{"slug": "architect-reduce-fable-tokens-by-80-fable-orchestrates-reviews-codex-builds", "title": "/architect: Reduce Fable tokens by 80%, Fable orchestrates/reviews, Codex builds", "summary": "Developer Dan McInerney released Architect Loop, an open-source tool that uses Claude Code as an architect and GPT-5.5 Codex as a builder to run cross-vendor AI development loops on flat-rate subscriptions. The system splits work into isolated lanes with acceptance gates, parallel builders, and judgment sessions, aiming to improve software development throughput without API token costs.", "body_md": "**Claude Fable is the architect — it designs every slice, freezes the\nacceptance gates, and judges the results. GPT-5.5 Codex is the builder and\nresearcher — it does all the engineering and all the web research, in\nparallel, unattended, for hours.** Two Claude Code skills that run this\ncross-vendor loop on the flat-rate subscriptions you already have — no API\nkeys, no token bills.\n\n```\ngit clone https://github.com/DanMcInerney/architect-loop\ncd architect-loop && ./install.sh        # Windows: .\\install.ps1\nnpm i -g @openai/codex@latest            # the builder (Codex CLI >= 0.133)\n```\n\n`./install.sh --project`\n\ninstalls to the current repo only instead of\nglobally. You need [Claude Code](https://claude.com/claude-code) on any paid\nplan and the Codex CLI signed into a ChatGPT plan.\n\n```\n/architect                                      # the build loop\n/architect-research <what you're considering>   # the research loop\n```\n\n`/architect`\n\nruns one work block: judge the last run, spec the next slice,\ndispatch builders. `/architect-research`\n\nis for when you're still deciding\n*what* to build — its cited report feeds the build loop's PRD.\n\nOne short Fable session per work block — judgment only, it never writes code:\n\n**Spec + gates first.** Fable specs a one-PR slice, splits it into 1–4 lanes with provably disjoint file sets, and commits the acceptance gates to`docs/gates/`\n\n*before*any builder starts. Gates are read-only; a builder edit to a gate file fails the slice automatically.**Parallel isolated builders.** One fresh`codex exec`\n\n(xhigh) per lane, each in its own git worktree. Builders must argue with the spec before building (silent compliance = defect), build only their declared files, and report raw results — they physically can't commit (the sandbox protects`.git`\n\n).**Fable judges and integrates.** It runs the gate commands itself (builder claims are hearsay), reads the diff against the spec's intent (passing tests ≠ mergeable work), then commits and merges passing lanes. Judgment happens in a fresh session — cross-context review measurably beats same-session review.**The repo is the only memory.**`docs/HANDOFF.md`\n\n(a short table of contents, pruned every session),`docs/gates/`\n\n,`docs/lanes/`\n\n, git history. Not in the repo = didn't happen.**Supervision built in.** Liveness checks on dispatched runs, stall triage (diagnose the child process tree, kill the narrowest thing), explicit timeouts on every long command.\n\nScout-first, like the production deep-research systems — no fixed lane taxonomy:\n\n**A cheap Codex scout maps the topic**(~10 searches): canonical terminology, the load-bearing systems and papers, the named people, the topic's natural fault lines. Skipped for comparisons and fact-finds.**Fable designs 3–6 topic-specific lanes** from the scout's map, drawing per-source-class tactics from a library (academic citation snowballing, dependents-not-stars repo evidence, emerging-vs-hype gating, production pattern mining, expert tracking) — checked for overlap and gaps before dispatch.**Parallel Codex researchers** run under hard budgets: search caps, ≤5 subjects per lane, saturation stop, strict findings discipline (URL + date- quote + confidence tag; NOT FOUND beats inference; no recommendations). Expert opinion runs as a second wave, roster-seeded by the first.