Greg’s working
AI workflow
Every model, skill, agent and gate actually wired on this machine, read from ~/.claude rather than from a wishlist. Cross-model review fires at session end, and no two seats on a panel share a family.
Full model roster33 models: tier, size, reasoning, what each is good at, pros and cons
What changed in v30
v30, 2026-09-16: TWO ACCOUNTS, ASTRA ON PLUS, A MEASURED PANEL. Claude Code now has a real profile switcher: claude = jesse Max 20x (own config dir ~/.claude-jesse, all MCPs symlinked), claude-jmf = admin@jmfield.com Max 5x; Orca panes stay jmf. Codex CLI 0.154 removed mcp-server, so mcp__codex__codex is dead: Codex work rides the codex plugin (/codex:rescue) or codex exec, and luna / terra / sol / gpt-6-astra all answer on the Plus plan at $0. Astra moved off the funded key and is now the FINAL cross-arch pass AFTER Fable on the ship gate (Fable on the 20x plan is the abundant one, Astra on Plus base is scarce, so it runs last, once). A 27-call gauntlet (3 diff sizes x 3 plants + decoy) reordered the panel: gemma4 9/9 at 4.5s takes seat 1, terra 2, glm 3, minimax 4 (9/9 but 48s on 13.5KB), mistral 5, qwen3.5 6 (new: Alibaba family, domain verify seat). luna off the panel (two OpenAI seats is an echo), Fable and Opus have cards, a full model roster page went up.
v29: BULK HTML MOVED TO GROQ. Eight lanes benched on one real landing-page spec (four Groq, four Ollama) with eight machine gates: Groq gpt-oss-20b 4.7s and gpt-oss-120b 8.8s, both 8/8, free tier; deepseek-v4.1-flash 12.5s on the flat Ollama plan; Kimi K2.7 passes only with thinking forced off and once returned nothing; glm-5.3-flash hit the token cap. Groq retired every Llama model plus Kimi K2 and qwen3-32b, so groq-mcp was repointed. Because gpt-oss is OpenAI-family, codex_terra is no longer cross-arch to HTML drafts; kimi_k3 on the ship gate is. kimi_k3 left the AUTOMATIC panel earlier (too slow for a 30s slice) and stays the manual cross-arch lane Fable consults; the panel table below reflects that. Cerebras key added, 402 until billing clears.
v28: THE GATED REVIEW CHAIN IS WIRED: the Ollama Pro lanes draft, then kimi_k3 → ollama_glm → ollama_mistral → ollama_minimax → codex_terra review, then Fable signs off. Five families, no per-call spend, same panel at internal, client and prod. The automatic hook runs seats 1 and 2; the rest fire on gated reviews and the ship gate. gpt-6-astra is manual only and never sits on an automatic panel. Moonshot, OpenCode paid, Gemini Pro and DeepSeek-direct are all unwired. Plan tiers differ: Claude Max 20x, Ollama basic Pro, ChatGPT basic Plus.
How one prompt moves through the stack
Input lands on the planner, the lead agent executes and integrates, and everything below is the routing table it draws from.
Reasoning, strategy, architecture, plans
Primary executor · Reviews all outputs · Integrates · Deploys
Research, blog posts, FAQs, comparisons, case studies
Small sibling of the bulk author, benched 4.7s for a full landing page on 2026-09-11. The fast_llama4 tool now serves it: Groq retired Llama 4 Scout, Llama 3.3, Kimi K2 and qwen3-32b in 2026-09.
First call for HTML/CSS from scratch (RULE #0.5): fastest lane, free tier, bursty. Fallback deepseek-v4.1-flash on Ollama. OpenAI family, so NOT cross-arch to Codex or Astra output; a wired free fallback reviewer for Claude or moonshot drafts only, unseated from panels because it cannot co-sit with the Codex seats.
Alibaba family on Groq, so a cross-arch second draft against gpt-oss at the same price. Benched 13.8s, 8/8 on 2026-09-11. fast_code and fast_ask default to gpt-oss-120b.
Draft lane, not a reviewer. Ollama's copy is a stale April build. The metered direct API is deleted and its MCP unregistered.
Bulk HTML fallback behind Groq gpt-oss (12.5s, 8/8 on 2026-09-11). Deep reasoning and agentic work. Use deepseek-v4.1-flash -- the old deepseek-v4-flash:0731 pin was superseded 2026-09; the bare v4.1 name is the current live build (confirmed 2026-09-11, vision/tools/thinking/cloud).
