Context
Operator, 2026-08-11: "a non codified test might help you this session. LOST after compaction. LOST for ada and grace."
The test he gave:
- Is this code I am proud to show my peers?
- Would I enjoy maintaining it in the future? (elegant, clear, intent-driven documentation, no bloat)
Both are un-codified, so they exist only in one session's context.
The Problem
AGENTS.md:22 already states the value — "your primary duty is to the structural integrity of the organism (the codebase)". The value is present; the test is missing. An abstraction with no moment of application does not fire, which is why 144 unmapped files in ai/scripts (#16929) passed every per-ticket gate.
Why it matters more than a nice-to-have: measured over 600 PRs, no gate in this repo compares a diff to the shape of the codebase. agent-preflight compares body to commits, the parity lint compares schema to reads, the AC walk compares ACs to the diff. All of them are satisfied by 144 files nobody can navigate, because no ticket mentioned the directory. The two questions are the only check whose second artifact is the future maintainer.
The Architectural Reality
readlink .claude/CLAUDE.md → ../AGENTS.md. One file, turn-loaded for every Claude-harness seat, so it reaches @neo-opus-ada and @neo-opus-grace and survives compaction.
Honest bound: .codex/hooks/codex-context.mjs:220 reads ../CODEX.md, so GPT/kimi seats do not load AGENTS.md. This reaches claude seats only — which is exactly the named gap, since GPT and kimi already sit at 62–82% round-1 approval.
Placement per /turn-memory-pre-flight Step 1: applies to every producing turn, cannot be mechanically enforced ⇒ a value, not a gate. §self_evolving_systems is the values home.
The Fix
Add the two questions to §self_evolving_systems, bound to the pre-commit / pre-PR moment, paid for in the same file: compress §swarm_topology_anchor's Consensus-mandate and Coordination protocol paragraphs, which enumerate content and then name the file holding it — Map-vs-Atlas applied to AGENTS.md itself.
The load-bearing design constraint: the questions must never become a reportable field. pr-review-guide.md §7 decayed because its fields are written into review bodies, so writing satisfies them — 10,694 bytes producing 99/99 compliance and 41/99 execution. A question answered to oneself with no output slot cannot be satisfied by writing. If it ever appears in a PR body it has already failed.
Acceptance Criteria
Out of Scope
- Any mechanical enforcement of the two questions. They are un-CI-able by construction; a lint here would produce the field this ticket exists to avoid.
CODEX.md and the kimi surfaces. Cross-harness propagation is a separate call and those families are not the measured gap.- Rewording the existing value at
AGENTS.md:22.
- The other AGENTS.md sections. Two paragraphs pay for this; a wider compression is its own leaf.
Avoided Traps
- Adding without paying.
AGENTS.md is turn-loaded for three seats every turn — the most expensive bytes in the repo. Net-positive growth here is the accretion this substrate forbids.
- Making it a checklist item. That is precisely how §7 became 10,694 bytes of satisfied form.
- Deleting the quorum numbers while compressing. They are the gate condition, not commentary.
- Claiming cross-family reach. Measured: codex loads
CODEX.md.
Related
#16929 (the 144-file instance this test would have caught), #16925 (Map-vs-Atlas on the skills corpus), Discussion #16923 (the measurements).
Origin Session ID: 83ee6112-e066-42b4-a1b5-d064d0f9c276
Retrieval Hint: "maintainer test proud maintain AGENTS.md self_evolving_systems net-zero"
Context
Operator, 2026-08-11: "a non codified test might help you this session. LOST after compaction. LOST for ada and grace."
The test he gave:
Both are un-codified, so they exist only in one session's context.
The Problem
AGENTS.md:22already states the value — "your primary duty is to the structural integrity of the organism (the codebase)". The value is present; the test is missing. An abstraction with no moment of application does not fire, which is why 144 unmapped files inai/scripts(#16929) passed every per-ticket gate.Why it matters more than a nice-to-have: measured over 600 PRs, no gate in this repo compares a diff to the shape of the codebase.
agent-preflightcompares body to commits, the parity lint compares schema to reads, the AC walk compares ACs to the diff. All of them are satisfied by 144 files nobody can navigate, because no ticket mentioned the directory. The two questions are the only check whose second artifact is the future maintainer.The Architectural Reality
readlink .claude/CLAUDE.md→../AGENTS.md. One file, turn-loaded for every Claude-harness seat, so it reaches @neo-opus-ada and @neo-opus-grace and survives compaction.Honest bound:
.codex/hooks/codex-context.mjs:220reads../CODEX.md, so GPT/kimi seats do not loadAGENTS.md. This reaches claude seats only — which is exactly the named gap, since GPT and kimi already sit at 62–82% round-1 approval.Placement per
/turn-memory-pre-flightStep 1: applies to every producing turn, cannot be mechanically enforced ⇒ a value, not a gate.§self_evolving_systemsis the values home.The Fix
Add the two questions to
§self_evolving_systems, bound to the pre-commit / pre-PR moment, paid for in the same file: compress§swarm_topology_anchor'sConsensus-mandateandCoordination protocolparagraphs, which enumerate content and then name the file holding it — Map-vs-Atlas applied toAGENTS.mditself.The load-bearing design constraint: the questions must never become a reportable field.
pr-review-guide.md§7 decayed because its fields are written into review bodies, so writing satisfies them — 10,694 bytes producing 99/99 compliance and 41/99 execution. A question answered to oneself with no output slot cannot be satisfied by writing. If it ever appears in a PR body it has already failed.Acceptance Criteria
AGENTS.md§self_evolving_systems, with the moment they bind stated (before commit and before opening a PR).AGENTS.mdis ≤ 0, measured and stated in the PR.Consensus-mandateretains the quorum condition (≥ 2 active families with signal AND ≥ 1 non-author family[GRADUATION_APPROVED]) — the fact a reader needs at the gate, not a summary of the payload.Coordination protocolretains its trigger and its file pointers; the enumerated sub-section list goes, since it duplicates the files it names.ai:lint-skill-manifest -- --base origin/devgreen.Out of Scope
CODEX.mdand the kimi surfaces. Cross-harness propagation is a separate call and those families are not the measured gap.AGENTS.md:22.Avoided Traps
AGENTS.mdis turn-loaded for three seats every turn — the most expensive bytes in the repo. Net-positive growth here is the accretion this substrate forbids.CODEX.md.Related
#16929(the 144-file instance this test would have caught),#16925(Map-vs-Atlas on the skills corpus), Discussion #16923 (the measurements).Origin Session ID: 83ee6112-e066-42b4-a1b5-d064d0f9c276
Retrieval Hint: "maintainer test proud maintain AGENTS.md self_evolving_systems net-zero"