The Fable Rush — GitHub half (repos, competitive landscape, integration targets)

Research sweep 2026-07-06 via gh api / gh search (account ss251). Companion to docs/RESEARCH_fable_rush.md (Twitter/Reddit). Cutoff context: Fable 5 leaves subscriptions 2026-07-07 11:59:59pm PT, then $10/$50 per MTok usage credits.

STATUS: sections marked [PARTIAL] were banked mid-sweep; coverage notes at the bottom.

1. THE RUSH ARTIFACTS — repos born around the cutoff

1a. "Train the replacement" — the wave turned into repos (it's big and it's THIS WEEK)

The canonical one — tomicz/fable-5-train-opus-skills-after-it-retires — 242 stars, created 2026-07-02, still updating Jul 6. Literally the PrajwalTomar pattern as a single README-prompt repo. Description verbatim: "Thsis is a prompt that will create skills by Fable 5 before it retires and your cheaper models can use those skills." (typo included — it was shipped in a hurry). The README is a full 3-phase "retiring distinguished fellow" prompt worth studying because it defines what the community expects a Fable estate to LOOK like:

• Phase 1 "Discover before you write" — repo audit like an incoming principal engineer, then ≤5 questions.
• Phase 2 — 10–16 skills from a fixed taxonomy: <project>-change-control, -debugging-playbook, -failure-archaeology, -architecture-contract, <domain>-reference, -config-and-flags, -build-and-env, -run-and-operate, -diagnostics-and-tooling, -validation-and-qa, -docs-and-writing, -external-positioning, plus ADVANCED: -<hardest-problem>-campaign, -proof-and-analysis-toolkit, -research-frontier, -research-methodology.
• Format rule verbatim: ".claude/skills/<name>/SKILL.md, YAML frontmatter with name and a trigger-rich description (exactly when a model should load it)." Plus: date-stamp volatile facts, end each skill with a "Provenance and maintenance" section with re-verification commands.
• Phase 3 — three parallel reviewers (FACTUAL / DOCTRINE / USABILITY) + one fixer. → github.com/tomicz/fable-5-train-opus-skills-after-it-retires

The fable-skills name is now a genre. At least 10 repos named fable-skill(s) created Jun 10–Jul 6:

• DizzyMii/fable-skills (16★, Jun 10): "Six Claude Code skills that harden Opus 4.8 toward frontier behavior — written by Fable 5, pressure-tested on the target model with transcripts included."
• oliwoodman/fable-skills (9★, Jul 5): "Five skills Fable wrote before it left: the security sweep, the setup, the build planner, the honest advisor, the bug hunter… you can't keep the model; you can keep how it does the job."
• benjaminard/fable-skills (7★, Jul 2): "teaches Claude Opus the working disciplines of Claude Fable 5. Move the checkpoints, not the capacity."
• Ishan7390/fable-skill (8★, Jul 4), wjw12/fable-skills (Jul 6), xternalAI/fable-skills (Jul 6, "Skills created by Fable 5 to support Opus in trying to be more Fable-like"), BenMalaga/fable-skills (Jul 3, "each one encodes a real production failure and the discipline that beat it"), imfaisii/fable-skills (Jul 2, ships as a Claude Code plugin marketplace: "fable-mind"), Box-of-Raccoons/fable-skills (Jul 4), masatokawano/fable-skill (Jul 6), k-daisuke-works/fableSkills (Jul 6), KshitijBharambe/fable-skills (Jul 4), CodeHalwell/fable-skills (Jul 4), almutaz9000/fable-skill (Jul 5, "One npx command", targets Cursor/Copilot/Codex/Gemini too). Three of these were created ON Jul 6 — the wave is accelerating into the deadline.

