Iterative Development

Iterative Development

Overview

Orchestrator for the iterative-development plugin. Drives the full autonomous lifecycle: extract requirements with proof obligations and behavior scenarios from human spec collateral, define a walking skeleton that passes its first journey scenario, then loop through audited sprints that continuously build a reusable behavior evidence corpus. Completion means the product has passing behavior evidence at the correct seam for every externally observable requirement — not just that stories are marked done. Every evaluative gate uses parallel adversarial review (PAR).

This is an alternative to superpowers:writing-plans → superpowers:subagent-driven-development for projects where the upfront-planning approach would lose the plot.

When to Use

• Spec is large, comprehensive, or ambiguous (10+ files, 100+ requirements)
• You need the product to be in a working, testable state at every iteration boundary
• You want an autonomous audited loop rather than a single upfront plan
• The writing-plans flow has lost the plot on this project before

Do NOT use for small, bounded projects — superpowers:writing-plans → superpowers:subagent-driven-development is simpler and more appropriate.

The Autonomous Loop

Bootstrap (first invocation)

1. Check docs/superpowers/iterations/ for existing state. If found, skip to Resume below.
2. Invoke extracting-requirements on the human-provided spec path.
   • Chunks the spec, classifies by taxonomy (journeys → E2E, domains → integration, etc.)
   • Dispatches parallel extraction subagents that produce stories with proof obligations AND behavior scenarios
   • Aggregates stories into per-epic files, scenarios into behavior-scenarios.md
   • Builds coverage ledger with both story AND scenario coverage
   • Produces docs/superpowers/iterations/requirements/, docs/superpowers/iterations/behavior-scenarios.md, docs/superpowers/iterations/behavior-corpus.md
3. Invoke scoping-the-simplest-core on the resulting backlog.
   • Defines the walking skeleton iteration (ITER-0000) + ordered follow-on iterations
   • Runs citation check + PAR scope review
   • Produces docs/superpowers/iterations/roadmap.md
   • Walking skeleton must close at least one journey scenario (not just compile)
   • Applies story splitting when stories have heterogeneous-dependency ACs

Main loop

while True:
    check_for_human_interrupt()

    if not roadmap has pending iterations:
        if last audit was clean:
            run final behavior-evidence audit (see below)
            if behavior audit clean:
                break  # done
            # else: audit found uncovered surfaces or weak evidence, new iterations added
        # else: audit found gaps, new iterations were added, continue

    run next iteration:
        - running-an-iteration (sentinel baseline → scope review → decompose code + evidence tasks → implementing-tasks → impacted + sentinel scenario runs → wrap up)
    
    audit:
        - auditing-progress (PAR paired auditors, three-tier: deep evidence + impacted behavior + sentinel corpus)
        - if gaps: append to backlog, revise roadmap, continue
        - if clean: mark last_audit_clean, continue

Final behavior-evidence audit

Before declaring the project complete, verify that the product has adequate behavior evidence — not just that all stories are marked done:

1. List every major user-facing surface from the original spec (settings panes, UI flows, CLI commands, journeys, etc.)
2. For each surface, verify that:
   • Corresponding stories exist AND are implemented
   • Corresponding scenarios exist AND have passing evidence at the correct seam
   • Journey scenarios that cross multiple surfaces are passing E2E
3. Check the behavior corpus index for completeness:
   • Every journey spec file has at least one JOURNEY-NNNN scenario
   • Every scenario has a non-TBD execution command
   • All sentinel scenarios pass
4. Flag any surface with:
   • No corresponding story (extraction under-scoped)
   • No corresponding scenario (evidence gap)
   • Evidence at a weaker seam than the requirement demands
   • Manual-residual scenarios that could be automated
5. If gaps found: create new stories/scenarios/iterations, continue the loop

The final question is: "Can the system point to passing behavior evidence for every externally observable requirement the spec describes?" Not: "Are the stories done?"

