Audit Feature Gates

Audit Feature Gates Skill

Audits feature flag isolation across 6 dimensions: config projection, import chain integrity, runtime gate consistency, tool/skill tag completeness, boundary coupling, and test marker coverage. Detects where disabled features leak through import chains, runtime bypasses, UI surfaces, and ungated callables.

When to Use

• User says "audit feature gates", "check feature isolation", "feature flag audit"
• After adding a new feature to FEATURE_REGISTRY to verify isolation is complete
• As part of the full-audit.yaml 4th parallel chain

Arguments

No arguments required. Reads FEATURE_REGISTRY from src/autoskillit/core/_type_constants.py to enumerate features.

Critical Constraints

NEVER:

• Modify any source code files
• Create files outside {{AUTOSKILLIT_TEMP}}/audit-feature-gates/
• Issue subagent Task calls sequentially — ALL 6 must be in a single parallel message
• Write output files before synthesizing ALL subagent results
• Run subagents in the background (run_in_background: true is prohibited)

ALWAYS:

• Use model: "sonnet" when spawning all subagents via the Task tool
• Issue all 6 Task calls in a single message to maximize parallelism
• Subagents must NOT create their own files — they return findings in response text only

Severity Semantics

• BLOCK — active leakage that would execute feature code when the feature is disabled
• WARN — isolation gap that may cause confusion or is one step from leakage
• INFO — coupling assessment, informational only

---

Workflow

Step 0: Pre-flight — Enumerate Features

Read src/autoskillit/core/_type_constants.py. Extract all entries from FEATURE_REGISTRY exhaustively — do not assume a fixed list. For each feature, note: name, lifecycle, import_package, tool_tags, skill_categories, default_enabled.

Step 1: Launch 6 Parallel Subagents (SINGLE MESSAGE)

Issue ALL 6 Task calls in a single message.

---

D1 — Config Projection (subagent):

Finding IDs: Assign each finding a unique ID using the scheme FG-D1-{seq:02d} where {seq} is a zero-padded sequence starting at 01 (e.g., FG-D1-01, FG-D1-02). Include the ID in every finding line. For inventory table rows, cite file:line for the config source (e.g., src/autoskillit/config/defaults.yaml:NN or src/autoskillit/core/_type_constants.py:NN for the registry entry).

For each feature in FEATURE_REGISTRY: do not output any prose between iterations — process all features and return findings as structured text only.

• Parse src/autoskillit/config/defaults.yaml features: section
• Parse .autoskillit/config.yaml features: section (if it exists)
• Compute resolved state: config_override ?? defaults ?? FeatureDef.default_enabled
• FLAG (WARN): EXPERIMENTAL feature with default_enabled=True in registry
• FLAG (WARN): EXPERIMENTAL feature enabled in project config while on a stable/main branch
• FLAG (WARN): expired sunset_date values (compare against today's date, if field present)
• Produce inventory table: ID | FEATURE | LIFECYCLE | DEFAULT | CONFIG | RESOLVED | RISK | file:line

Return findings as structured text. Do NOT create any files.

---

D2 — Import Chain Integrity (subagent):

Finding IDs: Assign each finding a unique ID using the scheme FG-D2-{seq:02d} (e.g., FG-D2-01). Every BLOCK and WARN finding must include file:line — this is mandatory, not optional.

For each feature's import_package: do not output any prose between iterations — return findings as structured text only.

• Grep all from {package} import and import {package} across src/ (excluding tests)
• Classify each import site:
• GUARDED: inside a function body, inside if TYPE_CHECKING:, or inside if is_feature_enabled(...):
• UNGUARDED: top-level import in a module that is not itself the feature's package
• FLAG (BLOCK) all UNGUARDED imports with file:line
• Ground-truth targets: tools_kitchen.py, _cook.py, _prompts.py, _fleet.py

Return findings as structured text. Do NOT create any files.

---

D3 — Runtime Gate Consistency (subagent):

Finding IDs: Assign each finding a unique ID using the scheme FG-D3-{seq:02d} (e.g., FG-D3-01). Every BLOCK and WARN finding must include file:line.

For each feature: do not output any prose between iterations — return findings as structured text only.

• Find all is_feature_enabled("{name}" call sites across src/
• Find all AUTOSKILLIT_FEATURES__{NAME} env-var reads (bypass paths)
• FLAG (BLOCK): env-var gate without a corresponding is_feature_enabled() in the same code path
• FLAG (WARN): _fleet_auto_gate_boot() calling mcp.enable() without then calling _redisable_subsets() (server/_lifespan.py)
• FLAG (BLOCK): tool handlers for feature-tagged tools without an in-handler is_feature_enabled() check (e.g., dispatch_food_truck in server/tools_execution.py)
• FLAG (WARN): session-type checks that enable feature functionality without verifying the feature flag

Return findings as structured text. Do NOT create any files.

---

D4 — Tool/Skill Tag Completeness (subagent):

Finding IDs: Assign each finding a unique ID using the scheme FG-D4-{seq:02d} (e.g., FG-D4-01). Every BLOCK and WARN finding must include file:line.

For each feature: do not output any prose between iterations — return findings as structured text only.

