- Stop cleanup interval when all auth sessions are removed (unregister,
cancel, cancelAll) so the Node event loop isn't held alive.
- getActiveSessionForProvider now returns the most recent session by
startedAt instead of the first map entry.
- HTTPS tunnel includes port in Host header when non-443.
forceValidChars() now ensures leading char is letter or underscore
(catches dot, hyphen, colon, slash — not just digits). registerTools()
always registers in mapping so collisions are detected regardless of
tool registration order.
sanitizeToolName() fallback now forces valid characters (replaces
unsupported chars with underscores, ensures leading letter/underscore)
instead of only truncating. registerTools() disambiguates colliding
sanitized names with numeric suffixes so restoreToolUse() always maps
back to the correct original.
- base-config-loader.ts moved from src/cliproxy/ to src/cliproxy/config/,
so __dirname relative path needs one more .. to reach config/ dir
- Update mock.module paths for proxy-detector and routing-strategy
in non-colocated test files
- Add src/**/__tests__/** to tsconfig.json exclude list
- Add ignores pattern for __tests__ in eslint.config.mjs
- Fix ui/src/lib/api-client.ts import path for provider-entitlement-types
- Remove stale build artifacts from ui/src/lib/ and tests/mocks/
Apply Copilot review feedback on PR #1113:
- getCodexWindowKind: require cadence alongside category; fall through
to label sniffing when missing instead of silently defaulting to 5h.
- getCodexWindowDisplayLabel (CLI): for category 'additional', fall back
through featureLabel -> window.label -> 'Additional' so partial cached
windows preserve user-visible context.
The Codex wham/usage API moved GPT-5.3 Codex Spark quota out of
code_review_rate_limit (now null) into a new additional_rate_limits[]
array. Parse the new field and surface it as Spark windows in the quota
view. Adds explicit category/cadence/featureLabel metadata on
CodexQuotaWindow so display logic does not depend on label sniffing.
The legacy label-sniffing path is preserved as a fallback so cached
windows from older versions still render correctly.
Anthropic OAuth usage returns utilization as a percent (0-100) while the
older policy-limits restrictions payload returns a 0-1 ratio. The shared
heuristic "value <= 1 means ratio" misinterpreted percent values like
1.0, rendering 1% Sonnet usage as 100% used. Pass the unit explicitly
from each payload branch so the same boundary value is interpreted
correctly in either format.
Tam Nhu Tran