Thundering real repositories: results
thunder repo <owner/repo | GitHub URL | folder> shallow-clones a public repository and reads every
supported file with the Thunder code reader. It is read-only: nothing is executed, imported, or built.
Each decision becomes a Thunder rule card, every other function becomes Thunder steps, and anything Thunder
cannot read is listed for a person.
thunder repo psf/requests
thunder repo https://github.com/spf13/cobra --json
thunder repo ./my-service
One public repository per supported language (2026-11-22)
Thunder reads a function in one of two ways. A decision (a function that chooses between fixed answers) becomes a Thunder rule card you can ask questions. Every other function is read as Thunder steps: a short plain list of what it does (checks, branches, loops, values it works out, services it calls, what it returns or fails with), with one level of nesting. A statement Thunder does not recognize stays as its original code.
| Language | Repository | Files | Rule cards (of decisions found) | Functions read as steps, at least 80% readable | Reader errors |
|---|---|---|---|---|---|
| Kotlin + Java | JakeWharton/timber | 12 | 0 of 3 | 194 of 243 (80%) | 0 |
| JavaScript | expressjs/express | 142 | 3 of 3 | 427 of 447 (96%) | 0 |
| Java | google/gson | 268 | 4 of 17 | 3462 of 3677 (94%) | 0 |
| PHP | guzzle/guzzle | 147 | 10 of 17 | 2544 of 2616 (97%) | 0 |
| TypeScript (React) | jaredpalmer/formik | 120 | 2 of 3 | 218 of 233 (94%) | 0 |
| C# | jbogard/MediatR | 161 | 3 of 4 | 830 of 862 (96%) | 0 |
| Python | psf/requests | 37 | 3 of 3 | 599 of 632 (95%) | 0 |
| Rust | rust-lang/log | 13 | 0 of 1 | 396 of 449 (88%) | 0 |
| Ruby | sinatra/sinatra | 147 | 19 of 25 | 684 of 895 (76%) | 0 |
| TypeScript | sindresorhus/ky | 98 | 0 of 1 | 1338 of 1370 (98%) | 0 |
| Go | spf13/cobra | 36 | 2 of 2 | 551 of 587 (94%) | 0 |
| Total | 1181 | 46 of 79 | 11243 of 12011 (94%) | 0 |
Reading all eleven repositories takes under 10 seconds. Example, from this repository's sample shop:
What issue refund does
= Works out decision from approve refund
◇ If decision is "Approve"
→ Calls Stripe
→ Writes to the database (insert): refunds
→ Announces an event: refund.issued
⇥ Returns decision
What this says, honestly
- Most real code is steps, not decisions. The reader's main blocker was "this statement does work besides choosing an answer" (about 34,000 lines across these repositories). Reading steps raised the share of functions Thunder can show in plain words from under 1% to 89%.
- Steps are read from code, not rules. They say what the code does in order; they do not say what it was meant to do, and they are labeled that way. Rule cards remain the only views you can ask questions of.
- TypeScript pass (2026-11-22): object literals, destructuring, and JSX no longer split statements; casts,
non-null and optional chaining, private fields, counters, inner classes, and collection chains
(
map/filter/find/some/every/reduce, LINQ, Rust iterators, Java streams) read as plain steps. Newline-terminated statements (Go, Kotlin, JavaScript without semicolons) are split correctly. - Weakest today: Ruby (76%) and Kotlin (80%, mostly in tests): blocks, DSL-style calls, and heredocs are often kept as original code.
Next compiler steps (in order of expected coverage gain)
- Chained and closure-heavy expressions in TypeScript, Kotlin, and Rust (
a.map(x => ...).filter(...)), read as "for each … keeps …" steps. - Comparisons between two values (
length > limit) and method calls with arguments (list.contains(x)) inside decisions, so more decisions become rule cards. - Early-return chains mixed with assignments, read as rule cards with "works out" steps.
- Calculations returned as values, which needs the Thunder Text calculation layer.
Each step is measured by re-running this table.