fix(quick-flow): use human-in-the-loop fallback for review without subagents

Sequential inline reviews in the same context suffer from anchoring
bias and context blowout. Instead, generate separate review prompt
files and ask the human to run each in a separate session.
This commit is contained in:
Alex Verkhovsky 2026-03-02 16:58:45 -07:00
parent f740cd2d17
commit c036e4cc60
1 changed files with 1 additions and 1 deletions

View File

@ -29,7 +29,7 @@ Do NOT `git add` anything — this is read-only inspection.
**One-shot:** Skip diff construction. Still invoke `{adversarial_review_task}` in a subagent with the changed files — inline review invites anchoring bias. **One-shot:** Skip diff construction. Still invoke `{adversarial_review_task}` in a subagent with the changed files — inline review invites anchoring bias.
**Plan-code-review:** Launch three subagents without conversation context. If no sub-agents are available, perform all three reviews inline sequentially. **Plan-code-review:** Launch three subagents without conversation context. If no sub-agents are available, generate three review prompt files in `{implementation_artifacts}` — one per reviewer role below — and HALT. Ask the human to run each in a separate session (ideally a different LLM) and paste back the findings.
- **Blind hunter** — receives `{diff_output}` only. No spec, no context docs, no project access. Invoke via `{adversarial_review_task}`. - **Blind hunter** — receives `{diff_output}` only. No spec, no context docs, no project access. Invoke via `{adversarial_review_task}`.
- **Edge case hunter** — receives `{diff_output}` and read access to the project. Invoke via `{edge_case_hunter_task}`. - **Edge case hunter** — receives `{diff_output}` and read access to the project. Invoke via `{edge_case_hunter_task}`.