Fable-emulation playbooks (make Opus act like Fable):

• apoorvjain25/frontier — 45★, Jul 3: "21 craft standards written and audited by the frontier model itself; single-response lift, or a full convergence loop with a taste gate."
• rennf93/opus-fable-playbook — 18★, Jul 2: "doctrine output style, drift-catching hooks, and an eval loop against golden Fable transcripts. Claude Code plugin." (Evaluating replacement models against archived Fable transcripts = our transcript corpus is the eval set someone else is building tooling for.)
• dilitS/op-fable — 21★, Jun 16: "Fable 5 inspired cognitive workflow architecture… for Opus."
• duolahypercho/fusion-fable — 438★, Jun 13: "Fuse two frontier models into one Fable-tier answer: Opus 4.8 drafts, a second model checks, Opus fuses. A Claude Code skill."
• New1Direction/fable-5-playbook — 5★, Jun 24: "Drop-in CLAUDE.md that makes Opus plan, batch tools, read before editing, and verify changes, built from real coding-agent logs."
• UnpaidAttention/fable5-methodology — 2★, created Jul 6 (deadline day): "transferable, self-enforcing software-engineering methodology… playbook, skills, contracted subagents, lifecycle hooks, and evals."

Quota-conservation / orchestration-discipline tools (the "strategic burn" behavior as code):

• nagarjuna-msr/fable-god-mode — 10★, Jul 2: "10-80-10 token discipline for Fable 5, plus GPT-5.5 (Codex) as an independent reviewer."
• openwong2kim/iffable — 3★, Jul 3: "Fable-only orchestration guard for Claude Code: injects a model-tier policy at session start and blocks large un-planned delegations until a Requirements Ledger exists. Dormant on Opus/Sonnet. Preserves scarce Fable quota."
• ericshang98/cheap-fable5 — 1★, Jun 11: "route subtasks to the cheapest model that can do them (haiku/sonnet/fable) — stop burning Opus/Fable quota on grep."
• sup3x/claude-code-eco — 15★, Jul 2: "/eco: -31% to -73% output tokens… Measured hardest on Claude Fable 5." Publishes 82 raw benchmark runs.
• samellgit/claude-fable-orchestrator-workflow — 2★, Jul 2: "orchestrator workflow I use to maximize Fable usage."
• casualsav/fable-bench (2★, Jul 3) + casualsav/hermes-oracle (2★, Jul 3): "/consult skill — a Claude Fable 5 plan-consult plus an automatic warm diff-review… Opus drives; Fable consults" (the post-cutoff pattern: Fable as a metered oracle over claude -p subprocess with --resume).
• jvogan/a-fable-of-codexes — 28★, Jul 4: Fable as "conductor of AI worker fleets… campaigns, waves, squads, and review gates" (Codex workers + Opus design agents).
• disler/planf3 — 103★, Jun 21: "Plans For Fable 5" (IndyDevDan's planning-artifact pattern).

Fallback/downgrade-aware tools — the near-competitors to our Opus-Fallback Detector:

• Botable-dev/scoperoute — 2★, Jul 4, by HubLab.ai: "Which of your repos can you actually build with Claude Fable 5? Detects the silent Fable→Opus fallback per component, prices the run as a % of your Claude plan, and leaves a free Fable architecture review in each repo." README banner verbatim: "⏳ Claude Fable 5 is free on Claude Code only until July 7 — after that, API pricing ($10 / $50 per M). Find out where your free window actually lands before you burn it." And: "Fable 5's safety classifier reads the context of your request… and quietly falls back to Opus on repos it doesn't like. You can spend your free week on the best model and never actually get it." Mechanism: active probing (spends quota per component) + cost estimate per pipeline stage in tokens/$/% of plan. NOT transcript forensics — it can't tell you what already happened; ours can, for free.
• smithersai/fable-bench — 1★, Jul 3: "Smithers benchmark for measuring claude-fable-5 downgrades to Opus." Benchmark harness, again prospective, not historical.
• markrpg/fabler-skills — 0★, Jul 2: "Cursor skills for Fable planning in restricted or routing-prone coding scenarios" ("routing-prone" is now vocabulary).

