Skip to content

v1.1.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 28 Jun 11:52
v1.1.0
509bcf3
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

m1nd v1.1.0 — code-intelligence correctness

A new attention runtime, a function-level call graph across languages, and a hardened cross-file resolver — so impact / why / focus give coding agents the right answer, measured rather than asserted.

What's new

  • focus — goal-conditioned attention runtime. Returns the minimal, budget-bounded working set for a goal, an honest ignored tail (every relevance-clearing node left out is counted, never silently dropped), and an answer-free sufficiency signal (sufficient / gathering / saturated). seek gained the same sufficiency envelope. When an X-RAY manifest resolves, seek/focus bias ranking by conformance (off-by-absence without one).
  • Call graph at function granularity (Rust + TypeScript). Calls are sourced from the enclosing function; free-function and method calls are followed; impact ranks code symbols above containers and production callers above test functions — no more file-containment noise on code.
  • Same-name function ids no longer collide — all 6 extractors. Function nodes used a line-less id, so same-named functions in one file collided and ~6.3% of functions were silently dropped from the graph. Fixed for Rust, TypeScript, Java, Go, Python, and the generic extractor.
  • Cross-file resolution. Proximity prefers same-file > same-directory > cross-crate, and Type::method() / module::func() calls bind to the impl owner via the call qualifier instead of an arbitrary same-name sibling.
  • scan honesty. total_matches_validated counts validation survivors (not the display limit), and documented mitigated findings are now visible at the default severity.

Why

m1nd's bar is to genuinely beat plain rg/Read in the agent inner loop — measured honestly. The graph-reasoning layer was only as good as the extractor + resolver; this release hardens both until that bar is met.

Verification

Every change was gated by an honest battery (m1nd vs rg on m1nd's own repo, ground-truth PASS/FAIL): it went from 10/12 to 28/28 with 0 losses to grep across this arc. cargo test --workspace green on Linux/macOS/Windows; clippy -D warnings + fmt clean.

Known limitation

x.method() calls on a local/field receiver still need receiver-type inference to disambiguate same-name targets — a tracked follow-up.

Install

npm install -g @maxkle1nz/m1nd · cargo install m1nd-mcp

Assets 5
Loading