uditgoenka__autoresearch/docs/development-roadmap.md

Development Roadmap

Strategic milestones and feature development phases for autoresearch.

Completed Phases

Phase 1: Core Skill Framework

• Base skill architecture with SKILL.md pattern
• Multi-agent orchestration capabilities
• Reference documentation system
• Scenario-based workflow definitions

Phase 2: Multi-Subcommand Pattern

• debug, fix, security, ship, scenario subcommands
• Chain integration between subcommands via handoff.json
• Interactive AskUserQuestion setup gates

Phase 3: Advanced Analysis Capabilities

• Codebase analysis engine
• Dependency mapping and component clustering
• Git integration for staleness detection

Phase 4: Predict Swarm Intelligence — 2026-03-18

• Multi-persona swarm prediction (5 default personas)
• Sequential debate protocol with anti-herd detection
• Adversarial Red/Blue team mode
• Budget enforcement and git-hash stamping

Phase 5: Learn + Probe Subcommands — 2026-03-21 / 2026-04-16

• /autoresearch:learn autonomous documentation engine (v1.8.0)
• 4 modes: init, update, check, summarize
• Diff-based targeting for update mode
• /autoresearch:probe requirement interrogation engine (v1.10.0)
• 8 personas, saturation-based termination

Phase 6: Multi-Platform GA — 2026-04-28

• OpenCode + Codex distributions (v2.0.0)
• YAML strict compliance across all SKILL.md files
• Security-hardened install and sync scripts
• 11 subcommands, unified version references

Phase 7: Modular Rebuild + Evals — 2026-05-22 (v2.1.0) — COMPLETED

• Thin SKILL.md routing table (41 lines, was 813 lines)
• 12 self-contained command files (94–120 lines each)
• 3 focused reference files (was 13)
• /autoresearch:evals new subcommand for TSV analysis
• scripts/transform.sh replaces sync-opencode.sh + sync-codex.sh
• Removed Python wrapper CLI and autoresearch-command-spec.json
• TSV # metric_direction comment for auto-detection
• 8 TSV status values (added keep (reworked), hook-blocked, metric-error)
• --evals / --evals-interval flags on all looping commands
• ~95% token reduction per invocation

Phase 8: Hook System + Product Improvement — 2026-05-22 (v2.1.1 / v2.2.0)

• 9-hook safety system (v2.1.1)
• /autoresearch:improve — product improvement research + PRD generation (v2.2.0)
  • 5 research categories with saturation-based termination
  • ICP-aligned tiered ranking (Must-have / Nice-to-have / Moonshot)
  • Per-feature PRD generation with evidence chains
  • Conditional auto-discover for zero-context codebases
  • Terminal emitter — outputs PRDs for external tools, not autoresearch re-entry
• Subcommand count: 12 → 13

Current Phase

Phase 9: Regression Stability Gate + Stabilization — 2026-06-19 (v2.1.4)

• /autoresearch:regression — layered stability gate, 14th family member
  • 8 dimensions, tiered HARD/SCORE verdict, green→red classification invariant
  • git-worktree baseline cache (SHA-keyed), statistical perf gate (Mann–Whitney U), hard-guarded forward-only data-migration
  • scripts/score-regression.sh (rubric + verdict) + tests/test-regression.sh (50 assertions, 10 golden fixtures)
  • Subcommand count: 13 → 14
• Guide updates reflecting v2.1.4 command file structure
• Scenario guide coverage for evals, improve, and chain workflows
• Community contributions and bug reports

Future Phases

Phase 10: Extended Integration

• GitHub Actions workflow templates
• CI/CD pipeline integration examples
• Web dashboard for TSV result visualization

Phase 11: Advanced Features

• Custom persona templates for predict and probe
• Domain-specific analysis profiles
• Multi-repo analysis support
• Historical trend analysis across runs

Success Metrics

• Token cost per invocation: <10K (achieved in v2.1.0 for typical use)
• Zero external runtime dependencies (achieved)
• All platforms supported from single canonical source (achieved in v2.1.0)
• Graceful degradation under bounded iteration limits (achieved)