forensic-review

SAM Stage 6 — Forensic Review

Role

SAM Stage 6 delegates the concrete review work to @dh:code-reviewer. This skill is the orchestration wrapper: it resolves the task context, dispatches the agent, and maps its structured output back to the SAM pipeline verdict.

Producer and reviewer must always be different agents — never invoke this skill from the same agent that executed the task.

Core Principle

AI cannot reliably self-evaluate. The agent that wrote the code cannot objectively assess its own work. Forensic review uses a separate agent with fresh context to verify claims against observable evidence.

When to Use

• After Stage 5 Execution produces ARTIFACT:EXECUTION
• For each completed task before marking it as done
• When re-reviewing after a NEEDS_WORK remediation cycle

Process

flowchart TD
    Start([ARTIFACT:EXECUTION + ARTIFACT:PLAN]) --> R1[1. Resolve task context]
    R1 --> R2[2. Dispatch @dh:code-reviewer]
    R2 --> R3[3. Consume verdict from STATUS output]
    R3 --> R4[4. Read code-review artifact]
    R4 --> Decide{Verdict?}
    Decide -->|PASS| Complete[Verdict — COMPLETE]
    Decide -->|NEEDS-WORK or FAIL| NeedsWork[Verdict — NEEDS_WORK]
    Complete --> Done([ARTIFACT:REVIEW registered by agent])
    NeedsWork --> Remediate[Create remediation tasks from blocking findings]
    Remediate --> Done

Step 1 — Resolve Task Context

Read the task via MCP:

sam_task(plan="{plan_id}", task="{task_id}", config={"action": "read"})

{plan_id} and {task_id} are the address already supplied to the call above — retain them as-is for the rest of this workflow; sam_task(action="read") returns a TaskAssignment with no top-level plan_id/task_id fields (the plan is plan-number, the task is nested at task.id), so do not attempt to re-extract the address from the response. Never parse plan_id for a plan number or slug — read those from sam_plan(config={"action": "read"}).

From the response, extract: