Coact Complete

COMPLETE a skill into an agent definition

A SKILL.md is procedural knowledge injected into the caller's turn. A subagent is a separate worker with its own context, persona, tools, model, and a defined return value. An agent is mostly a skill plus a thin "extras" envelope. COMPLETE synthesizes that envelope — mechanically, no LLM needed.

The two extras that actually matter:

1. the persona (system prompt / identity), and
2. the return contract (a JSON Schema so a manager can consume the output).

coact keeps the skill on disk as the single source of truth — the agent references the skill by name and never copies its body.

Look before you leap: plan first

plan_completion is a dry-run. It returns the proposed agent plus per-field provenance and warnings — every synthesized field and where it came from:

from coact import plan_completion

plan = plan_completion(".claude/skills/ux-analyst")
print(plan.render())     # name←skill, tools←inferred/coact:, model←policy, ...
plan.agent               # the AgentDefinition it would produce
plan.warnings            # e.g. "tools not declared; inferred from ..."

coact plan .claude/skills/ux-analyst      # same dry-run, in the terminal

When you're happy:

from coact import complete, emit_agent

agent = complete(".claude/skills/ux-analyst")          # -> AgentDefinition
print(emit_agent(agent, "claude-agents-md"))           # the .md, as text
emit_agent(agent, "claude-agents-md", dest=".claude/agents/")   # write it

coact complete .claude/skills/ux-analyst --dest .claude/agents

source accepts a Skill, a path to a skill dir / SKILL.md, or a key/name resolvable in the local store or .claude/skills/.

Make the lift reproducible: the coact: block

The standard name / description frontmatter doesn't carry tools, model, MCP, or a return contract. Add an additive, optional coact: block (every other tool ignores it). When present it wins over policy; when absent coact infers and reports what it guessed.

---
name: ux-analyst
description: Analyze a captured UX evidence bundle for usability issues.
coact:
  tools: [Read, Grep, Glob]        # else inferred from the skill (a warning is recorded)
  disallowed_tools: [Bash]         # narrow the allowlist
  model: sonnet                    # sonnet | opus | haiku | inherit  (else policy routes)
  memory: project                  # user | project | local  (opt-in)
  permission_mode: default
  skills: [shared-evidence]        # extra skills the agent references (point-don't-copy)
  consumes: evidence_bundle        # input contract name (signals interdependence)
  returns:
    schema_ref: ov.schemas:UxFindings   # XOR json_schema — resolved to canonical JSON Schema now
    description: Usability findings for the bundle
  mcp:
    - module: ov.analyzers              # python tools, exposed at REALIZE-time (backend='mcp')
      functions: [score_contrast, find_tap_targets]
  persona: |
    You are a meticulous UX analyst...  # else a template persona is synthesized
---

Validate it (also runs under skill validate once coact is imported):

from coact import validate_coact_block, parse_coact_meta
validate_coact_block(skill_frontmatter_dict)   # [] = valid/absent
parse_coact_meta(".claude/skills/ux-analyst")  # -> CoactMeta

Key behaviors to remember

• No LLM by default. The whole path is mechanical. To draft a richer persona, opt in explicitly: complete(skill, llm=...) (a callable(str)->str, an aw StepConfig, or a model name). Without llm=, nothing calls a model.
• Defaults are sensible & reported. Tools/model/memory come from the coact: block if declared, else from the injected CompletionPolicy (from coact import default_policy) — and provenance tells you which.
• schema_ref is resolved at completion time to a canonical JSON Schema; an unresolvable ref is recorded as a warning, never a crash.
• mcp tools are not embedded in the definition — they're exposed at realize-time (backend="mcp"); plan warns you when a skill declares them.

Next

• coact-realize — turn the completed definition into something that runs.
• coact-analyze — diff a skill against its agent to audit the extras.