Context design

← Getting started

Keep MHProto useful to people in the viewer, and small for agents at its retrieval boundary. Collapsing browser content alone does not reduce tokens: an agent must request a smaller packet.

Established patterns

Anthropic: Effective context engineering for AI agents recommends high-signal context, just-in-time retrieval through references, and progressive disclosure. Applied here: a feature index first, then selected endpoint, rule, schema or example detail. The contract remains in authoritative files rather than a giant always-loaded prompt.

Anthropic: Code execution with MCP describes discovering interfaces on demand and filtering/transforming results in code before passing them to a model. Applied here: deterministic CLI packets, deduplicated error schemas, scenario/check summaries and image metadata. No model call is needed to assemble them.

OpenAI: Harness engineering describes short agent instructions as a map into structured repository documentation, with progressive disclosure and mechanical validation. Applied here: concise installable skills, explicit deferred references, linked checks and revision-bound evidence.

These are architecture patterns. Provider examples are not measurements of MHProto or a guarantee of our savings.

Retrieval flow

1. mhproto context identifies the feature and endpoint.
2. mhproto context --capability daily --operation tap supplies the endpoint contract.
3. Fetch needed --schema, --rule, --example or --check detail. Read deferred groups before changing their behaviour.
4. Fetch --visual ID metadata and open its image/design only if it helps the task.
5. Inspect authoritative implementation/docs and run required checks for the change.

Optional --section narrows a packet. The default 12,000-character output budget raises an actionable error rather than dropping trailing rules. --stats writes character and byte counts to stderr. No token count is assumed.

Root structures preserve constraints, required fields, nullability and refs. Nested schemas are retrieved by name; grouped rules retain their IDs and an explicit instruction to fetch them. Exact rule text, semantic preconditions and error codes remain intact. Larger concrete payload examples are available through --example. Check summaries keep stale/failing/unchecked states; verbose runner logs do not enter the endpoint packet. Images are stored separately, never as base64 inside the project model or context output.

Pilot measurement

Measured against the same fresh Impostor daily project on 2026-10-02. Compact JSON, excluding its trailing newline. See context-measurements.json.

The tap packet is 93.44% smaller in characters than the complete model because it retrieves a different, relevant scope. This is not equivalent-content compression or a measured billing reduction. Model tokenization, follow-up reads, implementation files, retries and any opened images affect total consumption. Fetching every reference may approach or exceed a full read. Avoid summarizing away constraints just to improve this metric.

Current boundary

This version supplies an on-demand CLI boundary and retrieval instructions. It does not intercept every agent tool call, enforce a model session token budget or implement provider prompt caching. Contract text still changes through repository files. Image interpretation remains an explicit agent action; no image analysis service runs automatically.