Skip to content

ROADMAP.md

Latest commit

 

History

History
102 lines (68 loc) · 3.17 KB

ROADMAP.md

File metadata and controls

102 lines (68 loc) · 3.17 KB
Raw

Reflect Roadmap

git answers "what changed." reflect answers "why."

Current State (v4)

  • A Python CLI + Claude Code skill
  • Reads from Entire CLI + git on demand — zero intermediate storage
  • Replaceable harness generates context briefings
  • Two read paths: passive (context.md) and active (why/search dump raw evidence)
  • Session-start hook with freshness tracking

What works: The read path architecture is clean. Harness is replaceable. What's missing: Multi-tool wiring, community harnesses, team workflows.

Phase 1: Universal Read Path

Goal: Any AI coding tool gets the briefing automatically.

Auto-wire context.md into every instruction file format:

File Tool
CLAUDE.md Claude Code (done)
.cursor/rules/*.mdc Cursor
.github/copilot-instructions.md GitHub Copilot
.windsurfrules Windsurf
.clinerules Cline

Phase 2: Harness Ecosystem

Goal: Community-contributed harnesses for different use cases.

  • harness/default.py — recency-ranked, ships with reflect
  • harness/semantic.py — embedding-based retrieval
  • harness/summarizer.py — LLM-powered session summarization
  • harness/legacy-v3.py — implements the old confidence/decay model

Harness benchmarking: run two harnesses against the same evidence, compare context quality. This is the Meta-Harness optimization loop made accessible.

Phase 3: Trust & Provenance

Goal: Safe multi-source evidence ingestion.

When reflect adds adapters for PRs, CI logs, Slack threads — trust matters:

Source Trust Level
Entire session (human-in-the-loop) verified
Git commits inferred
PR descriptions inferred

Trust boundaries prevent low-quality evidence from flowing into instruction files. Deferred until multi-source adapters exist.

Phase 4: Team Workflows

Goal: The "why" is shared across the team.

  • GitHub Action on PR merge → captures decisions from PR description
  • PR review bot → surfaces relevant evidence for changed files
  • reflect onboard → generates comprehensive briefing for new contributors

Phase 5: Harness Optimization

Goal: The harness improves itself.

The Meta-Harness paper shows that searching over harness implementations outperforms hand-designed ones. With reflect's architecture:

  1. Evidence filesystem exists (Entire + git)
  2. Harness is a replaceable script
  3. Evaluation signal exists (session outcomes, retry counts)
  4. A proposer agent can rewrite the harness and measure results

This is the long game: reflect becomes the platform for automated harness engineering.

Priority Matrix

Priority Item Effort Impact
P0 Multi-tool instruction file wiring Small Every AI tool benefits
P1 Harness ecosystem + benchmarking Medium Community contribution
P1 Additional evidence adapters (PRs, CI) Medium Richer evidence
P2 Trust & provenance model Small Enables safe multi-source
P2 Team workflows (GitHub Action, bot) Medium Team-level memory
P3 Harness optimization loop Large Self-improving memory