Skip to content

LemonCrow v0.8.2

Latest

Choose a tag to compare

@github-actions github-actions released this 02 Nov 05:48

LemonCrow 0.8.2

0.8.1 was prepared but never published; this release is the first to carry its changes, listed under
[0.8.1] below.

Added

  • The pull request summary leads with the verdict. The first lines say what to do (merge, fix first, or a person
    must look), how much was reviewed (reviewed N of M files on a partial review) and whether human review is
    required, in words. A clean review reads as a normal result, not an empty one. Confidence Score: N/5 moved to
    --json.
  • Proof of work and independence in the summary. The summary lists the CI check runs and statuses reported for
    the exact head commit, labelled "reported by CI, not run by LemonCrow", and flags a "tests pass" claim in the
    description that no check backs. An independence line names the reviewer's model family, the author's when it is
    known exactly, and says independent, same family or unknown.
  • LemonCrow / human review status. A second commit status, from fixed signals in the change and never the
    model: pending with what it touches (auth, permissions and secrets, data and migrations, payments, a public API, a
    path your policy marks critical, a matched rule), otherwise success. An approval of exactly the current head by
    an owner, member or collaborator turns it success; an optional pull_request_review workflow shows that
    approval without waiting for the next push.
  • A default per-pull-request budget on every model route. max-usd-per-pr defaults to $3.00 and is enforced on
    every route; a route that reports no cost is held to a call count instead. Every Action exit, including a key
    failure, a budget stop and a fork skip, posts a LemonCrow Review status that names the reason.
  • lc review proof. For each open P0/P1 finding it asks the model for one minimal failing test, runs it on the
    head and the base commits, and reports supported, contradicted, inconclusive or could_not_run. It also runs
    the pull request's own new tests on the base ("red-on-base"). Results post as a LemonCrow Proof check run with a
    reply on each finding's thread. Linux runners only. See Proof.
  • lc review qa run. Replays a saved browser scenario from .lemoncrow/review.yaml in headless Chromium
    against a trusted preview of one commit, with one outcome per run, and can post a LemonCrow QA check run. No
    model is involved. See Web QA.
  • lc review watch-deploy. Reads Sentry (issues first seen in the release) and Prometheus (one error-rate
    query, before against after) for a deployed SHA, links what it finds to the pull request's files, and can comment
    on that pull request. Read-only; silence is not_observed or no_new_signal, never "healthy". See
    Production watch.
  • Change Story. A step-by-step walkthrough of how the change flows, in the Review Map and Brief, animating only
    Snapshot-verified edges, and as an SVG in the pull request summary. Diagram edges are drawn by exactness: solid for
    Snapshot-verified or observed, dashed for the current index or inferred.
  • The investigation cycle in the Review Brief. Blind spots (what the review could not see, over a typed
    investigation context), a Behavioral X-Ray from any Change Story step, an Assumption Explorer, a What-if impact
    explorer, "Challenge this claim" (claim, source, checks, a reproduction candidate, then the observed result,
    run only through an existing execution path with your consent), and an Investigate tab that ties them together.
  • Author provenance from Entire and Agent Trace. Entire-Checkpoint / Entire-Attribution trailers and Agent
    Trace records are read as exact provenance of which agent wrote what.
  • GitLab merge request bot (early access). lemoncrow review --publish-bot in a GitLab merge request pipeline
    posts one summary note, one discussion per finding and a LemonCrow Review external status. The template installs
    LemonCrow from the public repository at LEMONCROW_REF. See GitLab CI.
  • Multi-repository workspaces. A Runtime Lite root that holds submodules, or a folder of repositories, serves
    each member repository from its own index. lc workspace status|branch|commit|push|pr carries one change across
    them (submodules first, one linked pull request per pushed repository), and lc workspace graph|impact lists the
    edges between members (submodule, package, import, proto, OpenAPI, HTTP route, event topic, each with evidence and
    a confidence) and the members a change reaches. The Brief of a local lc review shows that reach; hosted Review
    never receives it.
  • Benchmark harness with a DIY baseline. lemoncrow dev eval review benchmark --reviewer lemoncrow,diy
    reviews the golden-comment corpus with LemonCrow and with a plain "review this diff" prompt on the same model,
    repeats each run, and reports raw per-PR true and false positives, confidence intervals, rerun stability and
    cost per PR. See Review benchmark.
  • Measured against a DIY prompt. On the 50-PR set, at thorough, LemonCrow posted half as many comments as a
    plain "review this diff" prompt on the same model, and more of them were right. The DIY prompt found more
    of the answer key at a seventh of the cost. LemonCrow scored precision 48.5%, recall 52.1% and F 50.3% at a
    median $0.54 a pull request. The DIY prompt scored 34.7%, 72.9% and 47.0% at $0.07. These are 49 pull requests,
    three runs each. LemonCrow's findings also agreed more often across reruns (Jaccard 0.66 against 0.48).
  • A new benchmark row. The README's LemonCrow row is now the 2026-11-01 run on all 50 pull requests: precision
    47%, recall 51%, F-score 49%, a median $0.56 and 8 model calls a pull request. The 2026-10-26 run had 43%, 48%
    and 46% at $0.52. Both runs use the claude CLI on Sonnet and LemonCrow's judge, and the raw report is under
    evals/review/published/runs/.

