agents/shared/ — Shared Agent Reference Files

agents/shared/ — Shared Agent Reference Files

Shared reference files extracted from duplicated content across multiple agent definitions. Agents load these via Read() directives in their Bootstrap Context section.

Architecture: Two-Track Self-Read

Rune has two distinct spawn patterns for teammates. The Self-Read approach handles both:

Pattern	subagent_type	Agent .md body loaded?	Read() in body works?	Used by
Track 1	rune:{category}:{name}	YES — body = system prompt	YES	strive, arc, mend, inspect
Track 2	general-purpose	NO — only prompt param	NO	appraise, devise plan-review

Track 1 — Custom subagent_type (body IS loaded)

Add ## Bootstrap Context as first content in agent .md body:

## Bootstrap Context (MANDATORY — Read ALL before any work)
1. Read `plugins/rune/agents/shared/communication-protocol.md`
2. Read `plugins/rune/agents/shared/context-checkpoint-protocol.md`

The agent reads shared files from disk at startup. Zero lead context burden.

Track 2 — general-purpose (body NOT loaded)

Prepend Read() directives in the spawn prompt composition:

const bootstrapDirectives = `
## Bootstrap Context (MANDATORY — Read ALL before any work)
1. Read plugins/rune/agents/shared/quality-gate-template.md
`
const composedPrompt = bootstrapDirectives + '\n' + agentBody

5-line bootstrap (~50 tokens) vs 400-line inline rules (~3000 tokens) = 95% lead context reduction per spawn.

Shared Files

Core Protocols

File	Lines	Track 1 Consumers (Read() in agent .md)	Track 2 Consumers (injected by orchestrator)	Content
communication-protocol.md	~36	rune-smith, trial-forger, mend-fixer, gap-fixer, verdict-binder, grace-warden-inspect, sight-oracle-inspect, vigil-keeper-inspect, ruin-prophet-inspect	—	Seal format, shutdown handling, SendMessage conventions, exit conditions
quality-gate-template.md	~37	rune-smith, trial-forger, grace-warden-inspect, sight-oracle-inspect, vigil-keeper-inspect, ruin-prophet-inspect	ward-sentinel, knowledge-keeper (via devise/arc orchestrators)	Confidence calibration, Inner Flame supplementary, self-review pass
context-checkpoint-protocol.md	~39	rune-smith, trial-forger	—	Adaptive reset depth, Seal summary requirements, context rot detection
iron-law-protocol.md	~31	—	—	Iron Law enforcement wrapper ("this rule is absolute"), agent-specific law statements remain inline
truthbinding-protocol.md	~96	rune-smith, trial-forger, mend-fixer, gap-fixer, ward-sentinel, pattern-seer, flaw-hunter, lore-analyst, goldmask-coordinator, grace-warden-inspect, sight-oracle-inspect, vigil-keeper-inspect, ruin-prophet-inspect	—	ANCHOR/RE-ANCHOR Truthbinding security framing, untrusted content handling, injection defense
finding-format-template.md	~62	ward-sentinel, pattern-seer, flaw-hunter, grace-warden-inspect, knowledge-keeper	—	Standardized finding output format, severity levels, evidence requirements

Phase-Specific Protocols

File	Lines	Track 1 Consumers (Read() in agent .md)	Content
phase-review.md	~149	ward-sentinel, pattern-seer, flaw-hunter, knowledge-keeper	Review-phase conventions, file scope, output format, review workflow
phase-work.md	~112	rune-smith, trial-forger	Work-phase swarm worker patterns, TDD cycle, ward checks, task lifecycle
phase-goldmask.md	~101	lore-analyst, goldmask-coordinator	Goldmask investigation patterns, risk scoring, lore layer conventions
phase-inspect.md	~120	grace-warden-inspect, sight-oracle-inspect, vigil-keeper-inspect, ruin-prophet-inspect	Inspect investigation patterns, requirement matrix, gap categories
phase-devise.md	~121	knowledge-keeper, scroll-reviewer, decree-arbiter	Devise/planning utility patterns, plan review conventions, enrichment workflow

Note: Track 2 consumers are general-purpose agents whose Read() directives are composed by orchestrator skills, not embedded in agent .md files. The validate-agent-shared-refs.sh script only validates Track 1 references. Track 2 validation is out of scope for the current script — broken Track 2 references are discovered at runtime.

Rules

1. DO NOT duplicate — if content exists in a shared file, do not inline it in agent .md
2. Extraction header — every shared file has a <!-- Source: --> comment listing origin agents and date
3. Agent-specific stays inline — ANCHOR/RE-ANCHOR blocks, Iron Law statements, and role-specific rules remain in each agent's .md file
4. Flat structure — no transitive includes; shared files do not Read() other shared files
5. Size cap — shared files should stay under 150 lines each
6. Bootstrap abort-on-failure — if any Read() fails, agents should STOP and report to team-lead

Enforcement Limitations

Bootstrap abort-on-failure (Rule #6) is prompt-level enforcement only. There is no hook, script, or programmatic check that verifies an agent actually stops when a Read() returns an error. The LLM may continue processing despite the error, especially under context pressure.

Accepted risk: Shared files are git-tracked and versioned. Accidental deletion or rename would be caught by PR review and the validate-agent-shared-refs.sh validation script (Check 1: SHARED-001 verifies file existence). The residual risk of an agent proceeding without shared protocol content is low-probability and bounded — agents would produce malformed output that downstream aggregation would reject.

Future hardening options (not currently implemented):

1. SubagentStart hook that validates shared file existence before the agent starts
2. PreToolUse:TaskUpdate hook that checks Bootstrap Read() completion before task claiming
3. PostToolUse:Read hook that tracks which shared files were successfully loaded