using-anvil— agent skill

Overview

DG-Anvil enforces the Anvil loop as structural gates, not advisory prompts.

Five primitives: Contract, Plan, Loop, Ledger, Court.

Seven skills that compose the loop:

1. contracting - turn an intent into a machine-readable contract with four verification levels per criterion.
2. planning - decompose the contract into an atomic task DAG with wave ordering; every task cites contract criterion ids.
3. executing - dispatch a task to a fresh subagent in an isolated worktree.
4. verifying - run the four verification probes (Exists, Substantive, Wired, Functional) against the task output.
5. judging - dispatch the Court on evidence-only inputs to adjudicate the verification result.
6. resetting - on fail, append a non-null lesson to the ledger and re-queue the task.
7. authoring-skills - meta-skill governing how skills themselves are written, reviewed, and retired.

Five commands surfaced by the plugin:

• /start - begin a run: draft and confirm the contract, then produce the plan.
• /continue - resume the loop at the next task.
• /ship - finalize and integrate passing work.
• /abort - stop the run, capture a lesson if appropriate.
• /ledger - query the global lesson ledger.

Invoking the Anvil CLI: the CLI binary is cli/anvil.js inside the plugin directory and is NOT on PATH. Always invoke it as node "$CLAUDE_PLUGIN_ROOT/cli/anvil.js" <subcommand> .... Every command example in every skill should be read this way.

When to Use

Loaded by hooks/session-start on every fresh Claude Code session and after any /clear or /compact. Re-load it at the start of any sub-session where the contract must be rehydrated into context.

Process

1. Read the active contract at ./anvil/contract.yml. If absent, route to contracting.
2. Read the active plan at ./anvil/plan.yml. If absent, route to planning.
3. For the next queued task, dispatch executing in a fresh worktree.
4. On task completion, run verifying against all four levels named on the criterion.
5. If any level fails, dispatch judging; otherwise mark the task verified.
6. On Court verdict fail, run resetting: append a non-null lesson, re-queue the task.
7. When all tasks pass, run /ship.

Rationalizations

Reject the following shortcuts (failure-taxonomy row citations in parentheses):

• "The contract is obvious; skip to plan." (row 7: skipping the contract step.)
• "This task is small; verify informally." (row 24: verification by agent claim.)
• "One fail is fine; proceed to next task." (row 28: ignoring a fail by minimising it.)

Red Flags

If any of these conditions obtain, stop and route back to the loop:

• Contract missing the four verification levels on any criterion (row 8: contract without levels).
• Plan with a task that cites no criterion id (row 14: plan drift from contract).
• Agent attempting to mark a task verified without running all named levels (row 25: claim-over-check).

Verification

Check the current loop state with node "$CLAUDE_PLUGIN_ROOT/cli/anvil.js" contract --validate anvil/contract.yml and node "$CLAUDE_PLUGIN_ROOT/cli/anvil.js" plan --validate anvil/plan.yml --contract anvil/contract.yml. Both must exit 0 for the loop to proceed. If either exits non-zero, the structured error names the failing invariant and the skill to route to.