fix(usage): surface active-provider fetch errors instead of dropping them

[junhoyeo]

Problem

fetch_all in crates/tokscale-cli/src/commands/usage/mod.rs spawned each active provider as fetch.call().ok() and collapsed the results with filter_map(|h| h.join().ok().flatten()).flatten(). Any provider that returned Err was silently dropped.

The clearest failure case is Sakana: sakana::fetch deliberately builds a helpful "Sakana session expired or invalid. Refresh SAKANA_SESSION_COOKIE…" error when the session cookie is stale. But because has_credentials() returns true (a cookie exists, it's just expired), the provider is considered active — yet it just disappeared from the output with no message. This contradicts both docs/providers/sakana.md and sakana::fetch's own carefully-worded error path. The same silent-drop affected every usage provider.

Root cause

Errors were discarded at the join/merge step (.ok() + filter_map), with no channel to carry them back to the user.

Fix

• Added fetch_all_with_errors() -> (Vec<UsageOutput>, Vec<ProviderError>). Each active provider's thread now returns (name, Result<...>); successes and errors are split by a new pure helper partition_results.
• pub fn run now calls fetch_all_with_errors and, in non-JSON mode only, prints one concise line per failing provider to stderr: "Sakana: <error> — skipped".
• JSON mode emits no warnings — stdout stays pure JSON and nothing is written to stderr, so --json consumers that also read stderr are not corrupted. (Note: pre-existing unconditional eprintln!s in minimax_tokenplan.rs already violate this, but that's out of scope here.)
• fetch_all is retained as a thin error-discarding wrapper (fetch_all_with_errors().0), so the TUI dashboard caller (tui/app.rs) is unchanged. No UsageOutput struct change was needed.

Approach / tradeoffs

I chose the least-invasive option from the brief: collect per-provider errors and surface them as a one-line stderr warning in non-JSON mode. Rejected a sweeping UsageOutput/struct redesign (carrying an error variant through serialization and the TUI rendering path) as unwarranted for the goal of making failures visible. The TUI continues to use fetch_all and ignores errors for now; fetch_all_with_errors is public so the TUI can adopt error display later without a signature change.

Tests

• Refactored the join/collect logic into a pure partition_results and added two regression tests:
  • partition_results_surfaces_provider_errors_instead_of_dropping_them — a failing Sakana provider's error (including the SAKANA_SESSION_COOKIE guidance) is surfaced, not discarded, while successful providers (including a Multi provider's several outputs) are preserved in order. This fails against the old drop-everything behavior.
  • partition_results_reports_no_errors_when_all_succeed — success path unchanged.
• cargo test -p tokscale-cli and cargo clippy -p tokscale-cli --tests are clean for this file (remaining clippy warnings are pre-existing in report.rs, untouched here).

Residual concerns

• Provider thread panics are skipped (join error has no useful message); this matches prior behavior of not crashing the command.
• No real network/auth round-trip is exercised — only the pure partition logic is unit-tested.

🤖 Generated with Claude Code

---

Summary by cubic

Surface fetch errors from active usage providers in the usage command instead of silently dropping them, so users know why a provider was skipped. Non-JSON runs now print one concise stderr warning per failing provider; JSON output is unchanged.

• Bug Fixes
  • Fixed silent drops by collecting per-provider Err instead of discarding them at join time.
  • Added fetch_all_with_errors() returning (Vec<UsageOutput>, Vec<ProviderError>) and a pure partition_results to split successes/errors.
  • Updated CLI run to print "Provider: — skipped" to stderr in non-JSON mode only; --json emits clean JSON with no warnings.
  • Kept fetch_all() as a thin, error-discarding wrapper so the TUI remains unchanged.
  • Added regression tests to verify errors are surfaced and success paths are intact.

^{Written for commit bcd1785. Summary will update on new commits.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
tokscale	Ignored		Jun 22, 2026 8:21am

[cubic-dev-ai]

No issues found across 1 file

_{Re-trigger cubic}