rounds

iterate with spawned agents until results stabilize.

prerequisite skills: spawn, coordinate, report

when NOT to use

before running multi-round review, ask:

1. •is there a single source of truth? if verifying against a spec or code, one agent reading carefully beats rounds of opinion-generation.
2. •do i have explicit exit criteria? "2+ clean rounds" is built in, but you still need to define what "clean" means.
3. •is the scope bounded? rounds on unbounded tasks produces sprawl. define what you're reviewing before starting.

rounds is for verification where first-clean can't be trusted. don't use it to generate opinions for reconciliation.

core rule: don't trust first clean

round 5 finding issues after round 4 was clean is common. run 2-3 verification rounds minimum after issues stop appearing.

round lifecycle

round N:
  1. spawn agents for task (parallel if independent)
  2. wait for reports
  3. if issues → spawn fix agents → wait → proceed to N+1
  4. if clean → proceed to N+1 anyway (verify the clean)
  5. repeat until 2+ consecutive clean rounds

injecting skills into spawned agents

tell spawned agents which foundation skills to load. example for code review:

load the review skill before evaluating. load the write skill for report quality.

adapt based on task:

• •code review → inject review, write
• •investigation → inject review, dig
• •docs review → inject review, write, document

ambiguous decisions

when agents find issues requiring product decisions ("remove feature X or implement it?"), pause and ask user. don't make product decisions unilaterally.

composition with spar

rounds can orchestrate parallel spar debates. see spar skill for the full composition protocol, interface contract, and termination conditions.

summary format

after completion:

list all changes made across fix phases.