v0.3.1

Highlights

Full-breadth model routing with a frontier escalation ladder. The cost-aware selector now ships a default capability ladder grounded in the live OpenRouter catalog (harvested-free / cheap → mid → frontier) and routes each task by quality-per-dollar across the full breadth of models. The strongest tier is reserved for the final attempt only — a hard task starts mid-ceiling and escalates to a frontier model as the safety net, not on the first try. The gate quality floor is unchanged, so routing only ever moves cost/latency, never shipped quality. Verified live: a task cheaper tiers stalled on was resolved by the frontier tier on its final attempt, for pennies.

Reproduction-first with pre-patch validation. The spec-as-tests workflow runs its generated reproduction test on the clean tree first and keeps it only if it actually fails there — a test that reproduces nothing is a false gate and is discarded. Engaged automatically on both SWE-bench execution paths (host and containerized), where the judged tests are hidden.

Two-timescale evolution — self-authored tool library (consolidation layer). A new ToolLibrary over the existing MAP-Elites + held-out promotion machinery: a runtime-invented tool is admitted to a persistent, best-per-capability library only if it passes the same held-out generalization gate that guards scaffold self-edits, and promoted tools seed future runs. Capability compounds across runs instead of being reinvented each task. Design: docs/two-timescale-evolution.md.

Test suite: 1,897 passing. Full changelog in CHANGELOG.md.

v0.2.2

Changed

• Enriched the MCP tool definitions for tool-definition quality: every parameter
  now carries a concrete example and constraints, each tool documents when to
  use it (and when not, and related tools), and titles describe the action with
  an explicit side-effects note.

v0.2.1

Fixed

• Use an absolute logo URL in the README so it renders on the PyPI project page
  (relative image paths only resolve on GitHub).

v0.2.0

Added

• SWE-bench evaluation harness (evaluation/swebench/) — run misterdev on
  real GitHub-issue tasks and grade the patch against the task's own hidden
  tests. Includes a Docker runner that executes the build/test gates inside each
  instance's official image, so a task runs in its exact environment.
• Complete reference sites — every external call site of a symbol being
  edited is surfaced up front, so a delete/rename/refactor updates them all in
  one attempt instead of chasing missed callers one build-error at a time.
• Dangling-reference gate — a deterministic pre-build check that rejects an
  edit which removes or renames a symbol while leaving references to it.
• Prompt caching — the stable context prefix is marked cacheable (Claude),
  so a task's retries re-read it at a fraction of the input cost; cache reads are
  priced accordingly in the budget and ledger.
• Smarter model ledger — hard-avoids models proven incompetent on a task
  cell, skips models proven too slow, warm-starts a cold cell from a model's
  global record (empirical-Bayes shrinkage), and reserves free models for the
  easiest tasks.
• Adversarial critic now also checks for symptom-vs-root-cause fixes and
  code duplication (DRY), and auto-enables for refactor/fix/integration tasks.

Fixed

• The symbol graph now refreshes after each task instead of going stale for the
  rest of the run.
• A completed task's status: completed is committed into its source markdown,
  so a finished devplan is no longer re-run.
• An acceptance-command manifest error no longer false-fails a task whose real
  gates already passed.
• A reverted task's untracked orphan files are cleaned up (bounded to files the
  task created).
• Multi-file edits apply atomically, rolling back on a mid-batch write failure.
• An out-of-credits (HTTP 402) response halts the run gracefully instead of
  crashing with a stack trace.

misterdev v0.1.0

First public release.

misterdev is an autonomous, extensible LLM build orchestrator: point it at a repo and a goal, and it plans, edits, and verifies code across languages, merging only changes that keep the build green.

Highlights:

• Autonomous polyglot build loop with correctness gates (build/test/lint/typecheck + optional critic/mutation/runtime/web/vision) and git rollback on regression.
• Dynamic, cost-aware model selection over OpenRouter/Anthropic with failover; parallel execution in isolated git worktrees.
• Pluggable tools, gates, and targets via Python entry points — pip install misterdev-plugin-x adds capability with zero core edits.
• Agentic MCP tool use, including remote gateways (e.g. Glama) with a tool allowlist.
• misterdev as an MCP server (misterdev-mcp) and a natural-language CLI — drive it in plain English, no flags to memorize.

Docs: see the README and docs/. Dual-licensed AGPL-3.0-or-later + commercial.