Resume (re-invocation with existing state)

All process state lives in artifact files:

• docs/superpowers/iterations/requirements/ (backlog with story status and proof obligations)
• docs/superpowers/iterations/behavior-scenarios.md (scenario cards with stable IDs)
• docs/superpowers/iterations/behavior-corpus.md (execution index)
• docs/superpowers/iterations/roadmap.md (iteration plan with status)
• docs/superpowers/iterations/iteration-log.md (completed iteration history)

On re-invocation: read roadmap.md, find the next pending iteration, and continue from there. There is no ephemeral in-memory state to recover. The command "continue iterative development with the existing plan" always works.

If the orchestrator crashed mid-iteration, the partially-completed iteration's git commits are preserved. On resume, the next un-started iteration picks up. If the in-progress iteration left the code in a broken state, treat it as a gap — the audit will catch it and add corrective work.

Human Interrupt Protocol

The loop runs without human intervention. The only way the human injects new information mid-run is by interrupting between iterations.

How it works:

• The human types the update into the chat session ("we dropped feature X", "the spec changed, re-read specs/foo.md", "add a new requirement for Y")
• The orchestrator notices the interrupt at the next iteration boundary — after the current iteration's audit completes, before the next iteration starts
• At the boundary: invoke extracting-requirements in incremental mode on the changed spec files, merge new/revised story cards into the backlog, revise the roadmap if changes invalidate downstream iterations, then resume

Guarantees:

• Changes during mid-iteration do NOT disrupt in-progress work. The current iteration completes first.
• The orchestrator never silently drops an interrupt. If ambiguous, ask for clarification before resuming.
• Existing story IDs are preserved across re-extraction. Removed stories flip to deferred, not deleted.

What does NOT trigger interrupt processing:

• The orchestrator does not poll the filesystem for spec changes
• The orchestrator does not ask "anything to change?" between iterations
• Human presence is not required at iteration boundaries

Progress Reporting

The autonomous loop may run for hours. Two progress mechanisms ensure visibility without requiring interruption:

1. Progress file: Write docs/superpowers/iterations/progress.md at each phase transition:

# Progress

**Phase:** implementing ITER-0003
**Task:** 4/7 (CleanupPipeline integration)
**Iterations:** 3/18 done, 15 pending
**Sentinel corpus:** 10/10 passing
**Last event:** 2026-04-11T14:23:00Z — Task 3 committed

Update this file at: iteration start, each task completion, iteration wrap-up, audit start/end. Overwrite (not append) — it's a snapshot of current state, not a log.

2. Git log: Every task produces a commit. The commit history is a detailed progress trail. A human can check git log --oneline for fine-grained status without interrupting the loop.

Skill Precedence

When running autonomously, this orchestrator takes precedence over interactive-gate skills (e.g., brainstorming which requires design approval before implementation). The iterative-development process has its own design gates (scope review, PAR) that replace interactive approval. Do not block on skills that assume a human is present to approve each step.

Escalation Policy

Catastrophe-only. The loop is autonomous. Human escalation is reserved for total failure — the plugin cannot make any forward progress at all.

These do NOT trigger escalation:

• A reviewer finding issues (those become fix work)
• An audit finding gaps (those become new iterations)
• An implementer reporting BLOCKED on a task (try: more context, more capable model, smaller task)
• Ambiguity in the spec (make a reasonable judgment call, document it in the iteration log)
• Difficulty or slow progress (keep going)

The orchestrator does NOT prompt "should I continue?" between iterations.

Skill Invocation Reference

Artifact Location

All plugin artifacts live in docs/superpowers/iterations/. Never modify the human's spec collateral.

Quality Gates

Every evaluative gate uses parallel adversarial review (PAR):

• Pre-iteration scope review (citation + scope-creep + boxing-in + scenario coverage + story splitting)
• Pre-iteration sentinel corpus baseline
• Per-task spec-compliance review with evidence quality check
• Per-task code-quality review with boxing-in + corpus contribution check
• Post-iteration impacted + sentinel scenario runs
• Per-sprint audit (deep evidence + impacted behavior + sentinel corpus)

See skills/shared/parallel-adversarial-review.md for PAR methodology.