Skip to content
Agentmart
Submit a tool

compass-runtime

Which command runs each stage, which agent owns it, where issue state lives.

Skill by ayeo-io

IndexedApache-2.0★ 1
Write a review

Install

Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.

npx agentmarket install skill/ayeo-io-compass-runtime

Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.

Get the weekly diff

The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.

We use your email only for this newsletter. Privacy

About

Which command runs each stage, which agent owns it, where issue state lives. Load when an issue begins.

Use cases: agent-building

No reviews yet. Installed it? Reviews come from verified installers only. How reviews work

Permissions

What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.

Tools
No tools found
Secrets
None asked for
Install scripts
None found

Network hosts it talks to 1

  • codeload.github.com · found in install command

Files it touches

  • ~/.claude/skills/compass-runtime

What's inside

Skill name
compass-runtime
Description
Which command runs each stage, which agent owns it, where issue state lives. Load when an issue begins.
License
Apache-2.0
Instructions
197 lines in SKILL.md · view on GitHub

Agent Skills spec issues

  • files: referenced but missing: scripts/multiagent.sh, scripts/integrate.sh

Files (5)

  • SKILL.md
  • archive-quote-manifest.json
  • text-sweeps.md
  • writing-voice-worked-example.md
  • writing-voice.md

SKILL.md (first lines)

# Compass - the stage map

The rules of behavior are in `compass-contract.md`, and the SessionStart hook
puts them in every session before you read this. They are not repeated here,
so the two cannot disagree.

What this file carries is the mapping a session needs once it is already
following the contract - which command runs which stage, what each writes,
which agent owns it, and where the result lands on disk.

If anything here conflicts with `docs/methodology.md`, the methodology doc
wins. First Compass session? `docs/five-minutes.md` has the mental model and a
worked example; `docs/safety-contract.md` says what Compass guarantees and
what it explicitly does not.

## The stages and their commands

| Stage | Command | Artifact it writes |
|---|---|---|
| Assess | `/compass:assess` | `delivery-approach.md` + the manifest's assessment |
| Define acceptance criteria | `/compass:define` | `acceptance-criteria.md` |
| Requirements review | `/compass:refine` | `requirements-review.md` (ends with the Definition of Ready) |
| Plan | `/compass:plan` | `technical-design.md` (+ `distribution-map.md` on parallel work) |
| Break down the work | `/compass:breakdown` | worktrees + subtask assignments |
| Implement | `/compass:implement` | code + the red and green records (named by binding) |
| Test & review | `/compass:verify` | `verification-report.md` (ends with the Definition of Done, a block `compass issue render` writes from the manifest's records) |
| Ship | `/compass:ship` | the integration commit + settled follow-ups |

`/compass:assess` itself carries only that light path. The full assess
procedure - setup, `--reassess`, steps 1 to 7 and the gate - is in
`approaches/assess-procedure.md`, which assess names for every heavier
approach and for `--reassess`.

When assess computes a **quick fix**, the eight rows above collapse into one
command: `/compass:quick-fix` carries the whole light path - assess, the one
scenario, red-green, the check, the commit - inlined in a single file, and it
is read alone. It is the same pipeline
at its lightest weight, not a way around it: if `compass approach evaluate`
returns anything heavier, the rows above are what runs.

Cross-issue: `/compass:status` (one issue or a flat list), `/compass:flow`
(the managed cross-issue view - advisory, never gating), and
`compass board render` (the same view as one page in the browser;
`docs/board.md`). Entry points for the consulted perspectives:
`/compass:intent` (product owner), `/compass:position` (marketer),
`/compass:design` (designer - produces the UI contract), `/compass:consult`
(multi-role decisions). `/compass:init` is optional, and it is not what
creates the project: the entry points above run `compass init` for you.

