sdlc-guided

sdlc-guided

The doorway into the human-in-the-loop agentic-sdlc flow. On hosts like Codex that have skills but no commands, this stands in for the retired sdlc:start command.

Stay a doorway

A doorway that starts doing the work is no longer a doorway, and a run whose artifacts were half-written before Phase 0 cannot be reasoned about. So this skill reads intent and delegates; it does not touch the repository.

Specifically, nothing here may:

• Make docs/superpowers/runs/<run-id>/, or any other run or task directory.
• Write meta.json, requirements.md, complexity.json, design.md, plan.md, events.jsonl, decisions.jsonl, work-item.md, or any ledger.
• Do a phase's work early — no intake, no sizing, no brainstorming, no planning, no branch guard, no building, no QA, no status repair.
• Improvise sdlc-engine because the host would not invoke it.

That last one is the tempting failure. If sdlc-engine cannot be reached, say startup is blocked and stop there. A hand-rolled approximation looks helpful and leaves behind artifacts nothing downstream can trust. Ownership of every artifact starts inside sdlc-engine Phase 0, and not one step sooner.

When the project owns implementation

Read .agentic/agentic-sdlc/config.json before delegating. If it names project orchestrators, pass their routing contract through to sdlc-engine: this package still owns intake, planning, QA and the delivery handoff, while implementation is delegated once to the orchestrator the project selected. Fleet workers are never called directly — going around the project's own orchestrator is how two pipelines end up believing they own the same change.

Inputs

• raw_input — whatever the user is asking for: prose, a tracker reference, or a path to a story file
• mode_flag — --greenfield when starting from nothing, otherwise unset
• escalate_on — risk flags that force a human decision; defaults to ["security", "breaking-change"]

Usage Examples