uditgoenka/autoresearch v2.2.0

v2.2.0 — Autonomous Goal-directed Orchestrator

on GitHub

Bare /autoresearch becomes a complete autonomous orchestrator — state a plain-language goal and it self-selects the subcommands, flags, and iteration counts needed to reach it, the way /ck:cook orchestrates implementation.

Added

• Orchestrator mode on bare /autoresearch — classifies a free-form goal, derives a concrete success predicate (exact command + expected output), confirms once, then loops the right subcommands until the predicate holds.
  • Dispatch: Metric:/Verify: present → classic metric loop (unchanged); free-form goal → orchestrator; nothing → setup wizard; --classic/--auto force the mode; mode printed in a banner.
  • 9 goal archetypes in two modes — Orchestration loop (ship-ready, optimize-metric, fix-broken, harden, build-feature, explore) loops until a mechanical predicate is met; Single-pass dispatch (document, what-to-build, decide-design) routes once to learn / improve / reason and self-terminates.
  • Bounded termination: plateau detection (5 cycles no net progress) + hard ceiling (50, override --max-cycles N); repeated unknown-units cycles route to BLOCKED.
  • Safety: never auto-approves ship/deploy/push; every derived command is safety-screened (and re-screened on resume); data-migration stays behind the anchored DB-URL allowlist.
• scripts/orchestrate.sh — deterministic routing seam (classify, next-hop, units, plateau, screen-cmd, verdict), mirroring scripts/score-regression.sh.
• tests/test-orchestrator.sh (97 assertions) + fixtures; references/orchestrator-routing.md; guide/autoresearch-orchestrator.md.

Changed

• Bare /autoresearch dispatch is now mode-aware; classic Metric-loop behavior is unchanged when Metric:/Verify: are supplied.
• Version 2.1.4 → 2.2.0 across all 3 plugin manifests and 5 SKILL.md mirrors. Command count stays 14 — the orchestrator overloads the root command, it is not a new subcommand.

Fixed

• screen-cmd hardened against two command-screening bypasses: path-qualified rm (/bin/rm, ./rm) and curl/wget piped to alt-shells/interpreters; parser pipes (curl | jq/grep) remain allowed.
• scripts/transform.sh — latent :regression colon-drift in the Codex/OpenCode adapters.

Gates: tests 252/252 (orchestrator 97, regression 50, hooks 105) · /autoresearch:regression STABLE (negative-control verified) · security PASS.

Full changelog: v2.1.4...v2.2.0