The only Google-family lane on the plan. Re-bench 2026-08-29: 3/3 planted defects, 0 decoys flagged, 0 errors, fastest reviewer we have. Seat 1: one of the two automatic-hook lanes. Gauntlet 2026-09-16: 9/9 plants, 0 false flags, 4.5s median at every size. Terse; unproven on multi-file traps.
Seat 3, gated reviews only; the automatic hook takes seats 1-2 (gemma, terra). Gauntlet 2026-09-16: 9/9 plants, 0 false flags, 10-16s. Agentic coding over long sessions. Verbose, strip its thinking block in loops.
Seat 4, gated reviews. Gauntlet 2026-09-16: 9/9 plants, 0 false flags, but 48s on a 12KB diff, too slow for the per-turn hook. MSA architecture, so a genuine cross-arch second opinion. 1M context and vision: the only Ollama route that takes images or prompts over 10k chars.
Unseated from the automatic panel 2026-09-08: ship-gate lane beside Fable and --lane kimi_k3 only. Reviewer only, never bulk (K2.7-code is the bulk Kimi). Moonshot family, served by Ollama Pro. Needs reasoning_effort: medium and max_tokens: 48000, or it returns empty.
The judge. Reviews every session wrap and every deploy; nothing ships without Fable AND a cross-arch stamp. Multi-hour marathons are its one real edge. Needs [paid-subagent-OK]. Never on a loop or cron.
The session lead: spec, decompose, integrate, decide. Never the bulk author (RULE #0.5). Deep reasoning, architecture and security-adjacent work route here, not to Fable.
Flipped to Codex 2026-09-16: no API key, no per-call cost. Heaviest OpenAI lane, so it is the FINAL cross-arch pass, run once after Fable on a clean diff (Fable rides the 20x plan, Astra the Plus base quota), the ship chain is k3 (optional) -> Fable -> Astra; reach it with --lane gpt6_astra. Never an automatic panel seat. Code, architecture and security. Never facts.
Seated 2026-09-13 as a review-tier lane, never an automatic panel seat: the shared weekly Plus allowance is the budget. Reach it with --lane codex_sol.
Seat 2, the default OpenAI-family reviewer and a genuine cross-architecture check against Claude-authored code (astra and sol are the manual review-tier siblings). Runs through Codex on the ChatGPT Plus sub: no API key, no per-call cost. Fires on the automatic panel (seat 2) and gated reviews; the ship chain itself is k3 (optional) -> Fable -> Astra.
Second OpenAI-family pass, off the panel since 2026-09-16 (terra holds the openai seat; two seats of one family is an echo). Bench 2026-09-07: terra 3/5, luna 3/5, each catching what the other missed, so run luna when terra passes and you want a second GPT look. --lane codex_luna.
Backend / SWE first choice as ollama_devstral. Bulk HTML only with thinking OFF, and third behind Groq gpt-oss and deepseek-v4.1-flash. Moonshot family: never a verifier for kimi_k3 or its own drafts. ollama_kimi is the same weights.
Seat 5 of six on the gated panel. Complex reasoning, large documents, vision. Also serves ollama_code.
Agentic reasoning, content, mid lane (fast)
High-throughput reasoning + long-running agents. Content and research lane; holds no review seat.
Fast/cheap agentic lane, quick tasks, classification, routing, lightweight reasoning.
Open multimodal, Alibaba family (the panel's only one). Reasoning, writing, analysis, vision. Seated 2026-09-16 as gated-panel seat 6 and the domain verify lane (laravel, android, ios, astro, voip, infra), replacing the disabled opencode qwen3.6. Off the automatic hook.
Text and image to video. Lite $0.40/8s is the default, about 125 clips per $50. Cheapest video lane.
Render before judging. DOM + accessibility tree + screenshot, not curl. Default for any UX, design audit, redesign, or rendering task per CLAUDE.md.
Live SERP, keyword volume, backlinks, on-page Lighthouse, business listings, AI visibility (LLM mentions). 80+ tools, the canonical SEO data source.
Google Stitch, UI generation from prompts. Design systems, screen variants, project scaffolding. Convert output to semantic CSS before shipping.
fal.ai image generation + edit endpoints, fallback when cc-nano-banana (Gemini) is unavailable or for specific fal-only models.
