diff --git a/docs/plans/context-architecture-protocols-plan.md b/docs/plans/context-architecture-protocols-plan.md
new file mode 100644
index 00000000..c5a9156a
--- /dev/null
+++ b/docs/plans/context-architecture-protocols-plan.md
@@ -0,0 +1,786 @@
+# Implementation Plan: Context Architecture Optimization (Protocol Partitioning)
+
+> Generated by `/PACT:plan-mode` on 2026-01-15
+> Status: PENDING APPROVAL
+> **Last Updated**: 2026-01-15 (Second peer review feedback incorporated — agent files scope added)
+
+<!-- Status Lifecycle:
+     PENDING APPROVAL → APPROVED → IN_PROGRESS → IMPLEMENTED
+                    ↘ SUPERSEDED (if replaced by newer plan)
+                    ↘ BLOCKED (if unresolved conflicts)
+
+     Transition Ownership:
+     - PENDING APPROVAL → APPROVED: User (explicit approval)
+     - APPROVED → IN_PROGRESS: Orchestrator (when /PACT:orchestrate starts)
+     - IN_PROGRESS → IMPLEMENTED: Orchestrator (after successful completion)
+     - Any → SUPERSEDED: plan-mode (when creating replacement plan)
+     - Any → BLOCKED: plan-mode (when unresolved blocking conflicts)
+-->
+
+## Summary
+
+Optimize PACT framework context loading by:
+1. **Compressing CLAUDE.md** from 306 lines to ~100-110 lines (preserving phase principles)
+2. **Partitioning pact-protocols.md** (1,085 lines) into 8 focused protocol files
+3. **Updating commands** to @-reference only the specific protocols they need
+4. **Updating agent files** (7 specialists) with targeted @-refs to new protocol files
+
+**Why protocol partitioning over skills**: PACT is a deterministic, structured workflow where we know exactly when each protocol is needed. Explicit @-references provide predictable, controlled loading that aligns with PACT's structured nature. Skills (auto-discovery) are better suited for optional, contextual guidance.
+
+**Framework-wide impact**: All 8 PACT commands benefit. The heaviest commands (`orchestrate`, `plan-mode`) see ~55% reductions. All commands benefit from 67% smaller CLAUDE.md. Expected savings: ~55% reduction in typical orchestration context load.
+
+**Supersedes**:
+- PR #68 draft plan
+- `context-architecture-skills-plan.md` (skills approach reconsidered)
+
+---
+
+## Specialist Perspectives
+
+### 📋 Preparation Phase
+**Effort**: Low-Medium
+
+#### Research Completed
+- [x] Anthropic context engineering guidance reviewed ([source](https://www.anthropic.com/research/building-effective-agents))
+- [x] CLAUDE.md size guidelines established (working assumption: < 60 lines ideal, < 300 max — derived from context engineering principles, not explicit Anthropic documentation)
+- [x] Current context load measured (CLAUDE.md ~4K, pact-protocols.md ~10K tokens; estimated via ~10 tokens/line approximation)
+- [x] Skills vs @-references trade-offs analyzed
+- [x] PACT workflow determinism assessed (favors explicit references)
+- [x] Agent files audited for protocol references (**21 refs found across 7 files** — require updates)
+- [x] rePACT.md dependencies verified (self-contained, no @-refs needed)
+
+#### Questions Resolved
+- **Skills vs Protocols**: Protocol partitioning preferred for PACT's deterministic workflows
+- **Algedonic location**: Keep separate algedonic.md file (already concise, safety-critical)
+- **VSM distribution**: Keep vsm-glossary.md as central reference
+- **rePACT dependencies**: Command has conceptual references ("see Autonomy Charter", "Apply S2 coordination") but these are inline explanatory text, NOT @-file references. No changes needed.
+- **Agent file dependencies**: All 7 specialist agent files contain @-refs to `pact-protocols.md` for workflow handoffs and S1 Autonomy. These require targeted updates after partitioning.
+
+---
+
+### 🏗️ Architecture Phase
+**Effort**: Medium
+
+#### Components Affected
+| Component | Change Type | Impact |
+|-----------|-------------|--------|
+| `CLAUDE.md` | Compress | 306 → ~100-110 lines |
+| `protocols/pact-protocols.md` | Partition | 1,085 → ~180 lines (lean index) |
+| `protocols/delegation.md` | New | ~145 lines |
+| `protocols/s4-checkpoints.md` | New | ~140 lines |
+| `protocols/variety-management.md` | New | ~60 lines |
+| `protocols/s2-coordination.md` | New | ~110 lines |
+| `protocols/s3-audit.md` | New | ~60 lines |
+| `protocols/s3s4-tension.md` | New | ~65 lines |
+| `protocols/workflow-protocols.md` | New | ~105 lines |
+| `protocols/phase-guidance.md` | New | ~100 lines |
+| `commands/PACT/orchestrate.md` | Modify | Replace monolithic @-ref with targeted @-refs |
+| `commands/PACT/plan-mode.md` | Modify | Replace monolithic @-ref with targeted @-refs |
+| `agents/pact-*.md` (7 files) | Modify | Update @pact-protocols refs to targeted protocol files |
+
+#### Design Approach
+
+**Protocol Partitioning Architecture**:
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│                    CLAUDE.md (~100-110 lines)               │
+│  Always loaded: Mission, S5 Policy, Phase Principles,       │
+│                 Specialist list, Command list               │
+└─────────────────────────────────────────────────────────────┘
+         │
+         │ Commands @-reference only what they need
+         ▼
+┌─────────────────────────────────────────────────────────────┐
+│                  Partitioned Protocol Files                 │
+│                                                             │
+│  pact-protocols.md (~180 lines)                             │
+│  └─ Lean index + S1 Autonomy + Cross-Cutting Concerns       │
+│                                                             │
+│  delegation.md (~145 lines)                                 │
+│  └─ S5 Policy, Tool checkpoint, app code definition         │
+│                                                             │
+│  s4-checkpoints.md (~140 lines)                             │
+│  └─ S4 checkpoint protocol, environment model               │
+│                                                             │
+│  variety-management.md (~60 lines)                          │
+│  └─ Variety scoring, workflow selection, strategies         │
+│                                                             │
+│  s2-coordination.md (~110 lines)                            │
+│  └─ Pre-parallel checks, conflict resolution                │
+│                                                             │
+│  s3-audit.md (~60 lines)                                    │
+│  └─ Continuous audit, RED/YELLOW/GREEN signals              │
+│                                                             │
+│  s3s4-tension.md (~65 lines)                                │
+│  └─ Tension detection, resolution protocol                  │
+│                                                             │
+│  workflow-protocols.md (~105 lines)                         │
+│  └─ plan-mode, imPACT, comPACT protocol summaries           │
+│                                                             │
+│  phase-guidance.md (~100 lines)                             │
+│  └─ Test Engagement, Phase Handoffs, Backend/DB Boundary    │
+│                                                             │
+│  algedonic.md (~235 lines) — unchanged                      │
+│  └─ Emergency bypass, HALT/ALERT signals                    │
+└─────────────────────────────────────────────────────────────┘
+         │
+         │ Commands load only relevant protocols
+         ▼
+┌─────────────────────────────────────────────────────────────┐
+│  orchestrate.md                                             │
+│  └─ @delegation, @variety-management, @s4-checkpoints,      │
+│     @s2-coordination, @s3-audit, @phase-guidance, @algedonic│
+│                                                             │
+│  plan-mode.md                                               │
+│  └─ @variety-management, @workflow-protocols                │
+│                                                             │
+│  imPACT.md                                                  │
+│  └─ @algedonic, @workflow-protocols                         │
+│                                                             │
+│  comPACT.md                                                 │
+│  └─ @workflow-protocols (optional, for protocol reference)  │
+└─────────────────────────────────────────────────────────────┘
+```
+
+#### Key Decisions
+| Decision | Options | Recommendation | Rationale |
+|----------|---------|----------------|-----------|
+| Approach | Skills (auto-discovery) vs Protocols (@-refs) | Protocol partitioning | PACT is deterministic; explicit control preferred |
+| Protocol count | 6 files vs 8 files | 8 focused files | Complete content accounting shows 8 needed for target index size |
+| CLAUDE.md size | ~85 vs ~100-110 lines | ~100-110 lines | Preserves phase principles that guide behavior |
+| pact-protocols.md | ~100 vs ~180 lines | ~180 lines (lean index) | Realistic after complete content accounting |
+| Algedonic | Partition vs keep whole | Keep whole | Safety-critical, already concise, well-isolated |
+
+#### Interface Contracts
+
+**Targeted @-Reference Pattern**:
+
+```markdown
+# Current (loads entire 1,085-line file every time):
+See @~/.claude/protocols/pact-protocols.md for S4 Checkpoint Protocol.
