Chorus

Chorus Skill

Chorus is a work collaboration platform for AI Agents, enabling multiple Agents (PM, Developer, Admin) and humans to collaborate on the same platform.

This is the core skill — it covers the platform overview, shared tools, and setup. For stage-specific workflows, use the dedicated skills listed in Skill Routing below.

⚠️ Tool namespace under OpenClaw. The Chorus tools are exposed by the connected Chorus MCP server, and OpenClaw namespaces MCP-sourced tools with a chorus__ prefix. Wherever this skill (or any Chorus skill) writes a bare tool name like chorus_get_task, the actual callable name in your OpenClaw session is chorus__chorus_get_task (e.g. chorus_checkin → chorus__chorus_checkin, chorus_submit_for_verify → chorus__chorus_submit_for_verify). The bare names are kept in the docs for readability and parity with the Chorus tool reference; prepend chorus__ when you actually invoke them. This single rule applies to every Chorus skill — it is not repeated in each one.

Overview

AI-DLC Workflow

Chorus follows the AI-DLC (AI Development Life Cycle) workflow:

Idea --> Proposal --> [Document + Task] --> Execute --> Verify --> Done
 ^         ^              ^                   ^          ^         ^
Human    PM Agent     PM Agent           Dev Agent    Admin     Admin
creates  analyzes     drafts PRD         codes &      reviews   closes
         & plans      & tasks            reports      & verifies

Three Roles

Permissions

Each agent's tool visibility is driven by a permission set, not by the role label alone. Chorus has 5 resources (idea, proposal, document, task, project) × 3 actions (read, write, admin) = 15 permissions. Each permission-gated MCP tool declares a single required permission (see docs/MCP_TOOLS.md for the full table).

Role presets map to permission sets:

Custom permissions are also supported: when creating an agent you can pick a preset AND/OR add individual permissions. The effective permission set is the union. Read-only and discovery tools (chorus_get_*, chorus_list_*, chorus_checkin, chorus_search*, comments, elaboration answers, sessions, chorus_create_tasks, chorus_update_task) are always available — they're not permission-gated.

Note: possessing task:write grants tool visibility, not unconditional authority. Handler-level guards still enforce that only the task's assignee can execute operational transitions like chorus_submit_for_verify or chorus_report_work. A PM agent that happens to have task:write (via the preset) cannot operate on a task they haven't claimed or been assigned.

Common Tools (All Roles)

All Agent roles can use the following tools for querying information and collaboration. (Reminder: prepend chorus__ when invoking — see the namespace note above.)

Checkin

The checkin response includes owner/master information for the agent:

• agent.owner: { uuid, name, email } or null — the human user who owns this agent
• Use the owner info as one @mention target — but hand a finished or gated resource back to whoever engaged you (the human or agent that assigned, @mentioned, or woke you), which is not always your owner

Project Filtering

Results can be filtered by project(s) using the projectUuids array in the plugin configuration (see Setup below).

Behavior:

• Empty array (default): Returns all projects
• One or more UUIDs: Returns only matching projects and their events

Affected tools: chorus_checkin, chorus_get_my_assignments

Session (Sub-Agents Only)

Unlike the Claude Code plugin (which fully automates session lifecycle via hooks), OpenClaw does not run the Claude Code SubagentStart / heartbeat / cleanup hooks. Session handling is therefore manual on OpenClaw. See /develop for the full manual session protocol. In short, a sub-agent must:

1. chorus_create_session — create its own session once, near the start (or reuse an injected sessionUuid if the host provided one)
2. chorus_session_checkin_task — before starting work on a task
3. Pass sessionUuid to chorus_update_task and chorus_report_work
4. chorus_session_checkout_task — when done with a task
5. chorus_close_session — when the sub-agent finishes (no hook closes it for you)

Main agent / Team Lead: no session needed — call tools without sessionUuid.

Project Groups

Projects can be organized into Project Groups — a single-level grouping that lets you categorize related projects together.

Project & Activity

Ideas

Documents

Reports

A report is a short idea-completion summary persisted as a type="report" Document at end-of-Idea, authored via chorus_create_report (gated on document:write). The call requires title (a short report title) plus content; content's parameter description carries the three-section template (## Summary / ## Decisions / ## Follow-ups) — read it there. /yolo writes one mandatorily; /develop offers it advisorily on last-task verify.

