Step 2 - Review the draft
Goal: read what the agent proposed before you judge it. The draft lives under .alloy-output/survey/your-repo/draft/ - gitignored, uncommitted. Open it, or ask the agent to walk you through it file by file.
Each file has one job:
| File | What it decides |
|---|---|
source-entry.yaml | The heart of the draft: the scanners, the units definitions that group candidates into documentable units, and the coverage: block. Decides what gets documented and how it is grouped. |
rulepacks/<kind>.yaml + templates/<kind>.md | Only for new kinds: which files feed each kind's doc, and the doc's sections. |
manifest-categories.yaml | Where each kind's docs live in the corpus tree. |
units.yaml | Inventory tier only - repos where paths do not predict units, so units are listed explicitly. |
follow-ups.md | Read this first. The agent's open questions and judgment calls. A genuine ambiguity lands here rather than being guessed past. |
A unit definition reads roughly like:
yaml
units:
- kind: service # the unit type - fixes doc shape and rule pack
match: services/{name} # how raw candidates map to units
groupBy: nameRead the survey report too
The agent's report carries what the YAML cannot:
- Decompositions considered - the unit grain it chose, the alternative it rejected, and the coverage delta between them. The grain is the decision you most want to get right; the report is where you get to veto it.
- The coverage triage - every exclusion and its rationale, plus any UNKNOWN it could not resolve.
- The ontology triage - each kind's definition and the members it tested against it.
Don't fix anything yet. Step 3 is how to judge it; step 4 is where it changes.