You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
#2338 established that skills: preload does not reach dispatched discovery:researcher agents, and PR #2374 fixed the consequence: the agent now recovers by reading SKILL.md (+ context files) from disk, so the discipline runs either way.
That fallback creates a new problem: the recovered agent echoes the same preload_token a preloaded agent would, because the token is in the file it just read. A matching sentinel in the return payload no longer proves preload fired — it proves only that the agent saw the file by some route. Any parent or gate treating the token as preload evidence is unsound.
Evidence
Nine independent discovery:researcher dispatches on 2026-08-17 (CLI v2.1.232, this repo): 9/9 preload failures, 9/9 successful fallbacks, 9/9 matching tokens. Every agent disclosed the provenance — but only as free prose in open_questions, e.g.:
"The sentinel is echoed above because I read it from the skill file, not because it was preloaded. If the parent treats a matching token as proof the preload worked, that inference is wrong for this run."
Suggested direction
Make provenance structured, not prose: a preload: fired | fallback field in the return payload contract, graded by the parent, with the bare token demoted to file-identity evidence only.
Separately, the underlying preload failure from discovery:researcher skill preload does not reach dispatched agents #2338 is still live at v2.1.232 and now has a 9-run reproduction; worth deciding whether to reopen upstream investigation (agent-frontmatter skills: on the dispatch path) or record it as accepted-with-fallback.
What happens
#2338 established that
skills:preload does not reach dispatcheddiscovery:researcheragents, and PR #2374 fixed the consequence: the agent now recovers by readingSKILL.md(+ context files) from disk, so the discipline runs either way.That fallback creates a new problem: the recovered agent echoes the same
preload_tokena preloaded agent would, because the token is in the file it just read. A matching sentinel in the return payload no longer proves preload fired — it proves only that the agent saw the file by some route. Any parent or gate treating the token as preload evidence is unsound.Evidence
Nine independent
discovery:researcherdispatches on 2026-08-17 (CLI v2.1.232, this repo): 9/9 preload failures, 9/9 successful fallbacks, 9/9 matching tokens. Every agent disclosed the provenance — but only as free prose inopen_questions, e.g.:Suggested direction
preload: fired | fallbackfield in the return payload contract, graded by the parent, with the bare token demoted to file-identity evidence only.skills:on the dispatch path) or record it as accepted-with-fallback.Related
docs/topics/context-budget/FINDINGS.md— "Corrections owed to this repository" item 5 records the 9/9 observation