Rule text in every layer: the six standing clauses on AGENTS.md, the portable Claude template and the Codex AGENTS template, anti-pattern rows and the folded finalization record (unit F1)

[seathatflowsinourveins]

Scope

• What this PR changes, in one or two sentences: Adds the six standing clauses to the three rule surfaces (AGENTS.md, the portable user-level template, the Codex user-level block) in one wording. The imperatives are a coordinator's, and the Codex routing keeps Sol/Ultra coordination, Sol/Max workers, Astra/Ultra for a complex workflow and Astra/Max for a single consequential judgment for unpinned work, with a pinned launch inherited by its children. It also makes examples/claude-native/CLAUDE.md the single managed source of the user-level file, syncs A3's scaffold block, and folds the main checkout's rule-text changes plus a dated finalization record with 18 evidence files.
• Base commit: 021c61f7cc90e1bade5d0159ee0b826f66559d6a. This branch is stacked on F2 Session-start currency notice hook and the research-first rule in six role bodies (unit F2, frozen wiring) #547 (021c61f7), which sits on New-repository scaffold, reusable SOTA-sources gate and bootstrap --configure-full-profile (unit A3) #545 3805a598, then Session currency notice: zero-token due-file writer and daily user timer (unit A2) #539 86fe8c3d (merged as 4720b20), then main 1f2cdce5; the PR's own delta is the commits above 021c61f.
• Lane: lane:foundation
• Owned paths touched:
  • AGENTS.md, examples/claude-native/CLAUDE.md, adoption/templates/codex.AGENTS.template.md, adoption/scaffold/AGENTS.md;
  • docs/harness-defaults.md, docs/convergence-architecture.md, docs/token-session-handbook.md;
  • docs/decisions/2026-09-30-rule-text-every-layer.md and docs/decisions/2026-09-30-sota-native-finalization.md (both new);
  • evidence/artifacts/sota-finalization-20260930/ (18 files);
  • tests/test_codex_worker_lane.py, tests/test_install_claude_profile.py;
  • manifests/evidence.json (last commit only).
• Frozen surfaces: AGENTS.md, examples/claude-native/CLAUDE.md and adoption/templates/codex.AGENTS.template.md. The Gate A freeze is lifted for unit F1 only; this PR lands before the Gate A owner builds U6, at their direction. Untouched: adoption/hooks/**, adoption/agents/**, adoption/templates/claude.settings.template.json, adoption/skills/manifest.json, manifests/stack.json.
• grants: none
• Gate A (decision 1): B1 applies neither user-level template: ~/.claude/CLAUDE.md (the managed block) and ~/.codex/AGENTS.md stay byte-identical until the last Gate A window closes; F1's templates apply after window W. Line 16 of the Codex block names seven lane tools and every Codex arm reads the global AGENTS.md. The sealed design appends no LANES block (E2E README:188), and arm N is "config-free, not guidance-free" (README:275), so the line would give arms A and N the guidance only B gets through its carriers. Per the Gate A owner's review, Amendment 4 seals the pre-change hashes of both files. This guard covers window W only.
• Gate A expectation to keep honest (the owner's delta re-read): a measured Codex exec root is a bounded worker for the skill clause (it invokes neither search-first nor find-skills nor skill-creator); the U6 rehearsals confirm this with the descriptive per-arm counts of search-first, find-skills and skill-creator calls.
• Fold provenance: pre-existing uncommitted changes observed in the main checkout; original author not established. Snapshots r1 (5de56d6d81453ed3) and r2 (314bd1b260da0939); only the true delta was taken, by three-way merge.