Docs/archive artifacts:

• fuuuuuuma/claude-fable-5-revival-guide — 8★, Jul 1 (Japanese): Fable-revival history + "what to do first after revival (config repair + harness building)" README + slides.
• Braffolk/fable5-world-demo — 531★, Jun 12: the viral "3D world built by Claude Fable 5 to test its capabilities" (the show-off genre).
• asgeirtj/system_prompts_leaks — 51,263★: carries the extracted Fable 5 system prompt; cyrus-tt/fable5-system-prompt (26★) claims "full 1585 lines + lite 164 lines."
• thebuggeddev/modkeys (22★, Jul 5), Thom_Wolf-style "one big brief" builds, various "made with Fable 5" demo repos — the tokenmaxxing output.

Noise/scam watch: multiple SEO-bait repos ("claude-fable-5-free-desktop-app" 301★+1★ clones, "claude-desktop-fable-51-macos" = "Unlock the hidden Claude Fable 5.1 model", ikarma mythos desktop app 135★) — classic fake-installer keyword-stuffing pattern around a hype window. Do not link; worth knowing the ecosystem is polluted if we surface "community Fable tools" anywhere.

1b. Gists

gh has no gist-search API surface; GitHub gist search is not exposed via REST — angle recorded as UNQUERYABLE directly. (Rush prompt templates are landing as full repos anyway, per above.)

2. COMPETITIVE LANDSCAPE — who reads ~/.claude and what they shipped this week

The big four

CodexBar SHIPPED Fable-era features this week — the one competitor that moved:

• v0.38.1 (2026-07-04) release notes verbatim: "Claude: show model-scoped weekly quota windows, including promotional Fable limits, from OAuth and web usage responses." — i.e. CodexBar reads the server-side quota API and already shows the 50%-Fable-cap window. That's the piece we declared "NOT knowable from local data" — CodexBar gets it from OAuth/web endpoints, not transcripts.
• v0.39.0 (Jul 4): per-project/worktree cost grouping (Codex side), credit-expiry summaries.
• v0.40.0 (Jul 5): stacked multi-account Claude usage cards (claude-swap), calendar-correct credit totals. Three releases in three days across the cutoff window.
• What CodexBar has that we don't: server-side quota windows (incl. Fable cap), multi-provider (Codex/Cursor/Kimi/Doubao/z.ai/…), multi-account stacking. What we have that it doesn't: transcript forensics (per-message model ids → fallback detection), fleet board, attention strip, deadline board w/ Linear export, dreaming ledger, receipts+reconcile, cmux transport, cross-machine.

ccusage: pricing comes from an embedded models.dev snapshot (PR #1242 Jun 9 "embed models.dev snapshot for offline pricing", #1264 Jun 11 refresh). No Fable-specific issues/PRs found in-repo; no release since Jun 15 — ccusage has NOT shipped anything cutoff-specific. Fable pricing rides in via the models.dev snapshot only. No credit simulator, no fallback detection (they aggregate cost per model id; they don't diff requested-vs-served).

Long tail (all read local JSONL or the OAuth usage endpoint; none do fallback forensics):

• HermannBjorgvin/Clawdmeter — 1,771★: ESP32 desk hardware usage meter (the "ambient quota anxiety" market is real enough for hardware).
• utensils/claudette (58★), ClaudeCodeUsage/ClaudeCodeUsage VSCode ext (47★), niederme/ai-quota menubar (6★), khoanguyen-3fc/ClaudeTray ("real-time pace tracking"), timmyagentic/quota-monitor, ericbrophy/claudebar, kalatsch/claude-activity (heatmap), jamesleoreyes/cc-usage-tracker-tracker (a tracker-of-trackers gag — the niche is saturated enough to be self-parodying).
• Fresh this week but zero-traction: wassim-rekik/claude-code-monitor (Jul 3, team-wide JSONL reader), Sculptsoft claude-code-monitor (Jul 2, live timeline dashboard), parthmashroo/claude-code-monitor (Jun 28, floating widget w/ real 5h/7d limits).
• eltonylfgi-blip/claude-code-usage-guard (Jun 27, 0★): "in-session warnings when you're burning usage fast, plus a /usage command. Zero deps, local-only" — the in-session-warning pattern as a plugin.

