Provider throttling triggered it.
Retry code sustained it.
Ignored Retry-After · five zero-delay retries · queue 420 → 18,240
Turn @mentions into scoped, reversible agent work—inside the document.
Ignored Retry-After · five zero-delay retries · queue 420 → 18,240
Without a clear document history, context gets lost, decisions become confusing, and agents cannot do their best work.
Provider throttling triggered the incident. Retry middleware sustained the failure.
Every agent gets the same document history and provenance. Ratiflow then exposes only the tools allowed by company policy.
search_demo_coderead_demo_fileIn the Postmortem, select any safe passage, choose @Code, write the instruction, and run—no permission step.
Provider 429 throttling at 09:43 UTC was the external trigger. Retry middleware introduced in commit 7d3c9e1 ignored Retry-After and made up to five zero-delay retries.
Every agent gets the same document history and provenance. In this demo, hard-coded company policy maps @Code to Repository tools.
Full document history + provenance
The same decision trail is available to every agentCode → Repository
Hard-coded for this demo · organization-configured in practicesearch_demo_coderead_demo_fileMetrics · 6 toolsRepository · 7 toolsEditorial · 7 toolsRepository evidence separates the trigger from the retry amplifier; the new replacement is green for 30 seconds and remains restorable.
7d3c9e1 ignored Retry-After and made up to five zero-delay retries, driving traffic to 5.8× baseline and queue depth from 420 to 18,240. The retry regression sustained the failure.History keeps the asker, agent, runtime, evidence, revision lineage, and restore point attached to the document.
The document and comments still work; dynamic discovery and the managed relay fail closed.
Document editable
Comments available
Managed relay discovers assignment tools
TAB-BOUND SITE TOOLSDocument editable
Comments available
Managed relay unavailable by design
WEBMCP OFF · HUMAN MODESources: OpenAI Site Tools · WebMCP draft · OpenAI tool search
An agent running through an API composes each call; the browser discovers and executes WebMCP tools; Ratiflow enforces and records the result.
Sources: OpenAI Site Tools · WebMCP draft · OpenAI tool search
Today, tool execution depends on a live page. The next step is keeping context current and approved work durable.
Use case An agent can refresh the affected facts instead of starting over or continuing with stale context.
Engineering Typed resources plus change notifications let agents re-read only invalidated state.Use case A long-running task can survive navigation and still return a clear, reviewable result.
Engineering Worker-backed sessions carry delegated identity and scope, with idempotent receipts.Official discussion: current draft · resources #151 · progress #196 · service workers
See the full people-and-agents document workflow in the live app.
Slide 1 of 11: Ratiflow