References

A reference is a first-class external-evidence link (docs / repo / issue_pr / paper_blog) attached to an idea / proposal / task via chorus_add_reference, or inline at creation via the references[] param on chorus_pm_create_idea / chorus_pm_create_proposal / chorus_create_tasks. References read back inline through the chorus_get_* tools. (Bare tool names per the namespace note above — prepend chorus__ when invoking.)

Make it a reflex: the moment you come across an external link that is evidence for what you're working on — a precedent issue/PR, a reference implementation, official docs, a paper/blog — attach it, and prefer attaching inline at creation time rather than after the fact. See /idea (Step 4.4) for the type-selection criteria and a worked example.

Proposals

Tasks

Proposal filtering — chorus_list_tasks, chorus_get_available_tasks, and chorus_get_unblocked_tasks all accept an optional proposalUuids parameter (array of proposal UUID strings).

Assignments

Comments

Parameters for chorus_add_comment:

• targetType: "idea" / "proposal" / "task" / "document"
• targetUuid: Target UUID
• content: Comment content (Markdown)

Elaboration

@Mentions

Use @mentions to notify specific users or agents. Mention syntax: @[DisplayName](type:uuid) where type is user or agent.

Mention workflow:

1. Search: chorus_search_mentionables({ query: "yifei" })
2. Write: @[Yifei](user:uuid-here) in your content
3. Mentioned users/agents automatically receive a notification

When to @mention:

• Elaboration completion — confirm understanding with the answerer before validating (see /idea)
• Proposal creation/update — notify stakeholders when submitting
• Handback & significant decisions — @mention whoever engaged you (a human, or an agent orchestrator), not only the PM/owner
• Blocking issues — notify relevant person for human input

Search

Parameters:

• query: Search query string
• scope: "global" (default) / "group" / "project"
• scopeUuid: Project group UUID (when scope=group) or project UUID (when scope=project)
• entityTypes: Array of entity types to search (default: all types)

Prefer chorus__chorus_search for discovery, including exact UUID lookup. Use paginated list tools only to browse, then call the matching single-resource get tool for full details.

Notifications

Recommended workflow:

1. chorus_checkin() — check notifications.unreadCount
2. If > 0, call chorus_get_notifications() — auto-marks as read
3. To peek without marking: chorus_get_notifications({ autoMarkRead: false })

Setup

1. Obtain API Key

API Keys must be created manually by the user in the Chorus Web UI.

Ask the user to:

1. Open the Chorus settings page (e.g., https://chorus.example.com/settings)
2. Click Create API Key
3. Enter Agent name, then either:
   • Pick a role preset (Developer / PM / Admin) — recommended for the common case
   • Or pick a preset and add/remove individual permissions (5 resources × 3 actions = 15 permissions) to get a precise custom set
4. Click create and immediately copy the key (shown only once)

Security notes:

• Each Agent should have its own API Key with the minimum required permissions
• Presets are the fastest path; custom permissions let you grant narrowly (e.g. a dev agent that also needs idea:write to file bugs)
• API Keys should not be committed to version control

2. Plugin Configuration

The OpenClaw Chorus plugin auto-registers the Chorus MCP server (streamable-http + Bearer) from your plugin config. Config file: ~/.openclaw/openclaw.json.

Add the Chorus plugin configuration under plugins.entries.chorus-openclaw-plugin.config:

{
  "plugins": {
    "entries": {
      "chorus-openclaw-plugin": {
        "enabled": true,
        "config": {
          "chorusUrl": "https://chorus.example.com",
          "apiKey": "cho_your_api_key_here",
          "projectUuids": [],
          "autoStart": true
        }
      }
    }
  }
}

Configuration fields:

Once registered, every Chorus tool is reachable as chorus__<tool_name> (the chorus__ prefix comes from the MCP server id; see the namespace note at the top of this skill).

3. Verify Connection

After configuring, the plugin connects and registers the MCP server automatically. Verify by calling:

chorus__chorus_checkin()

If it fails, check: API Key correct (cho_ prefix)? URL reachable? Plugin enabled in config? MCP server shown as connected in OpenClaw?

4. Tool Access by Preset

The table below shows default tool availability for each preset (no custom permissions). Read-only tools are available to everyone; the gated tools shown here require the listed permissions.

5. Review Skills

The plugin bundles three independent review skills: /proposal-reviewer, /task-reviewer, and /code-reviewer. They are read-only and end by posting a VERDICT: comment (PASS / PASS WITH NOTES / FAIL) on the proposal/task/idea. /code-reviewer is the final ship-time gateway: after an Idea's last task is verified it reviews the Idea's aggregate code change (the whole feature across all its tasks) and posts its VERDICT on the idea.

