Rune Orchestration Skill

Guides designing distributed multi-agent systems that partition work across multiple language model instances to overcome single-agent context limitations.

Core Principle

Sub-agents exist primarily to isolate context, not to anthropomorphize role division.

The fundamental value lies in distributing cognitive load across separate context windows rather than simulating organizational hierarchies.

When to Use

Generic trigger conditions:

• Context window constraints from single-agent approach
• Complex tasks requiring specialized expertise in parallel
• Need for parallel processing of independent subtasks
• Any workflow summoning 3+ agents via Agent tool or Agent Teams

Agent Coordination Patterns

File-Based Handoff Pattern

Instead of supervisor synthesizing responses (telephone game), agents write directly to files:

Directory Purpose:

• Persistent (todos/): Findings that become actionable items
• Ephemeral (tmp/): Intermediate artifacts consumed by coordinator

The workflow: Agents write findings → Coordinator reads files → Synthesizes into Tome (TOME.md)

Structured output: Ash MAY also write companion JSON files ({ash}-findings.json) for CI/CD integration. Legacy: completion.json was used in early versions but is no longer written by built-in workflows. Use Seal metadata + TOME.md for workflow results. See Output Format for full specs.

Agent Output Formats

Each agent writes findings in a format matching its workflow type. All formats require mandatory evidence blocks.

• Report Format (Reviews, Audits): P1/P2/P3 severity with Rune Traces. See Output Formats
• Research Format (Plans): Knowledge synthesis with sources. See Output Formats
• Status Format (Work): Implementation progress reports. See Output Formats

Conflict Resolution Rules

When multiple agents flag the same code with different priorities:

Priority Hierarchy

ward-sentinel P1 > any other agent P1
any agent P1 > any agent P2
any agent P2 > any agent P3

Three Architectural Patterns

1. Supervisor/Orchestrator (Default)

Central coordinator delegates to specialists and aggregates results.

The Tarnished (orchestrator)
    ├── forge-warden   (backend review)
    ├── ward-sentinel  (security review)
    ├── pattern-weaver (quality patterns)
    ├── glyph-scribe   (frontend review)
    └── knowledge-keeper    (docs review)

Used by: /rune:appraise, /rune:audit

Mitigate telephone game: Use file-based handoff, not message passing.

2. Peer-to-Peer/Swarm

Direct agent-to-agent communication. Workers self-organize and race to claim tasks.

Used by: /rune:strive

Example use case: Multiple rune-smiths claiming independent tasks from a shared pool.

3. Hierarchical

Layered abstractions. Use for complex multi-phase workflows.

Strategy Layer (plan)
    └── Planning Layer (spec)
        └── Execution Layer (implement)

Used by: /rune:devise → /rune:strive → /rune:appraise

Agent Role Patterns

Defines how agents are organized for each workflow type, including conditional summoning rules and validation pipelines.

• Review Ash: Parallel specialists writing to tmp/reviews/{pr}/. See Role Patterns
• Audit Ash: Fan-out/fan-in to tmp/audit/{id}/. See Role Patterns
• Research Agents: Parallel exploration to tmp/plans/{timestamp}/research/. See Role Patterns
• Work Agents (Rune Smiths): Swarm workers to tmp/work/. See Role Patterns
• Conditional Ash: Summoned based on file types in scope. See Role Patterns
• Validation Agents (Truthsight Pipeline): Post-review verification. See Role Patterns

Token Economics

Guideline: Use multi-agent only when context isolation is necessary, not for simple division of labor.

Essential Failure Mitigations

1. Supervisor Saturation

Problem: Coordinator overwhelmed by aggregating many agent responses.

Solution: Glyph Budget — each agent writes to file, returns only file path + 1-sentence summary (max 50 words).

2. Coordination Overhead

Problem: Too much time on handoff protocols.

Solution: Clear file-based handoff with consistent format.

3. Lack of Convergence

Problem: Agents disagree indefinitely.

Solution:

• Time-to-live limits (max 2 iterations)
• Priority hierarchy for conflicts
• Human review for unresolved conflicts

4. Error Propagation

Problem: Early agent error cascades.

Solution:

• Validate outputs between agents
• Each agent reads source code directly, not previous agent's summary

Glyph Budget

Every agent in a parallel workflow:

1. Writes detailed findings to a file
2. Returns only: file path + 1-sentence summary (max 50 words)
3. Does not include analysis in the return message

See context-weaving skill for the full Glyph Budget protocol and pre-summon checklist.

Inscription Protocol

All Rune multi-agent workflows use Agent Teams (TeamCreate) + inscription.json.

Rune command (any agent count) → Agent Teams + inscription.json
Custom workflow (3+ agents)    → Agent Teams + inscription.json

Three steps:

1. Generate inscription.json before summoning agents/teammates
2. Inject required sections + Seal Format into each agent prompt
3. Validate output files after completion (circuit breaker → per-file → gap report)

Agents send structured Seal messages (key-value: file, sections, findings, evidence-verified, confidence, self-reviewed, summary). TaskList completed is authoritative; Seal is supplementary metadata.

Full spec: Inscription Protocol

Structured Reasoning Integration

Multi-agent workflows benefit from structured reasoning at key decision points.

Three reasoning checkpoints for the Tarnished:

Full protocol: Structured Reasoning

Compaction Recovery

On context compaction or session resume during any Rune workflow:

1. Read active team config: ~/.claude/teams/{team-name}/config.json
2. Read task list via TaskList() to discover team state
3. Read inscription contract: tmp/{workflow}/{id}/inscription.json
4. For arc workflows: read .rune/arc/{id}/checkpoint.json
5. Resume from the last incomplete task — do not re-read completed agent outputs
6. If team config is missing: the workflow was interrupted before team creation — restart

Session handoff template: session-handoff.md

References

• Output Formats — Report, Research, and Status output templates
• Role Patterns — Agent organization per workflow type, conditional summoning, validation pipeline
• Structured Reasoning — Reasoning principles for lead + teammate reasoning
• Inscription Protocol — Output validation for all multi-agent workflows
• Prompt Weaving — 7-section prompt template, context rot prevention, instruction anchoring
• Truthsight Pipeline — 4-layer verification spec
• Verifier Prompt — Smart Verifier prompt template
• Damage Control — 6 recovery procedures for agent failures (DC-1 through DC-6)
• Companion: context-weaving (Glyph Budget, pre-summon checklist)
• Review workflow: roundtable-circle skill