[pavelhov]

Summary

Ordinary model turns use their known output allowance by default, including existing configurations and requests with a smaller client cap. Native Fable 5.2 gets 128,000 tokens, removing the effort-dependent 32,768-token default that interrupted long coding turns.

The default model-max policy selects known capacity ahead of caller and configured totals. This can increase usage or lower an oversized requested total. Set outputTokenPolicy to request to honor caller caps and configured fallback budgets. Capacity metadata must match the provider destination. Unknown capacities use request and adapter fallbacks. Context compaction has its own maintenance budgets.

This also prepares package version 0.1.27 for the official signed updater release.

Validation

• Focused public-entry, adapter, configuration, and routing checks passed 383 tests.
• Typecheck, privacy scanning, GUI lint, and the documentation build passed. GUI lint reported its existing Combos hook warning.
• The full parallel suite passed after its four legacy truncation cases selected the request-policy opt-out. The required failed-file parallel rerun passed all 50 tests, including new JSON/SSE default-policy coverage.
• Isolated real Sparkle replacement and relaunch qualification is running with disposable keys and uniquely named fixtures on both architectures.

Security review

Reviewed native Messages mutation, destination-scoped capacity metadata, mixed-wire routing, private provenance fields, cancellation, and credential boundaries. Private routing fields are request-local and strict configuration validation rejects them on disk.

Release preparation uses the existing external Ed25519 trust anchor and pinned Sparkle 2.9.6 tools. Signing material uses owner-only files and stdin. Universal archive identity and signatures will be verified against the qualified source commit, downloaded draft assets, and the published updater discovery URL. The release is an ad-hoc, unnotarized preview.