• Cross-reference feature_def.tool_tags against TOOL_SUBSET_TAGS in src/autoskillit/core/_type_constants.py
• Grep skill bodies in src/autoskillit/skills_extended/ for feature-specific references
• FLAG (WARN): skills with feature references in body but missing the feature's category in frontmatter
• Verify _DISPLAY_CATEGORIES in cli/_cook.py applies feature-check filtering before displaying
• Verify list_recipes in server/tools_recipe.py filters kind: campaign when fleet is disabled
• FLAG (WARN): any run_python callable in skill_contracts.yaml whose package matches feature_def.import_package without a feature gate in the execution path

Return findings as structured text. Do NOT create any files.

---

D5 — Boundary Coupling (subagent):

Finding IDs: Assign each finding a unique ID using the scheme FG-D5-{seq:02d} (e.g., FG-D5-01). For WARN/BLOCK coupling table rows, cite file:line for the import site or field declaration (e.g., src/autoskillit/pipeline/context.py:NN).

For each feature:

• Grep src/autoskillit/core/ (IL-0) for feature-specific constants or imports beyond FeatureDef/FEATURE_REGISTRY
• Check src/autoskillit/pipeline/context.py for feature-specific fields unconditionally allocated on ToolContext
• Check src/autoskillit/config/settings.py for feature-specific config dataclasses parsed without a validation gate
• Check src/autoskillit/execution/headless.py for unconditional reads of feature config
• Check src/autoskillit/server/_factory.py for unconditional feature-object allocation
• Produce coupling table: ID | LAYER | FEATURE | COUPLING TYPE | SEVERITY | file:line

Return findings as structured text. Do NOT create any files.

---

D6 — Test Marker Coverage (subagent):

Finding IDs: Assign each finding a unique ID using the scheme FG-D6-{seq:02d} (e.g., FG-D6-01). Every WARN finding must include file:line (the test file missing the marker).

For each feature:

• Find test files importing from feature.import_package or referencing feature-specific symbols (Grep tests/ for the import_package name)
• Verify each such test file carries pytest.mark.feature("{name}") at file or class level
• Check tests/arch/test_feature_markers.py for per-feature enforcement lists
• FLAG (WARN): features with no marker enforcement list in test_feature_markers.py
• FLAG (WARN): test files with feature code but no feature("{name}") marker
• Ground-truth: check for test files with fleet code but no feature("fleet") marker; check if planner marker enforcement exists in test_feature_markers.py

Return findings as structured text. Do NOT create any files.

---

Step 2: Synthesize and Assemble Report

After all 6 subagents return, consolidate findings per dimension. Count BLOCK/WARN/INFO totals.

Ensure {{AUTOSKILLIT_TEMP}}/audit-feature-gates/ exists (mkdir -p). All paths below are relative to the current working directory.

Write report to: {{AUTOSKILLIT_TEMP}}/audit-feature-gates/feature_gate_audit_{YYYY-MM-DD_HHMMSS}.md

Report format:

# Feature Gate Audit

**Date:** {YYYY-MM-DD HH:MM:SS}  **Features audited:** {comma-separated list from FEATURE_REGISTRY}

## Config Projection

{D1 findings with BLOCK/WARN/INFO severity badges and FG-D1-NN IDs}

### D1 Remediation Checklist

- [ ] [{ID}] {one-line action to resolve the finding}

## Import Chain Integrity

{D2 findings with BLOCK/WARN/INFO severity badges and FG-D2-NN IDs}

### D2 Remediation Checklist

- [ ] [{ID}] {one-line action to resolve the finding}

## Runtime Gate Consistency

{D3 findings with BLOCK/WARN/INFO severity badges and FG-D3-NN IDs}

### D3 Remediation Checklist

- [ ] [{ID}] {one-line action to resolve the finding}

## Tool/Skill Tag Completeness

{D4 findings with BLOCK/WARN/INFO severity badges and FG-D4-NN IDs}

### D4 Remediation Checklist

- [ ] [{ID}] {one-line action to resolve the finding}

## Boundary Coupling

{D5 findings with BLOCK/WARN/INFO severity badges and FG-D5-NN IDs}

### D5 Remediation Checklist

- [ ] [{ID}] {one-line action to resolve the finding}

## Test Marker Coverage

{D6 findings with BLOCK/WARN/INFO severity badges and FG-D6-NN IDs}

### D6 Remediation Checklist

- [ ] [{ID}] {one-line action to resolve the finding}

## Summary

| Severity | Count |
|----------|-------|
| BLOCK    | N     |
| WARN     | N     |
| INFO     | N     |
| **Total**| N     |

Step 3: Print Handoff

Print to terminal:

[audit-feature-gates] Done.
  BLOCK: {N} | WARN: {N} | INFO: {N}
  Report: {AUTOSKILLIT_TEMP}/audit-feature-gates/feature_gate_audit_{ts}.md

---

Output Location

{{AUTOSKILLIT_TEMP}}/audit-feature-gates/
└── feature_gate_audit_{YYYY-MM-DD_HHMMSS}.md

Related Skills

• /autoskillit:validate-audit — validates the report produced by this skill
• /autoskillit:audit-arch — parallel audit chain in full-audit.yaml
• /autoskillit:audit-tests — parallel audit chain in full-audit.yaml
• /autoskillit:audit-cohesion — parallel audit chain in full-audit.yaml

Source & license

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

• Author: TalonT-Org
• Source: TalonT-Org/AutoSkillit
• License: MIT

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