Skip to main content
Sedum can remember how it found a browser control. On a later run, it checks that recipe against the current page. A unique, actionable match skips the locator model call. An absent, changed, ambiguous, or invalid recipe is a cache miss; Sedum uses the normal resolver and reports the reason. A cache miss is not an assertion verdict or automatic target healing.

​
When it runs

The local cache is enabled by default when sedum run starts inside a Git checkout during development. It is off by default in CI and always off outside Git. A fresh clone has no cache. --no-locator-cache disables it for one run. Targets that depend on a runtime placeholder bypass the cache. In CI, --locator-cache-ci or SEDUM_LOCATOR_CACHE_CI=1 opts in; an explicit --no-locator-cache wins. Opting in still requires a Git checkout. Set CI=0 or CI=false only when the environment is genuinely a development run.
sedum cache clear removes the cache and its digest key for the current worktree. The files are in that worktree’s Git metadata, outside the tracked tree; Sedum does not change the project’s .gitignore. A linked worktree has its own cache. There is no age-based expiry in 0.1. Incompatible formats or matching rules, corrupt data, and failed current-page validation cause misses and invalidation. Local files are not uploaded to a hosted cache.

​
Parallel lanes

Lanes and concurrent runs in one worktree share the cache safely:
  • Each entry is written under its own lock and moved into place atomically.
  • An invalidation removes an entry only if it still holds the recipe that lane saw.
  • A lock left by a crashed run is broken once it is more than ten seconds old.
If a lane cannot get an entry’s lock in about two seconds, the step records the cache reason conflict and keeps its model result, and the run summary counts the conflict. The test itself is unaffected.

​
What is retained

The stored key is a keyed digest of the full origin, path, query, fragment, operation, and normalized target sentence. For type steps, Sedum removes the value operand before making this key, so different values for the same field can reuse one recipe. The recipe contains bounded role, tag, input type, state, and structural path, plus keyed digests of permitted identity and nearby-label signals. It does not store the raw URL, sentence, labels, page text, current input values, or HTML. The 256-bit digest key stays beside the cache in Git metadata. Digests reduce accidental disclosure from a copied cache entry; a person who can read both the entry and key may still test guesses against them. Clear the cache when working with sensitive pages or when sharing the checkout’s Git metadata. The cache is separate from the classification cache in .sedum/. On a miss, the normal Resolver may send the step sentence and allowed page excerpts to the configured model provider; see TypeSafe provider and Clef provider. Locator recipes are provider-independent: a locally revalidated hit can be reused after changing providers, so that hit does not test the selected provider’s decision quality. Run results and evidence have their own privacy controls in run results. Each action’s RunResult.locator.cache records hit, miss, or bypassed, the reason, whether fallback called the locator model, and proven target-change evidence. Terminal output shows the same outcome. A changed or stale recipe remains visible even when the subsequent model-selected action succeeds. A miss with reason not_cacheable means the target has no signal that can safely find it again, for example an <input type="submit"> button whose name is its runtime value, so that step always uses the model. A run outside a Git checkout reports bypassed (outside_git) and prints a hint to run git init.