Holistic reviewer overhaul and context management improvements.
New:
holistic-reviewskill that routes tocodex:rescuefor dueling-model review when available, falling back to a direct subagent- Multi-persona holistic reviewer: panel of staff engineers (systems, distributed systems, security, infra ops, product) that discuss findings from different perspectives
- Tests-first review methodology: reviewer reads tests before implementation code to assess coverage and falsify assumptions
- Mandatory context compaction after each step's code review — prunes context no longer needed for remaining work
Changed:
- Implementation steps are now ordered for optimal context management (foundational pieces first, independent work early)
- Holistic reviewer now reviews across seven dimensions (added test coverage) instead of six
- Removed redundant final review from
execute-one-shotskill — holistic review is now handled by theholistic-reviewskill invoked from the command - PR description quality is now explicitly enforced with troubleshooting guidance for common errors
Fixed:
- Formatting cleanup in
autoresearch-createskill (whitespace, JSON readability)
Fixed:
- Minor post-release fixes
Fixed:
- Minor post-release fixes
Use codex for holistic reviewer
New:
- Uses the
/codex:rescuecommand along with theholistic-revieweragent to get a true clean room review via dueling models
Initial release of the blank-slate-review plugin.
New:
/blank-slate-review:startcommand -- validates optional scope and launches the review pipelineexecute-reviewskill -- orchestrates scout, file sampling, fan-out decision, and reviewer dispatchcodebase-revieweragent -- Opus-powered multi-persona panel with seven analysis dimensions, four severity levels, and holistic remediation synthesis
Initial release of the quick-plan plugin.
New:
/quick-plan:startcommand -- creates a branch, picks a plan file location, and launches the planning skilldevise-a-planskill -- structured 9-step planning process with specialist subagent dispatchsystems-performanceagent -- evaluates approaches for hot path efficiency, allocation pressure, and resource usagedistributed-systemsagent -- investigates codebase structure and evaluates cross-boundary interactionssecurity-engineeragent -- evaluates threat surface changes, trust boundaries, and escalation pathsinfra-operationsagent -- evaluates deployment impact, rollback safety, and observability gapsproduct-leadagent -- evaluates strategic fit, scope, and future flexibility
New:
/autoresearchcommand - start or resume experiment sessions with optional duration limit (/autoresearch 60for a 60-minute session)maxDurationMinutesconfig field - wall clock time limit for sessions, useful as a plan quota budget proxy- Graceful shutdown procedure - clean state commit and summary when time/iteration limits are reached
plugin.jsonfor proper plugin registration
Changed:
- Loop rules now check stop conditions (time, iterations) before each experiment
- "What's Been Tried" section in autoresearch.md gets updated on shutdown
New:
howto-qmdskill - local semantic search for markdown knowledge bases with keyword, vector, and hybrid search modes
Initial release of the tooling plugin.
New:
howto-ast-grepskill - structural code search and rewriting with ast-grep CLI patterns, metavariable syntax, and language-specific examples