autoresearch

Standalone AIDE-style autoresearch agent: tree-search over candidate solutions (draft/debug/improve operators) for any task directory with a task.md and score.py — ML engineering, heuristic optimization, or harness engineering. Runs the current best generation evolved by the rsi-loop outer loop (falling back to baseline/gen-000). Use when the user wants to solve an optimization task by iterative search, or invokes /rsi:autoresearch.

Aller à l'installation

Dépôt: cajias/claude-skills
Dernière activité de la source: 21 juillet 2026 à 10:21
Langue détectée de SKILL.md: anglais
Étoiles: 0
Forks: 0

Options d'installation

Le prompt qui vérifie d'abord la source est sélectionné par défaut. Vous pouvez passer à une commande directe ou télécharger une copie locale.

Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.

Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.

npx skills add https://github.com/cajias/claude-skills --skill autoresearch

La commande reste sur une seule ligne. Faites défiler horizontalement pour la vérifier avant de la copier.

Vous préférez une copie locale ? Téléchargez les fichiers actuellement disponibles dans SkillsMP.

Vérifiez les fichiers source

Lisez SKILL.md et les fichiers associés affichés par SkillsMP avant de décider de l'installer.

SKILL.md

Instructions source · Aperçu en lecture seule

Retour à la recherche Retour à l’installation

milestone-goal-prompt

Research (as an internal step) a GitHub/GitLab milestone's open issues in order to PRINT a copy-paste autonomous-loop prompt that drives the milestone to completion — behavior-driven (BDD) scenarios, adversarial gap-checking, a per-iteration Definition-of-Done gate (build + all tests + zero lint + /code-review + /security-audit + /ponytail:ponytail findings all cleared), specialized-agent selection with model tier scaled to task complexity, and a root-cause→harness-hardening loop. Use this whenever the user says "generate a goal prompt", "goal-driven prompt", "milestone loop prompt", "prompt to finish/complete the milestone", "clear-and-paste prompt for milestone N", "make me a prompt to drive /autoresearch", or otherwise wants a ready-to-paste block that autonomously completes a milestone's remaining issues. The deliverable is the prompt itself, not a summary. NOT for actually executing the work (this only GENERATES and PRINTS the prompt — the loop does the work), NOT for merely summarizing or listing a mile

2026-08-13

apm-marketplace-authoring-gotchas

Eight APM-marketplace-authoring gotchas discovered during real migrations from hand-authored .claude-plugin/marketplace.json to APM-managed and from CC-native plugin authoring. Use when: (1) running APM's packaging step (with or without `--dry-run`) or its marketplace-check command in a repo with an `apm.yml`; (2) wiring CI for an APM marketplace; (3) the CI staleness check (repack, then `git diff --quiet`) fails despite a clean local; (4) prettier reformats `marketplace.json` after APM's packer writes it; (5) APM exit code 1 on local-path-only marketplaces; (6) confused by `packages:` in apm.yml vs `plugins:` in marketplace.json; (7) commitlint rejects a commit with `footer-leading-blank` despite no apparent footer; (8) `apm compile -t claude` exits 0 with zero output and no compiled plugin appears; (9) husky+lint-staged pre-commit blocks the FIRST marketplace commit with markdownlint `MD040/fenced-code-language` errors on a SKILL.md file copied verbatim from `~/.claude/skills/`. The gotchas form a chain — f

2026-07-21

apm-plugin-distribution-paths

Canonical authoring layout for an APM (microsoft/apm) package, plus the decision matrix for whether you need a `marketplace:` block. Use when: (1) you've run `apm init` and are unsure where SKILL.md / agent .md / command .md files belong (answer: under `.apm/`, NOT at repo root); (2) you've authored CC-native files at `commands/agents/skills/` at repo root and wonder if that's "right" (it works as fallback layout #3 "Claude plugin", but it's not the canonical authoring path); (3) you're trying to choose between the three layouts APM recognizes (one-skill repo, multi-primitive `.apm/`, Claude plugin); (4) `apm targets` fails with "No such command" in CLI 0.12.1 (subcommand doesn't exist); (5) `apm validate` fails with "No such command" in CLI 0.12.1 (use APM's packaging command in dry-run mode instead); (6) wondering whether `apm compile -t claude` will generate output for a hand-authored CC-native plugin (it won't; it's a no-op for layout #3); (7) wondering if you need a `marketplace:` block in apm.yml (you d

2026-07-21

Tout voir dans ce dépôt

name

autoresearch

description

autoresearch

Runs the rsi-loop inner agent standalone: point it at a task directory containing task.md, score.py, and public/, and it executes tree-search autoresearch (parallel drafts → greedy debug/improve loop) under a token budget, returning the best-scoring solution.

Follow the procedure in commands/rsi-autoresearch.md: sandbox from public materials only (scripts/rsi-sandbox.sh), launch the generation's inner-agent.workflow.mjs with {sandbox, genDir, taskName, seed, policy}, report the best node's public score. Private scoring is outer-loop-only (RSI_OUTER_LOOP=1) and never enters the inner agent's context.

The generation is resolved through the run's best pointer (falling back to baseline/gen-000), so improvements the outer loop discovers (/rsi:step) flow to standalone users automatically. First verified run: see docs/experiments/m1-smoke-bin-packing.md.

Design details: docs/PLAN.md, §3 "autoresearch as a standalone skill".

Naming note: this is distinct from the third-party uditgoenka/autoresearch skill (a single-level metric hill-climb loop, evaluated in PLAN.md §5.1 as a candidate outer-loop chassis). The §5.2 experiment resolved in favor of the native chassis (Arm B, /rsi:step / /rsi:run); the third-party chassis was not adopted, so no rename is needed and this skill keeps its name. See docs/experiments/chassis-ab/.

autoresearch

Informations de source

Options d'installation

Vérifiez les fichiers source

Plus depuis ce dépôt

Plus depuis ce dépôt

autoresearch

autoresearch