Skip to content

v0.3.0

Choose a tag to compare

View all tags
@anshal21 anshal21 released this 10 Jul 04:28
v0.3.0
b0ac535
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Changelog

  • b0ac535 Designed banner hero + README truth fixes (#5)
  • af959fe Docs refresh + zero-config smart routing (#4)
  • 6cf7aad Router: out-of-the-box policies, battle-tested reconciliation, calibrated signals, one design doc (#3)
  • a067f46 Add whittle route: an opt-in local model router with vSR-aligned smart mode (#2)
  • 55ead8a Shape-preserving hook replacements, uncapped updatedToolOutput, 100k prose ceiling (#1)
  • 7108174 daemon: self-heal the Claude Code hook — re-assert the PostToolUse entry on startup + every 5min if it vanishes (we observed it get dropped externally, silently stopping compression); atomic settings writes; logged, never fights cleanup/stop; pinned by test
  • 295ecc0 README: 'How it plugs in' mechanism paragraph up top (PostToolUse hook + local daemon, write-time, local-only) so readers grasp how whittle attaches before the deep-dive sections
  • 8effa70 cli: whittle compress routes through the running daemon (matches the hook; prose works without WHITTLE_MODEL_URL), falls back to in-process when no daemon; -stats shows via=daemon|in-process
  • 59ca8ca setup: warn when binary dir is not on PATH (go-install users hit 'command not found' after setup); README notes the fix + that brew avoids it
  • 41b436c bench tier-3: add τ-bench eval (columnar structural gain +24-45% beyond compact JSON on multi-record results, all records reconstructed exactly); restructure README tier-3 as corroborating breadth (Salesforce 5k) + depth (τ-bench columnar) customer-service story
  • bfab73a README tier-3: lead with token/context reduction (the product metric, compounds per turn); keep the ~3% cost as an honest downstream nuance, not the headline
  • 59b8d77 bench tier-3: Salesforce customer-support real-world dataset (22% tool-output reduction, 0 measured info loss; 15,846/15,846 lossless + 0/120 blinded-judge panel; honest ~3% cost caveat) + README TL;DR
  • aa03fc8 license: strip leading blank line so GitHub detects Apache-2.0
  • 7d42b4d hygiene: SECURITY.md (fidelity/retrieval/fail-open as security-class), PR template with guarantee checklist
  • eeb829a README: stats payoff as hero gif, compress mechanism in See-it; strip em dashes from CLI output (228, incl. stats dashboard); demo tapes reproducible
  • 274d546 README: hero demo gif (vhs, reproducible from demo/whittle.tape) — noisy 144-line log compressed with both errors kept and noise marked
  • fa49699 README release-polish: badges, See-it before/after demo, architecture sketch, Verify-it-yourself, consolidated Known limitations, Contributing block
  • a5531b0 README audit fixes: 56KB figure, hook-cap qualifier on install claim, prose row labeled sidecar-dependent, systemd unit sample (CLI pointed at one that did not exist), version unified to 0.2.1 + goreleaser ldflags injection, MAE calibration harness checked in (bench/calibrate_tokens.py reproduces 8.2%)
  • 391d55b README: remove em dashes
  • 60e8b33 README: prose sidecar is installed/supervised by setup automatically — manual wiring applies only to library/serve-without-setup usage
  • e50af57 example: executable pin of the README library snippet (ExampleNew runs in go test/CI)
  • d249b6b README: Performance section cross-references the comparative latency (different corpora, different questions — absolutes vs side-by-side)
  • 9d45f94 bench: category-split reduction (conversation vs bulk arrays, derived from sidebyside.json) + in-process latency comparison; names-neutral prose
  • 69c143b bench: honest side-by-side on headroom's own generator data — headroom 41.8% (lossy defaults) vs whittle 36.5% (byte-exact lossless), 6/10 vs 4/10 rows; full methodology + reproduce script in SIDEBYSIDE.md
  • 5a27f19 bench: tier-2 corpus_headroom wired into runner (10 files frozen from headroom's own generators, seed 42, pinned commit, PROVENANCE) — same SHA-integrity + baseline gating
  • c3def32 benchmarks: tier-2 reflects the verified finding — headroom vendors no corpora; side-by-side will freeze their own generators' output (seeded, pinned, attributed)
  • d87343c README: three-tier Benchmarks section (synthetic headline / headroom side-by-side pending / real-world placeholder)
Assets 7
Loading