+
+# New (loads only ~140-line file when needed):
+See @~/.claude/protocols/s4-checkpoints.md for the checkpoint protocol.
+```
+
+**Protocol File Mapping** (what each command needs):
+
+| Command | Protocols Needed |
+|---------|------------------|
+| `orchestrate.md` | delegation, variety-management, s4-checkpoints, s2-coordination, s3-audit, phase-guidance, algedonic |
+| `plan-mode.md` | variety-management, workflow-protocols |
+| `imPACT.md` | algedonic, workflow-protocols |
+| `comPACT.md` | workflow-protocols (optional) |
+| `rePACT.md` | None (self-contained with inline guidance) |
+
+**Estimated Load per Command**:
+
+| Command | Current @-refs Load | New @-refs Load | Reduction |
+|---------|---------------------|-----------------|-----------|
+| orchestrate | ~1,320 lines | ~715 lines | **46%** |
+| plan-mode | ~1,085 lines | ~165 lines | **85%** |
+| imPACT | ~235 lines | ~340 lines | -45% (adds workflow-protocols for context) |
+
+---
+
+### 💻 Code Phase
+**Effort**: Medium
+
+> **Note**: This is AI tooling (.claude/ files), not application code. Per delegation rules, the orchestrator may implement these changes directly. No specialist delegation required.
+
+#### Complete Content Accounting
+
+Every section of `pact-protocols.md` (1,085 lines) is mapped to its destination:
+
+| Source Section | Lines | Destination | Notes |
+|----------------|-------|-------------|-------|
+| Header + VSM Quick Reference | 1-10 | `pact-protocols.md` (index) | Keep as intro |
+| S5 Policy Layer (Governance) | 13-145 | `delegation.md` | Includes Non-Negotiables, Delegation Enforcement, Policy Checkpoints, S5 Authority, S5 Decision Framing |
+| S4 Checkpoint Protocol | 148-217 | `s4-checkpoints.md` | Includes Trigger Points, Questions, Outcomes, Format |
+| S4 Environment Model | 220-291 | `s4-checkpoints.md` | Combined with S4 Checkpoint |
+| S3/S4 Tension Detection | 294-356 | `s3s4-tension.md` | Includes Indicators, Detection Phrases, Resolution |
+| S2 Coordination Layer | 359-465 | `s2-coordination.md` | Includes Pre-Parallel Check, Conflict Resolution, Anti-Oscillation |
+| S1 Autonomy & Recursion | 468-540 | `pact-protocols.md` (index) | Keep inline (~72 lines) - referenced by multiple protocols |
+| S3* Continuous Audit | 543-601 | `s3-audit.md` | Includes Audit Modes, Signals, S2 Coordination |
+| Algedonic Signals | 604-644 | `algedonic.md` | Already separate (~40 line summary here, full file exists) |
+| Variety Management | 647-706 | `variety-management.md` | Includes Dimensions, Score, Strategies |
+| PACT Workflow Family | 709-718 | `pact-protocols.md` (index) | Keep inline (~9 lines) - overview table |
+| plan-mode Protocol | 721-754 | `workflow-protocols.md` | Protocol summary |
+| imPACT Protocol | 757-773 | `workflow-protocols.md` | Protocol summary |
+| comPACT Protocol | 776-832 | `workflow-protocols.md` | Protocol summary |
+| Phase Handoffs | 835-842 | `phase-guidance.md` | Universal guidance |
+| Backend ↔ Database Boundary | 846-858 | `phase-guidance.md` | Specialist coordination |
+| Test Engagement | 861-941 | `phase-guidance.md` | CODE→TEST handoff |
+| Cross-Cutting Concerns | 945-954 | `pact-protocols.md` (index) | Keep inline (~9 lines) - brief checklist |
+| Architecture Review | 957-963 | `pact-protocols.md` (index) | Keep inline (~6 lines) - brief |
+| Documentation Locations | 967-1069 | `pact-protocols.md` (index) | Keep inline (~102 lines) - reference table |
+| Session Continuity | 1073-1079 | `pact-protocols.md` (index) | Keep inline (~6 lines) |
+| Related | 1082-1086 | `pact-protocols.md` (index) | Keep inline (~4 lines) |
+
+**Summary**:
+- Content partitioned to new files: ~785 lines (8 files)
+- Content remaining in index: ~180 lines
+- Content in separate algedonic.md: ~235 lines (unchanged)
+- Total accounted: 1,085 lines ✓
+
+#### Cross-Reference Inventory
+
+Internal references within `pact-protocols.md` that need updating after partitioning:
+
+| Location | Current Reference | New Reference | Notes |
+|----------|-------------------|---------------|-------|
+| Line 7 | `@~/.claude/reference/vsm-glossary.md` | Keep as-is | External reference |
+| Line 212 | "see Variety Management" | `@variety-management.md` | Cross-protocol ref |
+| Line 349-355 | S3/S4 tension refs S4 Checkpoints | `@s4-checkpoints.md` | Cross-protocol ref |
+| Line 435-436 | "see CODE → TEST Handoff", "see Phase Handoffs" | `@phase-guidance.md` | Cross-protocol ref |
+| Line 488 | "see S2 Coordination" | `@s2-coordination.md` | Cross-protocol ref |
+| Line 539 | `@~/.claude/commands/PACT/rePACT.md` | Keep as-is | External reference |
+| Line 586 | "unlike algedonic signals—see below" | `@algedonic.md` | Cross-protocol ref |
+| Line 610 | `@~/.claude/protocols/algedonic.md` | Keep as-is | External reference |
+| Line 694 | "/PACT:rePACT" | Keep as-is | Command reference |
+| Line 696 | "S3*" reference | `@s3-audit.md` | Cross-protocol ref |
+| Line 969 | "filesystem-context" skill | Keep as-is | External reference |
+
+**Total internal cross-references requiring updates**: 7
+
+#### Agent Files Cross-Reference Inventory
+
+All 7 specialist agent files reference `pact-protocols.md`. After partitioning, these need targeted updates:
+
+| Agent File | Reference Purpose | Current @-ref | New @-ref |
+|------------|-------------------|---------------|-----------|
+| `pact-architect.md` | Cross-Agent Coordination | `@pact-protocols.md` | `@phase-guidance.md` |
+| `pact-architect.md` | S1 Autonomy & Recursion | `@pact-protocols.md` | `@pact-protocols.md` (unchanged — stays in index) |
+| `pact-backend-coder.md` | Cross-Agent Coordination | `@pact-protocols.md` | `@phase-guidance.md` |
+| `pact-backend-coder.md` | S1 Autonomy & Recursion | `@pact-protocols.md` | `@pact-protocols.md` (unchanged) |
+| `pact-frontend-coder.md` | Cross-Agent Coordination | `@pact-protocols.md` | `@phase-guidance.md` |
+| `pact-frontend-coder.md` | S1 Autonomy & Recursion | `@pact-protocols.md` | `@pact-protocols.md` (unchanged) |
+| `pact-database-engineer.md` | Cross-Agent Coordination | `@pact-protocols.md` | `@phase-guidance.md` |