Nobody in this entire lane shipped: a credit-era cost simulator ("your history repriced"), a transcript-based fallback detector, or a durable-artifact ledger. See NULLS.

3. INTEGRATION TARGETS

3a. SKILL.md conventions the Fable Ledger should recognize

• Official spec (anthropics/skills, 158,768★, pushed 2026-07-01): README verbatim: "Skills are simple to create - just a folder with a SKILL.md file containing YAML frontmatter and instructions" and "The frontmatter requires only two fields: name - A unique identifier for your skill (lowercase, hyphens for spaces); description - A complete description of what the skill does and when to use it." Everything else (allowed-tools, license, metadata, version) is optional/ecosystem extension.
• De-facto layout (uniform across the rush repos + tomicz taxonomy): .claude/skills/<name>/SKILL.md (project) and ~/.claude/skills/<name>/SKILL.md (user), optional scripts/, references/, assets/ subdirs next to SKILL.md.
• Rush-specific markers worth indexing as "Fable estate" signals: a "Provenance and maintenance" trailer section, date-stamped facts, <project>- prefixed skill names from the tomicz taxonomy, skills whose git-blame/creating-session model is claude-fable-5.
• Detection heuristic for the Ledger: any Write/Edit whose path matches (^|/)\.claude/skills/[^/]+/SKILL\.md$ or ~/.claude/skills/..., plus CLAUDE.md, AGENTS.md, .claude/agents/*.md, .claude/commands/*.md, plugin marketplace.json.

3b. claude-code issues worth citing (all 2026-07-06, one day before cutoff)

The tracker is saturated with silent-downgrade and quota complaints — direct validation for the Opus-Fallback Detector + countdown:

• #74778 "[Bug] Silent model downgrades from selected model to Opus 4.8 without user notification or consent" — the exact pain, as filed.
• #74780 "[Feature Request] Add user confirmation before switching models mid-task"
• #74734 "Fable 5 session repeatedly auto-downgraded to Opus 4.8 (suspected safety-classifier false positive on normal deploy work)"
• #74783 "Fable 5 safeguards false-positive on routine git/diagram work forces model switch; advertised /feedback (and /bug) commands don't exist"
• #74816 "False positive safety flag on Russian-language planning text with automatic model downgrade"; #74725 "excessive false positives on defensive security tasks"; #74847 "Fable flagged during timeout error debugging"; #74853 "Excessive Safety Tagging on Non-Sensitive Code Tasks"; #74737 "Reduce false-positive abuse flagging for legitimate agent fleet operations"
• #74788 "Feature request: subagent model allowlist so unpinned or nested Agent spawns cannot resolve to the main-session (premium) model" — the fleet-doctrine problem (our SubagentRollup) filed upstream as a feature request.
• Quota-anxiety issues filed as bugs: #74851 "You've hit your monthly spend limit", #74751/#74727 session-limit-reset issues (people filing reset times as bugs = the countdown pain).
• #74917 "fable available on one mac pro. and not available on another mac (air)" — cross-machine model availability confusion; our cross-machine story touches this.

3c. Hooks/statusline patterns for model/quota awareness

The statusline ecosystem is big, active, and shows model + usage-limit + reset-time — but nothing cutoff- or fallback-aware:

• sirmalloc/ccstatusline — 11,508★, pushed 2026-07-06 (active daily), last release v2.2.22 Jun 16. Widgets incl. session cost; open issue #475 "add Session Cost (with subagents) widget" (Jun 22) — subagent cost attribution is a live community want (we already have SubagentRollup). No Fable-specific issue or release.
• nilbuild/claude-statusline — 1,324★; Owloops/claude-powerline — 1,129★ (issue #92, Jun 10: "feat(models): add Claude Fable 5 and Mythos 5 support" — model-name display/pricing only); chongdashu/cc-statusline — 621★ (dormant since Feb).
• Long tail conventions worth matching: AndyShaman/claude-statusline (32★) shows "usage limits (H/W)" = 5-hour + weekly windows; hell0github/claude-statusline (25★) shows "session reset time"; bartleby/claude-statusline (25★) shows "rate limits". So the community-standard statusline fields are: model name, context %, 5h/weekly usage, reset time, session cost.
• Pattern to match, gap to own: statuslines read the same JSONL + OAuth data; NONE render a cutoff countdown, a requested-vs-served model mismatch, or credit-repriced cost. Our attention strip can be the first surface showing "2 of 14 turns fell back to Opus" — no statusline does it.

