Skip to content

0024: Context and audit pages

Status: proposed for M4. Covers the web part of spec DX-4 (P1) and the HTML report of AUD-10 (P2).

Problem

ouncemcp context showed one definition at a time in the terminal, and ouncemcp audit had terminal, JSON and Markdown output only. Comparing tools, or sharing what a server costs, meant reading JSON.

Proposal

Both commands write a page with the report app from proposal 0020:

  • ouncemcp context server.py [tool] --html context.html [--call JSON].
    • Startup context against the budget, the discovery mode and why, the tokenizer and the counting scope.
    • Every tool with a bar showing where its tokens go (description, input schema, output schema, name and structure), its status and any budget override (BUD-4).
    • For the selected tool: the counted definition exactly as clients receive it in tools/list, and the fields that are sent but not counted (title, annotations, _meta), since hosts may show them to the model.
    • With --call: the result's content as the model reads it, its token count, the structured content and _meta.ouncemcp.
    • The tool name is optional with --html; when given, the page opens on it.
  • ouncemcp audit TARGET --html audit.html: startup context, tools with the same token bars, findings, the discovery recommendation, tools/list latency, sampled result sizes, resources and prompts, and any regressions from --compare.

ouncemcp.ui.views builds the data (context_data, audit_data, call_result).

Alternatives considered

  • A live web server for context. A page is enough for one server, can be attached to a review, and needs no port. ouncemcp dev can later serve the same page and refresh it on reload.

Impact on tokens and task success

None.

Migration notes

None. The terminal output of both commands is unchanged.