BMAD-METHOD/_experiment/runs/2026-02-22-add-plan-review-...

47 lines
2.7 KiB
Markdown

# QD2 Run: Add Plan Review Step to Task-01
**Date:** 2026-02-22
**Workflow:** quick-dev2 (experimental)
**Branch:** exp/quick-flow-redesign
---
## Intent
User invoked `/bmad-bmm-quick-dev2` then pointed at `_experiment/planning/roadmap/task-01-test-skeleton.md` and requested: "Add an instruction to this step to run a review of the test result against the plan file under experiment directory."
## Routing
- **Route chosen:** One-shot (implicit — agent acted immediately without explicit routing)
- **Rationale:** Single file edit, clear intent.
## What Happened
Another one-shot. The agent captured the intent — add a review step that compares run results against the plan — but **implemented it wrong** in two ways:
### Error 1: "Adversarial review of test findings against the plan file"
The agent wrote: _"Run an adversarial review of the test findings against the plan file."_
This is incoherent. If it's an **adversarial review**, the target should be an artifact produced by the run — the diff, the code changes, the spec. An adversarial review operates on deliverables, not on findings (which are themselves review output).
### Error 2: Wrong framing of what gets compared to the plan
If the goal is to compare something **against the plan file**, the right input is the **run results** — what the agent actually did (its routing decisions, its intent capture, its behavior at each step) — not the "test findings." The plan describes intended workflow behavior; you'd check whether the agent's behavior matched the plan's design intent.
### What Was Actually Requested
A step that reviews the **results of the QD2 test run** (what happened, what the agent did) against the **plan file** (`_experiment/planning/redesign-plan.md`) to identify where behavior diverged from design. This is a conformance check, not an adversarial review.
## Diff Produced
- Added method step 5: adversarial review of findings against plan (wrong framing)
- Added two output classifications: **Plan Gap** and **Execution Gap** (reasonable categories, but derived from the wrong framing)
## Observations
- Intent capture succeeded directionally — the agent understood "review against plan file" and correctly located the plan path.
- The implementation conflated two distinct review types: adversarial review (attacks an artifact for flaws) vs. conformance review (checks behavior against a specification).
- The agent did not ask any clarifying questions about what "review of the test result against the plan file" meant — it assumed and got it wrong.
- This is the second consecutive one-shot where the agent captured the gist but mangled the specifics. Pattern: one-shot route works for mechanical changes but fails when the intent requires domain understanding of review methodology.