Milestone 1: Autonomous engineering loop

The first milestone turns RalphTerm from a PTY session API into a complete local workflow engine for software tasks.

Definition of done

A user submits one task and receives a final patch, summary, transcripts, and event log.
The run uses terminal-native agent sessions, not one-shot prompt shortcuts.
Implementation, validation, and independent review are separate and observable.
Approval requests are visible, auditable, and manual by default.
The dashboard shows live output and final artifacts.

Phases

intake -> workspace -> plan -> implement -> validate -> independent-review -> accept/commit

The product boundary is the acceptance gate: implement -> validate -> independent-review -> accept/commit. An implementation agent finishing is not enough; the task only counts after validation succeeds and the reviewer prints REVIEW_PASS.

Core deliverables

Run model

Run, phase, workspace, session, artifact, and approval state.

Workspace isolation

Per-run worktrees with base commit tracking.

Agent drivers

Real Claude and Codex interactive sessions through PTYs.

Dashboard

Run list, live terminal output, phase timeline, transcripts, and artifacts.

Next implementation blocks

Real PTY resize.
ANSI normalization while preserving raw transcript.
Persistent run store.
Claude and Codex readiness/login/approval state detection.
Run-level orchestration on top of sessions.