\n\n**Fable verifies and writes.**≥2 independent sources per load-bearing claim, adversarial falsification searches, citations only from URLs actually fetched — then one author writes one decision-oriented report. Gathering parallelizes; synthesis never does.\n\nEach piece is there because evidence put it there (full citations in\n[DESIGN.md](/DanMcInerney/architect-loop/blob/main/DESIGN.md)):\n\n- Weak planners hurt more than weak executors — so the strongest model does the design, and builders get exhaustive specs.\n- Manager + worktree-isolated workers is the measured-best topology for shared-artifact software work; naive shared-file coordination collapses throughput.\n- Frozen external gates beat trusting the agent — but agents game visible tests and their passing PRs are frequently unmergeable, so the architect also reads the diff.\n- Memory files rot — so the handoff stays a short map, and detail lives in linked gate/lane files.\n- Every production deep-research system uses planner-designed decomposition, none uses fixed lanes — so research lanes are designed per topic, after a scout pass.\n\n| File | What it is |\n|---|---|\n|\n\n[skills/architect/SKILL.md](/DanMcInerney/architect-loop/blob/main/skills/architect/SKILL.md)[skills/architect/dispatch.md](/DanMcInerney/architect-loop/blob/main/skills/architect/dispatch.md)`codex exec`\n\ncommands, builder block, worktree fan-out, stall triage[skills/architect/research.md](/DanMcInerney/architect-loop/blob/main/skills/architect/research.md)[skills/architect/HANDOFF.template.md](/DanMcInerney/architect-loop/blob/main/skills/architect/HANDOFF.template.md)[skills/architect-research/SKILL.md](/DanMcInerney/architect-loop/blob/main/skills/architect-research/SKILL.md)[skills/architect-research/lanes.md](/DanMcInerney/architect-loop/blob/main/skills/architect-research/lanes.md)[tests/validate_skills.py](/DanMcInerney/architect-loop/blob/main/tests/validate_skills.py)**Do I need API keys?** No. Claude Code runs on your Claude plan; Codex CLI\non your ChatGPT plan.\n\n**What does a run cost?** Builder/researcher runs draw on your ChatGPT\nplan's 5-hour and weekly quotas; a multi-hour run is a meaningful fraction\nof a weekly window. Fable's architect sessions are minutes, not hours.\n\n**What if a builder wrecks things?** Nothing reaches a branch until the\narchitect's tamper, boundary, and gate checks pass — worktrees are\ndiscarded and re-dispatched from the freeze commit.\n\n**Can I watch a run?** Yes — every dispatch prints the builder block, so you\ncan paste it into an interactive `codex`\n\nsession with `/goal`\n\ninstead.\n\n**Why two skills?** Research-grade fan-out costs ~15× chat-level tokens — it\nshould be a deliberate act, not a side-effect of the build loop.\n\nMIT", "url": "https://wpnews.pro/news/architect-reduce-fable-tokens-by-80-fable-orchestrates-reviews-codex-builds", "canonical_source": "https://github.com/DanMcInerney/architect-loop", "published_at": "2026-06-12 20:33:22+00:00", "updated_at": "2026-06-12 23:12:45.641424+00:00", "lang": "en", "topics": ["ai-tools", "ai-agents", "ai-research", "large-language-models", "generative-ai"], "entities": ["Dan McInerney", "Claude Code", "GPT-5.5 Codex", "OpenAI", "Anthropic", "Architect Loop", "Codex CLI"], "alternates": {"html": "https://wpnews.pro/news/architect-reduce-fable-tokens-by-80-fable-orchestrates-reviews-codex-builds", "markdown": "https://wpnews.pro/news/architect-reduce-fable-tokens-by-80-fable-orchestrates-reviews-codex-builds.md", "text": "https://wpnews.pro/news/architect-reduce-fable-tokens-by-80-fable-orchestrates-reviews-codex-builds.txt", "jsonld": "https://wpnews.pro/news/architect-reduce-fable-tokens-by-80-fable-orchestrates-reviews-codex-builds.jsonld"}}