Dark Factory KB — 16 Expansion Gaps

The factory adopted Beads for task tracking using formula TOML files as the pipeline source of truth. The "Option B" decision chose formula-driven molecules over ad-hoc bead creation. kelly-web-build.formula.toml and kelly-research.formula.toml define all pipeline steps with dependencies, agent assignment, and workflow paths. The KB has Yegge's Beads concept and a gap analysis, but no article documenting how we actually use beads — how formula TOML works, how spawn-subphase.sh resolves beads, how the bddone wrapper works, or the dual-write requirement (DONE marker + bead closure).

The --auto-spawn flag on spawn-subphase.sh appends structured metadata (NEXT_STEP, CLOSING_BEAD, PIPELINE_DONE) after sessions_spawn output. On wake-up after sessions_yield, the Router reads this metadata and auto-chains the next step without operator intervention. The Kelly handbook Ch7 describes basic spawn/steer/kill, but the auto-spawn chain pattern is a significant evolution — it's only documented in AGENTS.md, not in the KB.

Several refinements beyond the handbook's basic RALPH protocol: (1) QA agents must test CRUD, not just READ — discovered when EDIT/DELETE were never tested; (2) Lessons must go into agent skill files, not just SELF_IMPROVEMENT.md — the Router's notes don't propagate to sub-agents automatically; (3) Parallel pipeline runs validated that RALPH works across concurrent pipelines. The existing ralph-protocol.md concept only describes the handbook version.

Two patterns for operator control: (1) Operator Queue (PIP-58) — when the Router hits a blocker, it logs to operator.md via operator-queue.sh add instead of just saying it in chat; (2) Operator Hold (PIP-17) — hold after {subphase-id} pauses pipeline at a specific point, continue {project} resumes. Neither is documented in the KB. Essential for safe automation — the operator needs control levers.

Steve Yegge's Gas Town is partnering with Kilo (a model-agnostic cloud platform) to make Gas Town available as a cloud service — not just local. 70+ engineers built factories at the first Gas City Hall workshop. This is a major shift from local-only dark factories to cloud-hosted ones. The existing Gas Town article in the KB predates this development.

Karpathy's LLM Wiki gist has been extended with OmegaWiki — a full research lifecycle platform. New operations: /discover (venue-ranked reading lists), /ingest (typed entries linked into graph), /ideate + /novelty (ideas with bibliographic grounding), /exp-design → /exp-run → /exp-eval (experiment workflow), /paper-plan → /paper-draft, /poster (conference poster generation), /rebuttal (reviewer response). The original LLM Wiki pattern is in the KB; the OmegaWiki extension is not.

Extends Karpathy's pattern with production lessons from building agentmemory (10K GitHub stars). Key additions: (1) Confidence scoring — every fact carries a score based on source count, recency, and contradictions; (2) Supersession — new information explicitly marks old claims as stale with version control; (3) Forge testing — synthetic queries to validate wiki accuracy; (4) What breaks at scale — practical failure modes. Directly applicable to dark-factory-kb maintenance.

Simon Willison reports that Anthropic and OpenAI have switched enterprise plans from seat-based to API pricing. His personal usage: $1,200/month Anthropic + $980/month OpenAI API equivalent for a $200 subscription. Companies are shocked at LLM bills. Anthropic rumored first profitable quarter. April 2026 is an inflection point for enterprise AI spending. Validates dark factory economics — minimizing token usage via compiled KBs, bounded context, and story-by-story spawns is economically essential.

Ch7 describes "named leads" (research-lead, project-lead, test-lead) which the factory has moved past. The current architecture — formula TOML + auto-spawn chains + bead tracking + operator queue — isn't captured as a standalone KB article. Not about Ch7 being "stale" (it accurately describes what it describes), but about the current pattern needing its own article. The gap is that no single article documents how multi-agent orchestration actually works today.

To avoid context overflow in long BUILD phases, the factory spawns one story at a time rather than handing the agent the entire sprint backlog. Each story gets its own subagent session with bounded context. The RALPH concept mentions "break tasks into smaller chunks" as a failure mode fix, but the story-by-story spawn as a default operating pattern is not documented as a standalone article. Sally loops on multi-storyboard spawns; Amelia does 24+ stories in separate sessions.

Agents must write both a DONE marker file AND call bd update --status in_review. Neither alone is sufficient — Beads tracks work state; DONE markers track artifact output. BUG-01 (DONE marker skip) was caused by agents completing work but forgetting the DONE file, which broke prerequisite checks on subsequent spawns. The dual-write requirement is in AGENTS.md but not in the KB.

kelly-web-build.formula.toml had output_path = "artifacts/planning/2.1-prd/" but actual DONE files go to planning/2.1-prd/DONE (without the artifacts/ prefix). All 30 workflow YAML frontmatters had the same bug in done_dir and prerequisites. Fixed in commit 31f00f4. Critical lesson for anyone authoring formulas — path conventions must match actual filesystem layout.

Subagent spawns were failing with model not allowed: fireworks/accounts/fireworks/models/minimax-m2p7 because the fireworks provider wasn't in the agents.defaults.models allowlist. Root cause: main session could use any model, but subagents were restricted. Workaround: use model=opencode-go/minimax-m2.7 directly instead of agentId-based spawns. Important troubleshooting pattern for anyone configuring subagent models.

Simon Willison's agentic engineering patterns guide has been significantly expanded since the KB ingest (Feb 23, 2026). New sections: "Using Git with coding agents," "Agentic manual testing," "AI should help us produce better code" (personal manifesto), and updates on anti-patterns. The guide now has 20+ patterns vs the ~15 at ingest time. The existing KB article is stale.

Deep dive on how memory works in Claude Code, Codex CLI, and OpenCode — where it breaks (context limit warnings, forgetting decisions from 3 hours ago), and how Karpathy's LLM Wiki + agentmemory address the problem. Relevant to Kelly's memory architecture but supplementary — overlaps with existing memory concept pages in the KB.

Carson scored the dark-factory-kb at 19/100 against Karpathy's LLM Wiki pattern. Key gaps: no Q&A loop (KB is write-only), no lint gates, no filed-back query outputs, no auto-generated index. The gap analysis report exists in the KB but the meta-lesson about KB maintenance patterns is not compiled as a concept. Low priority because the gap analysis already covers this.