The 15 user-scope servers in ~/.claude.json: chrome-devtools · claude-design · dataforseo · fal-image · figma · gemini-research · google-ads · groq-fast · magic · mobbin · ollama-pro · opencode · refero · stitch-design · ubersuggest. Project-scoped adds onepagelove; claude.ai connectors (Gmail, Drive, Calendar, Dropbox) and plugin servers (playwright, context7) are separate again. Codex is no longer an MCP server: Codex 0.154 removed mcp-server, so it is reached through the codex plugin and codex exec. Disabled and deliberately not listed above: mirage, xiaomi-mimo, glm-free, mistral-code. Unregistered 2026-09-07: kimi-direct (Moonshot API retired) and deepseek (metered direct API; Ollama Pro serves the same model as ollama_deepseek_pro). Both keys vaulted.
Screenshot, brief, or reference file
Claude Code skills invoked by Opus
Specialized agents spawned per task
Fast code gen + design research
Code review then ship
5-part contract before any worker spawns. No spec, no spawn.
Kotlin + Compose screens. The lead never bulk-authors.
Never review a DeepSeek draft with DeepSeek. Same family reads as agreement.
Invariant tests, then hardware. A subagent cannot see a screen.
Signer compared to the last shipped APK. Manifest and binary propagate independently.
Enforced by fable-ship-gate.py. A deploy or push without a stamp is refused, not warned.
Run after the CDN purge, never before: an audit against a stale edge measures the previous build.
--selftest proves all 27 mutations can fail
A client POC, SEO audit never ran. Schedule seo-audit + seo-geo + seo-local once content is final.
Every deployed page runs: taste → humanizer → frontend-design → impeccable → SEO suite. SEO skills are not optional, not batch-only, per-page on every ship.
blog-factcheck verifies every claim against cited sources via WebFetch. BLOCKS ship on NOT-FOUND, unverified entity/date/quote, rubric < 90, or P0 fabricated stat. Pair with blog-factcheck-fix, Ollama Pro + subagents rewrite, Claude orchestrates (10-25× cheaper than Claude rewriting).
For Alpine-heavy templates (phone.html, contacts/detail.html), open as affected user with DevTools open BEFORE commit. Server-side smoke tests miss silent x-text binding failures.
Kimi K3 is a ship-gate / manual review lane (off the automatic panel since 2026-09-08; gemma4 holds seat 1 since 2026-09-16), not DeepSeek. Parity rule: Opus / Sonnet / DeepSeek / GLM / Gemini-Pro tie on correctness, spend paid verify only for cross-arch echo-breaking + code craft, never correctness.
Render before judging. DOM + a11y + screenshot, not curl. Default for any UX, render, audit, or redesign task.
Remove AI-writing tells. Detects/fixes inflated symbolism, em-dash overuse, rule-of-three, AI vocab. Run on every Gemini/GPT-drafted copy before publish.
Distinct production-grade UI. impeccable flags AI-slop tells (side-stripe borders, gradient text, glassmorphism, hero-metric clichés). Pair for design audit + polish.
Senior UI/UX engineer. Editorial typography, gapless bento grids, strict GSAP scroll triggers, massive section spacing. Default for any web dev (per CLAUDE.md taste-skill rule).
seo-audit · seo-technical · seo-content · seo-geo · seo-google · seo-local · seo-maps · seo-schema · seo-sitemap · seo-cluster · seo-drift · seo-firecrawl · seo-rotation · seo-sxo · seo-image-gen · seo-programmatic · seo-competitor-pages · seo-ecommerce · seo-plan · seo-dataforseo · seo-page · seo-backlinks
Required for ALL image generation. Nano Banana (Gemini CLI) for blog images, thumbnails, icons, diagrams, illustrations, photos. Free local fallback (2026-06-07): sd.cpp (Metal + SDXL, local-image.sh) for bulk/uncensored stills, keeps fal=video, recraft=text-in-image.
Query + manage Google NotebookLM via CLI. Vault-to-Master sync after every session.
Mandatory after any frontend edit, screenshot via Playwright, Read the PNG before calling work done. Per CLAUDE.md rule #4.
Run after ANY code edit, deployment, or fix. Requires running verification commands and confirming output before claiming completion.
Write handoff doc so fresh agent continues in clean context window. Use for scope-creep split-outs, ~120k dumb-zone compression, planner→prototype→planner round-trip, or cross-CLI pass (Codex, Copilot, OpenCode). Triggers: "handoff", "spawn agent for", "fresh session".
ponytail least-code ladder (YAGNI → stdlib → native → one line) cuts the code; caveman cuts the prose. Run both. A 10-run bench in June 2026 measured ponytail at −28% LOC on realistic builds, same correctness, faster. Levels: light/full/ultra. Installed 2026-06-19.