A command spelling that a release held and 6.0.0 renamed worked through an
alias in 6.x. The aliases were removed at 7.0.0, so the old spelling now exits 2
with the parser's usual error. That error names no new spelling, except for
two removed verbs that add one line. `docs/upgrade-7-0-0.md` names them and lists
each removed spelling beside its new one. `governance/terminology.yml` names each retired word beside the
one that replaced it, and `docs/glossary.md` says the same in prose; the
current verbs are whatever `compass --help` lists.

**The binding decides the filename.** `compass tdd-red --scenario <id>` and
`compass tdd-green --scenario <id>` write `evidence/red-<id>.json` and
`evidence/green-<id>.json`; a run with no `--scenario` writes
`evidence/red.json` and `evidence/green.json` instead. Nothing else is touched, so
recording one scenario cannot overwrite the record another gate is citing -
and a reader knows where their evidence went without guessing.

## Which agent and which skill, per stage

- **Assess** - load the `adaptive-routing` skill; consider the `router`
  agent.
- **Define and refine** - load `bdd-specification`; on brownfield work whose
  behavior is not yet written down, also load `behavior-mapping`. The
  `spec-author` agent owns both stages.
- **Plan** - load `plan-authoring` (which optional design sections earn a
  place) and `governance-check` (how to check the design against the
  governance in force). The `planner` agent owns it.
- **Implementing in parallel** - load `worktree-multiagent`. The `orchestrator`
  agent coordinates; `builder` agents work, one per worktree, each loading
  `tdd-discipline`.
- **Test & review** - the `verifier` and `reviewer` agents run; load
  `evidence-gates`. Load `receiving-code-review` when answering their
  comments.
- **An unexpected test failure while implementing** - load
  `systematic-debugging`, and after three failed fixes re-assess rather than
  try a fourth.
- **Role-facing work** - load `intent-interview` and read its
  `role-translation.md`, which is how one set of
  acceptance criteria is read through five role perspectives. The
  `product-owner`, `product-marketer` and `architect` agents apply specific
  ones.
- `evidence-gates` carries `traceability.md`, read whenever an artifact is
  written.

The full set is in `agents/`. The files in `agents/` are the authority.

An agent has Bash when it runs a `compass` verb, and Skill when it loads a
skill. `product-owner` and `product-marketer` have no shell; their caller runs
their verbs. No agent edits `manifest.yml` by hand (`tests/test_agent_tools.py`).

## Worktrees and multiagent

Only the `orchestrator` agent creates worktrees (`scripts/multiagent.sh`) and
integrates them (`scripts/integrate.sh`). A `builder` works *inside* its
assigned worktree and never touches a sibling's. The approach's distribution
map says how many subtasks exist; policy can cap the count. On solo work there
is no worktree - work on the current branch.

## Where state lives

```
compass.yml                     The one file a person edits: settings, and the project's edits to the shipped default (written by /compass:init or compass policy migrate)
.compass/
├── state.yml                   What the CLI wrote: what initialized the project, and when
├── config.yml                  Projects from 5.x only: read until the project moves its settings with compass policy migrate; no 6.0.0 command writes it
├── legacy/, migration.yml      Written by compass policy migrate: the superseded files, and the digests of them
├── current-task                One-line pointer to the active issue
├── work/
│   └── <issue-slug>/            One directory per issue: its machine state
│       ├── manifest.yml         The manifest
│       ├── generations/<n>/     The configuration the issue runs against (written by compass approach evaluate --write)
│       ├── evidence/            red/green records + typed gate evidence

From ayeo-io/compass, Apache-2.0.

More from ayeo-io/compass

Signals

Stars
1
Forks
0
Open issues
23
Last commit (this folder)
Oct 20, 2026 (9 days ago)
Last release
Oct 20, 2026 (9 days ago)

From GitHub (fetched Oct 24, 2026).

Details

License
Apache-2.0
First seen
Oct 24, 2026
Last checked
Oct 24, 2026

More from ayeo-io

Cite this page

compass-runtime skill. Agent Mart. https://agentmart.dev/skills/ayeo-io-compass-runtime (last updated 2026-10-24)