Search
Posts
Manifest System + Mutation Testing: Two Ways to Find Out What Actually Works
Epic 31-B lands the publish side of the bot-manifest protocol in claude-code-slack-channel. Epic 31 closes with an audit. Mutation tests and a cyclomatic gate make the test pyramid stop lying.
ReadFour Releases in One Day: How the claude-code-slack-channel Security Sprint Actually Shipped
Epic 29-A, 30-A, 30-B, 32-B land in a single calendar day across v0.5.0 → v0.5.1 → v0.6.0 → v0.7.0 — a supervisor, a hash-chained audit journal, and a policy engine that never sees manifests.
ReadLLM-as-Reducer and the Case for Killing the AI Label
Two AI product lessons from the Braves dashboard post-game expansion: use the LLM as a reducer over noisy community signal, and pull the AI label off the UI.
ReadFour Primitives, Three Reviews: How a Contributor PR Reshaped a Roadmap
claude-code-slack-channel v0.4.0 shipped a contributor PR that forced a scoping conversation. Four roadmap primitives emerged. Three parallel reviews — architect, sitrep, devils-advocate — converged on a narrower v0.5.0 than the original issues proposed. This is a case study in how collaborative review reshapes plans.
ReadPre-Warmed Narratives, Stat Digests, and a Community PR
Pre-warming AI narratives so viewers never see spinners, feeding structured stat digests into the prompt so narratives quote real numbers, and a contributor shipping three plugins to the claude-code-plugins marketplace.
ReadAI Code Review Blind Test: Where 5 Bots Shine
Five AI bots reviewed 74 PRs blind. No Slack, no roadmap. Matched maintainer decisions 63.5%. The breakdown is the finding.
ReadThe 35x FLOPs Error That Peer Review Predicted
A 35x FLOPs correction in pre-filing patent artifacts validated the reviewers' unchecked-derivation warning. The day's other shipments show what systematizing against named failure classes looks like.
ReadRepo-Resolver: Typed Errors and Monorepo Detection
A shared repo-resolver package shipped into claude-runtime — ADR to integration in seven PRs, with typed error classes and transparent fallback that avoided a flag day.
ReadTwenty-One Documents and a Weak Reject: Building a Research Corpus for a Novel Search Architecture
4000+ lines across 21 documents — invention disclosure, 6-paper research series, competitive landscape, probability assessment, and toolchain evaluation for a search architecture that eliminates precomputed embeddings. A simulated peer review returned Weak Reject. That was the most useful document of all.
ReadGroq on Cloud Run, a Dep Bump, and a Star Refresh
Three quick fixes across three repos — a missing env var that broke production inference, a dependency bump, and a GitHub star count update.
Read