71 DESIGN.md files from Stripe, Figma, Apple, PlayStation, WIRED, VoltAgent/awesome-design-md
Codex CLI second opinion on ChatGPT Plus (gpt-5.6-terra default; gpt-5.4-mini is not callable on this plan), invoke when stuck >2 attempts or need architecture review
Weekly audit: broken symlinks, GitHub freshness, usage stats, duplicate detection
Session lifecycle, /start reads Memory+Obsidian+NotebookLM, /wrap commits+pushes+updates Obsidian
142 active skills, 151 stashed in skills-disabled/ to keep the session baseline small. Active set: humanizer, frontend-design, impeccable, taste-skill, full 24-skill SEO suite, Astro suite, cc-nano-banana, notebooklm, visual-review, verify-my-work. Plus plugin marketplaces (caveman, ponytail, obra-superpowers, interface-design, codex, memsearch).
ByteDance text-to-video via fal.ai, Pro (1080p) + Lite (720p). Live now.
251 audit rules across 20 categories, @seomator engine. CLI: seo-audit skill + openseo
47 AI citability criteria, llms.txt, AI bot rules, passage-level optimization for ChatGPT/Perplexity/Gemini
Semantic vector search across 1,238 curated memory files. ONNX bge-m3 embeddings (local, free). Auto-captures sessions via hooks. Backed by MEMORY.md multi-index (root + 5 sub-indexes: active / infra / apis-skills / feedback / hot).
memory-audit.py (orphans / empties / oversized) + reorganize-memory.sh (split >50-line files, fix index) + sync-memsearch-ifchanged.sh (AUTO re-index, do NOT manual sync-memsearch.sh) + weekly agent-os hygiene (Layer 5, report-only, Sat)
4 behavioral principles from Andrej Karpathy: Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution. + RULE #0.7 (Boris Cherny, 2026-06-07): Loops not Prompts · Pre-compute > Inference · Tokens over Headcount.
Semantic across vault + master notebook. Master notebook. CLI authenticated (Mac + Fedora). Vault → project notebook → Master after every session.
the Obsidian vault, Projects, Servers, Sessions, NotebookLM. Per-project file pattern. Check FIRST before asking.
Interactive reporting dashboard for CrawlHound, grade history, scan trends, site-wide metrics visualization
Downloadable PDF audit reports for CrawlHound, MrBotsworth, and gjapp, branded, shareable
Automated build + test + deploy pipeline on git.myseodesk.com, lint, SEO checks, rsync deploy on push
Subagent-heavy sessions. Top spawns: backend-developer, ollama_kimi, code-reviewer, all have sub-tier MCP twins.
Usage at >150k context. Long mixed-topic sessions, no /clear or /compact between unrelated tasks.
Skill invocation share. The 100+ installed skills (visual-review, redesign, audits) barely fire, same work runs through paid subagents instead.
Fires on every Agent tool call. Drift list (6 named agents): warns ≤65% context used, blocks >65%. code-reviewer + general-purpose in ALWAYS_BLOCK. Whitelist (Explore/Plan/gsd/audit/seo/feature-dev) silent. Override: [paid-subagent-OK] token logs + allows.
Three triggers: session start, topic-switch keywords, post-/compact detection (transcript msg-count drop >50%). Injects ~250 tokens of routing rules from ~/.claude/refs/delegation-cheatsheet.md.
Detects topic-switch keywords (“okay new task”, “switching to”, “moving on”) and appends a /clear reminder. 10-message debounce so it doesn’t fire on every consecutive switch.
Highest-leverage hook. 46 keyword triggers in ~/.claude/rules/skill-triggers.json (regex, priority-ranked). On match, injects an EXTREMELY_IMPORTANT system-reminder naming the skill, same mechanism as using-superpowers.
