Home / Blog / Release notes

Release notes

Finder 1.2.0: 20 ms identifier lookups, one-file Finder Kit, finder update

Finder 1.2.0 answers "where is X / who calls X" in about 20 ms from local state, with no embed or rerank call (1.1.0: 1.2–1.7 s). Every answer is 5.6× smaller (6.6k → 1.2k characters) and slightly more accurate (MRR 0.912 → 0.932 on 43 agent questions). The Finder Kit is now one skill file, and Apply in the dashboard installs it globally and removes the old multi-file kit.

Published October 24, 2026 · 5 min read

The numbers: 1.1.0 vs 1.2.0

Measured on our own Rust/TypeScript repo and a Flutter app, with the settings we use every day. The question sets have hand-checked gold files; the agent runs used headless Claude Code (Haiku), 32 runs per tool set.

Metric1.1.01.2.0
Identifier lookup (p50)1.2–1.7 s18–20 ms
Identifier lookup output900–2,500 chars, 15 files266–856 chars, 2–8 files
MRR / hit@3, 43 agent questions0.912 / 0.970.932 / 0.98
Characters per answer6,6391,178
Kit bench: primary file at #1 (20 questions)1012
Semantic question (p50)1,468 ms1,360 ms

Against grep/glob in the same agent, Finder found the right file in the top 3 on 32 of 32 runs (grep/glob: 30 of 32), in 2.44 turns instead of 3.72, at $0.046 instead of $0.059 per question.

Why lookups got faster and answers got smaller

  • Identifier fast path. A short question whose snake_case, CamelCase or a::b names are all defined is answered from the local index: definitions first, then only the chunks that use the names. No provider call, so no provider cost.
  • FCP/2 output. One line per file with symbol names, paths relative to a root header, and deduplicated caller/callee lines. Agents read less and quote file:line directly.
  • BM25 is back on for kit requests. Questions containing :, quotes or brackets were query-syntax errors that silently turned keyword search off. They are now plain terms with identifiers kept whole.
  • Short questions blend rankings. The reranker's order is blended with retrieval order for short questions (MRR 0.833 → 0.920 on 27 of them); long questions are unchanged.

Finder Kit v7: one skill file

The kit used to install 13 files: a router, six phase skills, a pack reader and five shared contracts. It is now a single finder-kit/SKILL.md that covers the whole loop: gate, sharp requests, reading packs, explore, frame, spike, debug, plan, implement, test, verify and review.

  • Spike before planning. When a plan rests on runtime behavior that reading code cannot settle, the agent runs a small throwaway probe and cites its output instead of guessing or asking you.
  • A debug phase. Reproduce first, fix the first place the system went wrong, and stop to test the premise after two failed fixes.
  • Clean upgrade. Apply removes the old multi-file kit when its files still carry the Finder marker. Skills you wrote yourself are never touched.

Dashboard: Apply is global

The MCP settings page no longer asks for a scope. Apply writes your user-level config once, and that one install covers every configured repo; each workspace call carries the active repository root. The finder setup CLI still supports repository installs.

Smart index guard

Approving a folder that holds many projects used to index all of them into one slow, mixed index. Finder now refuses broad folders and says why, so the agent asks you for the project instead of retrying:

  • a drive root such as /Volumes/MyDrive or /mnt/data;
  • a parent of your home folder (/Users) or a system folder (/opt, /private/var);
  • a personal folder such as ~/Documents, ~/Desktop or ~/Downloads;
  • a folder that is not a project itself but contains two or more git repositories, such as ~/code.

Monorepos are not affected: a folder with its own .git or build manifest (package.json, Cargo.toml, pubspec.yaml, ...) is a project. Container paths like /app and /workspace stay indexable.

finder update

Upgrading no longer needs the install script. finder update reads the published checksums, downloads the newest build for your platform, verifies its SHA-256, swaps the binary in place and stops the old server so the next start runs the new version. It finds that server by its executable path, so a Finder started from PATH is stopped too, and nothing else is.

finder update

What did not improve

A semantic question still makes two provider calls (embed ~0.4 s, rerank ~0.6–0.9 s), so it stays around 1.4 s. Skipping the rerank for short questions is faster but misses paraphrased questions, so it is opt-in rather than the default.

Upgrade

1.2.0 is the first release with finder update, so this one time, run the install script:

curl -fsSL https://downloads.skillissuelab.com/install.sh | sh

Then open the dashboard and press Apply again on each agent to install the new kit. From now on, finder update is enough.

Frequently asked questions

Is 1.2.0 available for Linux and Windows?
Not yet. 1.2.0 ships for macOS (Apple Silicon) first; Linux and Windows keep their current build until the next cross-platform release.
Do I need to re-index?
No. The fast path and FCP/2 output use the index you already have.
Why was my folder refused?
It is a drive root, a personal or system folder, or a folder of several repositories. Approve the project folder inside it instead; each project gets its own index.
What happens to my old Finder Kit files?
Apply deletes the old kit files that carry the Finder marker and installs the single new skill. Files without the marker, such as skills you wrote, stay untouched.

Put the right code in your agent's context

One command installs Finder and wires it into your harness. Start with 40 free credits.

$ curl -fsSL https://downloads.skillissuelab.com/install.sh | sh