+| `pact-database-engineer.md` | S1 Autonomy & Recursion | `@pact-protocols.md` | `@pact-protocols.md` (unchanged) |
+| `pact-test-engineer.md` | Cross-Agent Coordination | `@pact-protocols.md` | `@phase-guidance.md` |
+| `pact-test-engineer.md` | S1 Autonomy & Recursion | `@pact-protocols.md` | `@pact-protocols.md` (unchanged) |
+| `pact-preparer.md` | Cross-Agent Coordination | `@pact-protocols.md` | `@phase-guidance.md` |
+| `pact-preparer.md` | S1 Autonomy & Recursion | `@pact-protocols.md` | `@pact-protocols.md` (unchanged) |
+| `pact-preparer.md` | S4 Environment Model | `@pact-protocols.md` | `@s4-checkpoints.md` |
+| `pact-n8n.md` | Cross-Agent Coordination | `@pact-protocols.md` | `@phase-guidance.md` |
+| `pact-n8n.md` | S1 Autonomy & Recursion | `@pact-protocols.md` | `@pact-protocols.md` (unchanged) |
+
+**Summary**:
+- 7 refs need updating to `@phase-guidance.md` (Cross-Agent Coordination)
+- 7 refs remain as `@pact-protocols.md` (S1 Autonomy stays in lean index)
+- 1 ref needs updating to `@s4-checkpoints.md` (pact-preparer's Environment Model ref)
+- All `@algedonic.md` refs: No change needed (already separate file)
+
+**Total agent file cross-references requiring updates**: 8 (7 to phase-guidance + 1 to s4-checkpoints)
+
+**Grand total cross-references requiring updates**: 15 (7 internal + 8 agent files)
+
+#### Files to Create
+
+| File | Content Source | Est. Lines |
+|------|----------------|------------|
+| `.claude/protocols/delegation.md` | pact-protocols.md lines 13-145 | ~145 |
+| `.claude/protocols/s4-checkpoints.md` | pact-protocols.md lines 148-291 | ~140 |
+| `.claude/protocols/variety-management.md` | pact-protocols.md lines 647-706 | ~60 |
+| `.claude/protocols/s2-coordination.md` | pact-protocols.md lines 359-465 | ~110 |
+| `.claude/protocols/s3-audit.md` | pact-protocols.md lines 543-601 | ~60 |
+| `.claude/protocols/s3s4-tension.md` | pact-protocols.md lines 294-356 | ~65 |
+| `.claude/protocols/workflow-protocols.md` | pact-protocols.md lines 721-832 | ~105 |
+| `.claude/protocols/phase-guidance.md` | pact-protocols.md lines 835-941 | ~100 |
+
+#### Files to Modify
+
+| File | Changes |
+|------|---------|
+| `CLAUDE.md` | Compress from 306 → ~100-110 lines (preserve phase principles) |
+| `.claude/protocols/pact-protocols.md` | Reduce to ~180-line lean index |
+| `.claude/commands/PACT/orchestrate.md` | Replace single @pact-protocols with targeted @-refs |
+| `.claude/commands/PACT/plan-mode.md` | Replace @pact-protocols with @variety-management, @workflow-protocols |
+| `.claude/commands/PACT/imPACT.md` | Add @workflow-protocols reference |
+| `.claude/agents/pact-architect.md` | Update Cross-Agent Coordination ref → @phase-guidance.md |
+| `.claude/agents/pact-backend-coder.md` | Update Cross-Agent Coordination ref → @phase-guidance.md |
+| `.claude/agents/pact-frontend-coder.md` | Update Cross-Agent Coordination ref → @phase-guidance.md |
+| `.claude/agents/pact-database-engineer.md` | Update Cross-Agent Coordination ref → @phase-guidance.md |
+| `.claude/agents/pact-test-engineer.md` | Update Cross-Agent Coordination ref → @phase-guidance.md |
+| `.claude/agents/pact-preparer.md` | Update Cross-Agent Coordination ref → @phase-guidance.md, S4 Environment Model ref → @s4-checkpoints.md |
+| `.claude/agents/pact-n8n.md` | Update Cross-Agent Coordination ref → @phase-guidance.md |
+| `pact-plugin/protocols/` | Mirror partitioned files for plugin distribution |
+| `pact-plugin/agents/` | Mirror agent file updates for plugin distribution |
+
+#### Protocol File Structure
+
+Each partitioned protocol file follows this structure:
+
+```markdown
+# {Protocol Name}
+
+> **Purpose**: One-line description
+> **Used by**: List of commands that @-reference this file
+> **See also**: Related protocol files (with @-refs)
+
+---
+
+## {Content sections extracted from pact-protocols.md}
+
+...
+```
+
+#### CLAUDE.md Proposed Structure (~100-110 lines)
+
+```markdown
+# MISSION (7 lines)
+Mission statement, motto, VSM note
+
+## S5 POLICY (18 lines)
+Non-negotiables table (4 rows), policy checkpoints summary (6 lines)
+→ "See @~/.claude/protocols/delegation.md for enforcement details"
+→ "See @~/.claude/protocols/algedonic.md for emergency bypass"
+
+## INSTRUCTIONS (6 lines)
+Core directives (read CLAUDE.md, apply PACT, delegate, update)
+
+## GUIDELINES (15 lines)
+Context management, git workflow, S3/S4 modes (brief)
+→ "See @~/.claude/protocols/s3s4-tension.md for tension resolution"
+
+## PACT PHASES (20 lines)
+Phase names with 2-3 key principles each (preserved from current)
+→ PREPARE: Documentation First, Context Gathering, Dependency Mapping
+→ ARCHITECT: Single Responsibility, Loose Coupling, Interface Segregation
+→ CODE: Clean Code, DRY, KISS, Error Handling, Security Mindset
+→ TEST: Coverage, Edge Cases, Integration, Regression Prevention
+
+## SPECIALIST AGENTS (10 lines)
+Agent list with emoji identifiers
+
+## ORCHESTRATION COMMANDS (15 lines)
+Command list with one-line descriptions
+→ "See @~/.claude/protocols/variety-management.md for workflow selection"
+→ "See @~/.claude/protocols/workflow-protocols.md for protocol details"
+
+## DOCUMENTATION LOCATIONS (8 lines)
+Phase → Output Location table (condensed)
+
+## PROTOCOL REFERENCE (12 lines)
+Quick index of partitioned protocol files with trigger phrases:
+- delegation.md → "Before Edit/Write on any file"
+- s4-checkpoints.md → "At phase boundaries"
+- variety-management.md → "Task assessment, workflow selection"
+- s2-coordination.md → "Before parallel agent invocation"
+- s3-audit.md → "During CODE phase for quality signals"
+- s3s4-tension.md → "When detecting operational vs strategic conflict"
+- workflow-protocols.md → "For plan-mode, imPACT, comPACT details"
+- phase-guidance.md → "For handoffs, test engagement, boundaries"
+- algedonic.md → "Emergency escalation to user"
+```
+
+#### pact-protocols.md New Structure (~180 lines)
+
+```markdown
+# PACT Protocols (Lean Index)
+
+> **Purpose**: Central index for PACT protocol documentation.
+> **Note**: Protocol content has been partitioned into focused files for targeted loading.
