refactor(poller): share background job scheduling by jzila · Pull Request #1703 · kenn-io/agentsview

jzila · 2026-09-10T17:49:38Z

Claude usage collection will also require polling. This moves pricing refresh
onto a shared scheduler so both jobs can use the same timing and retry behavior.

Jobs are configured at startup, with jitter, cooldown, capped retry backoff,
Retry-After support, and manual triggering. Pricing remains the only job; it
refreshes at startup and serializes writes with resync. State stays in memory,
with no new configuration, API routes, or tables.

Review internal/poller and cmd/agentsview/pricing_schedule.go.

roborev-ci · 2026-09-10T18:04:44Z

roborev: Combined Review (`9f570c1`)

Verdict: Changes require fixes for 2 findings.

Medium

internal/config/config.go:1743-1745: AGENTSVIEW_POLLER_ENABLED is loaded before the config file, but an explicit [poller].enabled value unconditionally overwrites it, so the documented environment override cannot take effect when the file specifies enabled. Preserve an environment-set value when applying TOML, or apply the environment override after loading the file; add a file-plus-environment precedence test.

Reported by: codex
internal/cursorusage/job.go:22-26,51-58: Cursor polling always uses a fixed 24-hour lookback, even when the configured cursor-usage interval is longer, creating gaps where events are permanently missed. Make the lookback at least the configured interval plus scheduling jitter, or persist and use the last successful fetch boundary so consecutive polls overlap.

Reported by: codex

Reviewers: 2 done | Synthesis: codex, 11s | Total: 9m28s

roborev-ci · 2026-09-10T19:29:49Z

roborev: Combined Review (`39818e3`)

No issues found.

Reviewers: 2 done | Synthesis: codex | Total: 4m50s

Add internal/cursorusage.Job, registered on the internal/poller Scheduler when cursor_admin_api_key is configured, so the archive stays current between manual `agentsview usage cursor` runs. Refactor the CLI command onto the same FetchAndStore helper the Job uses, so on-demand and scheduled fetches share one implementation and the existing cursor_usage_events dedup key holds across both. Fix roborev-ci finding on upstream PR kenn-io#1703 (kata vtk4): the poll's lookback window was a fixed 24h regardless of the configured interval, so an interval configured longer than 24h would permanently miss events landing in the gap between polls (the interval is 30m by default today, but [poller.intervals] can override it once the [poller] config lands). Add resolveLookback, which keeps the existing 24h default but grows the lookback to interval-plus-a-margin whenever interval exceeds it, so consecutive on-schedule polls always overlap.

…] config Add a [poller] config section (master enabled switch plus per-job interval overrides) plumbed through the same load/env/TOML spots as cursor_admin_api_key, and wire pricing refresh and the Cursor usage poll onto it via resolvePollerInterval. Add an IdleNotifier interface to internal/poller (Options.KeepsDaemonAlive, wired to server.IdleTracker) so a future job can opt into keeping an otherwise- idle detached daemon alive; both current jobs default to false, so neither does today. Add `agentsview doctor pollers`, an offline diagnostic that reads poller_status directly from the SQLite archive, and GET /api/v1/system/pollers, a live status endpoint backed by the running Scheduler, plus the generated frontend client for it. Fix roborev-ci finding on upstream PR kenn-io#1703 (kata vtk4): loadEnv runs before loadFile, so an explicit config.toml [poller] section unconditionally overwrote whatever AGENTSVIEW_POLLER_ENABLED set, even when the file only restated the default. Track that the environment variable set Poller.Enabled and skip the file's value in that case, so the documented environment override actually wins. Covered by a table-driven test over all four true/false combinations of env and file values.