Origin Trial research note: observations on tool exposure, validation, and context cost

Read the issue’s listed experiments and linked article and journal, then compare the observations with the README questions named in the issue. The issue asks whether to contribute evidence to an existing discussion but does not identify a specific issue or define a deliverable; done would require maintainers to select a focused, reproducible evidence package.

Hello WebMCP maintainers,

I have followed this project since its early development and wanted to contribute only after understanding its architecture more carefully.

My initial idea was to propose a Context Continuity Package (CCP) that could help agents resume work across local and browser surfaces. After auditing the repository, activating the Origin Trial, building an external laboratory, and testing the hypothesis across several agent surfaces, I concluded that my original proposal was aimed at the wrong layer.

The experiment suggests that WebMCP should expose page capabilities, while continuity, retention, compaction, and forgetting remain responsibilities of the agent harness or product.

I am therefore not reporting a bug, requesting a new memory primitive, or proposing code in this issue. I am sharing the resulting observations because they appear relevant to several questions currently listed in the README.

What the experiment covered

• Origin Trial activation with the experimental flag returned to its default state.
• A read-only WebMCP tool returning a bounded, structured continuity packet.
• Input rejection, output validation, document lifecycle, navigation, cancellation, deregistration, BFCache, origin boundaries, and hostile-but-valid content.
• A matrix of 100 trial slots across Codex, ChatGPT/Chrome, Antigravity, and Edge.
• Fifty ablation trials examining which fields improved decisions.
• An Antigravity experiment separating the visible cost of continuity, tool definitions, capability discovery, and tool results.

Results that corrected my hypothesis

• Manual CCP recovered the expected next action in 19/20 completed trials, compared with 14/20 for a byte-matched free summary and 19/20 for full history.
• In completed Antigravity trials, CCP reduced average visible context by approximately 27% in bytes and 25.1% in proxy tokens.
• Bounded tool results reduced a different component by approximately 21.2% in bytes and 19.6% in proxy tokens.
• The theoretical combined reduction exceeded 50%, but progressive discovery did not pass the native pilot. The 50% operational target is therefore not demonstrated.
• Native discovery was irregular. Antigravity produced two server-confirmed invocations in five attempts, only one of which produced a usable resumption report. The observed ChatGPT/Chrome and Edge configurations produced no confirmed native invocation.
• Structural validation and same-origin exposure reduced some risks but did not establish truth, trustworthy provenance, or safe intent.

Connections with the current open questions

The experiment may provide implementation evidence relevant to:

• Cross-document tool responses: navigation, cancellation, document destruction, deregistration, and BFCache behavior.
• Built-in agent exposure: separating API presence, registration, exposure, discovery, invocation, and actual use.
• Input/output validation and outputSchema: strict empty-input enforcement, bounded structured output, stable rejection codes, and the limitations of schema validity.
• Skills integration: evidence that higher-level continuity and coordination belong primarily to the harness rather than to a WebMCP memory layer.
• Streaming and bounded responses: evidence that response size materially affects visible context, without yet establishing a streaming contract.
• Progress and observability: timeouts and discovery failures showed the importance of identifying exactly where an invocation chain stops.

Short and long versions

• Short DEV article:
  https://dev.to/jsb-securedme/i-tried-to-add-memory-to-webmcp-the-experiment-told-me-not-to-1bg8

• Companion podcast:
  https://drive.google.com/file/d/1CQRfdNqnumy41DzyaloCewxEa6HyiVKl/view?usp=sharing

• Full trilingual construction journal:
  https://doi.org/10.5281/zenodo.22049347

Question for the maintainers

Would any of these observations be useful in an existing open issue—particularly around built-in-agent exposure, schema validation, cross-document lifecycle, or tool-discovery cost?

If so, I would be happy to extract a small, focused, reproducible evidence package for the most relevant discussion. I do not want to introduce a broader proposal where a narrower contribution would be more useful.

Thank you for the work on WebMCP and for making these design questions visible.