Entroly 1.0.77
Entroly 1.0.77 improves context economics, observability, privacy-safe product
health, parser-backed code intelligence, and local-model discovery.
User-visible changes
- The proxy protects provider cache economics with a stable-prefix/live-zone
boundary and requires provider-observed paired baselines before presenting a
net savings claim.
- The dashboard reports auditable context-health signals for confusion, rot,
drift, cache reuse, tokens, and provider-observed cost without turning local
estimates into billed savings.
- Optional product-health telemetry remains opt-in, collects operational
performance rather than user content, and supports privacy-safe uninstall
feedback.
- Parser-backed symbol extraction adds exact representative coverage for Dart,
Erlang, Go, Haskell, Kotlin, Protobuf, R, Solidity, SQL, Svelte, Vue, and Zig.
A sealed benchmark audits all 41 mapped languages and passes 32 of 37
declaration-bearing samples (86.5%), compared with 20 of 37 on the pinned
pre-change baseline. C3, F#, Groovy, Nim, and OCaml remain explicit gaps.
- NVIDIA Nemotron 3.5 Lightning is documented through Entroly’s existing
loopback-only Ollama discovery and OpenAI-compatible proxy path. The standard
Ollama 1M-context listing and Apple-silicon MLX 256K listing remain distinct;
Entroly inspects installed model metadata instead of assuming one limit.
- The verified model registry now includes the public GA IDs for Gemini 3.6
Flash and Gemini 3.5 Flash-Lite, with official context, output, tools,
reasoning, and list-price metadata. Restricted or private-preview model
announcements remain visibly gated instead of receiving invented limits.
- QCCR now reserves dependency evidence by distinct source before repeated
calls into the same module. On the preregistered 20-task graph-lane release
gate at a 2,000-token budget, indirect dependency recall improved from 19/20
to 20/20 with zero execution errors.
Install or upgrade
pip install -U entroly==1.0.77
npm install -g entroly@1.0.77
Container users can pull ghcr.io/juyterman1000/entroly:1.0.77 after the
coordinated publication workflow completes.
Compatibility and migration
No data migration is required. Existing CLI, SDK, MCP, proxy, npm/WASM, and
OpenClaw entry points remain supported. Nemotron support uses Ollama’s local
OpenAI-compatible route; it is not a bespoke NVIDIA adapter.
Privacy and network boundaries
Entroly sends no analytics by default. Enabling product-health reporting does
not authorize collection of prompts, files, conversations, or retrieved
content. Local Ollama inference can keep model prompts on the device, but agent
tools and separately configured providers retain their own network boundaries.
Verification
entroly verify-claims
python benchmarks/language_symbol_coverage.py \
--baseline-ref 2eeecb8733103fe7234133f48b105f271662b219 \
--output benchmarks/results/language_symbol_coverage.json --check
The language result is one valid representative sample per mapped language. It
does not prove complete grammar coverage, call resolution, data flow, model
answer quality, or token savings.
Rollback and uninstall
Pin the previous packages with pip install entroly==1.0.76 and
npm install -g entroly@1.0.76, or remove them with pip uninstall entroly
and npm uninstall -g entroly. Container users can return to the immutable
1.0.76 image tag.