Security

  • Model prompts are redacted of secrets; text in the pull request body, linked issues and code comments that tries
    to instruct the reviewer is flagged and never changes a verdict, posts a link or runs a tool; links in model
    output are defanged.
  • An agent-CLI reviewer reads a snapshot with secrets masked.

Fixed

  • Model-API routes are given the output schema, and an answer missing a required field fails instead of passing as
    empty. Ollama gets the schema as its format and a named cause when it returns no answer.
  • The GitLab template installed LemonCrow from PyPI, where it is not published; it now installs the public
    repository at a pinned release from its lockfile.
  • The architecture delta behind the change diagram, the Change Story and the Review Map tied JavaScript and
    TypeScript files together through shared local names (actual, body), imports of the same library name and
    unexported declarations. On a four-file pull request that read as 227 changed relationships across unrelated
    files. A reference now reaches only another file's exported, file-level declaration in the same language, and a
    name the file defines itself stays its own.

GitHub Action

Pin the review Action to this release. Replace the <sha> of the line in your workflow with this commit:

- uses: lemoncrow-lab/lemoncrow@cabdaab7d40bf107306069369f38dddc95a69b6e  # v0.8.2

The setup guide is docs/setup/github-action.md, "Pin the Action".

Installation

Install Review with the hosted service:

curl -fsSL https://lemoncrow.com/install.sh | bash

Scripted product choices:

# Review only (cloud)
curl -fsSL https://lemoncrow.com/install.sh | bash -s -- --product review

# Runtime only (cloud; coding-agent registration remains opt-in)
curl -fsSL https://lemoncrow.com/install.sh | bash -s -- --product runtime

# Review + Runtime (cloud)
curl -fsSL https://lemoncrow.com/install.sh | bash -s -- --product both

# Review + Runtime with local services
curl -fsSL https://lemoncrow.com/install.sh | bash -s -- --product both --local

After a cloud install:

# New terminals are configured automatically. In the current terminal:
source "$HOME/.lemoncrow/shell-env.sh"
lc auth login
lcr

# If Runtime was installed with --product runtime or --product both:
lc runtime status
lc runtime enable codex  # optional; or claude / opencode / auto

Assets

  • hosted.sh — hosted LemonCrow Review installer
  • lemoncrow-review-client.tar.gz — architecture-neutral zero-dependency Review client
  • lemoncrow-review-engine.tar.gz — Review-only local engine + shared product-aware frontend; no Runtime/MCP implementation
  • runtime.sh — LemonCrow Runtime client installer (agent integration remains opt-in)
  • local_review.sh — install and compose the isolated local Review engine + Review client
  • local_runtime.sh — install serverless Runtime Lite; no local API server, frontend, daemon, or listener
  • lemoncrow-runtime-client.tar.gz — architecture-neutral zero-dependency Runtime client carrying the Lite kernel
  • uninstall_product.sh — unified backend for lc uninstall review|runtime|all
  • uninstall_review.sh / uninstall_runtime.sh — independent low-level product uninstallers
  • install.sh — unified public installer for Review, Runtime, or both; cloud or local
  • sessions.sh — session management utility for managed/full local distributions
  • lemoncrow-distribution-*.tar.gz — internal full local-mode bundles for supported platforms