: 22, missingReport: 0, reportScoreless: 159 — 159 scoreless rows point at reports whose headers predate the machine block or carry unparseable score formats; recoverable only by re-evaluation or a wider parser pass (noted in CAR-129).
• Deliberately not done: no commute-column nulls-last change (pre-existing Infinity-on-desc quirk), no retroactive re-pricing of old usage records, no .is-error-style highlighting for the scan filter forms (CAR-130 scoped to the shared mode-page submit helper + api.js hint).
• Tests 5749 (+36). No scanner sources touched (registry 121 = 116 EN + 5 RU · 116 adapters); no i18n keys (the CAR-130 hint reuses conn.down; the skeleton region reuses common.loading).