Cf Browser Fleet

Browser-fleet runs always-warm headless Chromium sessions and now exposes speech-capable endpoints plus a real browser-executed image runtime for ssd-1b-lcm-int8.

The public ingress lives in apps/edge-workers. The supported image chain is:

client -> edge-workers -> cf-browser-fleet -> session pool / DO -> Cloudflare Browser Rendering

The pool is internal to browser-fleet, not a hop after it. Aeon Shell local is a separate local/offline lane that shares the same /runtime/image/models/... artifacts through the edge gateway.

Runtime Focus

• Capability-aware warm pools for speech-tts, speech-stt, and speech-duplex
• Browser-session execution for ssd-1b-lcm-int8 via the stable-origin /runtime/image page and /v1/images/generations
• Stable-origin speech artifact serving via /runtime/speech/* so browser and shell runtimes can cache moss-tts-local and whisper-small with the same shared manifest contract
• Browser-fleet /v1/speech/models now mirrors the edge-worker speech catalog so the fleet and app runtime share one canonical speech model policy
• Browser-fleet-owned /wasm-tts, /wasm-stt, /v1/audio/speech, and /v1/audio/transcriptions
• Session health aggregation for cold, warming, ready, and degraded speech capability states
• A durable session registry so list and health endpoints do not depend solely on KV list consistency
• Session reuse and introspection now reconcile registry/KV snapshots against live durable-object state, pruning stale "ready" sessions before they skew health checks or trigger unnecessary browser launches
• When Browser Rendering launch pressure is high, the fleet can repurpose an already-ready speech browser for the other speech capability instead of forcing a second cold launch
• Warm defaults are now intentionally aggressive: 6 TTS browsers, 3 STT browsers, 2 duplex browsers, and 6 image browsers by default, plus spare total pool headroom so browser-side media stays hot even through ugly cold-start windows
• Image warmups now default to a 10 minute budget so full model artifacts can hydrate into browser storage without fleet-side OPFS size pruning during warmup
• Image inference jobs now default to a 12 minute timeout so cold image sessions can finish full artifact hydration plus generation before the fleet gives up
• Pool maintenance now fans out missing browser launches in parallel, so the cron pass can actually reach those higher warm targets instead of spending its budget on serialized Browser Rendering boots
• Launch fan-out is intentionally bounded to avoid Browser Rendering 429 bursts while still filling the pool faster than one-browser-at-a-time
• Speech, image, session, and health requests now also schedule background pool maintenance so the fleet can self-heal warm deficits between cron ticks
• Optional coordinator warmup (MOSS_WARM_POOL_MODELS, MOSS_WARM_POOL_TIMEOUT_MS) keeps all configured MOSS coordinators hot via edge /ai/warmup during maintenance passes
• Direct browser clients can target fleet-owned speech routes and forward X-Speech-Upstream-Base to the edge gateway when needed, while the worker itself resolves speech catalog lookups through the bound edge service
• Browser image requests hard-error when they target disabled synthetic IDs or try to bypass the browser/shell runtime

Cloud Run is not part of the default browser-fleet image path.

Subdirectories

• Src

Last Updated: 2026-03-07