[hummern]

Summary

Fixes #13198. reka-flash-3 always runs a reasoning phase, so a caller-set undersized max_tokens (e.g. 256) is consumed entirely by reasoning_content and the upstream returns empty content with finish_reason: length. A transparent OpenAI-compatible proxy forwards that reasoning-only, empty-content response to the client, which can surface as huge garbage output (e.g. Hermes Agent's forwarded reasoning onto Telegram).

ensureThinkingBudget() (#6912) already floors undersized budgets for reasoning models, but it only fired when the client explicitly opted into reasoning. reka-flash-3 reasons unconditionally.

Change

• Add RegistryModel.alwaysReasons (boolean): true when an upstream always reasons regardless of client opt-in. Set it on reka-flash-3.
• DefaultExecutor.ensureThinkingBudget() treats alwaysReasons === true as implicit reasoning opt-in, so a tiny caller max_tokens gets the 4096 floor.
• Regression tests (verified to fail without the fix).

Repro → fixed

max_tokens=256  → content:'', finish_reason:length          (before)
max_tokens=4096 → content:'TELEGRAM_TEST_123', finish_reason:stop  (after)

Tests

• tests/unit/clinepass-thinking-budget.test.ts — 9/9 pass; the new alwaysReasons cases fail without the fix (sabotage-verified).
• tests/unit/provider-models-config.test.ts — 12/12 pass.
• tests/unit/provider-models-route.test.ts — 59/59 pass.
• npx tsc --noEmit -p tsconfig.typecheck-core.json — clean (exit 0).
• Pre-commit hooks (prettier, eslint, any-budget, quality gates) passed.

Related

• fix(providers): max_completion_tokens ignored when thinking mode enabled (non-ClinePass providers) #6912, fix(command-code): floor tiny muse-spark output budgets so hidden reasoning cannot consume the whole budget #12497, feat(reasoning): opt-in min output budget floor for thinking models #12742 (same class: reasoning models burning tiny output budgets)

Reviewer note

The two eslint errors in the touched files (shared.ts:791, default.ts:5) are pre-existing baseline warnings on lines I did not modify, not introduced by this diff.

[diegosouzapw]

The alwaysReasons floor for reka-flash-3 is a clean, well-tested fix (9/9 passing, good
coverage of the non-clinepass and non-regression cases). One thing to flag before merge:
your branch carries an unrelated commit (c70e425, "Change hasFree from true to false for
pioneer.ai") that touches frontier-labs.ts — nothing to do with reka-flash, and it looks
stale against the current Pioneer AI notice text on the tip. Could you drop that commit (or
rebase it out)? If the Pioneer AI free-tier change is still accurate, happy to see it as its
own small PR with a current source.

[hummern]

Closing this PR — the alwaysReasons floor for reka-flash-3 was already merged upstream via #13531 (2d1a281), and the unrelated pioneer.ai commit (c70e425) has also landed in release/v3.8.51. The branch is now rebased clean against the current tip. Thanks for the review feedback!

[hummern]

Superseded — fix already merged via #13531.