Roundtable Circle Skill

Orchestrates multi-agent code reviews using Claude Code Agent Teams. Each Runebearer teammate gets its own 200k context window, eliminating single-context bottlenecks.

Architecture

7-Phase Lifecycle

Phase 0: Pre-flight     → Validate git status, check for changes
Phase 1: Rune Gaze      → git diff → classify files → select Runebearers
Phase 2: Forge Team      → TeamCreate + TaskCreate + inscription.json
Phase 3: Spawn           → Fan-out Runebearers with self-organizing prompts
Phase 4: Monitor         → TaskList polling, 5-min stale detection
Phase 5: Aggregate       → Spawn Runebinder → writes TOME.md
Phase 6: Verify          → Truthsight validation on P1 findings
Phase 7: Cleanup         → Shutdown requests → approvals → TeamDelete

Built-in Runebearer Roles (Max 5)

Plus Runebinder (utility) for aggregation in Phase 5.

Custom Runebearers (Extensible)

Projects can register additional Runebearers from local agents, global agents, or other plugins via rune-config.yml. Custom Runebearers join the standard lifecycle:

• •Wrapped with Truthbinding Protocol (evidence, Glyph Budget, Seal format)
• •Spawned alongside built-ins in Phase 3 (parallel execution)
• •Deduplicated using their unique finding_prefix in the extended hierarchy
• •Verified by Truthsight (if settings.verification.layer_2_custom_agents: true)
• •Aggregated into TOME.md by Runebinder

Max total: 5 built-in + up to 3 custom = 8 Runebearers (configurable via settings.max_runebearers). The cap exists because each Runebearer output (~10k tokens) consumes verifier context budget.

See custom-runebearers.md for full schema, wrapper prompt template, and examples.

Output Directory Structure

tmp/reviews/{id}/
├── inscription.json         # Output contract (generated Phase 2)
├── forge-warden.md          # Backend review findings
├── ward-sentinel.md         # Security review findings
├── pattern-weaver.md        # Quality patterns findings
├── glyph-scribe.md          # Frontend review findings (if spawned)
├── knowledge-keeper.md           # Docs review findings (if spawned)
├── TOME.md                  # Aggregated + deduplicated findings
├── truthsight-report.md     # Verification results (if Layer 2 enabled)
└── completion.json          # Structured completion summary

Audit Mode

/rune:audit reuses the same 7-phase lifecycle with one difference in Phase 0:

Phases 1-7 are identical. Same Runebearers, same inscription schema, same dedup, same verification. Audit file prioritization differs: importance-based (entry points, core modules) instead of recency-based (new files, modified files).

Audit-Specific: Truthseer Validator

For audits with high file counts (>100 reviewable files), a Truthseer Validator phase runs between Phase 5 and Phase 6:

Phase 5.5: Truthseer Validator
  1. Read all Runebearer outputs
  2. Cross-reference finding density against file importance
  3. Flag under-reviewed areas (high-importance files with 0 findings)
  4. Score confidence per Runebearer based on evidence quality
  5. Write validation summary to {output_dir}/validator-summary.md

The Validator ensures audit coverage quality by detecting:

• •Under-coverage: Critical files reviewed but no findings (suspicious silence)
• •Over-confidence: High finding counts with low evidence quality
• •Scope gaps: Files in budget that weren't actually read

See Validator Rules for confidence scoring and risk classification.

Phase 0: Pre-flight

# Check for changes to review
git diff --name-only HEAD~1..HEAD
# OR for PRs:
git diff --name-only main..HEAD

Abort conditions:

• •No files changed → "Nothing to review"
• •Only non-reviewable files (images, lock files) → "No reviewable changes"

Phase 1: Rune Gaze (Scope Selection)

Classify changed files by extension to determine which Runebearers to spawn.

See Rune Gaze for the full file classification algorithm.

Quick reference:

Phase 2: Forge Team

1. mkdir -p tmp/reviews/{pr-number}/
2. Generate inscription.json (see references/inscription-schema.md)
3. TeamCreate({ team_name: "rune-review-{pr}" })
4. For each selected Runebearer:
   TaskCreate({
     subject: "Review {scope} as {role}",
     description: "Files: [...], Output: tmp/reviews/{pr}/{role}.md"
   })

Phase 3: Spawn Runebearers

For each selected Runebearer, spawn as a background teammate:

Task({
  team_name: "rune-review-{pr}",
  name: "{runebearer-name}",
  subagent_type: "general-purpose",
  prompt: [from references/runebearer-prompts/{role}.md],
  run_in_background: true
})

Each Runebearer prompt includes:

• •Truthbinding Protocol (ANCHOR + RE-ANCHOR)
• •Task claiming via TaskList/TaskUpdate
• •Glyph Budget enforcement
• •Seal Format for completion

Seal Format

Each Runebearer writes a Seal at the end of their output file to signal completion:

---
SEAL: {
  findings: 7,
  evidence_verified: true,
  confidence: 0.85,
  self_reviewed: true,
  self_review_actions: "confirmed: 5, revised: 1, deleted: 1"
}
---

Then sends to lead (max 50 words — Glyph Budget enforced):

"Seal: forge-warden complete. Path: tmp/reviews/142/forge-warden.md.
Findings: 2 P1, 3 P2, 2 P3. Confidence: 0.85. Self-reviewed: yes."

