Context
The review skill says "You are NOT the person who wrote the fix or the tests. You are a fresh set of eyes." In practice — especially in speedrun mode — the same agent wrote the fix, the tests, and then reviews its own work. The review cannot be truly independent.
This was identified during testing of the skill invocation cleanup (PR #107). The agent's self-review was thorough and did catch some issues, but it inherently cannot challenge its own assumptions the way an independent reviewer would.
Problem
The review skill's "fresh eyes" claim is false when the same agent runs all phases. This is always the case in speedrun mode and common in controller mode too.
Possible approaches
1. Run review in a separate agent context
The dev-team workflow uses the Agent platform tool to spawn a Checker as a separate subagent (subagent_type: "Explore", read-only). The Checker gets its own prompt and does not share the parent agent's conversation context, providing genuine independence.
The bugfix controller/speedrun could do the same for review: spawn a separate agent that reads the artifacts but doesn't have the conversational context of having written the fix.
Trade-off: This uses platform-specific tools (Agent), which conflicts with the portability principle from the skill invocation cleanup (PR #107). A more portable approach would be to describe the intent ("run the review skill in an independent context") without prescribing the mechanism.
2. Acknowledge the limitation and set expectations
Add a note to the review skill or speedrun's ### review section: "In speedrun mode, this is self-review. It can catch mechanical errors and logical gaps, but cannot replace an independent reviewer. For high-risk fixes, recommend the user seek external review."
This is less ambitious but honest, and doesn't require platform-specific tools.
3. Hybrid — acknowledge now, implement independence later
Add the acknowledgment (approach 2) as a quick fix. Track approach 1 as a follow-on once there's a portable way to express "run this skill independently."
Related
Context
The review skill says "You are NOT the person who wrote the fix or the tests. You are a fresh set of eyes." In practice — especially in speedrun mode — the same agent wrote the fix, the tests, and then reviews its own work. The review cannot be truly independent.
This was identified during testing of the skill invocation cleanup (PR #107). The agent's self-review was thorough and did catch some issues, but it inherently cannot challenge its own assumptions the way an independent reviewer would.
Problem
The review skill's "fresh eyes" claim is false when the same agent runs all phases. This is always the case in speedrun mode and common in controller mode too.
Possible approaches
1. Run review in a separate agent context
The dev-team workflow uses the
Agentplatform tool to spawn a Checker as a separate subagent (subagent_type: "Explore", read-only). The Checker gets its own prompt and does not share the parent agent's conversation context, providing genuine independence.The bugfix controller/speedrun could do the same for review: spawn a separate agent that reads the artifacts but doesn't have the conversational context of having written the fix.
Trade-off: This uses platform-specific tools (
Agent), which conflicts with the portability principle from the skill invocation cleanup (PR #107). A more portable approach would be to describe the intent ("run the review skill in an independent context") without prescribing the mechanism.2. Acknowledge the limitation and set expectations
Add a note to the review skill or speedrun's
### reviewsection: "In speedrun mode, this is self-review. It can catch mechanical errors and logical gaps, but cannot replace an independent reviewer. For high-risk fixes, recommend the user seek external review."This is less ambitious but honest, and doesn't require platform-specific tools.
3. Hybrid — acknowledge now, implement independence later
Add the acknowledgment (approach 2) as a quick fix. Track approach 1 as a follow-on once there's a portable way to express "run this skill independently."
Related
workflows/dev-team/.claude/skills/dev-team/SKILL.md)