The problem, stated plainly
M1 through M6 are closed. The product works, has measured evidence, and installs in one command. The star count is the honest measure of how many people know it exists.
M7 owns external proof and had zero issues. This one names the work.
What is actually blocking
Not evidence. This repository already measures retrieval (81.7% vs 42.0% at the shipped budget), latency (496 ms at 100k commits), hook cost (102 ms p50), and now behaviour (M5, in flight). More measurement does not move this number.
What is missing is that no one has ever been told.
The work
1. One artefact a stranger can share. The self-audit page (#449) is the first candidate: a tool that catches its own false claims is a story, and stories travel where benchmark tables do not.
2. A README a cold reader survives — #450.
3. Somewhere to be found. The obvious surfaces for a Git-native agent tool: awesome-claude-code and equivalents, the Claude Code plugin directory, HN Show, the agent-tooling subreddits. Each needs a different first sentence and none of them wants a paper.
4. A demo that runs in the reader's head in five seconds. The current one is a correct SVG of console output. It shows the mechanism; it does not show the moment of "oh — my agent does this to me every week."
What this issue is not
A request to soften any claim, drop any caveat, or publish a number that is not measured. The evidence discipline is the product's differentiator and stays exactly as it is. This is about the sentence before the evidence, not the evidence.
Sequencing note
CDEB v1 — the confirmatory benchmark — is downstream of this, not upstream. Its corpus requires five external repositories that have used CommitLore during ordinary work, which requires users, which requires this. A benchmark cannot produce its own adoption.
The problem, stated plainly
M1 through M6 are closed. The product works, has measured evidence, and installs in one command. The star count is the honest measure of how many people know it exists.
M7 owns external proof and had zero issues. This one names the work.
What is actually blocking
Not evidence. This repository already measures retrieval (81.7% vs 42.0% at the shipped budget), latency (496 ms at 100k commits), hook cost (102 ms p50), and now behaviour (M5, in flight). More measurement does not move this number.
What is missing is that no one has ever been told.
The work
1. One artefact a stranger can share. The self-audit page (#449) is the first candidate: a tool that catches its own false claims is a story, and stories travel where benchmark tables do not.
2. A README a cold reader survives — #450.
3. Somewhere to be found. The obvious surfaces for a Git-native agent tool:
awesome-claude-codeand equivalents, the Claude Code plugin directory, HN Show, the agent-tooling subreddits. Each needs a different first sentence and none of them wants a paper.4. A demo that runs in the reader's head in five seconds. The current one is a correct SVG of console output. It shows the mechanism; it does not show the moment of "oh — my agent does this to me every week."
What this issue is not
A request to soften any claim, drop any caveat, or publish a number that is not measured. The evidence discipline is the product's differentiator and stays exactly as it is. This is about the sentence before the evidence, not the evidence.
Sequencing note
CDEB v1 — the confirmatory benchmark — is downstream of this, not upstream. Its corpus requires five external repositories that have used CommitLore during ordinary work, which requires users, which requires this. A benchmark cannot produce its own adoption.