Tags: Brandon168/donsetch

release: v3.2.3

release: v3.2.2 — ghost leak fix, profile collision, donsetch stop, f…

…uzz crash

fix(pi-ext): handle stream EPIPE when MCP server dies mid-request (WS…

…L crash)

release: v3.2.0 — search-legibility (snippet clipping, engine provena…

…nce, whitespace collapse) + CI clippy gap fix

revert: proxy unification — fetch/ghost stay direct-only

Fetch makes one request to one URL, no rate limiting. Routing
through proxies wastes bandwidth and hurts the TLS fingerprint
(residential proxies don't use our Chrome-true BoringSSL stack).
Proxies belong on search (many engines) and crawl (many pages,
same host) only.

Also: homebrew formula URL fix (asset names have no version
in filename) and CHANGELOG updated.

DonSeTch 3.0.0 — context warfare

style: cargo fmt

fix: Cyrillic search results mangled (dondai44423#28) + cached search…

… ignores max-results (dondai44423#29)

dondai44423#28: Search engine result pages were decoded with
from_utf8_lossy, producing replacement characters for non-UTF-8
encodings (Windows-1251 Cyrillic showed question marks). Search
now uses charset::decode with full detection pipeline.

dondai44423#29: rank::merge trimmed to max_results before caching. First
search with max=2 cached only 2 results; later search with max=10
got the stale 2. Merge now always produces 12 (cache ceiling),
response trims to max_results, cache stores full 12.

Also: merged dependabot bumps (boring 5.2.0, tokio-boring 5.2.0,
boring-sys 5.2.0, futures-util 0.3.34, download-artifact v8).

fix: crawl PDF 3s timeout (dondai44423#26) + MCP text content dropped…

… by Claude Code (dondai44423#27)

dondai44423#26: Crawl PDF extraction timeout was 3s (from PR dondai44423#23), far too
short for real PDFs. A 28 MB archive.org PDF takes ~70s. Bumped
to 300s (5 min). fetch was never affected.

dondai44423#27: Claude Code and VSCode drop text content when
structuredContent is present, showing agents only metadata.
Fix: all MCP responses now prepend a compact [meta] JSON text
block with essential fields (url, tier, verdict, content_ok,
thin, next_offset, tokens_est, lang, title, pdf_pages) before
the content. Clients that only show text now see both. Error
responses include next_action in text. CLI skips [meta] blocks
when extracting content for stdout/json output.

fix: cap pdf per_page array in MCP response (dondai44423#25)

The markdown was correctly paginated to max_chars, but the MCP
structuredContent included the full per_page array with one entry
per page. A 1032-page PDF produced 60K of per-page JSON alone,
blowing past the MCP response limit even with max_chars=400.

Cap per_page at 50 entries. Include a summary (total pages, OCR
pages, mean confidence) always. per_page_capped signals when the
detail was truncated.