Decant Orchestrate

Decant Orchestrate

Coordinate the other Decant skills without turning every request into a large process.

Preflight

1. Restate the requested outcome, constraints, and forbidden actions.
2. Inspect the repository and current git state before proposing edits.
3. If the CLI is available, run decant doctor and preview routing with decant route "" --json.
4. Treat decant run "" --dry-run as a preview. Run a mutating pipeline only when the user authorized implementation.

This skill helps choose a workflow. It does not change the model of the host agent's current task, whether that host is Codex or Claude Code, and does not make unsupported providers available.

Hard-deadline progress gate

Use this gate whenever the user gives an explicit deadline and the work is low risk. For a budget of 15 minutes or less, treat it as a fast lane and recommend medium-or-lower host effort before starting. This skill cannot change the host model or effort and cannot enforce a wall-clock timer.

1. Deliver a runnable primary implementation slice by 30% of the budget, capped at five minutes.
2. Run a narrow smoke check, then core tests.
3. Expand behavior and documentation only with the remaining time.
4. Stop expansion when the final 20% begins; use that time to verify and report honestly.

Tests, plans, or documentation alone do not satisfy the first progress gate. Security-sensitive, deployment, destructive, irreversible, or otherwise high-risk work never qualifies for this shortcut; use the normal risk-proportionate workflow without weakening its safety gates.

Choose the smallest path

• Read-only evidence gathering: use decant-research only.
• Ambiguous or high-impact change: use decant-spec, then decant-build.
• Approved, well-specified edit: use decant-build directly.
• Failure or regression: use decant-debug; edit only when a fix is authorized.
• Review request: use decant-review without editing.
• Publication or handoff: use decant-release after implementation checks pass.
• Skill creation or trigger tuning: use decant-skill-lab.

Skip stages whose output cannot change the decision. Add review depth as risk, blast radius, irreversibility, or verification difficulty rises.

Handoff contract

For every active stage, record:

• goal and non-goals;
• inputs and evidence locations;
• allowed mutations;
• result and unresolved risks;
• deterministic checks and semantic review;
• next owner or next action.

Separate observed facts, inferences, and proposals. Never call a stage complete from intention alone.

Safety boundaries

• Preserve unrelated work in a dirty worktree.
• Do not broaden authorization from diagnosis to repair, or from local changes to publication.
• Do not hide destructive actions behind a recipe, subagent, hook, or retry loop.
• Stop repeated repair attempts after three materially similar failures and reassess the hypothesis.
• Keep databases, daemons, worktrees, swarms, and GUIs optional rather than core requirements.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: minwoo19930301
• Source: minwoo19930301/decant
• License: MIT

Install and usage instructions live in the source repository linked above.