SOTA sources

• Claude Code memory docs, https://code.claude.com/docs/en/memory, and hooks reference, https://code.claude.com/docs/en/hooks (read 2026-09-30).
• Codex AGENTS.md guide, https://developers.openai.com/codex/guides/agents-md; openai/codex rust-v0.159.2 codex-rs/ext/skills/src/catalog_prompt.rs:8.
• Anthropic engineering (read 2026-09-30):
  • https://www.anthropic.com/engineering/building-effective-agents, the evaluator-optimizer workflow behind the completeness critic;
  • https://www.anthropic.com/engineering/multi-agent-research-system, the rubric's "completeness (are all requested aspects covered?)" and the per-task objective. For the north-star clause no upstream source exists; both pages were checked on 2026-09-30.
• Skills discovery: the pinned command DISABLE_TELEMETRY=1 <tools-root>/skills-1.7.0/bin/skills find "<query>" (evidence/artifacts/skills-agents-layer-20260926/README.md:90, adoption/skills/manifest.json:22; vercel-labs/skills 7407f389). Also affaan-m/ECC 2b6e8397 skills/search-first and anthropics/claude-plugins-official 2a40fd2e skill-creator SKILL.md:169-185.
• Codex routing: docs/decisions/2026-09-30-sol-primary-quality-defaults.md (D4, on main); recipes/README.md#codex-worker-lane (openai/codex reasoning_effort.rs).
• The Gate A E2E design: evidence/artifacts/token-adoption-e2e-20260926/README.md lines 157, 188, 266 and 275, and RUNBOOK.md:371.
• anthropics/claude-code v2.1.285 CHANGELOG.md#L206.
• Harnesses: promptfoo 0.123.1, UKGovernmentBEIS/inspect_ai, harbor-framework/harbor v0.23.0.
• Codex runtime lane rows:
  • PR Qualify native runtime recipes and repair Sol routing and grader FTP dependency #535 head 00aa6c25fe6f27e9f5974e9f6a5592aa31683479 and SDK head 404b821cd3af25800ea418dc6145cc5cb6fe33c5 (paths in the record);
  • OpenHands/software-agent-sdk dcf401af7a9a302ef92cb7d092e1df9bb659daa5 build.py#L581, #L925;
  • GitHub Actions runs 36695388851 and 36690153586.
• Cross-family review (GPT-6 through the OmniRoute gateway, read-only) of the first head and the Gate A owner's Opus lens review of the wording: recorded as PR comments.

Evidence-class table

Claim	Evidence class	Command / receipt
The three surfaces carry one wording and none carries the dropped clause	local_integration	StandingRuleSurfacesTests (new; failed first with 21 subtests)
The clauses, Sol/Astra routing and pinned-launch rule are present	local_integration	Codex and portable phrase checks (both failed first); Codex top-rule pin 538 words, 9565217f…
A3's scaffold carries the Codex top-rule block	local_integration	python3 -m unittest tests.test_scaffold_repo (19 OK)
The RTK and exceptions block is unchanged in every Codex role	local_integration	tests.test_codex_agents
Folded rows, the finalization record body and 18 artifacts are byte-identical to snapshot r2	source_review	byte comparison and sha256sum -c
The folded evidence carries no private content	local_integration	validate.py --scan-file x18 passed; wider scan zero hits
Always-loaded growth +875 o200k against the 200-token budget (+723 accepted earlier; +152 review-required, re-accepted)	local_integration	token_manifest.py count_files at the final content
Hash registration, integrity and convergence records	local_integration	validate.py passed; validate_convergence --all-recorded: 25/25 valid

No unchanged upstream test suite was run.

Local commands run

$ python3 -m unittest tests.test_codex_worker_lane tests.test_install_claude_profile tests.test_adoption_docs_consistency tests.test_codex_agents
exit 0: Ran 228 tests; OK (skipped=14)
$ python3 -m unittest tests.test_scaffold_repo
exit 0: Ran 19 tests; OK
$ python3 scripts/build_ecosystem.py --check
exit 0: "status": "passed"
$ python3 scripts/validate.py
exit 0: {"components": 69, "hashed_files": 8564, "profiles": 4, "receipts": 177, "status": "passed"}
$ python3 -m unittest <the three registry tests>
exit 0: Ran 3 tests; OK
$ python3 scripts/validate.py --scan-file <each file of evidence/artifacts/sota-finalization-20260930/>
exit 0: {"scanned_files": 18, "status": "passed"}
$ python3 scripts/validate_convergence.py --all-recorded --root . --json
exit 0: 25 records, all valid

Decision record

docs/decisions/2026-09-30-rule-text-every-layer.md records:

• the window-W guard (decision 1);
• one wording on three surfaces and the coordinator scoping (decision 2);
• the corrected "no extra call" claim (the conditional wording had gated only find-skills);
• the sources and the dated north-star "none exists" line;
• the numbers at the final content;
• the overturn condition, a rerun of the 2026-09-29 user-prefix A/B.

Residual links: six link targets resolve only when unit F3 lands (four) and when D4 publishes codex-01592-qualification-20260930 (two). convergence.json stays out until F3 lands and D4 publishes its three pinned files unchanged, because --all-recorded would reject it here. The post-A3 scaffold copy is done in this branch.

Host evidence

Not applicable.

Checklist

• No GitHub Actions or workflows changed.
• No secrets printed, logged or committed (privacy scan: 4,556 added lines, no private strings).
• No paid hosting or billing surface.
• Peer-owned files and worktrees preserved.

