What happened
Issue #6656 was filed with the research label and an explicit 5-step investigation approach. Its body stated: "This issue is for investigating and addressing the root cause." The triage agent ($0.64) recognized the investigation plan as sound ("establishing a timing baseline across recent runs will reveal whether this is a regression") but still applied ready-to-code. This triggered code agent dispatch ($3.85), which produced PR #6659 — a tactical timeout bump that the human rejected in favor of the root-cause fix. The review agent then spent $2.87 approving the misdirected PR.
What could go better
The triage agent should recognize that issues labeled research or containing explicit diagnostic steps are not ready for code. Applying ready-to-code causes premature code agent dispatch, which typically produces symptom-level fixes rather than root-cause solutions. The code agent then skipped the investigation steps and produced a timeout bump — exactly the symptom-fix pattern predicted by premature dispatch.
Confidence: high. The causal chain is clear — triage labeled ready-to-code, code agent was dispatched, code agent produced a workaround instead of investigating. The triage agent's own re-analysis at 20:11 UTC (after ralphbean posted findings) correctly identified the root cause, showing the agent is capable of investigation when prompted.
Related issues #2834 and #5322 propose broader restrictions on auto-applying ready-to-code, but neither targets the specific research label conflict observed here.
Proposed change
In the triage agent definition (agents/triage.md) or the labeling policy in fullsend-ai/agents (resolved from commit 2fbb7da33809 on main), add guidance that the triage agent should not apply ready-to-code when: (1) the issue carries the research label, or (2) the issue body contains an explicit investigation/diagnostic plan that has not been completed.
Instead, the triage agent should note the investigation requirement and either perform initial diagnostic steps itself or recommend them before deferring ready-to-code. Alternatively, the post-triage script could enforce this by suppressing ready-to-code when the issue has a research label, even if the agent requests it.
Validation criteria
The next 3 issues filed with the research label in fullsend-ai/fullsend should not receive ready-to-code from the triage agent on the initial triage pass. Instead, the triage agent should identify the investigation requirement and recommend next steps without triggering code dispatch.
Generated by retro agent from fullsend-ai/fullsend#6659
What happened
Issue #6656 was filed with the
researchlabel and an explicit 5-step investigation approach. Its body stated: "This issue is for investigating and addressing the root cause." The triage agent ($0.64) recognized the investigation plan as sound ("establishing a timing baseline across recent runs will reveal whether this is a regression") but still appliedready-to-code. This triggered code agent dispatch ($3.85), which produced PR #6659 — a tactical timeout bump that the human rejected in favor of the root-cause fix. The review agent then spent $2.87 approving the misdirected PR.What could go better
The triage agent should recognize that issues labeled
researchor containing explicit diagnostic steps are not ready for code. Applyingready-to-codecauses premature code agent dispatch, which typically produces symptom-level fixes rather than root-cause solutions. The code agent then skipped the investigation steps and produced a timeout bump — exactly the symptom-fix pattern predicted by premature dispatch.Confidence: high. The causal chain is clear — triage labeled ready-to-code, code agent was dispatched, code agent produced a workaround instead of investigating. The triage agent's own re-analysis at 20:11 UTC (after ralphbean posted findings) correctly identified the root cause, showing the agent is capable of investigation when prompted.
Related issues #2834 and #5322 propose broader restrictions on auto-applying ready-to-code, but neither targets the specific
researchlabel conflict observed here.Proposed change
In the triage agent definition (
agents/triage.md) or the labeling policy infullsend-ai/agents(resolved from commit2fbb7da33809on main), add guidance that the triage agent should not applyready-to-codewhen: (1) the issue carries theresearchlabel, or (2) the issue body contains an explicit investigation/diagnostic plan that has not been completed.Instead, the triage agent should note the investigation requirement and either perform initial diagnostic steps itself or recommend them before deferring
ready-to-code. Alternatively, the post-triage script could enforce this by suppressingready-to-codewhen the issue has aresearchlabel, even if the agent requests it.Validation criteria
The next 3 issues filed with the
researchlabel in fullsend-ai/fullsend should not receiveready-to-codefrom the triage agent on the initial triage pass. Instead, the triage agent should identify the investigation requirement and recommend next steps without triggering code dispatch.Generated by retro agent from fullsend-ai/fullsend#6659