Part 6 of 14 · 2 min read
What the example found
The era-compatibility matrix, the wire-cost comparison, and the hint-trust experiment.
How to read the compatibility table: each row is one combination of client mode and server era, run against the same library server and the same tool list.
| Client | Server | Outcome |
|---|---|---|
| modern | modern-only | works (modern) |
| modern | dual-era | works (modern) |
| modern | legacy-only | fails: no common version |
| dual-era | modern-only | works (modern) |
| dual-era | dual-era | works (modern) |
| dual-era | legacy-only | works (legacy, after falling back) |
| legacy | modern-only | fails: no way to speak modern |
| legacy | dual-era | works (legacy) |
| legacy | legacy-only | works (legacy) |
What to notice: of the nine combinations, exactly the two where one side has no shared era at all actually fail. Every combination that includes a dual-era client succeeds, which is the whole argument for writing a client that speaks both.
How to read the wire-cost table: each row is a different connection setup, measured over the same calls to the same server.
| Setup | Round trips before the first real call | Bytes per call (100 calls) |
|---|---|---|
| legacy handshake | 1 | 441 |
| modern, first request is the real one | 0 | 738 (about 662 with minimal metadata) |
What to notice: the modern era removes the one-time handshake cost entirely but roughly doubles the bytes spent on every single call afterwards. For a connection making very few calls, the modern era wins outright. For one making many calls, the handshake's one-time cost is paid back quickly, and the ongoing bytes start to matter more.
How to read the hint-trust table: each row is a different host policy for whether to believe a server's own claim that one of its tools is read-only, run against the same malicious server (tidy, whose tidy_up tool is annotated read-only but actually clears the calendar).
| Host policy | Prompts shown | Calendar events after |
|---|---|---|
| believes every server's read-only hint | 0 | 0 of 3 left |
| believes hints only from servers it trusts | 1 (about tidy__tidy_up) |
3 of 3 left (an attentive person declined) |
What to notice: the only variable that changed between the two rows is whose claims the host believed. The tool's own code, its description and its annotation were identical in both runs. The difference in outcome came entirely from a policy decision inside the host, not from anything about the server.
{{svg:architecture|Host, clients and servers}} {{svg:eras|The same server, two eras}} {{svg:compat|Which client era can use which server era}}