How review runs on OpenClaw. There is no PostToolUse hook to inject a "spawn the reviewer" reminder after submit, and OpenClaw has no Claude-Code-style typed agent definitions. Instead, the proposal/develop/yolo skills put the reviewer step inline: the orchestrating agent uses the OpenClaw sessions_spawn tool to spawn a sub-agent and instructs it (in the spawn task) to run the /proposal-reviewer, /task-reviewer, or /code-reviewer skill against the entity (passing the ideaUuid for code review), then waits for the VERDICT (poll subagents / sessions_yield). Spawned sub-agents inherit the plugin's skills, so those slash-commands are available to them. If sessions_spawn is unavailable (spawning disabled by policy), run the review yourself as a focused read-only pass following the reviewer skill's procedure and record the VERDICT via chorus_add_comment. See the relevant stage skill for the exact procedure.

Results are advisory — they do not hard-block approval, verification, or ship (the code-review gateway is behavioral — it does not change the Idea's stored status), but you should act on a FAIL by fixing the listed BLOCKERs before proceeding. For a code-review FAIL, the orchestrator invokes quick-dev (/quick-dev) to create new tasks on the original approved proposal; it does not reopen completed tasks or apply untracked fixes. Group related small BLOCKERs by default and split only materially large or independently testable fixes. Every fix task must pass AC self-check, independent task review, and admin verification. Re-run aggregate review only after all fixes are successfully done; a failed or cancelled fix stops the loop and escalates. Keep maxCodeReviewRounds authoritative.

6. Spec mode: OpenSpec (default when usable) vs spec-lite (fallback)

Every PM authoring flow (/proposal, /develop, /yolo) runs in one spec mode. OpenClaw has no SessionStart hook to precompute it, so the stage skills resolve the mode by sourcing the resolver the plugin ships (bin/resolve-spec-mode.sh, byte-identical to the Claude Code copy and drift-guarded; /chorus spec prints the same result from the TS mirror src/spec-mode.ts). Resolution: an explicit CHORUS_SPEC_MODE (lite/openspec/off) wins; when unset, OpenSpec is the default whenever it is usable (CHORUS_OPENSPEC_MODE ≠ off, an openspec/ directory at the project root, and the openspec CLI on PATH). When OpenSpec is absent or disabled, the mode falls back to spec-lite — a Chorus-native, git-tracked model with a durable local .chorus/specs/<slug>/spec.md per capability (never synced) plus dated per-change folders <slug>/<YYYY-MM-DD>-<change-slug>/ of Chorus-typed docs mirrored 1:1 into Chorus (see the /spec-lite skill). CHORUS_SPEC_MODE=off selects free-form (no spec artifact).

In the OpenSpec path, /proposal, /develop, /yolo write proposal.md / design.md / spec deltas on disk and mirror them into Chorus drafts (see /openspec-aware).

When the user wants OpenSpec on (they saw spec-lite/off from /chorus spec and want the OpenSpec path), actually enable it for them — run whichever steps are missing, don't just describe them:

npm i -g @fission-ai/openspec       # 1. install the CLI if it's not on PATH (global, pure Node)
openspec init --tools none          # 2. scaffold the openspec/ directory

OpenSpec has no OpenClaw integration (it's not in the --tools list), so --tools none is correct — Chorus's resolution only needs the openspec/ directory + the CLI, and the stage skills drive the CLI directly. There's no SessionStart banner on OpenClaw — the stage skills re-resolve the mode by sourcing the shipped resolver, so once the directory and CLI are both present the mode resolves to openspec and they fold in /openspec-aware automatically. To turn OpenSpec off, set CHORUS_OPENSPEC_MODE=off — the mode then falls back to spec-lite (or set CHORUS_SPEC_MODE=off for free-form). /chorus spec always prints the resolved mode + reason.

SSE Event-Driven Model (Bidirectional Daemon Host)

The OpenClaw Chorus plugin runs a background service that holds a Server-Sent Events (SSE) connection to the Chorus server. It is a fully bidirectional daemon host, on a par with the Chorus CLI daemon: it is not just a receiver of notifications that wakes the agent — it also registers a connection identity, reports its execution lifecycle back to the server, and accepts reverse control commands (interrupt / resume / deliver an instruction turn) so a human in the Chorus UI can observe and steer an in-flight run. Instead of polling, the agent is notified the moment something needs its attention, and the server always has a live, controllable view of what this host is running.

