Manifesto 2026-06-15

What we shipped with Claude Fable 5: 42 PRs in 4 days

For four days our autonomous agent fleet ran on Claude Fable 5. Here is the honest build log: 42 pull requests across 10 repositories, from a full stolen-funds pipeline to a shared risk-signal catalogue, plus what the model was actually good at.

RektRadar is built by an autonomous agent fleet: a cockpit of Claude Code agents, each in its own git worktree, that pick up issues, open PRs, get reviewed, and auto-merge. The model behind those agents rotates as Anthropic ships new ones.

For a short window - June 10 to 13, 2026 - that model was Claude Fable 5. Then it rotated out (we are on Opus 4.8 now). This is the honest build log of those four days.

The headline numbers, straight from git:

  • 42 pull requests merged, across 10 repositories, in 4 days.
  • 42 commits co-signed Claude Fable 5 in the same window.

No cherry-picking - that is every PR that landed on main while Fable 5 was the engine. Here is what they added up to.

1. A stolen-funds pipeline, end to end

The biggest thread was answering one question with real on-chain data: how much ETH have scam tokens actually drained? That shipped across the whole stack in those four days:

  • global-analytics got a stolen-funds aggregation across Uniswap V2 + V3 + V4 scam pairs.
  • log-ui surfaced it as a card (and the auth layer was opened so the number is public).
  • rektradar-app proxied the aggregate so the public app could read it.
  • The landing page got an all-time “ETH drained” counter under the live rug counter.
  • And it became a blog post: $35M in four months, drained by Uniswap V2 scam pairs - chart inlined at the top, custom cover and all.

One question, one number, wired from the SQL aggregation all the way to a counter on the homepage and a published article. That is the kind of cross-repo thread an agent fleet is good at: the same feature touched five repositories without anyone hand-carrying context between them.

2. The risk-signal catalogue became a shared package

RektRadar scores tokens against 100+ risk signals (honeypot, blacklist, mass-deployer, sybil-funded holders…). Each one has a /signals/<slug> page for SEO and education. The problem: the catalogue lived as a copy in two codebases that had drifted, which left 404s on 73 signal slugs in the app.

In this window the fleet:

  • extracted the catalogue into a shared npm package, @mik3fly-lab/risk-flags,
  • made both the site and the app consume it (killing the 73 404s),
  • aligned the /signals pages with what the pipeline actually emits instead of an aspirational list,
  • added per-flag occurrence counts to each page,
  • promoted the holder-concentration flags from roadmap to active once the data backed them,
  • and added a dozen missing signal pages plus tag + severity filters with search.

This is the unglamorous work that keeps a content surface honest: one source of truth, consumed everywhere, matching reality.

  • A bilingual /glossary/ section (EN/FR) explaining market cap, liquidity, volatility, spread, leverage and friends - with a scam-detection angle, not a textbook one.
  • Branded /r/ short links for the glossary, plus evergreen /r/x reply links (landing, blog, signals) so social traffic is clean and attributable.

4. Imminent-rug Telegram alerts

On the detection side, the mempool watcher learned to persist privileged-call alerts (the on-chain moves that precede a rug) and hand them off to a Telegram broadcaster, backed by a new privileged_alerts table. That shipped alongside a blog post documenting four new scam signals, including the imminent-rug warning.

5. A UI refresh and mobile fixes

The app got a visual pass: palette, corner radii, a one-click address-copy component, clickable signals, spacing - plus a real mobile hamburger drawer instead of hiding nav links, and a breakpoint fix for tablet overflow.

6. The boring half: cleanup

Roughly a quarter of the PRs were housekeeping, which is exactly what you want a fleet to spend idle cycles on: dead-code removal (39 items in one repo, 11 in another), SonarQube auto-fixes, and dependency bumps across factory-watcher, graph-crawler, scoring-engine and token-forensics.

What Fable 5 was good at

A few honest observations from driving it for four days:

  • Broad, parallel, cross-repo threads. The stolen-funds feature and the catalogue extraction both spanned 4-5 repos. The fleet kept the shape consistent across all of them without a human re-stating the plan each time.
  • Content + plumbing in the same breath. It is comfortable writing the SQL aggregation and the blog post and the OG cover that ships it - the seam between “code” and “content” is where a lot of solo-founder time usually leaks, and it closed it.
  • Cleanup as a default. Given slack, it removed dead code and fixed Sonar findings rather than inventing scope.

None of this is hands-off magic. Every PR went through CI and an automated review gate before auto-merge; the human in the loop is steering, not typing. But four days, 42 PRs, 10 repos is a real snapshot of what a small operation can ship when the model behind the fleet is strong.

Fable 5 has since rotated out. The work it left on main did not. That is the point of building in public: the changelog outlives the model.