KetraTermPersistent storage
The standalone app owns its configuration files. IntelliJ uses IDE-managed application/project settings. Library embedders choose their own persistence.
Locations
| Data | Location |
|---|---|
| Standalone preferences | Resolved config.toml; see path resolution. |
| Malformed standalone configuration backup | <config filename>.broken beside the configuration; see recovery behavior. |
| Standalone completion learning | command-completion-learning-v3.tsv beside the configuration. |
| IntelliJ completion learning | ketraterm/command-completion-learning-v3.tsv under the IDE system directory. |
| IntelliJ tab state | Project workspace storage; see tab restoration. |
Raw terminal output is not saved automatically. Explicit output export is a separate user action.
Completion learning
Persistence is opt-in and requires completion to be enabled. Standalone uses
suggestion_learning_persistence_enabled under [behavior]. Product settings
controls are currently hidden; see completion availability.
A product starting with persistence disabled neither loads nor writes the learning file. Enabling hydrates the fixed file once. Disabling prevents new writes but does not clear in-memory learning or cancel a file operation already in progress. An unreadable/rejected file blocks ordinary writes until an explicit learning reset or a new coordinator lifecycle.
Dirty state is checkpointed every 30 seconds and on shutdown. Products allow up to 500 ms for final persistence, so shutdown is not an unconditional durability guarantee. Files are bounded; replacement is atomic where the filesystem supports it, with ordinary replacement as a fallback. The persistence module defines worker, failure, and lifecycle contracts.
File format
Version 3 starts with KetraTerm_COMMAND_COMPLETION_LEARNING<TAB>3.
Ranking rows precede replay rows:
| Row | Stored fields |
|---|---|
R |
Command digest, profile, working directory, usage/success/failure/acceptance/dismissal counters, and last-used time. |
H |
Matching digest and context plus policy-approved successful command text. |
Text fields use unpadded Base64URL. Unsupported versions, malformed rows, and inconsistent ranking/replay pairs reject the whole file. Derived matching data is recomputed, not persisted.
Privacy
Replay filtering excludes leading-whitespace commands, multiline/control-bearing or oversized text, malformed UTF-16, and recognized credential patterns. A host can further restrict replay with a predicate. Commands rejected for replay may still contribute opaque ranking evidence; malformed UTF-16 is not learned.
Approved replay contains command text. Base64URL is encoding, not encryption; deterministic digests can be guessed from candidate commands. Filtering is best-effort and cannot identify every secret. Treat the learning file as sensitive local command-derived data.