@rhei-agent/continual-harness
v0.1.0
Published
Versioned state and refinement primitives for continual harness learning
Maintainers
Readme
@rhei-agent/continual-harness
Versioned, inspectable state for RHEI's continual harness learning loop.
The package owns pure refinement operations and crash-safe persistence. It does not own model calls, session scheduling, or tools. The coding-agent package can therefore run planning at a turn boundary and apply the resulting proposal only after the active turn is quiescent.
The base system prompt is intentionally immutable. Refinements may create, update, or delete supplemental prompts, memories, skills, and sub-agent definitions. Every accepted refinement stores enough information for an exact rollback.
Scoped memory
FileMemoryStore provides a separate, private JSON store for durable memories;
it does not read or write session JSONL. Records are either global or tied to
the store's configured projectId. Every record has a typed source, trust
state, approval state, timestamps, version, and JSON metadata.
New memories default to pending and are never returned by recall() until an
explicit approve(id) call. Rejected memories are never returned and can only
be removed with forget(id). Project records cannot be read, updated, or
forgotten through a store configured for another project. Use a path under the
agent data directory for global memory and a path under .rhei for project
memory; the store creates the directory with mode 0700 and the file with mode
0600.
Recall currently uses DeterministicMemoryTextIndex, a dependency-free exact
token matcher over persisted records. It is deterministic and auditable, but is
not SQLite FTS or semantic retrieval. MemoryTextIndex is injectable so an FTS
implementation can be added later after dependency and migration policy are
reviewed.
All memory operations take an OS-process lock at <path>.lock, in addition to
in-process serialization. The default lock wait is 5 seconds; locks older than
30 seconds are treated as abandoned, with heartbeats keeping active operations
fresh. Configure lock.timeoutMs, lock.staleMs, and lock.retryDelayMs for
different bounds. The store reloads and validates the file only after acquiring
the lock, so separate processes cannot overwrite a fresher snapshot.
