/dr-auto

Overview

/dr-auto is the autonomous-execution meta-command introduced in Datarim v2.19.0+. It activates the eight FB-rules (FB-1 through FB-8 — see the Autonomous Agent Operating Rules mirrored in your workspace AGENTS.md; the machine-readable policy is dev-tools/rules/fb-rules.yaml), the L1 Inline Resolution Rule (close small gaps in-line rather than asking the operator), and the autonomous-ops scope by default through the environment variable DATARIM_AUTO_MODE=1 + the per-task marker file datarim/.auto/<TASK-ID>.mode (the legacy single datarim/.auto-mode-active is still honoured as a fallback).

It works as a subagent orchestrator: for each pipeline stage it spawns the matching Datarim agent (planner, architect, developer, reviewer, compliance) via the Agent tool, summarises that subagent’s result, and decides the next stage. The run stops at a successful /dr-compliance — which now writes the reflection internally — and does not run the final /dr-archive; archival stays an explicit operator step. At the terminal step the orchestrator prints one line stating how compliance resolved (the verdict, or a skip-by-design reason at complexity levels with no compliance stage) before the call-to-action, so the compliance pass is visible rather than implied. Re-entering an already-completed stage (after a review finding routes work back) is normal and updates that stage’s artefact rather than failing. Every dispatched stage runs under a closed evidence loop: the orchestrator advances only when the strict live-evidence check passes against the task’s acceptance contract; missing or stale evidence is never treated as green.

Before spawning any stage subagent, the orchestrator re-asserts the auto-mode marker (a mandatory pre-dispatch gate) so a stale or mismatched marker never silently carries into a spawned agent. Narrow L1 documentation-only changes (single markdown file, no structural split, no runtime behaviour) skip the full /dr-qa stage in favour of a lightweight inline style and cross-reference check, recorded as a qa-stub artefact that still counts as the QA gate for that class.

Two modes

• Continue mode — /dr-auto {TASK-ID} resumes a known task from its last checkpoint and runs it end-to-end through whatever pipeline stages remain.
• Bootstrap mode — /dr-auto "<free-text brief>" kicks off a fresh task from a free-text operator brief, runs /dr-init then routes through the appropriate pipeline (L1 through L4) without further per-step prompts.

Usage

# Continue: resume the named task from its last checkpoint
/dr-auto <TASK-ID>

# Bootstrap: kick off a fresh L1..L4 task end-to-end
/dr-auto "Bump skill count on home page after <TASK-ID> migration"

Canonical contract

The full contract — Ladder definitions, FB-rule enforcement points, hard-gated action list, the L1 Inline Resolution Rule decision tree, and the auto-inline-log schema — lives in skills/autonomous-mode/SKILL.md. Stage-specific suppression hooks are documented per-command (for example, /dr-do TDD red→green transitions resolve through Ladder L1 before L5).

Reversible vs irreversible

Reversible actions proceed automatically (file edits, local test runs, branch creation, Git commits, push to a feature branch, internal documentation updates). Before asking, the orchestrator resolves each action kind against the space’s autonomy.policy (resolve-space-autonomy.sh gate): operational actions such as production deploys, secret rotation or public messages ask the operator when the policy says operator or cannot be resolved. The always-gated floor never executes automatically in any space: finance or legal actions, irreversible database operations without a verified backup, git-history deletion, force-pushes that drop commits, and Supreme Directive violations.

Reporting the result

Apply Human Outcome Reporting before a human-facing update or result. Explain the user-visible outcome, the original conditions, checks and evidence, and open limitations. A stage result does not mean the entire task is complete. Keep exact protocols and technical records intact. To explain again without continuing work, use /dr-explain.