Full spec: Inscription Protocol

See references/runebearer-prompts/ for individual prompts.

Phase 4: Monitor

Poll TaskList every 30 seconds:

while (not all tasks completed):
  tasks = TaskList()
  for task in tasks:
    if task.status == "completed":
      continue
    if task.stale > 5 minutes:
      warn("Runebearer {name} may be stalled")
  sleep(30)

Stale detection: If a task has been in_progress for > 5 minutes:

• •Check teammate status
• •Default: proceed with partial results
• •Gap will be reported in TOME.md

Phase 5: Aggregate

After all tasks complete (or timeout), spawn Runebinder:

Task({
  team_name: "rune-review-{pr}",
  name: "runebinder",
  subagent_type: "general-purpose",
  prompt: "Read all findings from tmp/reviews/{pr}/. Write TOME.md..."
})

The Runebinder:

1. •Reads all Runebearer output files
2. •Deduplicates findings (see references/dedup-runes.md)
3. •Prioritizes: P1 first, then P2, then P3
4. •Reports gaps from crashed/stalled Runebearers
5. •Writes tmp/reviews/{pr}/TOME.md

Phase 6: Verify (Truthsight)

If verification is enabled in inscription.json:

Layer 0: Inline Checks (Lead Agent)

For each Runebearer output file, run grep-based validation:

# Required structure checks
grep -c "## P1" {output_file}      # P1 section exists
grep -c "## P2" {output_file}      # P2 section exists
grep -c "## Summary" {output_file} # Summary section exists
grep -c "SEAL:" {output_file}      # Seal present

# Evidence quality checks
grep -c "Rune Trace" {output_file} # Evidence blocks exist

Circuit breaker: If 3+ files fail inline checks → systemic prompt issue. Pause and investigate.

Layer 1: Self-Review (Each Runebearer)

Already performed by each Runebearer before sending Seal (embedded in prompts). Review the Self-Review Log section in each output file.

Layer 2: Smart Verifier (Spawned by Lead)

Spawn conditions: Roundtable Circle with 3+ Runebearers, or audit with 5+ Runebearers.

Task({
  subagent_type: "general-purpose",
  model: "haiku",
  description: "Truthsight Verifier",
  prompt: [from rune-orchestration/references/verifier-prompt.md]
})

The verifier:

1. •Reads each Runebearer's output file
2. •Samples 2-3 P1 findings per Runebearer
3. •Reads the actual source files cited in Rune Traces
4. •Compares evidence blocks against real code
5. •Marks each: CONFIRMED / INACCURATE / HALLUCINATED
6. •Writes {output_dir}/truthsight-report.md

Circuit breaker: 2+ HALLUCINATED findings from same Runebearer → flag entire output as unreliable.

completion.json

After verification, write structured completion summary:

{
  "workflow": "rune-review",
  "identifier": "PR #142",
  "completed_at": "2026-02-11T11:00:00Z",
  "runebearers": {
    "forge-warden": { "status": "complete", "findings": 7, "confidence": 0.85 },
    "ward-sentinel": { "status": "complete", "findings": 3, "confidence": 0.90 },
    "pattern-weaver": { "status": "partial", "findings": 2, "confidence": 0.60 }
  },
  "aggregation": { "tome_path": "tmp/reviews/142/TOME.md", "total_findings": 12 },
  "verification": { "layer_0_passed": true, "layer_2_hallucinated": 0 }
}

Full verification spec: Truthsight Pipeline

Phase 7: Cleanup

1. SendMessage(type: "shutdown_request") to each Runebearer
2. Wait for shutdown approvals
3. TeamDelete()
4. Read TOME.md and present to user

Error Handling

Cancellation

/rune:cancel-review triggers:

1. •SendMessage(type: "broadcast", content: "Review cancelled by user")
2. •SendMessage(type: "shutdown_request") to each teammate
3. •Wait for approvals (max 30s)
4. •TeamDelete()
5. •Partial results remain in tmp/reviews/{pr}/

/rune:cancel-audit triggers the same cancellation flow with tmp/.rune-audit-* state files. Partial results remain in tmp/audit/{id}/.

References

• •Rune Gaze — File classification algorithm
• •Circle Registry — Agent-to-Runebearer mapping, audit scope priorities, focus mode
• •Smart Selection — File-to-Runebearer assignment, context budgets, focus mode
• •Task Templates — TaskCreate templates for each Runebearer role
• •Output Format — Raw finding format, validated format, TOME format, JSON output
• •Validator Rules — Confidence scoring, risk classification, dedup, gap reporting
• •Runebearer Prompts — Individual Runebearer prompts
• •Inscription Schema — inscription.json format
• •Dedup Runes — Deduplication hierarchy
• •Companion: rune-orchestration (patterns), context-weaving (Glyph Budget)