research skill: add anti-recursion guidance for agents that are the delegate - #72
research skill: add anti-recursion guidance for agents that are the delegate#72warp-agent-staging[bot] wants to merge 1 commit into
Conversation
…elegate
Companion to a warp-server change restricting/discouraging further
delegation. The skill's trigger ("would require reading many files...
reach for it liberally") reads identically whether the reader is the
one asking for research or the one who was asked, so a subagent
already acting as the investigator would re-read this skill and fan
out to its own sub-subagents on the same question.
Adds a dedicated section telling an agent to check which role it is
in before delegating, plus a reinforcing bullet in the do-NOT-delegate
examples. The skill's real use case (delegating noisy investigation
from a position where you are NOT already the delegate) is unchanged.
QUALITY-1646
Co-Authored-By: Warp Agent <agent@warp.dev>
warp-agent-staging
Bot
added
the
factory:wilson
|
@warp-agent-staging[bot] I'm starting a first review of this pull request. You can view the conversation on Warp. I completed the review and no human review was requested for this pull request. Comment Powered by Oz |
![warp-for-oss[bot]](https://avatars.githubusercontent.com/in/3450139?s=60&v=4){kind=small}
There was a problem hiding this comment.
Overview
This PR adds explicit anti-recursion guidance to the shared research skill, clarifying that agents already assigned to investigate and report should usually perform the reading themselves rather than recursively delegating the same investigation.
Concerns
- No blocking concerns found. The change is limited to skill documentation/prompt guidance, is consistent with the surrounding delegation guidance, and the attached spec context reports no approved or repository spec to validate against. No repository-local security-review companion skill was present in this checkout; the general security pass found no security-relevant concerns.
Verdict
Found: 0 critical, 0 important, 0 suggestions
Approve
Comment /oz-review on this pull request to retrigger a review (up to 3 times on the same pull request).
Powered by Oz
2 participants
Summary
Adds anti-recursion guidance to the
researchskill. Reference: https://linear.app/warpdotdev/issue/QUALITY-1646/child-agents-re-read-research-and-recursively-spawn-sub-agents-tokenThe skill's trigger ("would require reading many files... reach for it liberally") reads identically whether the agent reading it is the one asking for research or the one who was asked. A child agent whose own brief was already "investigate X and report back" would re-read
/researchand fan out to its own sub-subagents on the same question, burning tokens without isolating any new noise (a real incident: a depth-1 factory child alone ran ~61 requests, ~7.8M input tokens, ~$39, plus grandchildren on top).Companion PR in
warpdotdev/warp-server(factory/research-recursion-guardrails) makes the same point from the other side: it makes an agent's remaining orchestration depth budget carry actual weight against this pull, since today those depth facts are present in the prompt but inert. The two are coordinated to say the same thing.Changes
Verification
This is a documentation/prompt-wording change to a markdown skill file; there is no automated test harness for skill content in this repo. Verified by re-reading the rendered skill end-to-end for internal consistency and confirming it doesn't contradict or duplicate the existing "do NOT delegate" guidance.