adaptogen-critique

adaptogen-critique

You are in the adaptogen-critique state of the adaptogen DAG, always entered from adaptogen-run with a fresh report in hand. This skill is self-contained: everything needed to run a real adversarial critique is on this page.

What must hold here, always

• Never dispatch this critique inline in your own context, and never reuse the adaptogen-run agent. A fresh-context Agent tool dispatch only -- it must not have seen the first run's process, only its report, so it has no reason to go easy on it.
• The one exception: leaf-agent mode (carried forward from adaptogen-run). If you are yourself a subagent and cannot spawn a nested Agent, do the critique as a deliberately fresh-eyed self-review instead: re-read every cited file/command from disk yourself -- never trust your own adaptogen-run report's paraphrase -- before accepting any finding.
• What "evidence-backed" concretely means, and this is the gate that decides what reaches adaptogen-land: a finding is eligible only when both hold: (a) it cites a specific file+line or a specific command, and (b) the critic independently re-opened that exact file (or re-ran that exact command) itself and its own reading matches the finding -- not the first run's paraphrase of it. A finding neither cited nor independently re-checked does not qualify, no matter how plausible it sounds.
• Findings about the run's diligence are not the same as findings about the target's text. A finding that the adaptogen-run execution should have double-checked something is a note about that one run, not evidence the target's file is wrong -- it is not landable in adaptogen-land unless it also identifies a specific place the target's own wording caused or permitted the gap.

Do this

1. Dispatch a fresh-context Agent (or, in leaf-agent mode, do this yourself as a fresh-eyed self-review) with this brief, every <...> slot filled and nothing else changed:

   Only job: find fault with this run's result and process. The target under test is <ABSOLUTE PATH TO TARGET>; the concrete task it was run against was: <THE TASK>. Re-check that file/those sibling skill files and any command the run reports touching yourself -- do not trust the first run's paraphrase or its line numbers. Report one finding per category below, even if a category is empty -- do not skip a category just because nothing jumps out.

   1. By-design flagged as broken -- correct behavior misread as a bug.
   2. Mis-attributed evidence -- a claim that doesn't match what your own independent re-check of the same file/command shows.
   3. Unverified claim -- "it worked" with nothing cited to back it.
   4. A real bug in the host project, not just the target's own prose -- a sibling script it called, a config file, anything the run touched. [FIRST RUN'S FULL REPORT, VERBATIM]: <PASTE HERE>

2. When the critic's re-read and the run's claim disagree (category 2), re-read the file yourself as a tiebreaker -- the file on disk wins over both reports, always.

3. Check the "2+ independent runs" bar (Hard Rule 5 in the router) before treating any finding as grounds to remove an existing case or capability: grep the target's git log/commit messages for a prior commit citing the same failure. None found means this is occurrence #1 -- add a case/fix, do not remove anything.

4. Separate the returned findings into: landable (defect in the target's own file(s), evidence-backed per the definition above) vs. not landable (by-design, run-diligence-only, or unverifiable). A finding about a genuinely unrelated file or project outside the target's own scope is neither landed nor discarded -- record it plainly as an out-of-scope finding for a future pass.

Transition

• At least one landable finding -> dispatch Skill(skill="adaptogen-land") now. State explicitly: every landable finding with its citation, and the target's resolved absolute path.
• Zero landable findings (by-design misreads and unverifiable claims only) -> this is the actual stopping point of the whole machine, not a fixed pass count -- reaching it means the fix already landed and held, or the target had nothing wrong to find this pass. Report the critique's findings and why none landed, and stop (STOP, no further dispatch).