autoresearch

v2.2.3

Eval-driven improvement, plus the harness to drive it

Two parts form one loop. An edit → eval → keep/discard engine improves code, prompts, or docs. It scores each change with a shell command, an LLM judge, or both. It shows progress in a live Artifact dashboard. A harness builder measures project health in six categories. It creates feedback loops, evals, sensors, and context management advisories. It then fixes the top-ranked issue through the same loop.

Install

$ claude plugin install autoresearch@21-breakincode

Commands

/autoresearch:harness-build Menu-driven scaffolder for harness components: feedback loop, eval loop, sensor, or context-mgmt advisory. Writes Tier-1 artifacts into your project's .claude/.
/autoresearch:harness-check Scan project health across code quality, tests, runtime, architecture, scriptability, and harness completeness. Produces a scored harness report with impact-ranked improvements
/autoresearch:harness-improvement Execute improvement loop on the top-ranked harness issue: auto-generates eval from probes and spawns the autoresearch:experimenter agent
/autoresearch:improve Iteratively improve any artifact using an edit-eval-keep/discard loop with live dashboard

Changelog

2.2.3 2026-09-24

fix state the eval-metrics guard in /autoresearch:improve once, with its reason, instead of five times with rising emphasis.

2.2.2 2026-09-19

feat publish the live improvement dashboard as a private Claude Artifact. The dashboard now uses inline SVG and no external assets.

2.2.1 2026-09-19

fix rewrite prose across plugin files to clear the simple-english linter. No content or meaning changed, only sentence shape. "harness" stays: standard term for this plugin.

2.2.0 2026-08-10

feat keep/discard no longer requires git. The experiment loop now snapshots target files into .autoresearch/snapshot/ instead of committing and checking out, so /autoresearch:improve runs against any directory, not just a git repo.
fix stops polluting real repos with per-iteration commits. Nothing read that history. The dashboard sources reasoning and diff_summary from experiments.json.
breaking iterations no longer record a commit SHA in experiments.json (iteration number already identifies them).
test tests/test_snapshot.sh: 12 assertions, including that a discard reverts to the last kept state rather than the baseline.

2.1.0 2026-08-04

refactor collapse experiment-loop Rules into steps
refactor source improve.md libs via ${CLAUDE_PLUGIN_ROOT}

2.0.0 2026-06-28

feat merge harness plugin into autoresearch (single plugin)

1.2.2 2026-06-27

feat glass + depth dashboard redesign

1.2.1 2026-06-27

feat redesign eval dashboard (Linear/Vercel-grade visuals + motion)

1.2.0 2026-06-07

refactor move harness commands and probes to standalone harness plugin
refactor convert harness-check and harness-improvement to deprecation shims

1.1.1 2026-06-06

fix infer metric direction, fix improvement %

1.0.0 2026-05-28

feat initial release of autoresearch plugin