The autonomous-mode Skill — Running Long Tasks Without Constant Check-ins

Long tasks have an interruption problem. The agent pauses to ask whether a config value is correct, then pauses again to confirm a file path, then again to pick between two approaches that are obviously equivalent. Each pause breaks flow and burns time. The autonomous-mode skill addresses this by giving the agent a structured procedure for answering its own questions — and a hard stop for the small class of actions that genuinely require human approval.

What it does

The skill activates when three conditions are all true: the environment variable DATARIM_AUTO_MODE=1 is set, the marker file datarim/.auto-mode-active exists and parses as valid YAML, and the task_id field inside that marker matches the current task ID. The marker expires after 24 hours, so a forgotten environment variable from a previous session cannot silently activate the mode.

Once active, the agent must walk a five-level Question Suppression Ladder before making any operator prompt. Level 1 is a codebase grep or file read — the cheapest resolution, under one second. Level 2 is a runtime probe such as curl localhost/healthz or git log -1. Level 3 is a lookup in the operator's MEMORY.md feedback notes. Level 4 delegates the question to an external LLM via coworker ask when the answer requires reading many large files. Only if all four levels fail to give an unambiguous answer does the agent escalate to Level 5 — the operator prompt.

Parallel to the ladder, the L1 Inline Resolution Rule governs small gaps discovered mid-task. A fix that touches one file, changes at most 50 lines, and does not alter any API contract can be applied inline without operator approval. Larger changes go to the backlog. Each inline fix is logged to an append-only file that surfaces in the archive step.

One concrete example

During a long /dr-do run, the agent needs to know which port the database is listening on. Under the Question Suppression Ladder, Level 1 applies: grep the project's config files for the port setting. If found unambiguously, proceed. If two config files disagree, escalate to Level 2: probe the running service directly. The operator is never asked — the answer was available in the codebase or runtime the whole time.

When it loads

The skill is invoked by /dr-auto, which sets the environment variable and writes the marker. Every pipeline command — /dr-init through /dr-archive — checks for the active marker on startup and adjusts its behavior accordingly. Hard-gated actions (production deploys, secret rotation, public posts, force-push to main) never execute automatically regardless of mode; the ladder always escalates those to the operator at Level 5.

See how this fits the larger workflow in what Datarim is, or read the /dr-auto command reference for the activation contract.