Skip to content

v0.1.16 — gates that actually run

Choose a tag to compare

@ebursztein ebursztein released this 07 Aug 18:36
· 30 commits to main since this release
e663db9

Sprinty used to take the agent's word for it. item_done accepted gate_results as assertions, and spec: "true" passed every check in the system — it was the default in ~150 places across the test suite.

Breaking

  • Gates execute. item_done({ id, commit_id, changelog }). Sprinty runs the item's executable gates and records their real output; only manual gates take agent evidence. Specs that cannot fail (true, :, exit 0, bare echo) are rejected at declaration, while echo ok && npm test still passes.
  • 5s item gate budget. A gate that overruns is killed and marked timed_out — distinct from failed, because nothing was proven either way. It does not block the item; sprint_close runs it in full, and the response warns with both ways out.
  • Spikes removed. They were never registered as MCP tools, so no client could reach them.
  • help replaced by next_tools. The old static string was the largest field in most responses and built ids from a regex fallback, so on a fresh sprint it suggested calls that error with Unknown.

Fixes

  • sprint_close no longer closes over a subsprint that never got items. item_deprecate takes a subsprint id as the way out.
  • sprint_resume returns a bounded briefing: 1358 chars against the 7896 of overview+next it replaces.
  • Note ids are minted rather than derived from ledger position, so concurrent writers cannot collide. Existing ledgers keep their ids.
  • Ledger appends are O(1): 4000 appends went from ~11s to 135ms.
  • Dashboard static-asset guard no longer admits prefix-sharing sibling directories via %2f.
  • npm run test:coverage runs; it previously failed on healthy code.
  • item_get reports the files the commit actually touched and flags undeclared ones as scope drift.

223 tests, typecheck clean.