+
+## Protocol Files
+
+| Protocol | File | Trigger |
+|----------|------|---------|
+| Delegation Enforcement | @delegation.md | Before Edit/Write on app code |
+| S4 Checkpoints | @s4-checkpoints.md | At phase boundaries |
+| Variety Management | @variety-management.md | Task assessment, workflow selection |
+| S2 Coordination | @s2-coordination.md | Parallel agent coordination |
+| S3* Audit | @s3-audit.md | CODE phase quality signals |
+| S3/S4 Tension | @s3s4-tension.md | Mode conflict resolution |
+| Workflow Protocols | @workflow-protocols.md | plan-mode, imPACT, comPACT |
+| Phase Guidance | @phase-guidance.md | Handoffs, test engagement |
+| Algedonic Signals | @algedonic.md | Emergency escalation |
+
+## S1 Autonomy Charter (~72 lines)
+[Keep inline - referenced by multiple protocols]
+
+## PACT Workflow Family (~9 lines)
+[Keep inline - overview table]
+
+## Cross-Cutting Concerns (~9 lines)
+[Keep inline - brief checklist]
+
+## Architecture Review (~6 lines)
+[Keep inline - brief]
+
+## Documentation Locations (~70 lines, condensed from ~102)
+[Keep inline - reference table, compressed]
+
+## Session Continuity (~6 lines)
+[Keep inline]
+
+## Related (~4 lines)
+[Keep inline]
+```
+
+#### Implementation Sequence
+
+1. `feat(protocols): create delegation.md from pact-protocols.md` — Extract S5 Policy Layer
+2. `feat(protocols): create s4-checkpoints.md from pact-protocols.md` — Extract S4 content + Environment Model
+3. `feat(protocols): create variety-management.md from pact-protocols.md` — Extract variety content
+4. `feat(protocols): create s2-coordination.md from pact-protocols.md` — Extract S2 content
+5. `feat(protocols): create s3-audit.md from pact-protocols.md` — Extract S3* content
+6. `feat(protocols): create s3s4-tension.md from pact-protocols.md` — Extract tension content
+7. `feat(protocols): create workflow-protocols.md from pact-protocols.md` — Extract plan-mode, imPACT, comPACT
+8. `feat(protocols): create phase-guidance.md from pact-protocols.md` — Extract handoffs, test engagement, boundaries
+9. `refactor(protocols): convert pact-protocols.md to lean index (~180 lines)` — Update internal cross-refs
+10. `refactor(claude): compress CLAUDE.md to ~100-110 lines` — Preserve phase principles
+11. `refactor(commands): update orchestrate.md with targeted @-refs` — Replace monolithic ref
+12. `refactor(commands): update plan-mode.md with targeted @-refs` — Replace monolithic ref
+13. `refactor(commands): update imPACT.md with @workflow-protocols` — Add protocol ref
+14. `refactor(agents): update all 7 agent files with targeted @-refs` — Update Cross-Agent Coordination + S4 Environment Model refs
+15. `chore(plugin): sync partitioned protocols and agents to pact-plugin/` — Mirror for distribution
+
+---
+
+### 🧪 Test Phase
+**Effort**: Medium-High (5-6 hours manual validation)
+
+#### Test Scenarios
+
+| Scenario | Type | Priority |
+|----------|------|----------|
+| Content preservation verification (checklist) | Verification | P0 |
+| Cross-reference resolution (all 15 refs: 7 internal + 8 agent) | Verification | P0 |
+| Agent file @-ref resolution (all 7 agent files load correct protocols) | Verification | P0 |
+| `/PACT:orchestrate` loads correct protocols at each phase | Integration | P0 |
+| Delegation enforcement on app code edit | Functional | P0 |
+| Algedonic HALT signal bypasses to user | Functional | P0 |
+| Tool Checkpoint Protocol fires before Edit/Write | Functional | P0 |
+| `/PACT:plan-mode` loads variety-management, workflow-protocols | Integration | P0 |
+| `/PACT:imPACT` loads algedonic, workflow-protocols | Integration | P1 |
+| S3/S4 tension protocol accessible via targeted @-ref | Functional | P1 |
+| S4 checkpoint protocol accessible at phase boundaries | Functional | P1 |
+| `/PACT:comPACT` works with optional workflow-protocols ref | Integration | P2 |
+| `/PACT:rePACT` works without protocol refs | Integration | P2 |
+| Plugin sync matches .claude/ structure | Verification | P2 |
+
+#### Coverage Targets
+- Critical path (P0): 100% — Must pass for merge
+- Core workflows (P1): 100% — Should pass for merge
+- Edge cases (P2): Best effort — May defer
+
+#### Content Preservation Verification
+
+Create `docs/testing/partition-verification.md` during implementation:
+
+```markdown
+# Protocol Partitioning Verification Checklist
+
+## Section Coverage
+| Original Section | Destination File | Lines | Verified |
+|------------------|------------------|-------|----------|
+| S5 Policy Layer | delegation.md | 13-145 | [ ] |
+| S4 Checkpoint Protocol | s4-checkpoints.md | 148-217 | [ ] |
+| ... (all sections from Content Accounting table) |
+
+## Content Integrity
+- [ ] Every H2 section from original has corresponding location
+- [ ] Every code example preserved
+- [ ] Every table preserved
+- [ ] All @-references updated or preserved correctly
+
+## Cross-Reference Resolution
+| Ref # | Original | Updated To | Verified |
+|-------|----------|------------|----------|
+| 1 | "see Variety Management" | @variety-management.md | [ ] |
+| 2 | S3/S4 refs S4 Checkpoints | @s4-checkpoints.md | [ ] |
+| ... (all 7 cross-references) |
+
+## Summary
+- Sections verified: X of Y
+- Cross-refs verified: X of 7
+- Content gaps: [none / list]
+```
+
+#### Validation Approach
+
+**Tier 1: Static Verification (1 hour)**
+- File existence checks (8 new files)
+- Line count validation (sum ≥ original)
+- @-reference resolution (verify all 7 updated refs work)
+- Section heading mapping (checklist)
+
+**Tier 2: Behavioral Verification (2.5-3 hours)**
+- Execute `/PACT:orchestrate` on sample task
+- Execute `/PACT:plan-mode` on sample task
+- Test delegation enforcement (attempt Edit on .ts file)
+- Test algedonic escalation (simulate HALT condition)
+- Verify protocol content appears in agent responses
+
+**Tier 3: Regression Verification (1.5-2 hours)**
+- Compare pre/post behavior on sample tasks
+- Verify edge cases still work
+- Plugin sync validation (`diff -rq .claude/protocols pact-plugin/protocols`)
+
+#### Regression Baseline
+
+Before implementation, capture:
+- [ ] Output of `/PACT:orchestrate` on "add pagination endpoint" task
+- [ ] Output of `/PACT:plan-mode` on "implement user auth" task
+- [ ] Delegation enforcement behavior (attempt to edit app code)
+- [ ] Decision log format produced by CODE phase
+
+After implementation, compare and document any differences.
