Feature

Feature — Feature Flags & Progressive Delivery

Activate When

• User invokes /godmode:feature
• User says "feature flag", "gradual rollout"
• User says "kill switch", "A/B test"
• Deploying needs safer rollout or deploy/release decoupling

Workflow

Step 1: Flag Strategy Assessment

# Detect existing flag infrastructure
grep -r "launchdarkly\|unleash\|flagsmith\|split\|statsig" \
  package.json pyproject.toml 2>/dev/null

# Find existing flag usage
grep -rn "featureFlag\|isEnabled\|isFeatureEnabled" \
  src/ --include="*.ts" --include="*.py" 2>/dev/null | wc -l

# Check for stale flags (not evaluated in 30+ days)
grep -rn "featureFlag\|isEnabled" src/ \
  --include="*.ts" 2>/dev/null | head -20

FLAG STRATEGY:
  Current: None | Homegrown | LaunchDarkly | Unleash
  Flag needs: Release | Experiment | Ops | Permission
  Environments: Dev | Staging | Prod | Mobile

IF  baseline - 2%
  Support tickets  5% for 2min: disable flag
  IF P99 > 3x baseline for 5min: disable flag

FALLBACK:
  Locally cached values (5min TTL)
  Then hardcoded defaults
  Never crash because flag service is down

Step 5: Lifecycle Management

NAMING CONVENTION:
  enable_ (release)
  exp_ (experiment)
  disable_ (ops kill switch)

STALE FLAG DETECTION:
  Release at 100% for > 2 weeks: STALE
  Experiment concluded > 1 week: STALE
  Not evaluated in > 30 days: STALE
  > 90 days without cleanup date: STALE

THRESHOLDS:
  Max flags per service: 50 (audit if more)
  Evaluation latency:  2 variants

Step 7: Validation

Check: types categorized, naming enforced, owner assigned, cleanup dates set, kill switches defined, sticky bucketing, fallback behavior, audit trail, evaluation — at %"`

Key Behaviors

Never ask to continue. Loop autonomously until done.

1. Flags are temporary by default. Cleanup required.
2. Decouple deploy from release. Ship dark.
3. Kill switches are not optional.
4. Clean up flags. Max 2 weeks at 100%.
5. Sticky bucketing is mandatory.
6. Test both paths. On, off, and transition.
7. Default to off, fail to off.

HARD RULES

1. Never leave release flags at 100% for > 2 weeks.
2. Never nest flag checks. Keep flags independent.
3. Never use flags as permanent config.
4. Always test both flag-on AND flag-off paths.
5. Never evaluate flags in hot loops. Once per request.
6. Never expose server-side targeting rules to clients.
7. Never call experiment winner without sample size.
8. Always maintain an audit trail for flag changes.

Auto-Detection

grep -r "launchdarkly\|unleash\|flagsmith" package.json
grep -rl "featureFlag\|isEnabled" src/ --include="*.ts"

Output Format

Print: Feature: {flag} at {pct}% rollout. Error={err}%, latency_p99={lat}ms. Kill switch: {tested|untested}. Status: {status}.

TSV Logging

step	flag_name	flag_type	platform	status	details

Keep/Discard Discipline

KEEP if: metrics within guardrails AND no error spike
  AND kill switch tested
DISCARD if: error > baseline+0.1% OR latency > +10%
  OR conversion drops > 2%
On discard: roll back to previous percentage.

Stop Conditions

STOP when ALL of:
  - Flag at 100% with stable metrics for 7 days
  - Kill switch tested and verified
  - Cleanup scheduled (within 2 weeks)

Error Recovery

• SDK init fails: check API key, implement fallbacks.
• Flag not evaluating: check context (user.id passed?).
• Rollout errors: set to 0%, check logs, resume at 1%.
• Stale flags: run audit, remove code, delete branches.
• Conflicting flags: map deps, encode as prerequisites.

Source & license

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

• Author: arbazkhan971
• Source: arbazkhan971/godmode
• License: MIT

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