firehose> #llmops

Context-Independent Review

An agent is a bad reviewer of code it just wrote — not because of what it knows, but because of what it has already committed to. The source's diagnosis is about authorship, not capability: "agents are often really bad at editing code or improving code they've just written… because they wrote it, so they just think, 'okay, that's fantastic, that's fine.'" The fix is cheap and structural: don't ask the writing agent to review; spawn sub-agents whose context window never contained the authoring. Same model, same prompt quality — but the reviewer arrives at the diff without the trajectory of decisions that produced it, and so has nothing to defend.

This names a third independence axis, distinct from the two already in the vault. Authority-Independent Verification buys independence along rank (no node is exempt because it is senior). Cross-Model Independence buys it along provider lineage (a checker from the same model shares the author's blind spots). This page buys it along context: the reviewer is the same model and the same lineage, but has not seen — and therefore is not anchored by — the authoring history. The three are orthogonal and stack; context independence is the cheapest and the weakest.

The boundary matters, because the cheapness is seductive. A fresh window removes authorship anchoring; it does nothing about model blind spots. A same-model reviewer with a clean context will still fail to see the class of error its training cannot represent — it just won't be additionally motivated to bless it. So the honest reading is: context independence is a real, near-free improvement over self-review, and not a substitute for a genuinely uncorrelated checker. The source claims fresh-context sub-agents do "a much better job"; the vault should hold that as better-than-self-review, not as verification solved.

In the source's flow the reviewer is also given what to check against, which is what stops it being theater: /code-review runs on two axes — the diff against the original spec (catching things the tickets left unspecified or the agent forgot) and the diff against the repo's own standards docs, falling back to Fowler's code smells where a repo documents none. On screen the implement skill runs typecheck, test, and build, then announces: "I'll pin the fixed point and gather the diff, then spawn the two review sub-agents in parallel."

Claims


Linked from