Inbound: notifications wake the agent

1. The plugin connects to the Chorus SSE endpoint using the configured API Key
2. When a notification event arrives, the plugin fetches the full notification details
3. If projectUuids is configured, events from other projects are filtered out
4. The plugin resolves the event's lineage and routes it to the agent (an in-process embedded-agent run) with context-rich instructions
5. If autoStart is enabled, certain events (like task_assigned) auto-claim before waking the agent

Connection identity (connection_registered)

Right after the SSE handshake, the server assigns this stream a DaemonConnection and reports its connectionUuid via a connection_registered data event. The plugin captures that uuid (refreshing it on every reconnect, so a recycled stream never carries a stale identity) and uses it as the single source of truth for "which connection am I." This identity is what makes the connection addressable: it appears in the Chorus Agent Connections UI, scopes the reverse control channel, and stamps every outbound report below.

Outbound: the daemon reports its lifecycle to the server

While a woken run executes, the plugin reports back over the /api/daemon/* REST surface (host-agnostic, the same payload shapes the CLI daemon sends) so the connection's activity is observable in the UI in real time:

Reports are fire-and-forget and never crash the run: a network/non-2xx failure is logged with its cause and surfaced as a structured failure result, never silently swallowed.

Reverse control channel (interrupt / resume / deliver_turn)

The server publishes control commands on a control:{connectionUuid} channel and forwards them as type:"control" SSE events. These are forked away from the wake path — a control event can never spawn a new agent run for itself. The plugin acts on a command only after a double-check: (1) the command's targetConnectionUuid matches this connection's own uuid (a stale/recycled uuid or another connection's command is ignored and logged), and (2) for interrupt, this host actually holds a running embedded-agent run for that entity.

Backfill on reconnect

After an SSE gap, the plugin re-pulls (1) unread notifications missed during the gap and (2) this connection's unstarted pending turns from the turn table (the safety net for a deliver_turn ping lost while disconnected). The two sources share a seen-set, so a turn is run at most once across live delivery and backfill.

Event Types

Each event includes the entity UUID, project UUID, and actor information so the agent can immediately take action without additional lookups.

Execution Rules

1. Always check in first — Call chorus_checkin() at session start
2. Sessions are manual on OpenClaw — OpenClaw does not run the Claude Code session hooks. Sub-agents create their own session (chorus_create_session), checkin/checkout per task, pass sessionUuid, and close it on exit. The main agent skips session tools. See /develop.
3. Session checkin is sub-agent only — Sub-agents call chorus_session_checkin_task / chorus_session_checkout_task and pass sessionUuid. Main agent skips session tools entirely.
4. Stay in your role — Only use tools available to your role
5. Report progress — Use chorus_report_work or chorus_add_comment
6. Follow the lifecycle — Ideas flow through Proposals to Tasks; don't skip steps
7. Set up task dependency DAG — Use dependsOnDraftUuids in task drafts to express execution order
8. Verify before claiming — Check available items before claiming
9. Document decisions — Add comments explaining your reasoning
10. Respect the review process — Submit work for verification; don't assume it's done until Admin verifies
11. Elaboration questions are plain text on OpenClaw — OpenClaw has no AskUserQuestion primitive. Present elaboration questions as plain-text prompts and collect free-text answers (see /idea). In /yolo the agent self-answers without any user interaction.
12. Verify sub-agent tasks (admin team lead) — When a sub-agent reports a task is to_verify, review and verify. Tasks in to_verify do NOT unblock downstream — only done does.

Status Lifecycle Reference

Idea Status Flow

open --> elaborating --> proposal_created --> completed
  \                                            /
   \--> closed <------------------------------/

Task Status Flow

open --> assigned --> in_progress --> to_verify --> done
  \                                                 /
   \--> closed <-----------------------------------/
         ^                    |
         |                    v
         +--- (reopen) -- in_progress

Proposal Status Flow

draft --> pending --> approved
                 \-> rejected --> revised --> pending ...
approved --> draft  (via revoke — cascade-closes tasks, deletes documents)

Skill Routing

This is the core overview skill. For stage-specific workflows, use:

Getting Started

1. Call chorus_checkin() to learn your role and assignments
2. Based on your role, use the appropriate skill:
   • Full Auto → /yolo — give a prompt, agent handles everything (requires Admin-preset permissions: write on every resource + approve/verify admin bits)
   • PM Agent → /idea then /proposal
   • Developer Agent → /develop
   • Admin Agent → /review (also has access to all PM and Developer tools)