+
+#### Acceptance Criteria
+
+- [ ] All 8 new protocol files exist and are syntactically valid
+- [ ] pact-protocols.md reduced to 150-200 lines (lean index)
+- [ ] Compressed CLAUDE.md is 95-115 lines
+- [ ] Compressed CLAUDE.md contains phase principles (PREPARE, ARCHITECT, CODE, TEST)
+- [ ] Compressed CLAUDE.md contains SACROSANCT policy table
+- [ ] Compressed CLAUDE.md contains delegation enforcement summary
+- [ ] All 15 cross-references updated correctly (7 internal + 8 agent file refs)
+- [ ] All 7 agent files updated with correct targeted @-refs
+- [ ] `/PACT:orchestrate` executes without errors
+- [ ] `/PACT:plan-mode` executes without errors
+- [ ] Tool Checkpoint Protocol triggers on application code
+- [ ] Algedonic HALT scenario escalates to user
+- [ ] Content preservation checklist: 100% sections verified
+- [ ] Plugin sync: `diff` shows only expected differences (protocols and agents)
+
+---
+
+## Rollback Strategy
+
+### Rollback Triggers
+- Any P0 test fails after implementation
+- More than 2 P1 tests fail
+- Critical user-reported issues within 48 hours of merge
+- Content preservation checklist shows gaps
+
+### Rollback Procedures
+
+| Stage | Action |
+|-------|--------|
+| **Before merge** | Abandon PR, no rollback needed |
+| **After merge, before release** | `git revert <merge-commit>` |
+| **After plugin distribution** | 1) Revert in repo, 2) Publish corrected pact-plugin, 3) Communicate if impact significant |
+
+### Rollback Verification
+After rollback:
+- [ ] `pact-protocols.md` is original 1,085 lines
+- [ ] `CLAUDE.md` is original 306 lines
+- [ ] All 7 agent files restored to original @pact-protocols.md refs
+- [ ] No orphaned protocol files in .claude/protocols/
+- [ ] Plugin sync reverted (protocols and agents)
+
+---
+
+## Synthesized Implementation Roadmap
+
+### Phase Sequence
+
+```
+Phase 1: Create Protocol Files (Low Risk)
+├── delegation.md
+├── s4-checkpoints.md
+├── variety-management.md
+├── s2-coordination.md
+├── s3-audit.md
+├── s3s4-tension.md
+├── workflow-protocols.md
+└── phase-guidance.md
+         │
+         ▼ CHECKPOINT: Verify content extraction complete (checklist)
+         │
+Phase 2: Update pact-protocols.md (Low Risk)
+├── Convert to ~180-line lean index
+└── Update all 7 cross-references
+         │
+         ▼ CHECKPOINT: Verify index references resolve correctly
+         │
+Phase 3: Compress CLAUDE.md (Medium Risk)
+├── Restructure to ~100-110 lines
+├── Preserve phase principles
+└── Add protocol reference guide with trigger phrases
+         │
+         ▼ CHECKPOINT: Run full P0 test suite
+         │
+Phase 4: Update Commands (Medium Risk)
+├── orchestrate.md (targeted @-refs)
+├── plan-mode.md (targeted @-refs)
+└── imPACT.md (add @workflow-protocols)
+         │
+         ▼ CHECKPOINT: Command validation
+         │
+Phase 5: Update Agent Files (Low Risk)
+├── All 7 agent files: Cross-Agent Coordination → @phase-guidance.md
+└── pact-preparer.md: S4 Environment Model → @s4-checkpoints.md
+         │
+         ▼ CHECKPOINT: Agent file @-ref validation + regression comparison
+         │
+Phase 6: Plugin Sync (Low Risk)
+├── Mirror protocols to pact-plugin/protocols/
+└── Mirror agent updates to pact-plugin/agents/
+         │
+         ▼ Final PR review
+```
+
+### Commit Sequence (Proposed)
+
+1. `feat(protocols): partition S5 policy layer into delegation.md`
+2. `feat(protocols): partition S4 checkpoint protocol into s4-checkpoints.md`
+3. `feat(protocols): partition variety management into variety-management.md`
+4. `feat(protocols): partition S2 coordination into s2-coordination.md`
+5. `feat(protocols): partition S3* audit into s3-audit.md`
+6. `feat(protocols): partition S3/S4 tension into s3s4-tension.md`
+7. `feat(protocols): partition workflow protocols (plan-mode, imPACT, comPACT)`
+8. `feat(protocols): partition phase guidance (handoffs, test engagement)`
+9. `refactor(protocols): convert pact-protocols.md to lean index`
+10. `refactor(claude): compress CLAUDE.md to ~100-110 lines`
+11. `refactor(commands): update orchestrate.md with targeted protocol refs`
+12. `refactor(commands): update plan-mode.md with targeted protocol refs`
+13. `refactor(commands): update imPACT.md with workflow-protocols ref`
+14. `refactor(agents): update all 7 agent files with targeted @-refs`
+15. `chore(plugin): sync partitioned protocols and agents to pact-plugin/`
+
+---
+
+## Cross-Cutting Concerns
+
+| Concern | Status | Notes |
+|---------|--------|-------|
+| Security | ✅ Ready | Delegation enforcement and algedonic signals preserved |
+| Performance | ✅ Focus | ~55% context reduction for heavy commands |
+| Predictability | ✅ Improved | Deterministic @-refs vs skill auto-discovery |
+| Maintainability | ✅ Improved | Smaller, focused files easier to update |
+| Rollback Safety | ✅ Added | Explicit rollback triggers and procedures |
+
+---
+
+## Open Questions
+
+### Require User Decision
+
+None — approach validated through discussion and peer review.
+
+### Require Further Research
+
+- [ ] **Documentation Locations section**: Currently ~102 lines; can be compressed to ~70 lines during implementation
+- [ ] **Optimal protocol granularity**: 8 files proposed; may consolidate if testing reveals unnecessary splits
+
+---
+
+## Risk Assessment
+
+| Risk | Likelihood | Impact | Mitigation |
+|------|------------|--------|------------|
+| Content lost in partitioning | Low | High | Content accounting table + verification checklist |
+| Cross-references break | Low (was Medium) | Medium | Cross-reference inventory with explicit mapping |
+| Commands miss needed protocols | Low | Medium | Protocol file mapping table |
+| Plugin sync drift | Low | Medium | Mirror protocols in same PR; diff verification |
+| CLAUDE.md too compressed | Low (was Medium) | Medium | Revised target ~100-110 lines; phase principles preserved |
+| Regression in behavior | Low | Medium | Baseline capture + post-implementation comparison |
+
+---
+
+## Scope Assessment
+
+- **Overall Complexity**: Medium (Variety Score: 8)
+- **Estimated Files**: 8 new protocol files, 12 modified files (5 original + 7 agent files)
+- **Specialists Required**: None (AI tooling — orchestrator may implement directly)
+- **External Dependencies**: No — internal refactoring only
+- **Estimated Time**: 6-7 hours implementation + 5-6 hours validation
+
+---
+
+## PACT Framework Coverage
+
+This plan addresses context management across the **entire PACT framework**.
+
+### Complete Command Inventory
+
+| Command | Lines | Current @-Protocol Load | New @-Protocol Load | Plan Action |
+|---------|-------|-------------------------|---------------------|-------------|
+| **orchestrate.md** | 480 | ~1,320 lines | ~715 lines | ✅ Targeted @-refs |
+| **plan-mode.md** | 441 | ~1,085 lines | ~165 lines | ✅ Targeted @-refs |
+| **imPACT.md** | ~100 | ~235 lines | ~340 lines | ✅ Add workflow-protocols |
+| **rePACT.md** | 250 | 0 | 0 | ⚪ No changes needed (verified) |
+| **comPACT.md** | 190 | 0 | ~105 (optional) | ⚪ Optional workflow-protocols ref |
+| **peer-review.md** | 27 | 0 | 0 | ⚪ No changes needed |
+| **wrap-up.md** | 35 | 0 | 0 | ⚪ No changes needed |
+| **log-changes.md** | 27 | 0 | 0 | ⚪ No changes needed |
+
+### rePACT Dependency Verification
+
+**Finding**: `rePACT.md` contains conceptual references to protocols ("Apply S2 coordination", "see Autonomy Charter") but these are **inline explanatory text**, not @-file references. The command is self-contained with its protocol details inline. **No changes needed** — verified correct.
+
+### Agent Files Audit
+
+**Finding**: Grep search of `.claude/agents/` for `@.*pact-protocols|@.*protocols/` returned **21 matches across 7 agent files**. Each agent references `pact-protocols.md` for Cross-Agent Coordination and S1 Autonomy rules. Additionally, `pact-preparer.md` references the S4 Environment Model. **8 cross-references require updates** (7 to @phase-guidance.md, 1 to @s4-checkpoints.md). See "Agent Files Cross-Reference Inventory" section for full mapping.
+
+### Universal Benefit: CLAUDE.md Compression
+
+| Entry Point | Current CLAUDE.md | After Compression | Savings |
+|-------------|-------------------|-------------------|---------|
+| All commands | 306 lines (~4K tokens) | ~100-110 lines (~1.3K tokens) | **67%** |
+
+### Per-Command Context Load (Before/After)
+
+| Command | Current Total | After Plan | Reduction |
+|---------|---------------|------------|-----------|
+| **orchestrate** | ~2,100 lines | ~1,300 lines | **38%** |
+| **plan-mode** | ~1,830 lines | ~720 lines | **61%** |
+| **imPACT** | ~640 lines | ~555 lines | **13%** |
+| **comPACT** | ~496 lines | ~315 lines | **36%** |
+| **rePACT** | ~556 lines | ~360 lines | **35%** |
+| **peer-review** | ~333 lines | ~137 lines | **59%** |
+| **wrap-up** | ~341 lines | ~145 lines | **57%** |
+| **log-changes** | ~333 lines | ~137 lines | **59%** |
+
+*Note: Totals include CLAUDE.md + command file + @-referenced protocol files.*
+
+---
+
+## Token Savings Analysis
+
+| Scenario | Current Load | After Refactor | Savings |
+|----------|--------------|----------------|---------|
+| Session start (CLAUDE.md) | ~4K | ~1.3K | 67% |
+| /PACT:orchestrate invocation | ~15K | ~8K | 47% |
+| /PACT:plan-mode invocation | ~12K | ~3K | 75% |
+| /PACT:comPACT invocation | ~2K | ~1.5K | 25% |
+| Typical full orchestration | ~45K | ~22K | **51%** |
+
+---
+
+## Comparison: Protocol Partitioning vs Skills
+
+| Aspect | Skills (Superseded Plan) | Protocol Partitioning (This Plan) |
+|--------|--------------------------|-----------------------------------|
+| **Loading behavior** | Claude decides when to load | Deterministic, controlled |
+| **Predictability** | Variable (auto-discovery) | Consistent (explicit @-refs) |
+| **Fit for PACT** | Suboptimal (PACT is structured) | Optimal (matches PACT's determinism) |
+| **Implementation** | New skill files + frontmatter | Partition existing content |
+| **Risk** | False positive/negative loading | None (explicit control) |
+| **Estimated savings** | ~73% | ~51% |
+
+**Trade-off**: Protocol partitioning sacrifices ~22% potential savings for predictability, verifiability, and alignment with PACT's structured workflow nature.
