dely — Agent Instructions

This repository contains the dely package: the delivery skill and its automation-first control protocol for coding agents, on Claude Code, Codex CLI, Cursor Agent CLI and the other harnesses harnesses.json marks supported.

Source of truth

• The workflow contract is skills/delivery/SKILL.md.
• Settled, open, and rejected decisions are recorded in docs/decisions.md.
• Installation and onboarding instructions are in README.md.
• The closure gates below are the whole structural check. There is no test suite and no CI workflow; nothing runs on a pull request automatically.
• Live verification is probe/checklist.md, run by a separate agent session against a candidate installed from a git archive snapshot. When a delivery requires that checklist, it runs before the review dispatch, so the reviewed head is the verified head; if remediation changes what a worker launch does, the affected rows rerun before the re-review. This ordering is a project rule for this repository, not part of the portable protocol in SKILL.md.
• CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md, and .github/ (issue and pull-request templates) own the human contribution, conduct, security, and PR contract; they need no Dely or Orca install to follow.

Repository artifacts are written in English.

Workflow

• Default branch: main.
• Bounded or Architectural work invokes the delivery skill; a Spike is investigation only and starts no delivery run.
• Durable decisions live in docs/decisions.md; transient plans live in docs/_plans/.
• Maintenance logging is machine-local and opt-in at ~/.dely/log.jsonl (JSON Lines, one object per line); no project-owned log file is tracked. Dely never creates ~/.dely/: a missing directory means nothing is written and nothing is created. The log records failed and abandoned deliveries too.
• A self-update runs its release phase through the frozen installed plugin version, because Control is already using it when the plan starts. Candidate changes in this checkout take effect for the next delivery, not the one shipping them. Plugin caches and any live worker hook wiring are refreshed only between plans.
• A delivery that changes anything under skills/, harnesses.json, extensions/ or package.json advances the version in both plugin manifests, package.json, and the version gate below, within that same delivery.

Phase dispatch

Name the model and reasoning effort on every dispatch.

Dely

Bounded or Architectural work invokes dely:delivery; Spike starts no delivery run.

.dely/pins.json is this repository's deployment selection, not the portable protocol. Per-harness facts — status, Control wake, trust, discovery, model and effort flag support, permission defaults, forbidden headless forms, the instructions-file rule, and measured notes including for deferred harnesses — live in harnesses.json at the repository root.

This project adds no phase-implied sandbox. Native Internet access, closure gates, result writes and coordinator completion stay available.

Closure gates

Run from the repository root:

git diff --check

jq -e . harnesses.json plugin.json .claude-plugin/plugin.json .claude-plugin/marketplace.json .codex-plugin/plugin.json .cursor-plugin/plugin.json package.json >/dev/null

jq -e '.harnesses | all((.status == "supported" or .status == "deferred") and (.controlWake == "background" or .controlWake == "waker" or .controlWake == "unsupported") and ((has("modelPin") | not) or .modelPin == "spec-unchecked") and ((has("effortInModel") | not) or (((.effortInModel | type) == "string") and .effortInModel != "" and .modelFlag == true and .effortFlag == false and (has("modelPin") | not))))' harnesses.json >/dev/null

node --check skills/delivery/scripts/dely.js
git ls-files -z '*.sh' 'skills/delivery/scripts/dely' | xargs -0 -n1 bash -n

test "$(jq -r .version .claude-plugin/plugin.json)" = 0.27.2
test "$(jq -r .version .codex-plugin/plugin.json)" = 0.27.2
test "$(jq -r .version package.json)" = 0.27.2

test "$(jq -r 'has("type")' package.json)" = false

test ! -e skills/delivery/references/harnesses.md

git grep -nE 'adoptCommand|adoptedPermission|adoptPath|readAdopts|writeAdopts|recordAdopt|takeAdopt|closeAdopted|closeCreated|waitQuiet|lastOutputAt|launchKind|hasGate|GATES|pinWhy|ELECTRON|harnessCell|shownModel|waitShownModel|pinCheckMs|PIN_CHECK|pin_not_applied' -- skills/ && exit 1 || true

git grep -Ei 'pace.?id' -- . ':!docs/_plans' && exit 1 || true
git grep -E '(^|[^A-Za-z0-9])[A-Z][0-9]+[a-z]?([^A-Za-z0-9]|$)' -- . ':!docs/_plans' && exit 1 || true

rows=$(jq -r '.harnesses[] | . as $h | (if $h.status == "deferred" or $h.controlWake == "unsupported" then "✗" elif ($h.limits // "") | test("(^|[.] )As Control,") then "⚠" else "✓" end) as $c | (if $h.status == "deferred" then "✗" elif ($h.limits // "") | test("(^|[.] )As a worker,") then "⚠" else "✓" end) as $w | "| \($h.name) | \($c) | \($w) | \($w) | \(if ($h.limits // "") == "" then "—" else $h.limits end) |"' harnesses.json) || exit 1
printf '%s\n' "$rows" | while IFS= read -r row; do grep -qxF -- "$row" README.md || { echo "README lacks: $row"; exit 1; }; done && { test "$(grep -cE '^\| [^|]+ \| (✓|⚠|✗) \| (✓|⚠|✗) \| (✓|⚠|✗) \|' README.md)" = "$(jq '.harnesses | length' harnesses.json)" || { echo "README role rows ≠ harnesses.json entries"; exit 1; }; }

Control is ✗ when status is deferred or controlWake is unsupported, ⚠ when limits has a sentence starting "As Control,", otherwise ✓; Implementer and Reviewer are ✗ when deferred, ⚠ when limits has a sentence starting "As a worker,", otherwise ✓. A role sentence starts the limits value or follows ". ".

The first three gates prove repository shape and syntax only. A Bounded or Architectural change to runtime behaviour must also name a focused instrument that distinguishes the changed behaviour from its failure mode, and a change to what a worker launch does is verified by running probe/checklist.md, not by these gates.

The version gate is a literal pin and the only check that keeps the two manifests from splitting: a skills/ change that forgets one manifest ships different protocol text under the same version string.

The absence command distinguishes a deleted rail from documentation that merely says it is deleted: it fails if the retired skills/delivery/references/harnesses.md returns. The identifier grep is the same instrument applied to code: it fails on a deletion that removes a definition and leaves a caller.

The disclosure grep is lexical. It catches named consumer identifiers; it does not catch a quoted consumer path, a branch name, or a session id. A green result is not a proof of non-disclosure.