open-ended

Use this for tasks with no fixed end, such as audits, gap hunts, "find all X",
or "improve X". One session anchors on its first ideas. Fresh agents miss
different things, so repeat rounds until they run dry.

Arguments: the task, --rounds N (most rounds, default 6), and --angles N
(agents per round, default 3).

0. Set up

1. Read sub-agents/SKILL.md and
   roles.md.
2. Define done. Write the scope (paths, features, configs, sources) and one end
   check that says what "covered" means. If the task is too vague to scope, ask
   one question.
3. Pick angles: distinct lenses that fit the task, one line each. For a code
   audit: correctness, error and edge paths, callers and sibling sites, tests.
4. If a ledger for this task exists, resume from it. Otherwise create
   <scratch>/open-ended/<task-slug>/ledger.md with the task, scope, end check,
   angles, and empty Found, Rejected, and Checked sections.

1. Run a round

1. Write the shared prompt once to <ledger-dir>/brief.md: task, scope, end
   check, ledger path, repository constraints, and the instructions below.
   Spawn one agent per angle, in one message, with "read /brief.md; your
   angle: X", using the absolute path. Use explorer for read-only search and worker when items need
   judgement or research. Tell each agent to:
   • read the ledger first;
   • write its result to <ledger-dir>/round-<k>/<angle>.md and return the path and item
     count;
   • return only items absent from Found and Rejected, each with source
     (file:line or URL), evidence, and the reason it is new;
   • re-raise a rejected item only with new evidence;
   • list the areas it checked;
   • for Claude session evidence, use agent-sessions (cost, errors,
     denials, friction) before writing a new transcript parser;
   • not edit, commit, or spawn agents.
2. From round 2, replace each angle that returned nothing with a new angle aimed
   at scope not yet in Checked.
3. Spot-check high-risk items yourself against the cited code or data. Give the
   rest to a fresh verifier agent that writes a verdict file. Move duplicates and
   unproven items to Rejected with a one-line reason.
4. Append verified items to Found with the round number. Append the areas to
   Checked.
5. When the task includes changes and they are authorized, apply the verified
   items now, so the next round sees the current state.
6. Record round k: X new, Y rejected in the ledger. Tell the user only if
   the round changes the plan.

2. Stop

Stop when any of these is true:

• Two rounds in a row add nothing to Found.
• The end check passes and the next round adds nothing.
• The round count reaches --rounds.

3. Report

Lead with the result: ran dry or stopped at the cap. At the cap, say
whether the last round still found items. Then give the Found items as a
list under the main rules' item limit, highest risk first. Add the
Rejected count, scope not yet checked, and the ledger path.