Enforces the “render before judging” rule. When a prompt looks like UX / design / audit / redesign work, injects a reminder to use mcp__chrome-devtools__new_page + take_snapshot + take_screenshot from the main session, not curl, not WebFetch, not subagent.
| drift agent | sub-tier mcp route | when in-agent is justified |
|---|---|---|
| backend-developer | ollama_devstral, ollama_deepseek_pro | 5+ files coordinated, or hard arch |
| ollama_kimi | fast_gpt_oss (HTML), ollama_code (review) | multi-framework full-stack work |
| fastapi-developer | ollama_deepseek_pro, ollama_devstral | complex async patterns + tests |
| code-reviewer | HARD-BLOCKED. Chain: kimi_k3 (DEFAULT) → ollama_glm → ollama_mistral → ollama_minimax → codex_terra | [paid-subagent-OK] override only |
| security-auditor | api-keys skill; security-auditor is a DISABLED agent | whole-system audit (5+ services) |
| database-administrator | sql-pro is a DISABLED agent. Use the mysql/postgres MCP directly | HA/replication infra, not just queries |
Apply when work is non-trivial: >50 LOC change, security-sensitive, or first-time deploy. Six distinct families, none metered. Gated panel order (from model-seats.yml): gemma4 (google), terra (openai), glm-5.3-flash (zhipu), minimax-m3 (msa), mistral-large-3 (mistral), qwen3.5 (alibaba). The automatic Stop hook runs seats 1-2 only; gemma finds every planted defect in 4.5s on a 13.5KB diff (gauntlet 2026-09-16), which is why it holds the per-turn seat. kimi_k3 is off the panel since c31c794: it is the optional first lane of the ship chain. The ship chain is kimi_k3 (optional) → Fable → gpt-6-astra FINAL.
The 40 that live in ~/.claude/agents. Anything under agents-disabled/ is left out, because it does not run.
| Plan Type | Lead Model | ~Tokens | ~Cost | Notes |
|---|---|---|---|---|
| Quick Ask | Ollama Pro (Nemotron/Chat) | 1-3K | $0 (Pro) | ollama_nemotron / ollama_chat, Tier 1 |
| Bug Fix | Sonnet 5 | 5-15K | ~$0.05-0.15 | Read file → diagnose → patch → verify |
| Single Feature | Opus 5 + Sonnet | 20-60K | ~$0.17-0.70 | Plan → subagent execute → /codex:rescue |
| Static Page Build | Opus + Groq gpt-oss | 30-80K | ~$0.17-0.50 | DESIGN.md → /stitch → Playwright check → deploy |
| Astro SSR Feature | Opus + ollama_devstral / ollama_code | 50-120K | ~$0.35-1 | /astro-ssr skill + full pre-deploy gate |
| Laravel API | Opus + ollama_devstral, verify codex_terra | 60-150K | ~$0.50-1.30 | Models + migrations + tests + PHPStan + review |
| Full Site (5-10 pages) | Opus + mix | 150-400K | ~$1.30-4 | Full 6-stage pipeline: PLAN→DESIGN→BUILD→SEO→CRO→QA |
| SEO Audit | Gemini 3.8 Flash | 10-30K | ~$0.10-0.30 | Schema + compliance + content + schema, mostly Gemini |
- → Use Ollama Pro (Tier 1) for drafts, analysis, code review
- → Use Groq gpt-oss-120b / 20b (free, 5-9s) for bulk HTML; deepseek-v4.1-flash on Ollama as the batch fallback
- → Use Gemini 3.8 Flash for research/content
- → NVIDIA NIM is archived. Groq is no longer a fallback: gpt-oss is the bulk HTML seat; its Llama, Kimi K2 and qwen3-32b lanes were retired by Groq in 2026-09
- → /compact regularly in long sessions
- → Each file read adds ~1-5K tokens
- → Large HTML files: 10-30K tokens each
- → Long conversations drift, use /clear or new session
- → Subagents get fresh context, use them for big files
- → CLAUDE.md loaded every session (~5K tokens)
- → Stuck after 2 attempts → /codex:rescue
- → Complex architecture → Opus (not Sonnet)
- → Critical prod deploy → the automatic panel ollama_minimax → codex_terra → ollama_glm → ollama_mistral, then the ship gate: kimi_k3 (cross-arch, manual lane) and Fable sign off. No metered lane.
- → Fresh frontend design take → Groq gpt-oss or ollama_glm; ollama_kimi only with thinking off. gpt-5.5 is DISABLED.
