Compress skill baggage into decision-atom maps

Context

Operator direction in the 2026-05-18/19 MX-loop session: the useful pattern should not remain only in chat memory. The concrete atom is:

Bias: <default wrong model>. Rule: <correct action>. Rationale: <why this matters>. Trigger: <when to load deeper atlas, if needed>.

This emerged from the pr-review guide friction: large skill workflow payloads carry incident archaeology and edge cases directly in runtime instruction substrate. Some of that content is load-bearing, but much of it is provenance / debugging atlas material. The desired shape is not another monolithic migration. It is a bounded epic that codifies the decision atom, measures existing baggage, pilots on the hottest skills, then batches the rest by risk and size.

Duplicate / adjacency sweep

No open equivalent epic was found.

Related but not equivalent:

• #10537 (closed) modularized pr-review-guide.md condition-gated audits. It solved one pr-review audit-extraction class, not the general incident-to-runtime-instruction compression problem.
• #11319 (closed) shipped recursive Map-vs-World-Atlas enforcement via skill manifest / per-file payload budget. It provides the mechanical substrate this epic should consume, but it did not define the Bias -> Rule -> Rationale -> Trigger atom or audit all skills for historical baggage.
• #11604 (open) captures codify-deferral discipline for premature skill bloat. This should become a child / first slice of this epic rather than a competing standalone root.
• #11603 (open) captures broader META bandwidth saturation. This epic is a focused remediation lane under that wider structural concern.

KB / local content sweep also surfaced the same hot-path skills repeatedly: pr-review, pull-request, and post-review lifecycle substrate.

The Problem

Skills have accumulated runtime payload as if every lesson must remain in the active workflow manual forever. The result is self-crippling:

• Incident provenance competes with current operating rules inside files that agents load during hot paths.
• Large guides invite context-pruning, half-reading, and re-derivation loops.
• Review and PR workflows become rule sinkholes: hard to scan, hard to retire, and hard to distinguish universal discipline from rare edge cases.
• Valuable lessons still need to survive, but not all lessons deserve runtime instruction slots.

The operator-calibrated distinction:

• Runtime instruction substrate: short, repeatable behavior shape. Example: Bias -> Rule -> Rationale -> Trigger.
• Atlas / provenance substrate: incidents, archaeology, examples, and debugging trails. They remain discoverable but are not dragged through every skill invocation.

The Architectural Reality

Existing substrate already points in this direction:

• ADR 0007: Map vs World Atlas compaction taxonomy.
• ADR 0008 / /create-skill: SKILL routers stay small; heavy payload belongs under references/; workflow files can themselves become Maps for sub-rule Atlases.
• #11319: recursive Map-vs-Atlas / per-file payload budget and trigger-aware workflow structure have shipped.
• #10537: condition-gated pr-review audit extraction proved the pattern for one class of review payload.
• #11604: premature codification needs deferral and consolidation discipline.

The missing primitive is a migration unit that keeps operational behavior intact while moving baggage out of runtime flow:

Bias: <default wrong model>. Rule: <correct action>. Rationale: <why this matters>. Trigger: <when to load deeper atlas, if needed>.

This preserves how a skill is used without carrying every historical incident inline.

The Fix

Create a bounded skill-baggage reduction program with four slices:

1. Codify the atom: amend ADR 0008 or a successor ADR, and update /create-skill, to define Bias -> Rule -> Rationale -> Trigger as the preferred runtime-instruction atom for converting lessons into skill substrate.
2. Inventory skill baggage: measure all .agents/skills/**/*.md workflow payloads for incident/provenance blocks, repeated edge-case stories, oversized sections, and absent trigger pointers. Produce a ranked audit with byte counts and candidate dispositions: keep, compress-to-trigger, move, rewrite, retire.
3. Pilot on hot paths: apply the atom to pr-review and pull-request first. These are high-frequency, high-cost, cross-family workflows and already have known bloat history.
4. Batch the rest: create child tickets by skill group / risk class after the pilot. Avoid a giant all-skills migration PR. Each child must preserve behavior, reduce or justify loaded bytes, and keep incidents discoverable in atlas/provenance files.

Contract Ledger Matrix

Early-filing rationale / divergence matrix

This ticket cites live session context and #11598/#11604 adjacency before every related discussion has fully stabilized. Filing is still correct because the operator explicitly identified the risk: without a ticket, this becomes lost knowledge.

Downstream amendments may be required after #11598 and sibling tickets converge. Expected refresh points: exact ADR target, incident-atlas location, and child-ticket decomposition.

Acceptance Criteria

• AC1: ADR 0008 or successor ADR codifies the Bias -> Rule -> Rationale -> Trigger atom, including when incidents belong in atlas/provenance instead of runtime workflow prose.
• AC2: /create-skill Lesson Promotion Path updated so new lessons route through the atom / atlas decision instead of blind payload append.
• AC3: Inventory ticket filed and completed for all .agents/skills/**/*.md workflow payloads, with byte counts, section counts, and disposition candidates.
• AC4: Pilot ticket for pr-review completed. Required evidence: behavior-preserving compression, byte impact, lint-skill-manifest pass, cross-family review.
• AC5: Pilot ticket for pull-request completed or explicitly deferred with evidence if pr-review pilot invalidates the atom.
• AC6: Batch rollout plan created from inventory results: child tickets grouped by risk/size, not one all-skills PR.
• AC7: Incident/provenance destination chosen and documented (learn/agentos/incidents/**, per-skill references/incidents/**, or equivalent), with retrieval hints from runtime atoms.
• AC8: #11604 either linked as a child or explicitly dispositioned as sibling if the team decides codify-deferral must remain independent.
• AC9: Each implementation PR is net loaded-byte reducing OR carries explicit future-decay mitigation per AGENTS.md substrate accretion defense.
• AC10: Cross-family review confirms the migrated skills remain usable without requiring agents to load historical incident books by default.

Out of Scope

• Rewriting every skill in one PR.
• Deleting incident history. Incidents move to better substrate; they are not erased.
• Creating new gates before the atom and inventory prove where enforcement belongs.
• Treating byte reduction as the only success metric. The real target is runtime cognitive-load reduction while preserving behavior.
• Body-pillar / v13 feature-bandwidth policy. That broader counter-bias belongs to #11603.

Avoided Traps

• Map versus world atlas violation: keeping a full incident story inline when one atom plus a trigger preserves the behavior.
• Baggage laundering: moving text to another always-loaded file and claiming compression.
• Three-to-five-shot impossible migration: trying to clean all skills in a single sweep without inventory or pilots.
• Premature codification recursion: adding another long rule about avoiding long rules. The atom must be short; detailed rationale belongs in atlas.
• Behavior loss through over-compression: if a rule stops firing because the atom is too terse, the migration fails even if bytes drop.

Related

• #11604 — skill bloat from premature codification / codify-deferral discipline.
• #11603 — substrate bandwidth saturation / META counter-bias mechanism.
• #11319 — recursive Map-vs-World-Atlas workflow trigger substrate.
• #10537 — pr-review condition-gated audit modularization precedent.
• Discussion #11598 — source META friction discussion.
• ADR 0007 — compaction taxonomy.
• ADR 0008 — skill anatomy and authoring contract.

Origin Session ID: 8591bc48-0ddc-48bf-aa47-58e53ea81a57

Retrieval Hint: query_raw_memories("Bias Rule Rationale Trigger skill baggage pr-review pull-request create-skill Lesson Promotion Path map world atlas incident provenance")