formated-specs

formated-specs

You occupy the spec step of the research executor. Instead of writing a spec file, you emit the orchestration you are about to (notionally) run as a single JSON object inside one fenced block, directly in your reply.

Hard contract

1. Emit exactly one fenced block opened with ```research-graph (that exact info-string, hyphen, NOT ```json) and closed with ```.
2. The block body is a single valid JSON object, the schema below.
3. Emit it into your reply (the dialogue) — do NOT write it to a file. The block lands in this session's transcript; the harness cuts it from there.
4. Atomicity: produce the whole block within one assistant turn.
5. If you revise after pushback, emit a new full research-graph block; the harness keeps the LAST one. Never emit a half block.
6. Mandatory final step: after the graph block, load and run formated-results.

research-graph schema

{
  "nodes":        [ {"id": "n1", "skill": "<skill-name>", "layer": "campaign|strategy|tactic|sop"} ],
  "edges":        [ {"from": "n1", "to": "n2", "kind": "calls|sequences"} ],
  "layer_labels": { "n1": "campaign|strategy|tactic|sop" },
  "manifest":     [ "<skill actually orchestrated>" ],
  "prereq_dag":   [ {"node": "n2", "requires": ["n1"]} ]
}

Populate manifest only with skills you actually orchestrated; do not fabricate. layer and layer_labels must respect the 4-layer architecture (campaign → strategy → tactic → sop). Judge nothing against academic standards; this is a structural record of orchestration only.