+
+---
+
+## Related Context
+
+- **Issue**: #64 - Context Architecture Review
+- **Supersedes**: PR #68 draft plan, `context-architecture-skills-plan.md`
+- **Research Basis**: [Anthropic Context Engineering Guide](https://www.anthropic.com/research/building-effective-agents), PACT workflow analysis
+- **Peer Review**: Incorporated feedback from pact-architect, pact-preparer, pact-test-engineer (2026-01-15)
+
+---
+
+## Next Steps
+
+After approval:
+```
+/PACT:orchestrate implement protocol partitioning context architecture optimization per approved plan
+```
+
+The orchestrator will reference this plan during execution. Since this is AI tooling (not application code), the orchestrator may implement changes directly without delegating to specialist coders.
diff --git a/docs/plans/context-architecture-skills-plan.md b/docs/plans/context-architecture-skills-plan.md
new file mode 100644
index 00000000..9ac49fe0
--- /dev/null
+++ b/docs/plans/context-architecture-skills-plan.md
@@ -0,0 +1,485 @@
+# Implementation Plan: Context Architecture Optimization (Skills-Based)
+
+> Generated by `/PACT:plan-mode` on 2026-01-15
+> Status: SUPERSEDED
+> **Superseded by**: `context-architecture-protocols-plan.md` (protocol partitioning approach preferred over skills)
+
+<!-- Status Lifecycle:
+     PENDING APPROVAL → APPROVED → IN_PROGRESS → IMPLEMENTED
+                    ↘ SUPERSEDED (if replaced by newer plan)
+                    ↘ BLOCKED (if unresolved conflicts)
+
+     Transition Ownership:
+     - PENDING APPROVAL → APPROVED: User (explicit approval)
+     - APPROVED → IN_PROGRESS: Orchestrator (when /PACT:orchestrate starts)
+     - IN_PROGRESS → IMPLEMENTED: Orchestrator (after successful completion)
+     - Any → SUPERSEDED: plan-mode (when creating replacement plan)
+     - Any → BLOCKED: plan-mode (when unresolved blocking conflicts)
+-->
+
+## Summary
+
+Implement skills-based lazy loading for the PACT framework by compressing CLAUDE.md from 306 lines to ~85 lines and creating three new protocol skills (`pact-delegation`, `pact-orchestration-modes`, `pact-variety-assessment`) to replace monolithic pact-protocols.md loading.
+
+**Framework-wide impact**: This refactoring benefits all 8 PACT commands, not just `/PACT:orchestrate`. The heaviest commands (`orchestrate`, `plan-mode`) see 67-71% reductions, while all commands benefit from 72% smaller CLAUDE.md at session start. Expected savings: ~73% reduction in typical orchestration context load (~45K → ~12K tokens).
+
+**Supersedes**: PR #68 draft plan (protocol-file approach replaced by skills-based approach per Anthropic research)
+
+---
+
+## Specialist Perspectives
+
+### 📋 Preparation Phase
+**Effort**: Low-Medium
+
+#### Research Completed
+- [x] Anthropic context engineering guidance reviewed
+- [x] CLAUDE.md best practices documented (< 60 lines ideal, < 300 max)
+- [x] Skills lazy-loading behavior understood (metadata at startup, full on activation)
+- [x] Current context load measured (CLAUDE.md ~4K, pact-protocols.md ~10K tokens)
+
+#### Remaining Gaps (Acceptable Risk)
+- [ ] Skill auto-activation keyword patterns — validate during implementation
+- [ ] Cross-skill dependency handling — test during Phase 5
+
+#### Questions Resolved
+- **Content classification**: Architect provided detailed partitioning (see Architecture Phase)
+- **Algedonic location**: Keep separate algedonic.md file (already concise, safety-critical)
+- **VSM distribution**: Keep vsm-glossary.md as central reference; skills link to it
+
+---
+
+### 🏗️ Architecture Phase
+**Effort**: Medium
+
+#### Components Affected
+| Component | Change Type | Impact |
+|-----------|-------------|--------|
+| `CLAUDE.md` | Compress | 306 → ~85 lines |
+| `pact-protocols.md` | Reduce | 1,086 → ~300 lines (keep S1/S2/S3* only) |
+| `commands/PACT/orchestrate.md` | Modify | Remove @protocol refs, add skill invocations |
+| `commands/PACT/plan-mode.md` | Modify | Remove @protocol refs, add skill invocations |
+| `commands/PACT/imPACT.md` | Modify | Remove @protocol refs |
+| `skills/pact-delegation/` | New | ~120 lines |
+| `skills/pact-orchestration-modes/` | New | ~180 lines |
+| `skills/pact-variety-assessment/` | New | ~150 lines |
+
+#### Design Approach
+
+**Three-Skill Architecture**:
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│                       CLAUDE.md (~85 lines)                 │
+│  Always loaded: Mission, S5 Policy, Specialist list,       │
+│                 Command list, Skill activation guide        │
+└─────────────────────────────────────────────────────────────┘
+         │
+         │ References by name (lazy-load on demand)
+         ▼
+┌─────────────────────────────────────────────────────────────┐
+│                      Protocol Skills                        │
+│                                                             │
+│  pact-delegation (~120 lines)                               │
+│  └─ Tool checkpoint, app code definition, recovery          │
+│                                                             │
+│  pact-orchestration-modes (~180 lines)                      │
+│  └─ S3/S4 modes, checkpoints, tensions, temporal horizons   │
+│                                                             │
+│  pact-variety-assessment (~150 lines)                       │
+│  └─ Variety scoring, workflow selection, strategies         │
+└─────────────────────────────────────────────────────────────┘
+         │
+         │ Auto-activated by keywords in commands
+         ▼
+┌─────────────────────────────────────────────────────────────┐
+│  Commands (loaded on invocation)                            │
+│  └─ No @pact-protocols refs; reference skills by name       │
+└─────────────────────────────────────────────────────────────┘
+```
+
+#### Key Decisions
+| Decision | Options | Recommendation | Rationale |
+|----------|---------|----------------|-----------|
+| Skill count | 3 focused vs 4-5 fine-grained | 3 skills | Clear boundaries, targeted loading, simpler maintenance |
+| CLAUDE.md size | 60-80 vs 80-100 lines | ~85 lines | Balance between compression and safety-critical content |
+| Algedonic handling | In CLAUDE.md vs skill vs separate | Keep algedonic.md | Already concise, safety-critical, well-isolated |
+| Protocol residue | Delete vs lean file | Lean (~300 lines) | Keep S1/S2/S3* for specialized cases |
+
+#### Interface Contracts
+
+**Skill Discovery Mechanism**:
+
+Skills in Claude Code are auto-discovered based on their `description` frontmatter. Claude reads all skill descriptions at session start and activates the relevant skill when conversation context matches.
+
+**Command Text Pattern** (how commands trigger skills):
+```markdown
+# Current (loads entire 1,086-line file):
+See @~/.claude/protocols/pact-protocols.md for S4 Checkpoint Protocol.
