Skip to content

dethrottled 1.1.2 — broader autonomous discovery and audited limits

Choose a tag to compare

@davidjangel davidjangel released this 02 Nov 13:04
· 68 commits to main since this release

Completes the coverage and reproducibility improvements assessed in the November 2 audit. Search accepts up to 50 results, research up to 30 sources and 10 additional queries, and explicit text reads up to one million characters. Defaults remain smaller; the limits page explains bounded jobs and continuation.

1.1.2 fixes automatic broad-search escalation. With cheap TLS discovery enabled, a request for 50 results no longer stops at ten: it invokes browser discovery, merges unique URLs and retains cheap results on browser failure. Diagnostics distinguish fulfilled and partial breadth. Old short browser-cache entries are bypassed for the changed semantics. Small requests retain the fast cheap path.

Before publication, an isolated API plus actual MCP SDK returned 50 unique URLs through the complete TLS/browser path in about 4.36 seconds (TLS 10, browser 50). This demonstrates additional discovery, not 50 independent facts or answer accuracy. The failed public 1.1.1 check is retained in the evidence. No general hostile-source breakthrough is claimed.

The audit changes include complete dependency locks for both adapters, a pinned multiarchitecture Python base, tested runtime dependency upgrades, backend-version diagnostics, compact MCP JSON preserving native payloads, and explicit documented limits. Four previously affected dependency groups have no known advisories in the checked updated locks. OpenTelemetry remains because FastAPI requires it; two unrelated major dependency upgrades remain held for compatibility review.

658 regression tests pass. All 14 CI checks pass, including native AMD64/ARM64 browser pagination, 21 fresh-stack contracts and restart restoration per architecture. SDK 1.30.0 stays pinned. The browser worker from 1.1.1 is unchanged; upgrading from older releases requires rebuilding it as well as the API. Host adapters need updated source/version metadata and the full adapter lock; reconnect MCP clients to refresh cached schemas.

Exact-commit CI · Whole-path evidence and negative results · Audit assessment · Limits and defaults

Production verification: public MCP returned 50 search results and a 68,405-character retained document; REST/MCP browser handoff, native clicks, replay and caller isolation pass. Bridge and backend both report 1.1.2; all rebuilt workers are ready, real AMD WebGL is verified, and retained SQLite databases pass integrity checks. PyPI installation/import succeeds and file hashes match the release build. Checksummed raw verification records are attached below.

Post-merge main CI also passes all 14 checks.