Skip to content

Latest commit

 

History

History
51 lines (41 loc) · 2.59 KB

ROADMAP.md

File metadata and controls

51 lines (41 loc) · 2.59 KB

Roadmap

Phased, not dated. Each phase is independently useful and shippable.

Phase 0 — the spec

  • Publish SPEC.md: trajectory, ledger, version formats.
  • Implementation-independent. This is the piece that could go to the AAIF.

Phase 1 — authoring + versioning

  • init, template, edit, render, lint.
  • log, diff, status, commit, revert, tag, blame.
  • No LLM required. Useful and adoptable on day one.

Phase 2 — the loop on one adapter

  • Claude Code capture via Stop hook + transcript_path.
  • learn, pending, promote, reject, savings.
  • The token-savings USP, demoable.

Phase 3 — multi-adapter capture + daemon

  • Project-local SessionEnd connectors: Codex, Claude Code, Cursor, and goose. Complete.
  • Rich provider-specific transcript normalization beyond Claude Code.
  • watch daemon with staleness detection.

Phase 4 — cross-CLI session tracking

  • Give every run a provider-qualified ID.
  • Correlate related runs under a stable logical task ID.
  • Capture session-start and session-end lifecycle events without duplicating completed reflection jobs.
  • Record timestamps, working tree revisions, changed files, final diffs, evaluation outcomes, wall time, model, provider, and available tokens.
  • Provide local sessions and sessions show inspection.
  • Add task-level progress comparisons.
  • Queue reflection outside latency-sensitive hooks with durable job results.
  • Add explicit recovery for interrupted/stale worker locks.
  • Require an evaluation command and confidence threshold for opt-in automatic promotion.
  • Compare equivalent benchmark tasks before and after a promoted rule using success rate, regressions, tokens, and completion time.
  • Keep raw transcripts local by default and support configurable redaction before any external reflector receives a trajectory.

Definition of done: a Codex, Claude Code, Cursor, or goose task can be traced from capture to proposal, promotion, and a later measured outcome without manual bookkeeping. Promotion remains gated by a human or evaluation policy.

Phase 5 — offline GEPA bridge

  • optimize -> gepa.optimize_anything, token cost in the objective.
  • Published before/after results.

Phase 6 — sharing

  • push / pull an AGENTS.md pack (OCI, KitOps-style).
  • Community template hub.

Governance

  • Apache-2.0 (matches AGENTS.md, MCP, goose under the AAIF).
  • Stable adapter + reflector interfaces so others extend without forking.
  • Semantic versioning. Spec is normative; CLI is one implementation.
  • Long-term: propose the spec to the Agentic AI Foundation.