Skip to content

README.md

Latest commit

 

History

History
56 lines (36 loc) · 3.14 KB

README.md

File metadata and controls

56 lines (36 loc) · 3.14 KB

Adapters

Each adapter maps the agent-agnostic kernel (kernel/) to one specific agent's conventions. The kit ships with a Claude-Code adapter today; stubs exist for Cursor, Amp, Codex.

Available adapters

Adapter Status Maps roles via Source of payload
claude-code/ Complete. reference adapter .claude/skills/ + .claude/agents/ + slash commands template/ at kit root (payload copied into target during Protocol A)
cursor/ Stub. not implemented TBD TBD
amp/ Stub TBD TBD
codex/ Stub TBD TBD

Adapter contract

An adapter is complete when:

  1. Every role in kernel/roles.md has a mapping documented in the adapter's README.md.
  2. The adapter ships a payload (equivalent of template/) that Protocol A can copy into a target.
  3. The primary agent-instructions file (equivalent of template/AGENT.md) executes the 12-step workflow in kernel/methodology.md §"The workflow".
  4. At least one target project has been unfolded with the adapter and passed the Bootstrap self-audit.

Writing a new adapter

Step 1. Read the kernel

Read kernel/README.md, kernel/methodology.md, kernel/roles.md, kernel/vocabulary.md end to end. The vocabulary is fixed; the roles are fixed; the methodology invariants are fixed. Everything else is adapter freedom.

Step 2. Map the roles

In your adapter's README.md, fill in the role-mapping table (see claude-code/README.md for the reference). Every role gets a concrete mechanism: "in this adapter, the verifier role is implemented as <path to agent file / configuration / skill>."

Step 3. Author the payload

Create a directory that Protocol A can copy into a target project. Mirror template/ but for your agent's conventions:

  • Primary instruction file (equivalent of template/AGENT.md)
  • Skill definitions (if your agent has skills)
  • Subagent definitions (if your agent has subagents)
  • The Blueprint skeletons from template/blueprint/. these are agent-agnostic; reuse verbatim
  • The scripts from template/skills/*/run.sh. these are agent-agnostic bash; reuse verbatim

Step 4. Update Protocol A

In the kit's root CLAUDE.md, Protocol A step 3, add a branch that detects your agent and copies the right adapter's payload. The kit should always detect at bootstrap time, not require the team lead to configure.

Step 5. Dogfood

Unfold a real target with your adapter. Ship at least one feature end-to-end through the workflow. File a retrospective in your target's specs/NNN-*/retrospective.md. Run tools/aggregate-retros.sh across your adapter's targets and add the corpus to docs/FIELD_NOTES.md. this is how the adapter earns its "Complete" status in the table above.

Why not just fork Claude-Code-flavored files

You could. But:

  • Every future methodology change (new role, new invariant) would then require N forks to update in lockstep.
  • Adapter contributors would have to reverse-engineer which parts are Claude-specific vs universal.

With kernel + adapter split, methodology evolves in one place (kernel/) and adapter changes are bounded to one directory. Contributors know exactly where their changes go.