tn-verification-gate

<critical_rules>

Pre-existing failures do NOT block. If a criterion failed at T0 and still fails at TN, that is pre-existing-fail — not a regression.

Only regressions block. A criterion that passed at T0 (exit 0) and fails at TN (non-zero) is regressed. Any regressed criterion sets verdict: FAIL.

Capture stdout and stderr in full. No truncation.

item_id is required. It is needed for both reading the T0 baseline via artifact_read and registering the TN artifact via artifact_register. Accepts either a GitHub issue number (integer) or a beads nanoid string (e.g., bd-a3f8). If not provided in the delegation prompt, return STATUS: BLOCKED immediately.

Register via MCP, not filesystem. Assemble TN YAML in memory and pass it as content= to artifact_register. Do not write to ~/.dh/ paths.

</critical_rules>

Step 1: Retrieve Both Inputs

You need two inputs, each from the operation that owns it:

1. T0 baseline — an artifact registered by the T0 agent, retrieved through the artifact operations
2. Plan record — SAM-owned state, retrieved through the plan operations to re-read acceptance-criteria-structured, the same way the T0 agent read it

Your delegation prompt carries item_id and a plan address (P{N}, or the task address P{N}/T{M} whose plan component is P{N}).

mcp__plugin_dh_backlog__artifact_read(item_id={item_id}, artifact_type="T0-baseline")
mcp__plugin_dh_sam__sam_plan(plan="P{N}", config={"action": "read"})

Task plans are SAM records, never artifact-registry content. Do not attempt artifact_read(item_id, "task-plan") — nothing registers that type, and the call returns no content.

Parse the T0 baseline content returned by artifact_read as YAML to extract the T0 results.

The plan read returns an envelope: plan, gaps, warnings, source_format, source_path. Every plan field sits inside plan, never at the top level. Take the criteria from plan.acceptance-criteria-structured and the slug from plan.feature; they are structured fields, so no parsing is needed. Each criterion carries criterion-id, description, check-command, expected-baseline, and expected-final. Criteria read as absent because they were looked for at the top level would verify nothing while still reporting a passing verdict.

Never read a plan by filesystem path. The plan lives in the configured backend, which may be remote, and a path read returns nothing in a worktree-isolated dispatch.

If artifact_read returns an error or empty result for type T0-baseline, return STATUS: BLOCKED with: "T0 baseline not found — artifact_read(item_id={item_id}, artifact_type='T0-baseline') returned no content. T0 agent must run first."