Skip to content

project-roadmap.md

Latest commit

 

History

History
69 lines (54 loc) · 4.5 KB

project-roadmap.md

File metadata and controls

69 lines (54 loc) · 4.5 KB
Raw

project-roadmap.md

Unscheduled product capabilities that need usage evidence or a separate design before implementation. Concrete selected work belongs in improvements.md.

Conflict detection depth

  • Full semantic conflict taxonomy: Direct violation (for example, an SDK called directly where a Provider Port was decided) and Pattern divergence. This requires AST/import-graph analysis beyond deterministic path/dependency rules.
  • Confidence scoring calibrated against a larger golden fixture set after real conflicts and false positives are available.
  • Architecture fitness-function integrations that can verify richer policies without presenting heuristic guesses as proof.

Harness parity

  • Automate the Codex CLI and Gemini CLI adapters' install-and-run verification — Done (2026-08-31). .github/workflows/test.yml's harness-parity job installs the real Codex CLI and Gemini CLI and runs preflight/init/validate from each one's installed snapshot on every push and pull request.
  • Extend harness-parity coverage beyond preflight/init/validate to check, search, graph, and create once a real regression in one of those commands under a specific harness demonstrates the gap matters.
  • Automate the Antigravity CLI (agy) adapter the same way once it has a package-registry distribution a CI runner can install non-interactively; today it has none, so adapters/antigravity/README.md's manual verification is the only signal.
  • Harness-specific hook support beyond Claude Code SessionStart when equivalent stable extension points exist — Evaluated, not pursued (2026-08-31). The precondition is now true: Codex CLI has a config-driven SessionStart/UserPromptSubmit hook system (~/.codex/hooks.json), and Gemini CLI ships gemini hooks migrate specifically to port Claude Code hooks over. But ADR Toolkit doesn't use a hook even on Claude Code today (it relies entirely on skill auto-discovery), and a hook that fires on every session regardless of relevance cuts against this project's own restraint principle (max 3 questions, judge what's significant, minimize interruption). The plausible use cases (nudge about an unfinished draft ADR, warn about a governed path) are already covered by deliberately invoking discover and check rather than an always-on hook. Revisit only if real usage shows people miss something that discover/check can't catch without a session-start nudge -- not just because the extension points now exist.

ADR navigation and scale

  • Test bulk performance of search and index under 500+ ADRs — Done (2026-09-02). tests/integration/test_bulk_adr_performance.py verifies search and index run in <0.5s over 500 synthetic ADRs without sharding or index breakdown.
  • Improve related-decision discovery beyond path/tag/keyword/body-substring only after real misses demonstrate the need for semantic retrieval.

Internationalization

  • Translate complete project documentation such as README and CONTRIBUTING after non-English contributors can review terminology and maintenance cost.
  • Add Traditional Chinese as zh-TW only through a separate catalog and review; keep zh defined as Simplified Chinese.
  • Evaluate bilingual ADR presentation only if teams need one decision rendered in more than one human language; do not create parallel sources of truth by default.

Public and enterprise governance

  • After the repository becomes public, apply and API-verify branch/tag protections, required checks, conversation resolution, force-push/deletion controls, and a documented bypass policy.
  • Add CODEOWNERS and mandatory independent review when the contributor model can actually satisfy it.
  • Organization-level rulesets, reusable workflows, RBAC, audit export, controlled taxonomy, exception governance, and adoption metrics.

Ecosystem integration

  • Pull-request review integration, initially as a GitHub Action and only later as a GitHub App if permissioned service behavior is justified.
  • C4 / arc42 Section 9 export.
  • Multi-repository decision discovery and graphing.
  • Central decision portal or web viewer.
  • Slack, Jira, and Notion integrations.
  • Vector-backed semantic search after deterministic search has measured misses.

Lifecycle research

  • Revisit whether retrospective should become a first-class status rather than metadata after enough retrospective ADRs exist to show a lifecycle need.
  • Define a narrow factual-correction policy for Accepted ADR metadata without weakening append-only decision history.