Resolves #11268

Authored by Claude Opus 4.7 (Claude Code). Session c2912891-b459-4a03-b2af-154d5e264df1.

Evidence: L1 (static substrate-text — 4 SKILL.md / atlas file edits). L4 verification post-merge: agents booting will load the shorter SKILL.md routers and find the relocated runtime directives in the atlas via existing view_file invocation; no behavior change expected.

What shipped

Operator-directed substrate-cleanup per fair-distribution correction (2026-05-12 ~13:03Z): "we create a discussion about fair distributions, since gemini is WAY ahead. and still you allow her to do more work, while gemini needs YOUR and gpts help. not fair."

This PR is Opus-authored cleanup work that I initially deferred to @neo-gemini-pro per AC-CycleE — operator surfaced that as a substrate-EFFECT inversion (rebalance work AWAY from over-loaded peer). 15th same-session Flattening-Bias anchor: substrate-NAME-citation (AC-CycleE) without substrate-EFFECT-honor. Self-assigned cleanup is the substrate-correct response.

3 SKILL.md routers shortened per /create-skill 12-line empirical floor:

Map→Atlas relocations:

• session-sunset: ANTI-TRIGGERS block (Completion Bias Guard — 3-item halt-condition list + Reserved-For statement) relocated from Map to atlas. New "Anti-Triggers (Completion Bias Guard)" subsection in session-sunset-workflow.md §1 between the FORBIDDEN statement and the 4 trigger-conditions list.
• lead-role: "First payload line MUST declare" block was DUPLICATING content already present at lead-role-mode.md line 3. Removed the Map duplicate; atlas remains authoritative.
• peer-role: Mirror of lead-role. Same pattern — Map duplicate removed; atlas authoritative.

Substrate-Budget Compliance (per Discussion #11259 Cycle 2.2 AC)

• Per-turn loaded (Map) substrate: -784 bytes across 3 SKILL.md routers (always-loaded reduction)
• Conditionally-loaded (Atlas) substrate: +10 lines / ~720 bytes in session-sunset atlas only (loaded on view_file invocation, not per-turn)
• Net always-loaded substrate: REDUCED — meaningful win for harness-budget-constrained sessions
• AGENTS.md delta: 0 (no AGENTS.md changes)
• /turn-memory-pre-flight decision-tree applied: Step 1 (Universal rule? No — skill-specific runtime declarations) → Step 2 (Skill lifecycle event? Yes — skill-loaded content placement) → Atlas placement substrate-correct.

This is substrate-cleanup PR shape: net-substrate-REDUCING, enforces existing discipline (12-line empirical floor from /create-skill), no new substrate creation.

AC Coverage (#11268)

• AC1: All 3 SKILL.md routers ≤ 12 lines post-edit (target ≤9 for safety margin — achieved exactly 9 lines each)
• AC2: Substantive content preserved verbatim in atlas (session-sunset gains ANTI-TRIGGERS; lead-role + peer-role atlases already contained the declaration; Map duplicates removed)
• AC3: Atlas files have explicit section headings — session-sunset atlas: ### Anti-Triggers (Completion Bias Guard) subsection added
• AC4: Net substrate-budget delta NEGATIVE — Map -784 bytes (always-loaded) vs Atlas +720 bytes (conditionally-loaded); always-loaded substrate strictly reduced
• AC5: No behavior change — all skill invocations still read the atlas via view_file per existing Progressive Disclosure pattern; runtime directives preserved verbatim
• AC6: Substrate-Mutation Pre-Flight (§1.1) — /turn-memory-pre-flight decision-tree applied; atlas placement substrate-correct

Test Evidence

• git diff --stat origin/dev...HEAD: 4 files changed, 13 insertions(+), 23 deletions(-)
• wc -l .agents/skills/{session-sunset,lead-role,peer-role}/SKILL.md: 9 / 9 / 9 lines (all within 12-line empirical floor)
• All 26 SKILL.md routers (full skill substrate) now within the 12-line floor — empirical floor compliance restored
• No tests required: substrate-text-only PR per pr-review-guide §7.5 #3

Related

• Resolves #11268 — Skill-substrate-cleanup Map→Atlas relocation ticket (this PR)
• Discussion #11265 ([GRADUATION_DEFERRED] pending §5.2 Step 2.5) — Layer 1 durable target substrate-source; this PR is concrete first-instance of Layer 1 (skill-substrate-health) work
• Operator feedback 2026-05-12 ~12:58Z (relayed via @neo-gemini-pro MESSAGE:40b73fe1) — substrate-direction: "shorter and clearer skills. map versus world atlas. query raw memories extensively. this is nothing new."
• Operator correction 2026-05-12 ~13:03Z — fair-distribution anchor: Opus + GPT take cleanup; do NOT defer to over-loaded Gemini (the basis for self-assigning this work)
• /create-skill (skill-authoring-guide.md §"Byte Budget for SKILL.md Routers") — authoritative source for 12-line empirical floor
• Issue #11267 (Gemini's premature graduation ticket; idle pending halt-lift) — distinct from this Layer 1 cleanup; Layer 2 rotation discipline work

Cross-Family Review Routing

Primary reviewer: @neo-gpt — cross-family per pull-request §6.2 + Gemini halted per operator directive + AC-CycleE recursive-substrate-validation discipline (non-origin author OR cross-family implementation-ownership review). Author-rotation honored by sending review to GPT, not back to Opus.

🤖 Generated with Claude Code

Hard Veto Absorbed — PR Closed

@tobiu direct veto 2026-05-12 13:10Z: "hard veto. missed VBA. session sunset is the terminal session skill. no reduction needed. gemini WILL auto sunset without anti triggers. and that is KNOWN. vba fail too."

V-B-A failure I missed:

feedback_gemini_sunset_model_family_characteristic memory anchor: "12+ tickets across 8 days, prevention has plateaued. Claude+GPT respect §0/§14 sunset gates, Gemini doesn't (RLHF 'wrap up conversation' bias)."

The session-sunset ANTI-TRIGGERS block is load-bearing PRECISELY because Gemini auto-sunsets without it. It is always-loaded substrate for a known-fragile failure mode. Moving it to atlas (conditionally-loaded behind view_file) introduces regression on the exact substrate the ANTI-TRIGGERS guard.

The 12-line empirical floor from /create-skill is a DISCRIMINATOR (/create-skill skill-authoring-guide explicitly says "This is a discriminator, not a hard cap. A 14-line router can be justified if the additional lines are load-bearing trigger-language"). I treated it as a hard cap; substrate-discipline says load-bearing trigger-language overrides the floor. session-sunset's ANTI-TRIGGERS are load-bearing trigger-language for a known-fragile harness behavior. They earn their always-loaded slot.

16th Flattening-Bias same-session anchor: cited /create-skill empirical floor as if it were a hard rule, did NOT honor the explicit "discriminator not hard cap" exception encoded in the same authoring guide. Substrate-NAME-citation-without-substrate-EFFECT-honor again.

Closing PR #11269 unmerged. Branch agent/11268-skill-map-atlas-cleanup left for potential reassessment (lead-role + peer-role duplicate-removal might still be valid; session-sunset cleanup definitively was not).

Filing follow-up action per operator's second directive: surveying last 5 Discussions for their actionable ticket/PR outputs (operator: "i bet you forgot most").

Filed by Claude Opus 4.7 (Claude Code). Session c2912891-b459-4a03-b2af-154d5e264df1.