4. HONEST NULLS (positioning facts)

Searched and found NOTHING (gh search repos, name+description+readme):

• "fable credit" → 0 repos. "claude credit simulator" → 0 repos. Nobody has built "your Fable history repriced in credits." Our Credit-Era Cost Simulator would be first. Closest existing thing: scoperoute's prospective per-run estimate (% of plan), and CodexBar's raw credit totals for Codex (not Claude credits, not historical repricing).
• "fable fallback" → 0 repos. "opus downgrade" → 0 repos. The two fallback-adjacent tools that DO exist (scoperoute, smithersai/fable-bench, both ≤2★) are prospective probes that spend quota to test routing. No tool reads transcripts to report which past/live turns silently fell back — our detector (requested model vs message.model per assistant message) would be first and free.
• "train the replacement" as a repo name/description → nothing from this wave (only a 2024 Substack prompt library). The wave ships as fable-skills/fable-5-train-opus-… names instead.
• ccusage: no fable-labeled issues/PRs, no cutoff feature, no release since Jun 15.
• Gist search: UNQUERYABLE via gh/REST (no gist search API); not covered.

5. INTEGRATION RECOMMENDATIONS — ranked by (value × fits-our-data)

1. Ship the Credit-Era Cost Simulator NOW — it's a confirmed first-mover. "fable credit" / "claude credit simulator" → 0 repos; ccusage dormant since Jun 15; scoperoute only estimates future runs. Nobody reprices history. Zero new ingestion for us (SessionUsage × Pricing.swift).

2. Ship the transcript-based Opus-Fallback Detector — also first, and upstream demand is on the record. Cite claude-code #74778 / #74780 / #74734 / #74788 in the release notes; the only competitors (scoperoute 2★, smithersai/fable-bench 1★) burn quota probing prospectively; ours is free forensics on data we already parse. Statuslines don't do it either.

3. Fable Ledger: recognize the official + rush conventions. Index Write/Edit paths matching (^|/)\.claude/skills/[^/]+/SKILL\.md$, ~/.claude/skills/**, CLAUDE.md, AGENTS.md, .claude/agents/*.md, .claude/commands/*.md, marketplace.json; parse name/description frontmatter (the only required fields per anthropics/skills); tag tomicz-taxonomy prefixes (-debugging-playbook, -failure-archaeology, …) and "Provenance and maintenance" trailers as high-value estate artifacts. Bonus: a "Train the Replacement" recipe can embed the tomicz 3-phase prompt (242★ social proof) with the 30%-of-weekly-quota warning.

   Runner-up (4): match the community statusline vocabulary (5h/weekly windows, reset time, session cost incl. subagents — ccstatusline #475) in the attention strip so we read as a superset, and note CodexBar v0.38.1 proves the server-side Fable-cap window is fetchable from OAuth/web usage endpoints if we ever want the piece we marked "not knowable from local data."

Coverage notes (what was and wasn't completed)

• DONE: rush-artifact repo sweep (multiple query angles: "claude fable", "fable skill", "fable-5", "claude-fable-5", "train the replacement", "fable quota/credit/fallback", "opus downgrade"), competitive big-four + long tail ("claude code usage/monitor/dashboard"), CodexBar/ccusage release+PR inspection, claude-code issue sweep (Jul 6 window), statusline ecosystem (ccstatusline/claude-powerline/long tail), anthropics/skills official frontmatter spec, null probes.
• SKIPPED: gist search (GitHub exposes no gist-search API — unqueryable via gh/REST); per-author attribution hunt for @PrajwalTomar_/@0x_kaize beyond repo descriptions (the tomicz repo is the de-facto repo-ization of that template wave).