name: spec-implementation
version: 2.0.0
description: Implement one or more specifications. Single-spec or multi-spec (from a git diff / PR / branch). Each spec gets its own implement + completeness + quality pipeline; a shared e2e loop covers all implemented specs together with a mock-agent policy and a small release-only exception set.
inputs:
- name: spec_path
description: Single-spec mode. Path to one spec file (or directory) to implement. Relative to the repository root. Leave empty when using spec_ref.
type: string
default: ""
- name: spec_ref
description: Multi-spec mode. A reference whose changed *.spec.md files will each be implemented in parallel. Any of - a PR URL or number, a branch name, a commit SHA or range (base..head), or a path to a .diff/.patch file. Leave empty when using spec_path.
type: string
default: ""
- name: spec_title
description: Human-readable title for the workspace. Shows up in the index heading.
type: string
default: Spec Implementation
- name: implementation_target
description: Agent that writes the initial implementation and closes completeness gaps for each spec. Typically your default high-capability code-writing target.
type: string
format: execution-target
default: claude-code[yolo]:anthropic:claude-opus-4-7
- name: review_targets
description: Two or more agents that independently review the implementation. Used in both completeness and quality reviews for every spec. More agents = more coverage, more cost.
type: array
items:
type: string
format: execution-target
default:
- claude-code[yolo]:anthropic:claude-opus-4-7
- codex[xhigh]:openai:gpt-5.5
- name: smart_target
description: Stronger agent that merges per-reviewer findings, writes the fix plan, and applies the fixes. Used by the coordinator, the completeness and quality aggregators, and the quality fixer.
type: string
format: execution-target
default: codex[xhigh]:openai:gpt-5.5
- name: review_passes
description: How many quality review/fix cycles to run per spec. 2 catches most regressions; 1 is faster; 3+ is for very high-stakes specs.
type: number
default: 2
- name: focus_areas
description: Optional review focus sections each quality reviewer must address (e.g., performance, security, error handling). Empty means a general correctness review.
type: array
items:
type: string
default: []
- name: e2e_writer
description: Agent that writes / extends end-to-end tests in the shared e2e loop.
type: string
format: execution-target
default: claude-code[yolo]:anthropic:claude-opus-4-7
- name: e2e_verifier
description: Agent that re-runs the standard e2e suite, audits the new tests, enforces the mock-agent policy, and lists remaining gaps for the next write pass. Should be a different agent from e2e_writer so the audit is independent.
type: string
format: execution-target
default: codex[xhigh]:openai:gpt-5.5
- name: e2e_passes
description: How many e2e write/verify ping-pong cycles to run.
type: number
default: 2
- name: e2e_test_root
description: Where the project's e2e tests live (e.g., e2e/, tests/e2e/). Leave empty to let the e2e agents discover the location.
type: string
default: ""
- name: mock_agent
description: Selector or test-side name of the mock agent that every standard e2e test must target. The mock returns canned outputs so the standard suite stays fast and deterministic.
type: string
default: mock
- name: release_only_marker
description: "Tag / attribute / filename suffix the project uses to mark e2e tests that hit real agent operations. These tests run only on release builds and the set must stay small (rule of thumb: one happy-path test per real-agent integration)."
type: string
default: release-only
- name: release_only_test_root
description: Optional path where release-only real-agent e2e tests live, if the project keeps them separate from the mock-driven suite. Empty means co-locate with the rest of the e2e suite and rely on the marker.
type: string
default: ""