Skip to content

ROADMAP.md

Latest commit

 

History

History
63 lines (52 loc) · 2.99 KB

ROADMAP.md

File metadata and controls

63 lines (52 loc) · 2.99 KB
Raw

Roadmap

📌 Sprachhinweis / Language Note: Englisch für technische Inhalte; Release-Planung wird von KI-Workern verarbeitet. English for technical content; release planning is processed by AI workers.

This document captures the long-term direction for the ai-issue-solver project. It is intentionally lean and strategic; concrete tasks live in BACKLOG/open.md.

Themes

  • Information architecture — every agent only gets the context it needs. See Release 0.7.0 planning discussion.
  • Agent-specific context routing — Planner gets more, Solver gets less, Reviewer gets the diff + rules, Watchdog gets cost + status.
  • Cost & process control over model benchmarking. See #16, #37, #38, #39.
  • Provider abstraction — GitHub-first, OpenRouter, native APIs as fallback. See REPO_PROFILE_PROVIDER.md.
  • Model override policy — model choice is explicit and overridable per run, with defaults kept in role/provider configuration. See MODEL_OVERRIDE_POLICY.md.
  • Knowledge lifecycle — every piece of information either gets kept / promoted / archived / deleted. Nothing lingers.

Release Anchors

  • 0.6.0 done — skills + rework + batch cost forwarding.
  • 0.7.0 done — Information architecture, role-based context routing, release-review process. Close-out: RELEASE_REVIEW_0.7.0.md. Release notes: RELEASE_NOTES_0.7.0.md.
  • 0.8.0 skipped — planned as Governance & Knowledge Release, but the scope (Handover Audit + Reviewer Runtime + Knowledge Dry Run) had no independent identity once meta-planning began. Pivoted.
  • 0.9.0 done — Solver Validation Release. It shipped validation infrastructure, PR rework, backward-split handling, hard stops for partial/reject patch failures, dynamic OpenCode/OpenRouter free-model discovery, and corrected benchmark classification from run reports. Close-out is recorded in CHANGELOG.md; the original plan remains in PLANNING_0.9.0.md.
  • 1.0.0 target — Runnable workflow app for AI-assisted repository work. The app coordinates transparent, reviewable steps around one repository: triage, planning, model selection, solver execution, observation, review, rework, recovery, merge preparation, and reporting. Product vision: PRODUCT_VISION_1.0.md.

Strategic Items (not yet in Backlog)

These need decomposition into BACKLOG/open.md items before they become executable:

  • Knowledge Manager as deterministic workflow (not LLM agent).
  • Watchdog as cron-driven skill (not LLM agent).
  • Per-role model routing as a config file (config/role_routing.yaml).
  • Documentation lifecycle automation (archive / promote scripts).

Out of Scope

  • New solver features before the documentation & context story is settled.
  • Model benchmarking without a cost-control baseline in place.