## TL;DR
Hash the source inputs and skip TypeDoc when nothing changed. The loop is the agent re-running generation unconditionally instead of checking whether inputs moved.

## Error
```text
(agent ran typedoc 12 times in one job; sources unchanged; no error text)
```

## Steps
1. Confirm the loop: check run logs for repeated identical TypeDoc invocations with no source changes between them. Expected: the repetition pattern visible.
2. Add a change gate: hash the entry points and tsconfig (e.g. a checksum file), and skip generation when the hash matches the last run. Expected: generation runs only on real changes.
3. Alternatively, gate on the build tool's own incrementality: only run the docs job when source files changed in the triggering commit or PR. Expected: the docs job is change-triggered.
4. Re-run the agent job on an unchanged tree. Expected: TypeDoc is skipped and the job finishes fast.

## When to use
- The agent re-runs doc generation in a loop
- Generation takes significant time on every iteration
- Sources are provably unchanged between runs

## When not to use
- TypeDoc fails (fix the failure, not the scheduling)
- Sources actually change each iteration (the runs are legitimate)
- The loop is in a different tool (apply the same hashing idea there)

## Tool compatibility
TypeDoc with any agent orchestrator. The hash-and-skip pattern works for Sphinx and JSDoc too.

### Variant phrasings
### loop detected across scheduled runs, not within one job
Persist the hash between runs (a state file or cache) so the next scheduled run can compare.

### agent re-runs because output timestamps change
Hash inputs, not outputs; output timestamps always change.

## Why it happens
Agents default to "generate then check", and without a change gate the generate step runs every iteration. TypeDoc is slow enough that the repetition dominates the job.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_qiqOVQwLYlbz5iyWGn0n8A