+
+# New (triggers skill discovery via keywords):
+When performing S4 checkpoints, apply the S4 Checkpoint Protocol.
+```
+
+The new text contains "S4 checkpoint" which matches `pact-orchestration-modes` description, causing Claude to load only that ~180-line skill.
+
+**Trigger Keywords by Skill** (embedded in description):
+| Skill | Keywords in Description |
+|-------|-------------------------|
+| `pact-delegation` | delegate, Edit, Write, application code, tool checkpoint, recovery |
+| `pact-orchestration-modes` | S3, S4, checkpoint, tension, phase transition, temporal horizon |
+| `pact-variety-assessment` | variety, complexity, comPACT, orchestrate, plan-mode, workflow selection |
+
+**Best Practice**: Commands should use natural language containing trigger keywords rather than explicit "invoke skill X" instructions. This allows Claude's discovery mechanism to work naturally.
+
+---
+
+### 💻 Code Phase
+**Effort**: Medium
+
+> **Note**: This is AI tooling (.claude/ files), not application code. Per delegation rules, the orchestrator may implement these changes directly. No specialist delegation required.
+
+#### Files to Create
+| File | Purpose | Est. Lines |
+|------|---------|------------|
+| `.claude/skills/pact-delegation/SKILL.md` | Delegation enforcement, tool checkpoint, recovery | ~120 |
+| `.claude/skills/pact-orchestration-modes/SKILL.md` | S3/S4 modes, checkpoints, temporal horizons | ~180 |
+| `.claude/skills/pact-variety-assessment/SKILL.md` | Variety scoring, workflow selection | ~150 |
+
+#### Skill Specifications (Frontmatter & Discovery)
+
+Skills are discovered through their `description` field, not @-references. Claude reads skill descriptions at session start and auto-activates based on context matching.
+
+**pact-delegation**:
+```yaml
+---
+name: pact-delegation
+description: |
+  Delegation enforcement rules and tool checkpoint protocol for PACT orchestrators.
+  Use when: about to edit application code, using Edit/Write tools, determining
+  what constitutes application code, or recovering from delegation violations.
+  Triggers on: delegate, Edit, Write, application code, tool checkpoint, recovery,
+  specialist agents, code ownership.
+---
+```
+
+**pact-orchestration-modes**:
+```yaml
+---
+name: pact-orchestration-modes
+description: |
+  S3/S4 operational modes, checkpoints, and tension resolution for PACT orchestration.
+  Use when: transitioning between phases, performing S4 checkpoints, detecting S3/S4
+  tension, assessing temporal horizons, or managing mode transitions.
+  Triggers on: S3, S4, checkpoint, tension, phase transition, temporal horizon,
+  inside-now, outside-future, mode, operational control, strategic intelligence.
+---
+```
+
+**pact-variety-assessment**:
+```yaml
+---
+name: pact-variety-assessment
+description: |
+  Variety scoring rubric and workflow selection for PACT task complexity assessment.
+  Use when: starting a new task, selecting between comPACT/orchestrate/plan-mode,
+  assessing task complexity, or applying variety attenuators/amplifiers.
+  Triggers on: variety, complexity, comPACT, orchestrate, plan-mode, workflow
+  selection, novelty, scope, uncertainty, risk, attenuate, amplify.
+---
+```
+
+#### How Skill Discovery Replaces @-References
+
+**Current pattern** (loads entire file):
+```markdown
+See @~/.claude/protocols/pact-protocols.md for S4 Checkpoint Protocol.
+```
+
+**New pattern** (triggers skill discovery):
+```markdown
+When performing S4 checkpoints, apply the S4 Checkpoint Protocol.
+```
+
+Claude's skill discovery matches "S4 checkpoint" against `pact-orchestration-modes` description and loads only that skill (~180 lines vs ~1,086 lines).
+
+**Key insight**: We remove explicit @-references and use natural language that contains the trigger keywords defined in skill descriptions. Claude handles the discovery automatically.
+
+#### Files to Modify
+| File | Changes |
+|------|---------|
+| `CLAUDE.md` | Compress from 306 → ~85 lines per structure below |
+| `.claude/protocols/pact-protocols.md` | Reduce from 1,086 → ~300 lines (keep S1/S2/S3* only) |
+| `.claude/commands/PACT/orchestrate.md` | Remove @pact-protocols refs, add skill references |
+| `.claude/commands/PACT/plan-mode.md` | Remove @pact-protocols ref, add skill references |
+| `.claude/commands/PACT/imPACT.md` | Remove @algedonic ref (keep algedonic.md reference) |
+| `pact-plugin/skills/` | Mirror new skills for plugin distribution |
+
+#### CLAUDE.md Proposed Structure (~85 lines)
+
+```markdown
+# MISSION (7 lines)
+Mission statement, motto
+
+## S5 POLICY (12 lines)
+Non-negotiables table (4 rows), policy checkpoint summary
+→ "Invoke pact-delegation skill for enforcement details"
+
+## INSTRUCTIONS (6 lines)
+Core directives
+
+## GUIDELINES (12 lines)
+Context management, git workflow, communication
+
+## PACT PHASES (8 lines)
+Phase names + one-liner principles
+→ "Invoke domain skills for detailed phase guidance"
+
+## SPECIALIST AGENTS (10 lines)
+Agent list with emoji identifiers
+
+## ORCHESTRATION COMMANDS (12 lines)
+Command list with one-line descriptions
+→ "Invoke pact-variety-assessment skill to select workflow"
+
+## DOCUMENTATION LOCATIONS (8 lines)
+Phase → Output Location table
+
+## SKILL ACTIVATION GUIDE (10 lines)
+When to invoke each protocol skill
+```
+
+#### Implementation Sequence
+1. `feat(skills): create pact-delegation skill` — Extract delegation content
+2. `feat(skills): create pact-orchestration-modes skill` — Extract S3/S4 content
+3. `feat(skills): create pact-variety-assessment skill` — Extract variety content
+4. `refactor(claude): compress CLAUDE.md to ~85 lines` — Restructure with skill refs
+5. `refactor(protocols): reduce pact-protocols.md to ~300 lines` — Keep S1/S2/S3* only
+6. `refactor(commands): update orchestrate.md skill references` — Remove @protocol loads
+7. `refactor(commands): update plan-mode.md skill references` — Remove @protocol loads
+8. `refactor(commands): update imPACT.md references` — Minor cleanup
+9. `chore(plugin): sync new skills to pact-plugin/` — Mirror for distribution
+
+---
+
+### 🧪 Test Phase
+**Effort**: Medium (4-6 hours manual validation)
+
+#### Test Scenarios
+| Scenario | Type | Priority |
+|----------|------|----------|
+| `/PACT:orchestrate` invokes correct skills | Integration | P0 |
+| `/PACT:comPACT` loads delegation skill | Integration | P0 |
+| Delegation enforcement on app code edit | Functional | P0 |
+| Algedonic HALT signal bypasses to user | Functional | P0 |
+| Tool Checkpoint Protocol fires before Edit/Write | Functional | P0 |
+| `/PACT:plan-mode` activates variety skill | Integration | P1 |
+| `/PACT:imPACT` loads triage guidance | Integration | P1 |
+| S3/S4 mode detection in responses | Functional | P1 |
+| `/PACT:rePACT` nested skill loading | Edge case | P2 |
+| Multiple skills loaded in single command | Edge case | P2 |
+
+#### Coverage Targets
+- Critical path (P0): 100% — Must pass for merge
+- Core workflows (P1): 100% — Should pass for merge
+- Edge cases (P2): Best effort — May defer
+
+#### Validation Approach
+
+**Phase 1: Smoke Testing (1-2 hours)**
+- Execute each PACT command once
+- Verify skill-specific guidance appears
+- Check delegation enforcement
+
+**Phase 2: Comprehensive Validation (3-4 hours)**
+- All P0/P1 scenarios
+- Before/after comparison for lost guidance
+- Edge case testing
+
+#### Acceptance Criteria
+- [ ] All new skill files exist and are syntactically valid
+- [ ] Compressed CLAUDE.md contains delegation enforcement rules
+- [ ] Compressed CLAUDE.md contains SACROSANCT policy
+- [ ] `/PACT:orchestrate` executes without errors
+- [ ] `/PACT:comPACT` executes without errors
+- [ ] Tool Checkpoint Protocol triggers on application code
+- [ ] Algedonic HALT scenario escalates to user
+- [ ] No command produces "skill not found" errors
+
+---
+
+## Synthesized Implementation Roadmap
+
+### Phase Sequence
+
+```
+Phase 1: Create Skills (Low Risk)
+├── pact-delegation/SKILL.md
+├── pact-orchestration-modes/SKILL.md
+└── pact-variety-assessment/SKILL.md
+         │
+         ▼ Smoke test skill activation
+         │
+Phase 2: Compress CLAUDE.md (Medium Risk)
+├── Restructure to ~85 lines
+└── Add skill activation guide
+         │
+         ▼ Validate delegation/algedonic preserved
+         │
+Phase 3: Update Commands (Medium Risk)
+├── orchestrate.md
+├── plan-mode.md
+└── imPACT.md
+         │
+         ▼ Full command validation
+         │
+Phase 4: Lean Protocol (Low Risk)
+└── Reduce pact-protocols.md to ~300 lines
+         │
+         ▼ Comprehensive validation
+         │
+Phase 5: Plugin Sync (Low Risk)
+└── Mirror to pact-plugin/
+         │
+         ▼ Final PR review
+```
+
+### Commit Sequence (Proposed)
+
+> **Note**: This sequence represents the intended final git history order, **not** the execution order. Phases 1-3 may have parallel work streams.
