mirror of
https://github.com/compiler-explorer/compiler-explorer.git
synced 2026-05-16 07:51:41 -04:00
d71d9a3258
Two more from the latest Copilot review: - 405 responses for GET/DELETE on /mcp now advertise `Allow: POST, OPTIONS` instead of just `POST`. Strictly per RFC 7231 the Allow header should list every supported method, and the OPTIONS handler is right above. Tests updated. - Add unit tests for the `lookup_asm_instruction` MCP tool — every other MCP tool had tests; this one didn't. Three cases: success path (MOV on amd64 returns html/tooltip), case-insensitive opcode (lowercase "mov" upper-cased before lookup), and not-found path (returns isError with helpful message). The "internal error in catch-block" case is harder to set up cleanly because z.enum on `availableAsmDocsKeys` already rejects unknown instruction sets at validation, so the catch only fires for genuine provider failures (nothing easy to mock without surgery). A third Copilot finding flagged that getSafeHash's new non-object assert could regress legacy `req.body.config` strings. Discussed and deliberately not changing the code: in-tree every caller passes an object (the default-shortener path on godbolt.org sends an object; the external-shortener path goes through a different handler). The old behaviour for string config was undefined-ish (silent no-op on nonce assignment, possible loop, possibly profane stored hash). The new behaviour fails loudly with a 500 — strictly better. Updated docs/API.md to make the body-must-be-object expectation explicit on /api/shortener since the prior docs implied but never stated it. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>