Repository navigation
Dethrottled 1.1.3 — measured source choices and dependable cooldowns
Search menus stay in source order after a 120-question source-choice study, with 80 development and 40 held-out questions. HTTP now defaults to ten choices, matching MCP. Agent instructions explain when to read, expand to 20 or rewrite around a missing fact; the 50-result ceiling remains available. Automatic Bing/Brave fusion and snippet reranking were not justified on the tested host and are not enabled.
This release fixes two upstream-control gaps: a sole/all-blocked engine selection no longer bypasses its cooldown, and the pinned Compose SearXNG honors longer Retry-After suspensions. MCP imports the API's shared workflow instead of maintaining a duplicate. Staged tests caught and reproduced stale instructions despite matching application versions.
The held-out DDG10 menu averaged 2.53 useful choices in the first three; equal-weight Bing/Brave RRF averaged 1.73. These are model-assisted menu judgements, not verified-answer accuracy. DDG20 improved estimated evidence coverage by 5 percentage points with more off-topic material. Bing browser/locale controls were more promising than its default scraped route, but need broader validation. Brave still returned 429 at the 30-minute recovery probe.
Validation includes 677 offline tests in the full API image (15 skipped), native AMD64/ARM64 default-Compose gates, 36/36 paced size checks, 9/9 bounded burst checks and 283 unique fetched URLs. Production on szbox reports 1.1.3 across API/MCP; authenticated public HTTPS MCP returned 10 default and 20 explicit results with matching workflow hashes. The existing shared helper also produced a busy-client 429, retained separately; a temporary independent test identity passed and was revoked afterward. No paid provider, ranking API or new runtime package was added.
Full report, methods and limitations · CI
Attached replay data includes frozen lists, labels, original provider ranks, selection lock, per-query comparisons and hashes. Ranking replay needs only Python's standard library. Credentials, browser state and bulk fetched third-party page bodies are excluded. The separate production verification records the exact deployed identity and functional checks.