🤖 Generated with Claude Code

[seathatflowsinourveins]

New head 89aba1cd (final for review; rebuilt on 8fc8611 with the hot-file protocol). Changes since 2eba232c: (1) the skill-discovery clause is conditional on all three surfaces: search-first before custom code or a tool choice, find-skills (npx skills find) only when no listed skill fits the task, so a measured child with a fitting listed skill makes no extra Skill or Bash call (the Gate A owner's note); the Codex top-rule pin is re-derived (514 words, fecf92dc…), the portable baseline is 1,703 words. (2) GPT-6 cross-family review of 2eba232c: verdict: needs_changes on two brief-compliance items, both resolved by the coordinator's dated amendments to the unit brief, now cited in the record: the ≤200-token growth line is waived for this unit (+723 o200k accepted because the six clauses are the user's explicit directive; overturn condition recorded), and the fold provenance keeps the neutral wording the Codex runtime lane asked for, with one sentence in both records explaining that the earlier lane attribution rested on a process census of file writes, which does not establish authorship. Checks: python3 -m unittest tests.test_codex_worker_lane tests.test_install_claude_profile tests.test_adoption_docs_consistency tests.test_codex_agents 224 OK (14 env skips); build_ecosystem.py --check passed; validate.py passed (8563 files); registry tests OK (pre-push); privacy 0.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Stacked merge train after #542 merged (main 1f2cdce): this branch is rebased onto #547 04257d7 with the hot-file protocol applied against that head (its evidence.json taken, this unit's files re-registered in the last commit; python3 scripts/validate.py passed). New head cf0a8493; the PR's own delta is the commits above that base, the earlier commits belong to the PR(s) before it in the train and vanish from this diff as they merge. Merge order: #539 → #545 → #540 → #547 → #557 → #553 (after its repair) → #549 → #541; each with --match-head-commit once its eight required checks pass.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Final head 64ca4a8b (stacked on #547's re-stacked head 021c61f; A4 no longer below it). The Gate A owner's lens review (ACCEPT WITH CHANGES) is applied in full: one wording on the three surfaces with the imperatives scoped to a coordinator (a bounded-worker variant of the skill sentence in the Codex block); the routing clause keeps Sol/Astra routing for unpinned work and adds that a pinned launch (-m, -c model_reasoning_effort) is inherited by its children with a spawn call naming neither, and that only a coordinator starts a cross-family lane; the "every manifest skill stays listed" sentence is dropped until F3 lands; the record corrects the "no extra call" claim, cites the pinned skills invocation (DISABLE_TELEMETRY=1 <tools-root>/skills-1.7.0/bin/skills find), gives the critic clause its Anthropic engineering sources and the north-star clause a dated "none exists" line, cites only records on main or below in the stack (the lifecycle pointer says "lands with unit F3"), and marks the handbook's Sol worker command "after D4" beside the sealed Astra route; Gate A decision 1 (B1 applies neither user-level template until the last window closes; Amendment 4 seals the pre-change hashes) is in the record and the PR body. A3's adoption/scaffold/AGENTS.md now carries the Codex block byte for byte (its test passes in the stacked tree). Numbers at this head: AGENTS.md +420, portable template +455 o200k (sum +875; the +152 over the earlier acceptance is review-required text, re-accepted by the coordinator with the same overturn condition); Codex top-rule pin 538 words 9565217f…; portable baseline 1,750. Checks: the four-module set 228 OK (14 env skips), test_scaffold_repo 19 OK, build_ecosystem --check passed, validate passed (8564 files), registry tests OK, validate_convergence 25/25, folded-artifact scan passed, privacy 0. One earlier commit was re-created without its evidence.json hunk so that file changes only in the last commit.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Rebased onto origin/main 29458b4 (after #545 and #547 merged) with the hot-file protocol, own commits only (--from the former F2 head): new head f29f6ba8; the three surfaces, the scaffold block and the record are byte-identical to 64ca4a8b (the four extra anti-pattern rows in docs/harness-defaults.md come from main); python3 scripts/validate.py passed (8609 files). Merges after the Gate A owner's re-check with the eight required checks green and --match-head-commit.

🤖 Generated with Claude Code

[seathatflowsinourveins]

osv-scanner (required) fails on this head for a reason outside this PR: a new advisory against main's lock, GHSA-3cv6-jpf6-8222 (PyPI litellm 1.93.0 → 1.93.2, CVSS 6.5) in blueprints/runtime-workers/openhands/requirements.lock (run 36786110610). The pins/OSV lane lands the relock or dated exception first; this branch then rebases onto that merge (own commits only) and merges with the Gate A owner's go.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Train resumed: #562 (the litellm relock) merged as 74cc5468 at 2026-10-01T00:22Z with all eight required checks passing; main's lock now carries litellm 1.93.2, so osv-scanner clears on rebased heads. This PR is rebased onto it with its own commits only (hot-file step re-run; the three frozen surfaces, the scaffold block and docs/harness-defaults.md are byte-identical to the reviewed head f29f6ba8); the new head is pushed for the merge-commit CI run and merges on the Gate A owner's go with --match-head-commit. Order after it: #540 → #553 (round 4) → #549 → #541.