decide v0.1.1 — isolated failures, explicit data roots

A corrupt provider response could cancel every worker and leave no summary. Version 0.1.1 sends decoding and other HTTP failures to that record's review queue while preserving the rest of the batch and its artifacts.

Setup change: DECIDE_ROOT must now be an explicit absolute path to an existing data directory. Missing or relative values fail before file access or provider requests. File and line IDs use forward slashes on all platforms.

Install the wheel directly, without cloning the benchmark archive:

uvx --python 3.11 --from https://github.com/alsoleg89/decide/releases/download/v0.1.1/decide_mcp-0.1.1-py3-none-any.whl decide-mcp

Set TYPESAFE_API_KEY and DECIDE_ROOT in your MCP client's server environment. Setup.

The README now leads with installation and the tool example. Reports add a dated Jev price source, exact McNemar tests and paired bootstrap intervals. For the main mini UX study, the accuracy gain is 6 percentage points, with a 95% paired interval of +4 to +8 points; exact McNemar p = 5.49e-9. These post-hoc diagnostics do not change frozen gates. All 14 historical arm reports retain the same costs and scores. The fixed 25-record baseline has not been optimized against larger batches.

Validation: 261 offline tests and 97% server coverage on macOS/Python 3.11 and Linux/Python 3.12; clean wheel installation and real MCP smoke checks on both. No paid model calls or GitHub Actions. Windows remains unverified.

The 30-second cooldown stop policy and lack of automatic resume remain unchanged. Documentation explains how to select failed records for a new batch and why timeout retries may create additional billable work.

MIT license, wheel, sdist and SHA256SUMS included. Detailed checks.

decide v0.1.0 — Your agent reasons. Jev sorts.

Your agent reasons. Jev sorts.

decide is an MCP tool for bulk decisions over files, log lines, JSONL records or inline items. Send a path and a rubric; receive a compact summary, a complete local audit trail and the cases that need closer review. Global and per-label confidence cutoffs let you optimize cost for the quality your task needs. Review rate is an outcome, not a fixed 5% target.

Measured results

On 1,000 new UX reviews, decide + GPT-4.1 mini reduced complete tool-loop inference cost 48.9% ($0.04993 → $0.02551), while improving accuracy 84.2% → 90.2%, macro-F1 0.7803 → 0.8577, precision 58.4% → 72.0% and recall 76.8% → 84.7%. The routing policy was frozen before these records were evaluated. Mini reviewed 123 records.

UX protocol, predictions and mistakes

On 300 OpenCV issues, a different policy saved 77.3% and improved accuracy and macro-F1, but reduced feature recall. These are task-specific sample results from a guided tool loop, not a universal quality guarantee or proof of the globally cheapest policy. Failed runs and losing comparisons remain published.

Developer results · All evaluations

Install

Requires uv, Git and a TypeSafe API key. Configure your MCP client to launch:

uvx --python 3.11 --from git+https://github.com/alsoleg89/decide@v0.1.0 decide-mcp

Set TYPESAFE_API_KEY and DECIDE_ROOT in the server environment. Codex and Claude setup.

Release verification

• MIT license included in both wheel and source distribution.
• Wheel built from the source distribution; packaged server and README match the release source.
• 257 offline checks passed on macOS/Python 3.11 and Linux/Python 3.12, with 97% server coverage.
• Fresh wheel installation and real stdio MCP checks passed on macOS/Python 3.13 and Linux/Python 3.12, with zero paid requests.
• GitHub Actions is not used. Windows remains unverified.
• Wheel, source distribution and SHA256SUMS are attached. No PyPI publication.

Verification details