Example from HELIX’s own docs. This generated page comes from docs/helix/. Use it to see the method in practice; start with the artifact-type catalog for reusable templates. Historical plans and reports may describe retired architecture.

Source identity (from 01-frame/features/FEAT-011-slider-autonomy.md):

ddx:
  id: FEAT-011
  status: draft
  depends_on:
    - helix.prd
    - FEAT-006
  review:
    self_hash: 0d31847adee3e2048695700a430e37b6c8c65c93c2ce94dc7d2cc11aa5f4f9eb
    deps:
      FEAT-006: 7517c7bf2db366dcdbae3ada995a5c0148955b332cf34d567180dff971d79489
      helix.prd: 2b22383538b33c6ecee57f43d85128dfef7d56254766b757aa36439e35f2bfc9
    reviewed_at: "2026-05-25T20:24:24Z"

Feature Specification: FEAT-011 — Slider Autonomy

Feature ID: FEAT-011 Status: Draft Priority: P1 Owner: HELIX maintainers

Overview

HELIX expresses an autonomy policy; a capable runtime supplies the agency. This feature defines that policy as a three-position spectrum — low, medium, high — that controls how often a HELIX workflow pauses for human confirmation without changing which activities run. The vocabulary and the per-level behavior already live in workflows/actions/input.md; this spec makes them a governed feature, extends them with concern-inference at high autonomy, fixes the resolution precedence, and records the two invariants the spectrum must never violate.

HELIX still ships no execution engine (PRD Non-Goals). Autonomy is a policy the methodology authors and a runtime honors. This spec restores the dangling references to FEAT-011 / TD-011 that survived the scope collapse (workflows/actions/input.md, CONTRACT-001, CONTRACT-002) and re-scopes the feature to the content-only methodology: there is no helix run, execute-loop, or slider-config file — the autonomy signal is carried in runtime-neutral artifacts (per-invocation argument, artifact frontmatter, or project policy), never in CLAUDE.md or a runtime-specific config.

Problem Statement

• Current situation: input.md defines low/medium/high semantics and cites “FEAT-011 / TD-011”, but those governing artifacts were removed in the collapse (823aa1ac), leaving the citation dangling. CONTRACT-002 still records a FEAT-011 review-dependency hash that resolves to nothing. Nothing states how the level is resolved, what high autonomy may infer, or what it may never skip.
• Pain points:
  1. Dangling authority — the autonomy vocabulary is used across actions but governed by no on-disk feature, so the behavior cannot be audited or ratcheted.
  2. Under-defined high autonomy — “create without prompts” does not say whether a high-autonomy run may select concerns when none are declared, which is exactly the gap that made the concerns library inert ([[FEAT-006]]).
  3. Ambiguous precedence — when a per-invocation level, an artifact frontmatter value, and a runtime default all exist, nothing says which wins.
  4. Risk of loop collapse — “more autonomy” is easily misread as “skip activities”, which would flatten the seven-activity loop the PRD forbids.
• Desired outcome: a governed autonomy spectrum with a defined resolution order, an explicit hard-stop invariant, an explicit never-collapse-the-loop invariant, and concern-inference defined for high autonomy.

Autonomy Levels (canonical semantics)

These reuse the input.md semantics verbatim in intent. They are the single source of truth; input.md and other actions implement them.

Requirements

FR-1: Three-position spectrum, default medium

Autonomy is low, medium, or high. The default when no level is supplied is medium. The vocabulary is fixed — actions and the skill must not introduce a parallel vocabulary (no ask-first/guided/yolo synonyms in normative text).

FR-2: Resolution precedence

A workflow resolves the active autonomy level in this order, first match wins:

1. Per-invocation override — the level passed with the request (e.g. an --autonomy high argument or an explicit instruction in the prompt).
2. Artifact frontmatter / project policy — an autonomy: value declared in the governing artifact’s frontmatter or in the project’s policy artifact (runtime-neutral markdown the project commits).
3. Runtime default — medium.

CLAUDE.md (or any runtime-specific instruction file) is not part of this chain. The autonomy signal must be expressible in runtime-neutral artifacts so the same project behaves identically across runtimes.

