Category B: Lightweight validation for trivial/reversible workflow steps

[sansari]

Parent: #405

Problem

Full deepwork:reviewer agents are spawned for low-risk, trivially reversible intermediate steps: config files, YAML decisions files, env file generation. These steps have no analytical content — errors are immediately obvious and trivially corrected by rerunning.

Evidence

From session logs (sansari-terminal-may12.txt):

Step	Review time	Tokens	Risk level
Finder decisions YAML (5 lines)	2m 10s	31.3k	Low — just rerun
Finder decisions YAML (first pass)	22s	8.8k	Low
Finder preparation / env file	45s	12.4k	Low — trivially reversible

A misconfigured YAML is caught immediately when the Finder binary runs. An env file with a wrong value produces an obvious error on the next step. Full reviewer agents are not the right tool for this class of problem.

Proposed fix

Annotate workflow step definitions with a review_level field:

steps:
  - name: finder_config
    review_level: low   # syntax/schema check only
  - name: finder_prep
    review_level: low
  - name: analysis_report
    review_level: high  # full reviewer
  - name: present_report
    review_level: high

review_level: low triggers a fast, non-agent validation (JSON/YAML schema check, required field presence) rather than spawning a full reviewer agent. review_level: high behaves as today.

Alternatively: derive review level from step position — steps that are not the terminal step and whose output is consumed by a downstream reviewed step can default to low.

Quality preservation

• review_level: high steps (final deliverables, analysis outputs with statistical claims) are unchanged.
• Low-level steps have their errors caught either by the next step's execution (immediate feedback) or by the downstream high-level review.
• No analytical content is reviewed at reduced fidelity.