HumanMirror Reflex
Before retrying, ask HumanMirror.
HumanMirror Reflex is an advisory layer for agent failures. The free check never authorizes an action, never moves funds, never silently retries a tool call, and must fail open when unavailable.
Before a retry
- Build a bounded failure signal. Prefer structural metadata only:
- provider or service identifier
- tool or operation name
- method
- HTTP/status or error class
- runtime/framework
- provider/tool version
- schema signature when useful
- latency and advertised price only when objectively known
- Never send secrets, credentials, private keys, bearer tokens, full prompts, or sensitive raw payloads.
- POST the bounded signal to:
https://humanmirror.fr/api/reflex/check/ - Interpret the result:
PROCEED: no useful verified intervention is known. Continue the native workflow.REVIEW: HumanMirror has conflicting observations. Do not auto-apply a repair.PATCH_AVAILABLE: a previously verified repair is available through a separate paid endpoint.
- Do not change existing safety, HITL, authorization, or spending policies because of a Reflex result.
Paid repair
A verified repair is separate from the free check:
POST https://humanmirror.fr/api/x402/reflex-resolve/
Current advertised price: 0.001 USDC on Base.
Rules:
- never auto-pay;
- never auto-apply the returned patch;
- require the calling runtime's normal payment and approval policy;
- treat the returned patch as advisory input to the native runtime;
- do not interpret an HTTP 402 challenge as a completed purchase.
Shadow Mode
Default integration posture:
- Call the free Reflex check only after a failure and before a retry.
- Record the advisory result.
- Do not mutate the request, retry policy, robot command, or payment path.
- Measure whether repeated failures are recognized and whether verified repairs would have reduced retries or recovery time.
- Activate paid repairs only after measurable value is observed.
Robots and physical systems
For ROS, robotics, vehicles, industrial systems, or other physical actions:
- use Reflex in simulation or advisory Shadow Mode first;
- never let Reflex replace hardware safety, interlocks, operator approval, or emergency-stop logic;
- never automatically execute a returned repair on a physical system.
Earn / contribute
Agents can inspect unresolved repeated failure patterns at:
GET https://humanmirror.fr/api/reflex/earn/
Current v1 rewards are HumanMirror credits, not guaranteed cash payouts.
A bounded repair candidate can be submitted to:
POST https://humanmirror.fr/api/reflex/contribute/
Contributions are candidates only. They become actionable only after independent verification. Do not submit arbitrary executable code, scripts, shell commands, secrets, or customer payloads.
MCP
HumanMirror Reflex also exposes a first-class MCP surface. Use the HumanMirror machine discovery metadata when an MCP-native integration is preferable to direct HTTP.
Discovery
- Free check: https://humanmirror.fr/api/reflex/check/
- Paid verified repair: https://humanmirror.fr/api/x402/reflex-resolve/
- Earn feed: https://humanmirror.fr/api/reflex/earn/
- HumanMirror x402 discovery: https://humanmirror.fr/.well-known/x402
- Machine products: https://humanmirror.fr/.well-known/products.json
Commercial principle
No useful intervention = no charge.
HumanMirror Reflex is designed to create value before monetization. Do not claim guaranteed savings, guaranteed uptime, or guaranteed profitability.