IncidentBridge

Delegate the vendor-support call, bring back corroborated ticket/ETA evidence, and return operator attention to the incident—without letting the phone close it.

Comment

The problem

During an outage, the same operator diagnosing the failure often has to stop, wait on a vendor-support line, repeat known incident context, collect a ticket/ETA/workaround, then carry those facts back into the incident workflow.

IncidentBridge delegates exactly that bounded interruption to CALL-E.

The business outcome is simple: reclaim operator attention while preserving incident authority.

The judge-visible outcome

IncidentBridge takes an approved incident brief and delegates one authorized vendor-support conversation. CALL-E returns structured evidence such as:

  • vendor acknowledgment
  • ticket ID
  • status
  • ETA
  • workaround
  • callback window
  • escalation state

That evidence is bound to the exact workflow, incident, call ID, and destination and corroborated against recipient-side transcript text.

Then the safety boundary matters:

the phone call can create evidence, but it can never declare recovery. incident_closed remains false until local health checks and a human incident commander verify the service.

Demo: https://youtu.be/x5YNuT4gNgA

Evidence console: https://yangyangnovelist-hub.github.io/incidentbridge-calle/

Impact calculator: https://yangyangnovelist-hub.github.io/incidentbridge-calle/impact-calculator.html

Why it matters

The first KPI is not a fabricated MTTR claim. It is directly measurable:

operator_minutes_reclaimed = manual_operator_minutes - approval_and_review_minutes

A real deployment can then calculate annual hours reclaimed from vendor-call frequency and delegation share.

The first safety KPIs are equally explicit:

  • unauthorized-call rate = 0 target
  • false-success rate = 0 target

Product experience

The operator console defaults to preview. A live call requires independent server-side and operator gates, including explicit authorization, an exact E.164 allowlist, live-call enablement, and typed confirmation.

The ideal workflow is:

incident starts → operator approves bounded vendor call → CALL-E calls → ticket/ETA evidence returns → incident timeline updates → operator keeps debugging

The next public proof target is one fully consented synthetic success call that produces a privacy-safe consented-live-success.json artifact.

Technical implementation

IncidentBridge uses the published calle-ai==0.2.0 Python SDK on the real execution path.

Key production-oriented controls include:

  • durable SQLite reservation before dispatch
  • stable idempotency identity
  • strict CALL-E result schema
  • call/workflow/incident/destination binding
  • recipient-side ticket corroboration
  • privacy screening and recursive redaction
  • fail-closed routing on low confidence, wrong desk, incomplete or mismatched evidence
  • ambiguous provider outcome → no blind automatic retry
  • loopback-only browser live mode with server-side credential isolation

The implementation has 33 passing tests, 93.80% total coverage, a 90% enforced coverage gate, Ruff-clean CI, and integration tests that exercise the published CALL-E SDK through loopback HTTP.

Independent upstream validation

CALL-E maintainers reviewed the contribution in PR #132 and identified two substantive security blockers: incomplete privacy validation and insufficient binding/corroboration of returned evidence. Both were fixed and converted into regression tests before the PR was merged into the official CALL-E community repository on August 17, 2026.

Source: https://github.com/yangyangnovelist-hub/incidentbridge-calle

What makes it different

IncidentBridge is not an internal on-call escalation bot and not a synthetic phone monitor. It targets the external dependency step during an incident: contact the vendor, bring back evidence, and keep recovery authority local.

Real-world validation boundary

IncidentBridge has already exercised the real CALL-E provider boundary. A real run reached voicemail and correctly failed closed without blind redial. The complete success route is inspectable deterministically, and the project includes a one-shot consented live runner for a privacy-safe public success artifact.

The next upgrade is not more feature surface. It is stronger real-world proof: one authorized success run and measured operator time reclaimed.

Why it is worth building

Incident response is already cognitively expensive. IncidentBridge removes one repetitive interruption without pretending a phone conversation is operational truth.

Built With

Share this project:

Updates