git:20261102.feb8b13 to git:20261103.fb62564

---

name: agent-compass

description: Choose owner, model and effort, outbound authorization, advisor use, and continuation for delegated work. Reads local usage and effective account policy, profiles the task with the configured System-One scorer, then allocates a compatible worker. Use for independent worker or advisor work; ordinary local execution needs no routing.

---

# AgentCompass

## Execution ownership

The main session executes serial work directly by default, including ordinary investigation, edits and verification. Duration, tool-call count and reading multiple files are not reasons to delegate. Delegate when independent work benefits from parallelism, isolation of a large context, a specialist capability or quota offloading. The main session is the only coordinator. Before writing any brief, list the separate tasks in the work (one deliverable each, with what it waits on): keep the tasks coupled to the conversation, write one brief per independent task, and roll them together; never hand several tasks to one worker to coordinate. Start a task that needs another's result after that result returns. Parts that each wait on the one before are one serial task: do it yourself or roll it as one brief. Continue related work with the existing worker while available and within its context limit. If routing itself is broken, diagnose and repair it directly. A delegated worker (`ROUTER_WORKER=1` or a `Parent session agent ID:` contract) executes its assignment without further delegation unless explicitly appointed coordinator.

## Dispatch

To delegate work to a subagent, worker or advisor, run one command; the router picks the model, launches the worker and returns what to do next:

```bash

agentcompass roll --task "<what the worker should do>"

```

Long or multi-line assignments: `--task -` reads stdin, `--brief <file>` reads a file; the worker always gets your text unchanged. Several single-task briefs go in one call, `--brief a.md --brief b.md`: they are classified at the same time and routed one after another (`status: batch`, one result per brief). Do not call `create_agent`, `Agent`, `spawn_agent`, `invoke_subagent` or a CLI yourself.

`agentcompass` is the only entry and works from any directory; do not run files under `scripts/`. Installed package (`uv tool install` / `pip install`): type it as shown. Source checkout without an install: `python -X utf8 <checkout>/agentcompass roll ...` (the same arguments; your global agent rules spell the absolute path). `agentcompass -h` lists the commands, `agentcompass <command> -h` shows the options of one.

Read `status` in the result and do exactly `next_action`: `launched` = start nothing else (`next_action` says how the result arrives), `pending` = run `next_argv`, never roll again, `split` = the classifier reads the brief as several separate tasks (`task_count`): split it yourself, roll one brief per independent task and do the coupled ones yourself; if it truly is one assignment, roll it again with `--keep-whole "<why>"` (the daily audit reviews these reasons), `self` / `native` / `blocked` / `error` = follow `next_action`. The full result is at `details` (or print it with `--verbose`).

Optional, only when true (`agentcompass roll -h`): `--hours <n>`, `--local-only "<dependency>"`, `--keep-whole "<reason>"`, `--file <path>`, `--fast`, and for a user-named model `--provider/--account/--model/--effort --basis "<user quote>"`.

More: [roll reference](references/roll-reference.md) (result variants, Dot/Grok/CLI/native) · [router internals](references/router-internals.md) (when to delegate, plan and record, scoring, launch contract, reference index).