[jaymar921]

Small models can now change part of an existing file. On the edit benchmark, llama3.2:1b went from 0 of 12 to 12 of 12 on the suite used during development, and from 0 of 12 to 9 of 12 on the later Python suite, with no damaged file in any 1.2.0 run. All benchmark numbers are from machine B on 2026-10-09.

What changed

• Ollama context window: nothing set num_ctx, so Ollama dropped the start of long prompts. Each model now gets one window, 8k on the constrained tier and 16k on the full tier. The new setting is hirayacoder.ollama.contextLength.
• edit_file: changes part of a file and goes through the existing writeFile path, so the permission gate, diff, content guards and audit log all still apply.
• Dictated edits for small models: on the constrained tier, the extension finds the lines itself and asks for the rewrite in plain words. It checks the answer before writing anything.
• Memory and translator: file-change notes now come from the diff, and memory lists what each changed file provides.
• Panel: two-row header, context use on each reply, and a What it remembered panel.
• Tooling: new benchmark, tools/bench-edit.js.

Full notes are in CHANGELOG.md.

Worth checking in review

Dependency overrides. package.json now pins glob ^14.0.3 and minimatch ^11.0.5, scoped to Mocha and @vscode/vsce. Current Mocha 12 still asks for a vulnerable glob, so upgrading would not have fixed it. npm audit went from 13 dev-only advisories to 0. Nothing here ships, and the pins can go once both tools depend on fixed versions.

3B models. llama3.2:latest gets no benefit by default: it stays on Tier A and passed 0 of 8 and 2 of 8. Forced to Tier B it passed 16 of 16. The default line has not moved, because the other benchmarks have not been re-run with it moved. The changelog recommends the override.

Testing

Run for this PR, on Windows:

• npm run lint: 0 errors, 85 warnings, all reviewed in the SAST report
• Unit tests: 1709 passing, 2 pending
• node tools/bench-edit.js --selftest: passed, with a real Python on PATH (the python command on this machine is the Microsoft Store stub, and the inventory suite fails with it)
• npm audit: 0 vulnerabilities
• npm run package: builds hirayacoder-1.2.0.vsix

Not run:

• Integration tests were not run locally, so CI is the first run on all three platforms.
• No model was run this time, because none is installed on this machine. The benchmark figures were not re-run, and this final state has not been exercised against a live Ollama.

Tagging

The version is already 1.2.0 and the changelog is dated 2026-11-06. After merging, tag v1.2.0 on main. If the merge lands on a later day, update the changelog date first.

Each file is its own signed commit, so please merge with a merge commit rather than squashing if you want to keep that history.