Source Pass Harvester

Source passes turn external inspiration into local evidence. They are a fetch/cache/classify/transfer loop, not just agent-written notes.

Workflow

1. Search local evidence memory first:

npm run docs:perf:search -- <topic terms>
npm run research:list

2. Define the question: what weakness or workload are we trying to learn about? If the target project still lacks target-owned contracts, corpus seeds, or benchmark fixtures, use target-evidence-designer first.

3. Add or update a repeatable fetcher and source config:

npx evidence-kit add-source-fetcher --name <topic>
npm run research:fetch -- <topic>

Fetchers should cache external repos under research/repos/<topic>/, raw data or metadata under benchmarks/data/<topic>/, and write research/repos/<topic>/manifest.json.

4. Inspect fetched material for test practice, fuzzer structure, benchmark fixture shape, failure handling, package boundaries, startup budgets, and corpus curation.

5. Transfer ideas into local artifacts:

   • tests,
   • fuzzer corpus cases,
   • benchmark fixtures,
   • package-boundary checks,
   • iteration/research notes.

6. Keep runtime dependencies out unless the benchmark proves they are worth it.

7. Record rejected ideas explicitly.

Decision Filter

Accept an idea only when it is:

• generic enough for the project contract,
• backed by a local test or benchmark,
• not hard-coded to the source project's domain names,
• not adding hidden state or dependencies without a clear contract.

Output

Create or update a research/*-sources.md note and an iterations/*-source-pass.md note. Include source names, URLs, commits or hashes, local cache paths, manifest path, commands, accepted transfers, rejected transfers, and artifact paths.