- → No-per-call cross-arch verify → fast_gpt_oss (free tier) or Codex CLI on Plus
- → Sub-tier model fails tool calls → switch to Sonnet
| Role | Model | MCP Server | Use Case | Cost |
|---|---|---|---|---|
| Planner | Opus 5 | native | Reasoning, architecture, strategy | paid |
| Lead | Sonnet 5 | native | Executes tasks, reviews, deploys | paid |
| Research | Gemini 3.8 Flash | gemini-research | Research, comparisons, blog posts, FAQs | free |
| Fast 20B | GPT OSS 20B replaced Llama 3.3 70B, retired by Groq | fast_ask model=openai/gpt-oss-20b | Fastest lane on the stack: 4.7s for a full landing page, 8 of 8 gates | FREE |
| Qwen | Qwen 3.6 27B replaced Llama 4 Scout, retired by Groq | fast_qwen / fast_kimi (both repointed) | 131k context, 13.8s on the same landing spec, 8 of 8. qwen3.8-27b needs the org toggle | FREE |
| Cross-arch | GPT OSS 120B | fast_gpt_oss | Bulk HTML author since v29 (see HTML row). Off every panel: as OpenAI-family it cannot cross-check its own drafts or terra's | FREE |
| HTML | GPT-OSS 120B / 20B bulk seat since v29 | fast_gpt_oss (Groq) | Bulk HTML. Benched 09-11 on a real landing spec: 20B in 4.7s, 120B in 8.8s, both 8 of 8 gates, free tier. Fallback deepseek-v4.1-flash (Ollama). Kimi K2.7 only with thinking off. | FREE |
| Content | Nemotron 3 Super 120B | ollama_nemotron | Agentic reasoning, content | PRO |
| Ollama Pro | Mistral Large 3 675B replaced qwen3-coder 480B | ollama_code | Review panel seat 5, cross-arch vs the google, openai, zhipu and msa seats above | PRO |
| Ollama Pro | Qwen3.5 397B | ollama_chat | Reasoning, writing, analysis, vision, thinking | PRO |
| Ollama Pro | Mistral Large 3 675B | ollama_mistral | Complex reasoning, large documents, vision | PRO |
| Ollama Pro | DeepSeek V4.1 Flash 158B | ollama_deepseek | Deep reasoning, thinking, agentic tasks | PRO |
| Ollama Pro | Kimi K2.7 Code replaced devstral-2 123B | ollama_devstral | SWE coding, multi-file editing (surviving specialist) | PRO |
| Ollama Pro | GLM 5.3 Flash | ollama_glm | Agentic coding, sustained over 100s of rounds | PRO |
| Stakes | Panel (after egress policy) |
|---|---|
| INTERNAL | ollama_gemma → codex_terra → ollama_glm → ollama_minimax → ollama_mistral → ollama_qwen |
| CLIENT | ollama_gemma → codex_terra → ollama_glm → ollama_minimax → ollama_mistral → ollama_qwen |
| PROD | ollama_gemma → codex_terra → ollama_glm → ollama_minimax → ollama_mistral → ollama_qwen |
| Key | Family | Model | Egress |
|---|---|---|---|
codex_luna | openai | gpt-5.6-luna | ok |
codex_sol | openai | gpt-5.6-sol | ok |
codex_terra | openai | gpt-5.6-terra | ok |
fable_5 | anthropic | claude-fable-5-1 | ok |
gpt6_astra | openai | gpt-6-astra | ok |
gpt_oss | openai | openai/gpt-oss-120b | ok |
kimi_k3 | moonshot | kimi-k3 | ok |
ollama_deepseek_pro | deepseek | deepseek-v4-pro | ok |
ollama_devstral | moonshot | kimi-k2.7-code | ok |
ollama_gemma | gemma4:31b | ok | |
ollama_glm | zhipu | glm-5.3-flash | ok |
ollama_kimi | moonshot | kimi-k2.7-code | ok |
ollama_minimax | msa | minimax-m3:cloud | ok |
ollama_mistral | mistral | mistral-large-3:675b | ok |
ollama_qwen | alibaba | qwen3.5:397b | ok |
opus_5 | anthropic | claude-opus-5 | ok |
qwen36 | alibaba | opencode/qwen3.6-plus | ok |
+ 52 hook commands across 15 events (incl. agent-spawn-gate, stop-cross-review, fable-ship-gate, delegation-table, skill-trigger) · 1,238 memory files in MemSearch · NotebookLM master notebook · Obsidian vault per-project files
Rules that live in code, not in prose, because prose loses. 52 hook commands across 15 events. These are the ones that BLOCK rather than remind:
Anchor a build to screens humans actually designed, instead of to model defaults. The house rule is that a gate measures the ABSENCE of known slop; it cannot supply a point of view, so the reference comes first.
Where the work actually went: the client-facing build and quality lane.