FR-3: Concern-inference at high autonomy

When autonomy is high and docs/helix/01-frame/concerns.md does not exist, the workflow infers the concern selection from the product’s nature (e.g. a web app implies a frontend + accessibility concern; a CLI implies a tech-stack concern) and writes concerns.md, recording each inferred concern as an assumption. At low/medium, concern selection stays interactive per [[FEAT-006]]. Inference never silently overrides an existing concerns.md.

Exclusive slots are filled deterministically. For each needed exclusive slot (the functional positions defined in workflows/concerns/slots.yml — frontend-framework, language-runtime, etc.; see [[FEAT-006]]), high-autonomy inference fills the slot by resolution order, first match wins:

1. Operator override — the slot’s value in docs/helix/01-frame/concerns.local.yml (read before concerns.md exists), if present.
2. Shipped default — the slot’s value in slots.yml defaults:.
3. Recorded assumption — if neither supplies a filler, record an assumption to revisit; never make a silent pick.

The chosen filler and its source (“concerns.local.yml override”, “slots.yml default”, or “assumption”) are recorded in concerns.md. A web app must fill the frontend-framework slot — leaving it empty (raw-served HTML with no supported framework) is the failure this requirement exists to prevent; the shipped default react-nextjs takes effect with no operator configuration. This slot-filling only fills absent selections; it never overrides an existing concerns.md.

FR-4: Hard-stop invariant (all levels)

Autonomy raises or lowers the pause threshold, never the stop floor. Regardless of level, a workflow must stop and surface to a human when:

• two higher-or-equal-authority artifacts truly contradict (a physics-level conflict the authority hierarchy cannot reconcile);
• the next action is destructive or irreversible (data loss, history rewrite, external publish) and was not explicitly authorized;
• a decision is required that only a human can make (product direction, an external contract change).

High autonomy proceeds through resolvable conflicts (recording assumptions); it never proceeds through a hard stop. A runtime preserve/abort outcome is an execution result handed back to HELIX, not a license to bypass a hard stop.

FR-5: Never-collapse-the-loop invariant (all levels)

Autonomy changes checkpoint density only. It must never collapse the seven-activity loop (discover → frame → design → test → build → deploy → iterate) into a single generic prompt, and must never skip an activity that the work requires. A high-autonomy run executes the same activities a low-autonomy run would; it simply pauses less often between and within them. This is the feature-level binding of PRD Non-Goal “HELIX will not flatten the seven-activity loop into one generic prompt.”

Acceptance Criteria

Edge Cases and Error Handling

• Conflicting precedence sources — if an artifact frontmatter value and a project policy value disagree, the nearer governing artifact’s frontmatter wins; record the divergence as an alignment finding.
• High autonomy with an existing concerns.md — never overwrite it; inference only fills an absent selection.
• Unknown level token — treat any value other than low/medium/high as unset and fall through to the next precedence source.

Success Metrics

• Zero dangling FEAT-011 / TD-011 references in the repo after restore.
• Autonomy behavior is auditable: alignment review can classify an autonomy precedence or loop-collapse violation as a finding.
• High-autonomy runs select concerns without a human prompt while never bypassing a hard stop (measured by the FEAT-011-AC4 / AC5 scenarios).

Constraints and Assumptions

• HELIX ships the policy, not the agency. The runtime executes; HELIX states how often it should pause and what it may never skip.
• The autonomy signal lives only in runtime-neutral artifacts. No CLAUDE.md dependency, no slider-config file, no CLI flag owned by HELIX.

Dependencies

• helix.prd — autonomy first-class goal (Non-Goals) and the seven-activity-loop guarantee.
• [[FEAT-006]] — concern selection that high autonomy infers.
• ADR-003 — the autonomy-spectrum decision record (ADR-003 depends on this feature).

Out of Scope

• A runtime autonomy UI or config schema (runtime-owned).
• Per-step autonomy beyond the three positions (no numeric slider).
• Replacing human judgment on hard-stop decisions.

References

• Autonomy Spectrum ADR
• TD-011: Slider Autonomy Implementation
• CONTRACT-002: HELIX Execution-Document Conventions
• Input action — implements these semantics