Quintet ICM Staged Folder Pipelines

Orchestrates multi-stage workflows across Quintet CLI agents following the Interpretable Context Methodology (ICM) (Van Clief & McDermott, arXiv:2603.16021). Numbered folders carry sequencing, L2 CONTEXT.md contracts carry context scoping, intermediate files in output/ carry state, and every boundary serves as an editable human check surface.

Core Invariants Enforced

1. One Folder, One Stage: Each stage does one job (01_debate, 02_spec, 03_implement, 04_review).
2. Layered Context Contracts: Every working stage and worker has an explicit CONTEXT.md defining Inputs, Process, Outputs, and Human Check.
3. Every Output is an Edit Surface: Intermediate files in output/ are plain markdown/JSON that humans can edit before downstream stages run.
4. Filesystem as State Machine: Status is derived by scanning artifacts on disk rather than relying on unstructured logs.
5. Token Discipline: Stage and worker contracts stay strictly within the 2,000–8,000 token budget.

Workflow

1. Initialize Pipeline

Choose a template (debate-build or review-fix) and initialize:

QBIN="${CLAUDE_PLUGIN_ROOT:+$CLAUDE_PLUGIN_ROOT/bin/quintet}"
[ -n "$QBIN" ] && [ -x "$QBIN" ] || QBIN="$HOME/.gemini/config/plugins/quintet/bin/quintet"
[ -x "$QBIN" ] || QBIN="$HOME/.gemini/extensions/quintet/bin/quintet"
[ -x "$QBIN" ] || QBIN="$(command -v quintet 2>/dev/null || echo quintet)"

"$QBIN" pipeline init my-feature --template debate-build --goal "Design and implement user rate limiting"

2. Inspect Pipeline Status

Check current stage and completed artifacts:

"$QBIN" pipeline status my-feature

3. Advance with Human Check Gates

Run the active stage:

"$QBIN" pipeline advance my-feature

The pipeline pauses at human check gates so the user can inspect or edit output/consensus.md or output/findings.json in place before downstream stages run. Use --auto to advance automatically once verified.

4. Audit with the Cold-Agent Walk Test

Validate contract integrity and state derivability:

"$QBIN" walk my-feature

A passing Walk Test proves that any newly spawned agent or human can understand the run's goals, inputs, boundaries, and outputs without reading raw terminal logs.