dev-decide 5.1.0

5.1.0 (25 October 2026)

Added

• Instant routing hint. A small local server (hookd.py) answers Claude Code's HTTP hook in about 5 ms instead of
  starting Python on every prompt (about 300 ms on Windows). A SessionStart hook starts it when needed. The old
  per-prompt command hook remains available (install.py --router-hook=command).
• Routing by meaning and area. Without a trained route model, matching blends words, the bundled embedding model
  and technical areas (postgres, column, schema count as database). On held-out requests hint precision went from
  80.5% to 88.1%, and false hints with one unrelated skill from 19.5% to 3.2% (tools/eval_router.py).
• Example agents and their route model. Eight subagents in dev-decide-kit/example-agents/ (backend-reviewer,
  db-migrate, frontend-dev, debugger, security-auditor, test-writer, devops-deploy, docs-writer). When they are
  installed, the shipped route model (trained on them) answers routing locally; it is never used for other agents.
• Web page: public demo token filled in for visitors, bring-your-own Anthropic key, one example per decision type,
  right/wrong feedback that the learning loop uses, batch mode (one input per line), history kept in the browser,
  and a warning when the daily Claude budget is nearly used.
• Server: DEV_DECIDE_DEMO_TOKEN (local answers only, nothing about demo requests is stored or learned),
  X-Anthropic-Api-Key (the caller's own key for that request only: never stored, logged or counted),
  DEV_DECIDE_LOCAL_ONLY, a per-client rate limit, a daily model-call cap with an 80% warning, security headers,
  /v1/presets, and richer /health.
• Learning that survives restarts. statesync.py saves what a server learned (models and hashed labels, never
  input text) to a private git repository with a deploy key, and restores it on start.
• GitHub Actions. dev-decide-triage.yml labels new issues and comments; dev-decide-pr.yml rates each pull
  request, keeps one comment up to date and fails the check on block.
• Per-class confidence thresholds. Training measures, per answer class, the confidence its answers need to reach
  the target; weak classes need more before they are answered locally.
• tools/benchmark.py (label, train with calibration, evaluate, write the results table), tools/smoke.py
  end-to-end checks, CI on Linux and Windows with Python 3.8 and 3.12.

Changed

• Triage models retrained with 149 hand-labeled short reports (checked blind by Sonnet) and 772 new generated issues.
  issue_type: short reports never trained on went from 35% to 88% right (96% of its local answers right), 95.3% on
  the held-out test split. severity and domain are now enabled (local answers 94% and 100% right on examples
  neither run trained on). Repositories holding an older starter model get the new ones on their next decision;
  models you trained or the learning loop promoted are never replaced.
• Rule-based guesses no longer call every short report noise.
• The project root is found without starting git (faster on Windows).
• Starter models and a documented settings file (.claude/dev-decide.json) are installed on first use.

Fixed

• Rolling back to the oldest archived model crashed.
• Secrets written as KEY=value without quotes (as in .env files) were not detected; a diff too large to scan
  fully was reported as having no secrets.
• A Claude Code plan or usage limit message was taken for a model answer (and cached by the kit).
• A gate could block on rough local guesses alone; such rules now ask for review, while rules and models still block.
• Reinstalling while the routing-hint server ran failed on Windows (a memory-mapped file); the installer now stops it first.
• Gates passed on answers that were unsure; .pyc files were reviewed as part of a change; a COMMIT -x argument
  could be read as a git option; and the local API answered other websites when no token was set.

Install or update in Claude Code:

/plugin marketplace add pawanptu/dev-decide
/plugin install dev-decide@dev-decide

Full guide: setup_deploy.md. Every number above is reproducible with python3 tools/benchmark.py.

🤖 Generated with Claude Code