status

Status — Orchestration Dashboard

When This Skill Runs

• User wants to understand current project state
• Before planning next steps
• After recovering from a crash or long pause
• Start of a new orchestrator session (first action should be status)
• Routine check-in during development

Prerequisites

• Orchestration state must be initialized (run /agent-atelier:init first if not)

Allowed Tools

• Read (state files and gate files)
• Bash (git root detection, directory listing)
• Glob (scan open gates directory)

Execution Steps

1. Detect root via git rev-parse --show-toplevel.

2. Read state files. Read these JSON files directly — no CLI needed:

   • .agent-atelier/loop-state.json
   • .agent-atelier/work-items.json
   • .agent-atelier/watchdog-jobs.json

3. Scan open gates. Glob both .agent-atelier/human-gates/open/ and .agent-atelier/human-gates/resolved/. If a gate ID appears in both (orphaned from an interrupted resolve), treat it as resolved — resolved/ is authoritative.

4. Derive next action. Determine what the orchestrator should do next based on state:

   • Open gates with blocked WIs → "resolve gate HDR-NNN or work unblocked items"
   • ready WIs exist → "dispatch builder for WI-NNN"
   • Active candidate → "await validation / dispatch VRM"
   • All WIs done → "loop complete — report to user"
   • Only pending WIs with unmet deps → "unblock dependencies first" If loop-state.json already contains a next_action, prefer that over inference.

5. Render the dashboard. Present a human-readable summary. Omit zero-count statuses and empty sections to keep output compact. Flag expired leases and budget warnings prominently.

Example Dashboard

Team:             atelier-sparrow-7
Mode:             IMPLEMENT
Active Spec:      docs/product/behavior-spec.md (rev 7)
Active Candidate: CS-003 (WI-014, WI-015) on candidate/WI-014 @ abc1234
Next Action:      orchestrator → dispatch_vrm_evidence_run → WI-014

Work Items (8 total):
  pending:               2  (WI-018, WI-019)
  ready:                 1  (WI-015)
  implementing:          1  (WI-014 — lease expires in 47min)
  candidate_validating:  1  (WI-016)
  blocked_on_human_gate: 1  (WI-012 — HDR-002)
  done:                  2  (WI-010, WI-011)

  !! EXPIRED LEASE: WI-014 — lease expired 12min ago

Candidate Queue (1 pending):
  CS-004 (WI-017) — queued 2026-04-08T13:00:00Z

Open Gates (1):
  HDR-002: "API shape: preserve current or allow breaking change?"
    Blocking: WI-012  |  Created: 18h ago
    Recommended: Option A

Watchdog:
  Last tick: 35min ago  |  Open alerts: none

Exit Codes

Code	Meaning
0	Success
3	State files not found (not initialized)
4	Runtime failure (malformed JSON in state files)

Output Contract

Returns JSON to stdout with these top-level keys:

Key	Type	Description
team_name	string/null	Session-scoped team name from loop-state
mode	string	Current loop mode (DISCOVER, BUILD_PLAN, IMPLEMENT, etc.)
active_spec	string	Path to the active behavior spec
active_spec_revision	number	Spec revision number
next_action	object	Derived or persisted next action
active_candidate_set	object/null	Currently active candidate set
candidate_queue	array	Queued candidate sets awaiting the active slot
open_gates	array	IDs of open human gates
work_items_summary	object	total, by_status (counts), expired_leases (IDs)
watchdog	object	last_tick_at, open_alerts count

When presenting to a human user, render the readable dashboard format shown in the execution steps. Diagnostic messages go to stderr.

Input Conventions

Status takes no arguments and no payload. It is a read-only operation.

Idempotency

Status is inherently idempotent — it reads state without modifying it. Running it multiple times produces the same output for the same state.

Error Handling

Condition	Exit Code	Action
State files missing	3	Suggest running /agent-atelier:init
JSON malformed in any state file	4	Name the broken file; suggest deleting it and re-running /agent-atelier:init to regenerate (init is idempotent and only creates missing files)
Gate directory missing	0	Render dashboard without gates section; note the absence
Watchdog file missing but others present	0	Render available sections; warn that watchdog state is absent
All WIs empty (initialized but no work yet)	0	Render dashboard with "No work items yet" and suggest next step

Constraints

• Status never writes to state files. It is strictly read-only.
• When a session restarts or context is compressed, run status first to restore situational awareness before taking any action.
• Prefer the persisted next_action from loop-state.json over inferred values. Only infer when the field is absent or stale.