Skip to content

ROADMAP.md

Latest commit

 

History

History
84 lines (68 loc) · 4.55 KB

ROADMAP.md

File metadata and controls

84 lines (68 loc) · 4.55 KB
Raw

Roadmap

v0.1, repository foundation

  • Define public and private artifact boundaries.
  • Add run metadata, workflow states, command capture, timeouts, redaction, and local health checks.
  • Add a provisional engineering skill and provenance format.
  • Add tests and continuous integration.
  • Run the first Codex-to-Claude task against a disposable fixture repository.

v0.2, first complete local loop

  • Add configurable Codex and Claude CLI adapters behind EngineeringAgent.
  • Prove the Codex adapter against a disposable failing fixture.
  • Clone a pinned target commit into an isolated private workspace.
  • Give the primary agent the issue and repository instructions.
  • Pass repository state and machine-captured evidence to the reviewer.
  • Allow one bounded revision cycle. Live evidence recorded 2026-09-02.
  • Re-run verification and stop at READY_FOR_HUMAN_REVIEW.

The acceptance test is one recorded run in which two real agent processes complete the loop. Mocks alone do not count. Met on 2026-09-02 by private run 20260901T201921Z-0b85ed, recorded in docs/runs/0003-two-model-fixture.md.

v0.3, the knowledge flywheel

  • Decide the flywheel architecture, artifacts, and promotion gates.
  • Add the failure and success taxonomy and the weighted learning channels.
  • Add the retrospective schema, template, and mailman retrospective.
  • Add the lesson registry with gated states and rule provenance.
  • Ingest a completed retrospective back into its JSON record.
  • Version the engineering skill and record that version at init-run.
  • Build the skill regression suite over historical issues.
  • Aggregate categories across runs.

The acceptance test is one rule promoted into SKILL.md through the registry, with the evidence, the validation, and the provenance that the gates require. Not met. No lesson has entered the registry yet.

v0.4, first real external issue

  • Capture a GitHub issue into the run record without leaking its accepted fix.
  • Build primary and reviewer prompts from the captured issue.
  • Install a target repository's dependencies outside its working tree.
  • Export a reviewable patch, summary, and draft pull request text.
  • Carry one live two-model run on an external repository with real dependencies.
  • Record live evidence for the revision cycle and at least one blocking path.

The acceptance test is one recorded run against a repository Mailman does not own, from issue URL to exported patch, with no hand-written prompt. Met on 2026-09-02 by private run 20260902T051904Z-f0cd07 against ayukhno/autosound-tcc issue #4, recorded in docs/runs/0004-first-external-issue.md. The revision cycle, the second review, and a BLOCKED ending were recorded live the same day in docs/runs/0005-revision-and-blocking-paths.md.

v0.5, a pull request that could be merged

  • Record each target project's contribution and AI policy as data.
  • Gate a submission on diff hygiene, so noise never reaches a maintainer.
  • Draft policy-shaped pull request text and a human accountability brief.
  • Record a duplicate search of a target's open and closed pull requests.
  • Put earlier attempts, and why they were rejected, into both agent prompts.
  • Prove the fix fails without its source change, as a recorded command. Met 2026-09-03 by private run 20260903T194455Z-140c59, whose reproduce record shows the defect at the base commit and nothing after.
  • Carry one live run on a permissive, recognizable target to a ready submission. Met 2026-09-03 by private run 20260903T052426Z-ad8196 against encode/starlette, recorded in docs/runs/0008-starlette-3497-submission-ready.md. The first of these submissions to reach a merge is pmorissette/ffn #330 on 2026-09-06, from run 20260903T194455Z-140c59, recorded in docs/runs/0009-ffn-330-first-merge.md.

The acceptance test is a submission package a maintainer of that project would recognize as following their own written rules, prepared without contacting them. Opening the pull request stays manual and needs explicit human approval.

Later, only after real-run evidence

  • Sanitized public run export.
  • Historical issue replay without accepted-fix leakage.
  • Usage and cost normalization across adapters.
  • GitHub pull request preparation after human approval.
  • Comparison reports across agents and prompt versions.

Dashboards, autonomous issue selection, vector databases, and hosted orchestration are out of scope until recorded runs show a concrete need.