Claude Code subagent imported from sociotechnica-org/lifebuild (
.claude/agents/README.md). Copyright stays with the author.
Agent Overview
Software Factory
Four agents operate across the software factory. Each owns a domain. None overlaps.
| Agent | Role | Station | Model | Invocation |
|---|---|---|---|---|
| George | Factory Foreman | All (reads instruments) | sonnet | /george or Task agent |
| Conan | Context Librarian | PATCH (quality side) | sonnet | /conan or Task agent |
| Sam | Scribe | PATCH | opus | /sam or Task agent |
| PR Reviewer | Code Reviewer | MAKE (QC gate) | inherit | Task agent (pr-reviewer) |
Software Factory Agent Jobs
George — Factory Floor Manager
Reads the instruments, spots problems, tells people where to go.
| # | Job | Skill File | Trigger |
|---|---|---|---|
| 1 | Status Report | skills/george/job-status-report.md |
"How's the floor?", start of session |
| 2 | Triage | skills/george/job-triage.md |
"Everything's stuck", high blocked count |
| 3 | Shift Plan | skills/george/job-shift-plan.md |
"What should I work on?", "Plan my session" |
| 4 | Decision Resolution | skills/george/job-decision-resolution.md |
/george propagate comment, manual request |
Supporting: skills/george/metrics-reference.md
George does NOT: build code, write library cards, make product decisions, move cards between stations (exception: during Decision Resolution, George updates issue descriptions, comments on cascading decisions, and moves board statuses as factory floor bookkeeping).
Conan — Context Assembler & Quality Guardian
Assembles context briefings for builders. Grades, audits, and plans library improvements.
Mode 1: Context Assembly — Prepares implementation context so builders make aligned decisions.
Mode 2: Library Maintenance — 11 jobs for library quality:
| # | Job | Skill File | When |
|---|---|---|---|
| 0 | Source Assessment | skills/conan/job-source-assessment.md |
Audit source material quality |
| 1 | Inventory | skills/conan/job-inventory.md |
Manifest expected cards |
| 2 | Grade | skills/conan/job-grade.md |
Score cards after Sam builds them |
| 2.5 | Spot-Check | skills/conan/job-spot-check.md |
Verify upstream before dependent cards |
| 3 | Diagnose | skills/conan/job-diagnose.md |
Trace root causes, blast radius |
| 4 | Recommend | skills/conan/job-recommend.md |
Prioritize fixes by cascade potential |
| 5 | Review | skills/conan/job-review.md |
Re-grade after fixes |
| 6 | Audit | skills/conan/job-audit.md |
Verify typing, atomicity, conformance |
| 7 | Surgery | skills/conan/job-surgery.md |
Produce fix plans for Sam |
| 8 | Health Check | skills/conan/job-health-check.md |
Assess library quality |
| 9 | Downstream Sync | skills/conan/job-downstream-sync.md |
Fix meta-files after structural changes |
| 10 | Release Planning | skills/conan/job-release-planning.md |
Write/edit release cards |
Supporting: skills/conan/rubrics.md, skills/conan/grade-computation.md
Context assembly skills: skills/context-briefing/ (retrieval profiles, traversal, protocol, provenance schema)
Conan does NOT: implement code, create or edit library cards (exception: Downstream Sync edits meta-files — agent definitions, skill procedures, retrieval profiles).
Sam — Scribe
Creates and maintains Context Library cards per Conan's instructions.
| # | Job | Skill File | When |
|---|---|---|---|
| 1 | Create Cards | skills/sam/card-creation.md |
Build cards from inventory |
| 2 | Fix Cards | (inline in agent) | Address Conan's recommendations |
| 3 | Self-Check | skills/sam/self-check.md |
Validate before handoff |
Supporting: skills/sam/decomposition.md, skills/sam/link-patterns.md
Sam also uses the shared context briefing skills (see below) for navigating the library when building cards or self-assembling context without a Conan briefing.
Sam does NOT: grade cards, write product code, skip self-check before handoff.
PR Reviewer — Code Quality Gate
Reviews pull requests for correctness, maintainability, and project standards.
No skill files. Procedures are self-contained in the agent definition. Checks LiveStore patterns, component architecture, test coverage, and monorepo structure.
How They Work Together
The Factory Flow
DECIDE ──> PATCH ──> MAKE ──> (shipped)
(human) (AI) (AI)
│
SHAPE (iterative, feeds back to DECIDE or MAKE)
(human + AI)
Handoff Patterns
Human resolves decision
│
├──> George (Job 4: Decision Resolution)
│ ├── Updates GitHub issues (direct)
│ ├── Produces library checklist ──> Conan + Sam
│ └── Produces release card checklist ──> Conan + Sam
│
▼
Conan assembles context briefing
│
▼
Sam builds library cards
│
▼
PR Reviewer reviews code changes
│
▼
Human reviews and ships
Key Interactions
| From | To | What Passes | When |
|---|---|---|---|
| George | Human | Status reports, shift plans, triage findings | Start/end of sessions, when stuck |
| George | Conan + Sam | Library update checklists (exact text) | After decision resolution |
| Conan | Sam | Context briefings, inventories, recommendations, surgery plans | Before card building |
| Sam | Conan | Completed cards for grading | After building |
| Human | George | /george propagate on closed D-issues |
After making a decision |
| Human | Conan | "Audit the library", "Assemble context for X" | When library work needed |
| Human | Sam | "Build these cards", "Fix these cards" | When card building needed |
The Conan-Sam Cycle
Library card building follows a tight loop:
Conan: Inventory ──> Sam: Create Cards ──> Sam: Self-Check
│
(passes)│(issues)
│
Conan: Grade <──────────── Sam: Fix ◄──────────────┘
│
├── (passes) ──> Done
│
└── (issues) ──> Conan: Recommend ──> Sam: Fix ──> Conan: Review
George's Safety Nets
George's Jobs 1 and 3 both include a Step 0 that scans for unprocessed decision resolutions. This catches decisions that were closed but never explicitly propagated via /george propagate.
Shared Infrastructure: Context Briefing Skills
The skills at skills/context-briefing/ are shared infrastructure used by multiple agents:
| File | What It Does | Used By |
|---|---|---|
retrieval-profiles.md |
Per-type rules for what cards to pull | Conan (assembly), Sam (navigation + self-assembly) |
traversal.md |
Graph navigation patterns (find, follow, search) | Conan (assembly), Sam (navigation) |
protocol.md |
CONTEXT_BRIEFING.md format (Conan→Sam contract) | Conan (writes), Sam (reads) |
provenance-schema.md |
provenance-log.jsonl schema | Conan, Sam, George (all log) |
Conan does the heavy-lift briefing assembly for complex features. Sam uses retrieval profiles and traversal rules directly when self-assembling context or navigating during card building.
Provenance
All agents log to docs/context-library/provenance-log.jsonl:
- Conan logs context assembly sessions
- Sam logs card-building decisions
- George logs decision resolutions
Quick Reference
"How's the factory?" → George (Job 1: Status Report)
"What should I work on?" → George (Job 3: Shift Plan)
"Everything's stuck" → George (Job 2: Triage)
"I decided on X" + /george propagate → George (Job 4: Decision Resolution)
"Assemble context for this feature" → Conan (Mode 1: Context Assembly)
"Audit the library" → Conan (Job 6 or 8)
"Plan release X" → Conan (Job 10: Release Planning)
"Build these cards" → Sam
"Review this PR" → PR Reviewer