+
+1. `feat(skills): create pact-delegation skill with tool checkpoint protocol` — Delegation enforcement content
+2. `feat(skills): create pact-orchestration-modes skill with S3/S4 guidance` — Operational modes content
+3. `feat(skills): create pact-variety-assessment skill with scoring rubric` — Variety management content
+4. `refactor(claude): compress CLAUDE.md from 306 to ~85 lines` — Core restructuring
+5. `refactor(protocols): reduce pact-protocols.md to S1/S2/S3* sections (~300 lines)` — Protocol cleanup
+6. `refactor(commands): update PACT commands to reference skills instead of protocols` — Command updates
+7. `chore(plugin): sync new skills to pact-plugin/ directory` — Distribution sync
+8. `docs: update context architecture documentation` — Document new structure
+
+---
+
+## Cross-Cutting Concerns
+
+| Concern | Status | Notes |
+|---------|--------|-------|
+| Security | ✅ Ready | Delegation enforcement and algedonic signals preserved in accessible locations |
+| Performance | ✅ Focus | ~73% token reduction: ~45K → ~12K per orchestration |
+| Accessibility | N/A | Not frontend work |
+| Observability | ⚠️ Monitor | Track skill activation patterns during validation |
+
+---
+
+## Open Questions
+
+### Require User Decision
+
+None — all key decisions resolved during synthesis.
+
+### Require Further Research
+
+- [ ] **Skill auto-activation reliability**: Validate trigger keywords work as expected during Phase 1
+- [ ] **Cross-skill dependencies**: Test if `pact-variety-assessment` correctly informs orchestration mode selection during Phase 3
+
+---
+
+## Risk Assessment
+
+| Risk | Likelihood | Impact | Mitigation |
+|------|------------|--------|------------|
+| Skill auto-activation fails | Medium | High | Careful keyword selection; fallback instructions in CLAUDE.md |
+| Critical guidance lost in compression | Low | High | Before/after comparison; keep delegation/algedonic explicit |
+| Commands break after @ref removal | Medium | Medium | Staged rollout; test each command after update |
+| Plugin sync drift | Low | Medium | Mirror skills in same PR; verify with diff |
+| Content duplication across skills | Low | Low | Clear ownership; cross-references not duplication |
+
+---
+
+## Scope Assessment
+
+- **Overall Complexity**: Medium (Variety Score: 9)
+- **Estimated Files**: 3 new skill files, 6 modified files
+- **Specialists Required**: None (AI tooling — orchestrator may implement directly)
+- **External Dependencies**: No — internal refactoring only
+
+---
+
+## PACT Framework Coverage
+
+This plan addresses context management across the **entire PACT framework**, not just `/PACT:orchestrate`.
+
+### Complete Command Inventory
+
+| Command | Lines | @-Protocol References | Plan Action |
+|---------|-------|----------------------|-------------|
+| **orchestrate.md** | 480 | 6× pact-protocols + 1× algedonic | ✅ Remove @refs, add skill invocations |
+| **plan-mode.md** | 441 | 1× pact-protocols | ✅ Remove @refs, add skill invocations |
+| **imPACT.md** | ~100 | 2× algedonic | ✅ Simplify algedonic refs |
+| **rePACT.md** | 250 | None (self-contained) | ⚪ No changes needed |
+| **comPACT.md** | 190 | None (self-contained) | ⚪ No changes needed |
+| **peer-review.md** | 27 | None (self-contained) | ⚪ No changes needed |
+| **wrap-up.md** | 35 | None (refs skill) | ⚪ No changes needed |
+| **log-changes.md** | 27 | None (self-contained) | ⚪ No changes needed |
+
+**Key insight**: Only three commands (`orchestrate`, `plan-mode`, `imPACT`) have @-protocol references that cause heavy context loading. The remaining five commands are already lightweight and self-contained.
+
+### Universal Benefit: CLAUDE.md Compression
+
+The CLAUDE.md compression (306 → ~85 lines) benefits **every command and every session**:
+
+| Entry Point | Current CLAUDE.md | After Compression | Savings |
+|-------------|-------------------|-------------------|---------|
+| All commands | 306 lines (~4K tokens) | ~85 lines (~1K tokens) | **72%** |
+
+### Per-Command Context Load (Before/After)
+
+| Command | Current Total | After Plan | Reduction |
+|---------|---------------|------------|-----------|
+| **orchestrate** | ~2,100 lines | ~700 lines | **67%** |
+| **plan-mode** | ~1,830 lines | ~530 lines | **71%** |
+| **imPACT** | ~640 lines | ~320 lines | **50%** |
+| **comPACT** | ~496 lines | ~275 lines | **45%** |
+| **rePACT** | ~556 lines | ~335 lines | **40%** |
+| **peer-review** | ~333 lines | ~112 lines | **66%** |
+| **wrap-up** | ~341 lines | ~120 lines | **65%** |
+| **log-changes** | ~333 lines | ~112 lines | **66%** |
+
+*Note: Totals include CLAUDE.md + command file + any @-referenced protocol files.*
+
+### Why Lightweight Commands Still Benefit
+
+Commands like `comPACT`, `rePACT`, and `peer-review` have no @-protocol references, but they still load CLAUDE.md at session start. The 72% reduction in CLAUDE.md size means:
+- More context budget available for actual task work
+- Reduced risk of hitting context limits during extended sessions
+- Faster session initialization
+
+---
+
+## Token Savings Analysis
+
+| Scenario | Current Load | After Refactor | Savings |
+|----------|--------------|----------------|---------|
+| Session start (CLAUDE.md) | ~4K | ~1K | 75% |
+| /PACT:orchestrate invocation | +15K (protocols) | +2K (targeted skills) | 87% |
+| /PACT:comPACT invocation | ~2K (no protocols) | ~1K | 50% |
+| /PACT:plan-mode invocation | +10K (protocols) | +2K (targeted skills) | 80% |
+| Typical full orchestration | ~45K | ~12K | **73%** |
+
+---
+
+## Related Context
+
+- **Issue**: #64 - Context Architecture Review
+- **Supersedes**: PR #68 draft plan (protocol-file approach)
+- **Research Basis**: Anthropic Context Engineering Guide, Skill Authoring Best Practices
+
+---
+
+## Next Steps
+
+After approval:
+```
+/PACT:orchestrate implement skills-based context architecture optimization per approved plan
+```
+
+The orchestrator will reference this plan during execution. Since this is AI tooling (not application code), the orchestrator may implement changes directly without delegating to specialist coders.