Design partners: reproduce exact one-use approval on a real agent tool

[AjnasNB]

Goal

Recruit independent developers to reproduce Maqam's exact-approval proof and apply it to one real agent tool.

Start here

npx -y maqam@0.3.2 demo approval

The bundled demo should show:

1. the first request requires approval;
2. altered input is rejected;
3. the exact approved input executes once;
4. replay is rejected;
5. source-linked evidence is emitted.

Design-partner task

Integrate one consequential tool through ToolGateway?for example a file write, HTTP mutation, deployment command, issue update, or browser form submission?and report:

• Node and operating-system version;
• time from install to first governed run;
• the policy and approval boundary used;
• whether altered-input and replay tests failed closed;
• any unclear documentation or unsafe default;
• a minimal reproduction for failures, with credentials and private data removed.

Acceptance criteria

• Clean-registry install works on a supported Node 22, 24, or 26 runtime.
• A denied or approval-required request is demonstrated.
• Exact input executes once.
• Changed input and replay do not execute.
• Evidence can be inspected without exposing secrets.
• Findings are posted as a comment or separate minimal issue.

This is early 0.x software. Maqam governs only operations routed through registered adapters; it is not an OS sandbox and does not make an unsafe tool safe.

[AjnasNB]

Thanks for volunteering—yes, the free design-partner proof is exactly the contribution we are looking for.

Please proceed with one narrowly scoped, reversible tool: either a write into a temporary directory or an HTTP mutation against a disposable test endpoint. Route it through ToolGateway and publish a sanitized reproduction that includes:

• repository/commit and exact commands;
• Node 22, 24, or 26 plus OS;
• time from clean install to first governed run;
• the policy and exact approval boundary;
• first-request, changed-input, approved-once, and replay results;
• evidence/receipt linkage to the A2A task ID;
• any documentation gaps or unsafe defaults you found.

Please do not include credentials, private data, production targets, or a destructive command. The acceptance criteria in this issue remain the review bar.

To keep the scope clear: this issue is an open-source design-partner evaluation and creates no paid commitment. If the public proof is reproducible and useful, we can discuss recurring regression fixtures separately, with their scope and terms reviewed independently.