orch-delivery

Orch Delivery

This skill is the wire contract and presentation layer for a Delivery run. Every decision (what routes where, what is allowed next, what a gate result means) comes from orch run <verb>; your job is to construct honest requests, run the verb, and present its result faithfully.

The JSON pattern every verb call follows

1. Write the request JSON to a scratch file in the OS temp directory, outside the repository (never inside the working tree or a worktree).
2. Run orch run <verb> < <scratch-file> and capture stdout. PowerShell (5.1 and pwsh 7) rejects < ("The '<' operator is reserved for future use."); use the pipe form instead: Get-Content -Raw <scratch-file> | orch run <verb>. On Windows PowerShell 5.1 specifically, that pipe alone silently corrupts non-ASCII (em dashes, §) into ?; guard it with $OutputEncoding = New-Object System.Text.UTF8Encoding $false; Get-Content -Raw -Encoding UTF8 <scratch-file> | orch run <verb>.
3. Exit 0: parse stdout as the verb's JSON result.
4. Non-zero exit: the engine refused. Present the stderr message verbatim — never paraphrase, never retry blind, never work around it. Only a revised request (different facts, different approval, a resolved precondition) is a valid next step.

orch run status --json never reads stdin — call it bare.

Selection-bearing wire versions are closed: StatusDoc 2, GateDoc 2, Dispatch 4, Escalate 2, Review 3. Reject any other schema_version before reading or submitting a Selection.

PlanDoc construction

Build a PlanDoc (schema_version: 1) honestly. Routing is derived entirely from the facts you declare; there is no field to choose a model or effort yourself. "Adjust agent routing" at the gate always means: revise the facts that were wrong and resubmit — never hand-edit a routed selection.

{
  "schema_version": 1, "host": "claude", "title": "...", "summary": "...",
  "issues": [{
    "id": "issue-slug", "title": "...", "objective": "...",
    "acceptance_criteria": ["..."], "type": "feature",
    "area_labels": ["..."],
    "facts": {
      "read_only": false, "unusually_difficult": false,
      "risk_domains": [],
      "downgrade": {"mechanical": false, "low_risk": false,
                     "fully_specified": false, "unsurprising": false}
    },
    "depends_on": [], "wave": 1, "required_tests": ["..."],
    "tests_ci_does_not_run": ["..."],
    "usage_class": "medium"
  }]
}