Fusion Operations

Eliminate redundant roundtrips by combining multiple operations into single server calls

Fusion operations are optimizations that eliminate redundant DOM walks, serializations, and WebSocket roundtrips. They combine what would be multiple server calls into a single call, reducing latency and token usage.

All fusion fields are optional protocol extensions — zero breaking changes. Old clients and servers simply ignore the new fields.

The 7 Fusion Operations

1. Navigate + Observe

Fuse navigation with observation in one server call instead of two.

Without fusion (2 calls)

bap goto https://example.com
bap observe

The server calls page.goto() then runs handleAgentObserve in-process, returning both the page state and interactive elements in a single response.

2. Act + Post-Observe

Get an observation after executing actions, without an extra roundtrip.

Without fusion (2 calls)

bap act click:text:"Next"
bap observe

In MCP, pass postObserve: true in the act tool parameters.

3. Act + Pre-Observe

Capture page state before the action executes. Useful for diffing before/after states.

In MCP, pass preObserve: true in the act tool parameters. The server calls handleAgentObserve before the step loop executes.

Pre-observe runs BEFORE steps execute (fixed in v0.6.0). Earlier versions had a bug where pre-observe ran after all steps completed.

4. Incremental Observe

Only return changes since the last observation instead of the full element list.

Full observation

bap observe
# Returns all 50 interactive elements

The server snapshots previous element refs before updating the registry, then diffs to produce a changes object. In MCP, pass incremental: true to the observe tool.

5. Response Tiers

Control how much data the observe response includes:

Tier	What is included	Use case
full	Interactive elements + accessibility tree + screenshot	First observation of a page
interactive	Interactive elements only (skips tree and screenshot)	Mid-workflow checks
minimal	Elements stripped to {ref, role, name, selector, tagName}	Tight token budgets

bap observe --tier=minimal
bap act click:e3 --observe --tier=interactive

6. Selector Caching

Each ElementRegistryEntry stores a cachedCssSelector. When resolving a selector, the server tries the cached CSS path for a direct lookup before falling back to semantic resolution. This eliminates DOM traversal for repeat interactions.

Selector caching is automatic. No flags needed — it kicks in after the first successful resolution of any element.

7. Speculative Prefetch

After handleAgentAct, if the last step was a click or navigation, the server fires off an observe call 200ms later. The result is cached in ClientState.speculativeObservation (URL-matched, 5-second TTL). If the agent requests an observation within that window, the cached result is returned instantly.

Fusion in the TypeScript SDK

// Navigate + observe
await client.navigate("https://example.com", { observe: true });

// Act with pre/post observe
await client.act(steps, {
  preObserve: true,
  postObserve: true,
});

// Incremental observe
await client.observe({ incremental: true });

// Response tiers
await client.observe({ responseTier: "minimal" });

Fusion in MCP Tools

// navigate with fused observe
{ "tool": "navigate", "arguments": { "url": "https://example.com", "observe": true } }

// act with post-observe
{ "tool": "act", "arguments": { "steps": [...], "postObserve": true } }

// incremental observe with response tier
{ "tool": "observe", "arguments": { "incremental": true, "responseTier": "interactive" } }

Impact

Fusion operations can significantly reduce the number of server calls in multi-step workflows:

Scenario	Without fusion	With fusion	Reduction
Navigate + observe	2 calls	1 call	50%
Login form (fill + fill + click + observe)	4 calls	1 call	75%
Browse 5 pages with observation	10 calls	5 calls	50%

In CLI mode, the --observe flag is the single most impactful fusion. Add it to goto and act commands whenever the agent needs to see the page after the action.