/autoiter

Collie Autoiter

Run an iterative test → observe → fix → re-run loop in Collie-style.

Completion Promise

This command uses ralph-loop. Completion signal: <promise>Collie: AUTOITER DONE</promise>

The completion signal can only be output when the skill's state machine reaches a terminal state (max iterations exhausted, goal achieved, or budget exceeded).

Absolutely no false completion reporting allowed (ralph-loop note: ONLY when statement is TRUE - do not lie to exit!)

Intelligent Exit Policy

stop-steps-counter.js hook (fires on every Stop event) detects:

• Same error appears consecutively 3 times → escalate WARN "loop_on_same_error"
• 5 consecutive steps with no file changes → escalate WARN "no_progress"

Skill Stage 6 stop conditions (evaluated after each iteration) detect:

• --max-iterations cap reached (SC1)
• Quality threshold met or convergence reached (SC2/SC3)
• Token budget exhausted (SC4)

Arguments

• <task> — required; the task description / goal for the iterative loop
• --max-iterations N — optional; maximum number of loop iterations, default 5
• --budget-tokens M — optional; token budget cap, default unlimited
• --mode interactive|queued — optional; default interactive

Task Prompt

When starting, inject this as the working prompt (substitute $ARGUMENTS with the actual arguments):

Your task: $ARGUMENTS

Invoke Skill('collie:autoiter') and pass all parsed arguments. The SKILL's §3.5 state machine will automatically call _state.projectId() to determine the current project ID, then check ~/.collie/autoiter/{project-id}/current-run to decide between fresh-start and resume. No additional logic is needed in this command file.

When the skill returns a terminal status, output: <promise>Collie: AUTOITER DONE</promise>