SuperAPI AI Directory

synthesis-agent-conformance

Agents & models

View repository

Updated 10-22 16:15

About this skill

Synthesis Agent Conformance

Treat cross-agent portability as a continuously tested system, not a file-count comparison. This skill audits five planes (source, installed, live, continuity, capability) with the v5 tools, repairs from source, and keeps always-loaded instructions small enough for every harness.

Binding rules

  1. Parity is tested, never inferred. Matching inventories or version labels are not parity; a self-report is a claim, so compare installed bytes and read each harness's own answers.
  2. Name the plane when facts conflict. Runtime state is current behavior; source is what the next deployment produces.
  3. UNKNOWN never becomes PASS. A check that could not run is reported as unknown, with what it needed.
  4. Never write a harness's trust state. Codex /hooks trust and Muse hook approval are the person's decisions; report what needs approving.
  5. Repair from source, never in an installed copy or cache. Keep shared behavior agent-neutral and put harness differences in adapters.
  6. AGENTS.md is canonical; CLAUDE.md is exactly @AGENTS.md. Private skills go to ~/.claude/skills and ~/.agents/skills, never ~/.codex/skills.
  7. Respect Codex's budgets. Instructions stay under project_doc_max_bytes minus 4 KiB, and the skill catalog within 2% of context, with specialists explicit-only and the router routing.
  8. Follow a generator-backed artifact only when it is verifiably the generator's output. A hand-made decision packet skips what the generator guarantees.
  9. When a desktop client cannot read a protected file, find the engine responsible and report UNKNOWN until an observed check shows otherwise.

Contents

Procedure

  1. Source. In the source checkout: python3 -m pytest -q tests/test_source_lint.py. Every failure names the file or skill to fix.
  2. Installed. synthesis doctor (add --latest to compare with the published release). The first line says healthy or how many problems; each later line is ok, warn, fail or info, a check name, and what it found. Fix every fail from source, then rerun.
  3. Live. Start a fresh session in each harness and confirm its SessionStart context names the active project. Ask the person to trust changed hooks in Codex's /hooks and to run muse plugins approve synthesis-skills when doctor lists pending hooks.
  4. Continuity. synthesis use <project> and synthesis brief in one harness; open another harness with no pasted transcript and confirm the same phase, plan and next action. For another Mac, synthesis handoff pushes the records in this session's claims.
  5. Instructions. For each folder doctor lists under instruction adapters, make AGENTS.md canonical and CLAUDE.md exactly @AGENTS.md; never overwrite a divergent pair blindly, merge it in source first.
  6. Close. Record genuine harness differences as boundaries with evidence; report parity only when every plane passed.