survey

Survey

1. Scope

If $ARGUMENTS is non-empty, that string is the scope. Skip hotspot inference. Tell the user the chosen scope in one line, then go to Context.

Otherwise run, from the repository root:

bash "${CLAUDE_PLUGIN_ROOT}/scripts/hotspots.sh"

Read the scope: line.

• scope: hotspot — the first three ranked paths are the scope (hotspots.sh prints up to five).
• scope: wide — the whole tree is the scope; keep the ranked list as hints. If history is thin (window below 10 or an empty log), say in the one-line scope message that the whole tree is being surveyed because history is thin.

If hotspots.sh exits 1 (not a git repository) and $ARGUMENTS is empty, stop. Ask in plain prose for a direction. That is an input question, not permission.

Tell the user the chosen scope in one line before anything else runs: the direction, the hotspot paths, or "whole tree".

2. Context

Before dispatch, read CONTEXT.md if it exists and any ADRs under docs/decisions/ or docs/adr/, the directories /adr:adr looks in first. They constrain what a candidate may contradict.

3. Explore

Print the one-line scope and how many explorers will run. Dispatch them in one message, at most three.

For each scope path (or one explorer for the whole tree), call:

Agent(subagent_type: "deepen:explorer", prompt: "Report path: /tmp/viz/<project>/deepen-explorer-<timestamp>-<n>.md\nScope: <path or whole tree>\n")

<project> is $(basename "$PWD"). Timestamp the files so a stale report from an earlier run cannot be mistaken for this one. Explorers run in parallel. Print nothing else until they return.

If a report file is missing after dispatch, say which scope produced nothing and continue with the others.

4. Read

Do not Read an explorer report whole. For each report path you dispatched in step 3 that exists, take only:

sed -n '/^## Summary/,/^## Notes/p' /tmp/viz/<project>/deepen-explorer-<timestamp>-<n>.md

That range stops at ## Notes (inclusive of the heading, exclusive of the trail). Never load ## Notes.