Feature Specification: FEAT-001 - HELIX Supervisory Control
Source identity (from
01-frame/features/FEAT-001-helix-supervisory-control.md):
ddx:
id: FEAT-001
depends_on:
- helix.prd
status: superseded
superseded_by: helix.prdSUPERSEDED — This feature defined
helix-runas HELIX’s supervisory autopilot and treated tracker-first execution as a core product capability. The current PRD (helix.prd) reverses that scope: HELIX is a portable methodology and artifact catalog; CLI, execution loop, and runtime orchestration are out of scope and belong to the adopting runtime (DDx, Genie, etc.). This document is retained for historical context only and must not govern new HELIX work.
Feature Specification: FEAT-001 - HELIX Supervisory Control
Feature ID: FEAT-001 Status: Specified Priority: P0 Owner: HELIX maintainers
Overview
HELIX supervisory control is the capability that lets helix-run act as an
autopilot controller for repository work. It keeps the workflow moving across
requirements, designs, tests, implementation, review, and metrics until human
judgment is actually needed.
Problem Statement
- Current situation: HELIX can execute bounded commands, but the control model is still easy to read literally unless the user restates the workflow contract.
- Pain points: Users must manually decide when requirements changes imply alignment, when specs should trigger polish, and when ready work should move into implementation.
- Desired outcome:
helix-runadvances the weakest ready layer autonomously, escalates only when human input is needed, and keeps the workflow coherent across interactive and autopilot use.
Requirements
Functional Requirements
helix-runmust act as HELIX’s supervisory autopilot.helix-runmust choose the least-power next bounded action that restores progress.- Requirement or specification changes must trigger alignment, planning, or polish when downstream work is affected.
- Ready governed work must trigger bounded build work.
- Users must be able to enter any layer of the loop interactively without breaking autonomous continuation later.
- Interactive refinement performed while
helix-runis active must be reflected at the next safe execution boundary. - HELIX must stop and ask for guidance when authority, approval, or product judgment is missing.
- The supervisory loop must keep doing useful work by staying focused on an active epic, retrying difficult work with bounded exponential backoff, and absorbing small adjacent changes when they are clearly part of the current slice.
- The supervisory loop must produce a blocker report and expose enough
structured state for
helix statusto explain what the run controller is doing and why it stopped.
Non-Functional Requirements
- Performance: Control-loop decisions should be quick enough to feel continuous in an interactive session.
- Security: Supervisor actions must remain bounded and authority-ordered.
- Scalability: The control model should work for small repos and larger tracker-backed queues.
- Reliability: Autopilot decisions must be deterministic enough to test with the wrapper harness and workflow docs.
User Stories
US-001: Steer HELIX autopilot [FEAT-001]
As a HELIX operator
I want helix-run to keep advancing work until human input is required
So that I do not have to manually decide every activity transition
Acceptance Criteria:
- Given a repository with vision and PRD, when HELIX can safely continue,
then
helix-runadvances the next bounded layer without asking for a activity name. - Given a user-requested functionality change, when it affects downstream artifacts, then HELIX routes to alignment, evolve, or design before build resumes.
US-002: Intervene directly anywhere [FEAT-001]
As a HELIX operator I want to invoke align, evolve, design, polish, build, review, status, or backfill directly So that I can focus on the highest-impact area that needs my judgment
Acceptance Criteria:
- Given a user invoking a specific layer directly, when they do so, then HELIX performs that action without breaking the supervisory model.
Edge Cases and Error Handling
- Missing or conflicting authority artifacts must stop autopilot instead of guessing.
- Open issues with changed specs must be polished before build resumes.
- A selected execution issue that changes materially before claim or close must be revalidated instead of being processed from stale state.
- If the tracker or workflow contract is unhealthy,
helix-runmust stop and surface guidance. - If an epic is in focus and one child becomes intractable after bounded retries, HELIX must block the epic explicitly and report the blocker rather than silently drifting to unrelated work.
Success Metrics
helix-runcan continue autonomous progress across multiple activities without explicit activity instructions.- Users spend less time manually orchestrating activity transitions.
- Trigger correctness is observable in deterministic tests and workflow docs.
Constraints and Assumptions
- HELIX remains bounded and tracker-first.
- The tracker is the steering wheel for day-to-day execution control, and the workflow contract plus skill descriptions must preserve that model.
- Cross-model verification is preferred when review or critique automation is configured.
Dependencies
helix.prdADR-001- Workflow contract docs
- Skill descriptions and CLI wrapper behavior
Out of Scope
- Generic autonomous coding beyond the HELIX workflow contract
- Replacing product judgment with guessed decisions
- Parallel issue tracking systems
Open Questions
- Which tracker metadata mutations and supersession relationships need to become first-class CLI operations for issue refinement workflows?