[Lutherwaves]

Stacked on #36. A fix at the cause, a change with nothing in it that the task did not need, and the same flow on five harnesses.

What a person can now do

• Fix a bug at its cause. The new debug skill reproduces the failure, states the root cause in one sentence, pins it with a failing test and fixes the cause. It changes one thing at a time and stops after three fixes that did not hold. build loads it when a test fails.
• Get a change without slop. Before the last test run, build removes what its own change did not need: a comment that restates the code, a guard for a case that cannot happen, a helper with one caller, an option nobody asked for, filler in prose. review lists the same things under a new "Slop" heading. Slop is not a finding and does not touch the gate.
• Use builder on Grok Build. grok is a harness with its own action map and install steps.
• Get the right instructions on Codex. The session start printed the Claude Code action map there. The hook now tells Grok Build, Claude Code and Codex apart by the variable that each one sets for a plugin hook.
• Follow a handoff on every harness. A stage hands over with builder:<name>. On Cursor CLI, Grok Build and opencode a skill has its plain name, so the session start now says what builder:<name> means there.
• Run the security score without the hook. The skill says where the helper binary is when no session-start hook ran, and to stop when it is not there.
• Get an architecture review in the plan. For work on an API, a service boundary, a data store or an agent tool, plan runs an architecture lens (skills/plan/architecture.md) and review checks the change against the same questions. The project's own rules win over it. Tenet 9 now says what a good guard does.

Proven

• go test ./..., bash scripts/validate.sh and bash tests/test_validate.sh pass.
• With PLUGIN_ROOT set the hook prints the Codex map, with GROK_PLUGIN_ROOT the Grok Build map, with neither the Claude Code map.
• Two of the new tests fail when I remove the rule they guard (the Slop heading, the builder:<name> line of a map).
• The sample test of the debug fixture fails before any fix.

Not proven

• No session ran on Codex, Cursor CLI, Grok Build or opencode. The two variables come from the Codex and xAI documents.
• Six of seven actions on Grok Build are marked "not verified". Grok Build has no eval runner and no row in the support table.
• No eval ran: the owner ruled out paid evals. The debug fixture is written and not graded.

The design record is docs/specs/2026-11-12-parity-design.md. Four of its rulings are proposed, not confirmed.

🤖 Generated with Claude Code