Inspiration

Yahoo Pipes (2007) let anyone wire feed sources and operators together on a canvas, watch the items at every wire in a debugger pane, and publish the result as a stable RSS URL. Yahoo shut it down in September 2015 with no export format and no successor, and thousands of embedded feed URLs went dark. Zapier and IFTTT fire per item and don't publish a stable output feed, so nothing today gives you "stream in, stream out" with a live preview and a feed URL you can subscribe to. Revenant resurrects that idea: the same canvas-and-wires model, with pipes stored as plain JSON files you own and every outbound fetch checked against SSRF.

What it does

  • Visual editor (React Flow): drag sources and operators onto a canvas, connect ports, configure each node, and preview the items at every node in the Ghost Inspector (debugger pane).
  • Eleven modules: Summon (Fetch Feed / Fetch JSON with JSONPath mapping), Banish (Filter), Arrange (Sort), Exorcise (Dedupe), Séance (Union), Transmute (Regex / Rename), Truncate, Rewind (Reverse), Raise Output. Every themed name is shown with its plain name.
  • Every saved pipe is served as a Haunting at /h/<id>.xml, .atom and .json, with ETags (304 on revalidation), a 5-minute cache and single-flight execution, so a popular feed runs once per window.
  • The engine is a pure TypeScript library with injected fetcher and clock; pipes are *.pipe.json files, so they outlive the server.
  • Outbound fetches refuse private, loopback, link-local, metadata, CGNAT, reserved and IPv6-transition addresses, pin the validated IP for the connection, and re-check every redirect.

How we built it (how Kiro was used)

Kiro surfaces. Everything ran in Kiro CLI, because the build machine is a NixOS laptop. Steering and the spec were written on Nov 5 with CLI 2.3 (V2 agent). The build ran on Nov 7 with CLI 3.1 (V3 engine) in spec mode (--mode spec --trust-workspace), one Kiro session per task in tasks.md, on Claude Sonnet 5.5. We did not use the IDE, Web or Crew.

Spec-driven development. The revenant-mvp spec has 12 EARS-style requirements, a design with 23 numbered correctness properties, and 22 tasks, each tagged with the requirements and properties it covers. Reviewing the spec before any code existed paid off: the first drafts used the current time as a fallback date (breaking determinism), returned 500 for a missing pipe, validated only the first DNS answer for SSRF, and left an open "Not implemented" stub. All of that was fixed in the spec (KIRO-LOG #3 and #5), not in code. Compared with vibe coding, each task arrived with a definition of done (tests, lint, build), so tasks 2–22 ran back to back without hand-holding.

Steering. Three always-on files (product.md: users, the themed-naming rule, non-goals; tech.md: pinned stack; structure.md: package boundaries) plus two conditional files: testing.md loads only for *.test.ts(x) (Vitest + fast-check conventions, "Property N" comments), and security.md loads only for server code and fetch-* modules (SSRF, ReDoS and secret rules). The conditional split kept security rules in context exactly where the fetcher and sources were written, without padding every other task.

Agent hooks.

  • test-on-save (PostFileSave, command): runs vitest related for every saved TypeScript file under packages/*/src. It ran 81 times during the build, so Kiro got test feedback on each file save, not only at the end of a task.
  • module-docs-sync (PostFileCreate, agent prompt): when a new engine module appears, Kiro updates docs/module-reference.md, adds the themed-label row to the product steering and makes sure a property-test file exists. It ran 4 times and wrote the module reference.
  • Lesson: the first version used the trigger id FileCreate, which CLI V3 silently ignores. We found it by streaming the session events (--output-format stream-json), where hook runs appear as hook_update events, and had Kiro change it to PostFileCreate.

Property-based testing. All 23 design properties are implemented as fast-check properties (16 test files, 100 runs each), next to 528 tests in total. PBT caught a real bug in task 8: fast-check generated an RSS feed with zero items, and the parser treated the empty <channel> as missing and threw, so a valid empty feed would have broken the whole pipe. Kiro fixed the parser ("empty feeds are valid and yield no items"), and the property now guards it.

Review loop. After each batch, the build was re-checked (lint, tests, build) and smoke-tested against the real app: live Hacker News through the API, ETag/304, and the editor driven in a headless browser. Two findings went back to Kiro as follow-up prompts: extra SSRF ranges (NAT64, 6to4, Teredo, reserved IPv4) and a UI pass (show the Haunting URLs with copy buttons; stop the canvas from zooming to 2× on the first drop).

MCP, powers, custom agents. Not used.

Cost. The whole build used 132.75 of the 4,000 Kiroween credits, about 1.5 hours of wall time for 22 tasks and 3 follow-ups.

AI disclosure. The entrant directed the project. Claude (Anthropic) acted as the orchestrator: it wrote the concept and the prompts, reviewed Kiro's output and ran the smoke tests. Kiro wrote all application code, tests, the README and everything under .kiro/. KIRO-LOG.md records every Kiro call, its purpose and its credit cost.

Challenges we ran into

  • CLI V3 on NixOS: the V3 engine downloads its own Node binary, which NixOS can't run as is, so Kiro runs inside an FHS wrapper (steam-run).
  • Hooks failing silently (wrong trigger id) until we read the event stream.

What's next

Server-side hosting of the built editor, more sources (HTML scraping with selectors), scheduled refresh and per-pipe secrets.

Built With

Share this project:

Updates

Submission history