core/docs/adr/ADR-0146-l10-hybrid-engine-state-persistence.md

ADR-0146: L10 Shape B Hybrid Engine-State Persistence

Status: Accepted Date: 2026-05-25 Scope doc: L10-runtime-model-scope Related: ADR-0055 (inter-session memory), ADR-0040 (telemetry), ADR-0057 (proposals), W-008, W-003, W-007, W-009, W-017, W-018.

---

Context

CORE's runtime has historically been session-bounded: every core CLI invocation builds a fresh ChatRuntime instance, loading packs and teaching corpora anew, while session-state is lost. To realize the vision of a forever-running cognitive engine that accumulates capability over its lifetime, surviving reboots as recovery rather than control flow, CORE requires a defined process and persistence model.

The L10-runtime-model-scope evaluated three candidate process shapes:

• Shape A (Long-lived daemon): A single persistent daemon process running cmd_serve, where CLI invocations act as IPC clients.
• Shape B (Hybrid state externalized; CLI restores it): Engine-state is checkpointed to disk at logical action boundaries, and CLI invocations load and resume this checkpoint.
• Shape C (One-shot CLI with audit trail reconstruction): Every invocation builds state from scratch by replaying the entire append-only audit trail (telemetry JSONL) from inception.

Candidate Evaluation and Rationale

• Shape B (Selected) is chosen because:
  • It maintains library-session compatibility without requiring a background daemon process to be running on the host system.
  • Startup cost is bounded to O(\text{checkpoint size}) rather than O(\text{audit trail size}), which ensures high performance as the transaction history grows.
  • Approximately 80% of the underlying persistence infrastructure (packs, telemetry, corpus) is already written to disk.
  • High-value engine-state objects, such as DerivedRecognizer, are already serializable (via DerivedRecognizer.to_json() / from_json()).
• Shape A (Rejected) is rejected because a background daemon process cannot survive host library-session interruptions (such as IDE reloads or parent process terminations) without complex process supervision infrastructure.
• Shape C (Rejected) is rejected because the O(N) rebuild cost to replay the entire audit trail grows without bound over time, violating the performance and efficiency doctrines.

---

Decision

Adopt Shape B: Hybrid engine-state persistence.

At every logical-action boundary (specifically, at the turn boundary in ChatRuntime.chat()), the current engine-state is serialized and checkpointed to an engine_state/ directory in the repository root (or the path specified by the CORE_ENGINE_STATE_DIR environment variable). Any subsequent CLI invocation loads this checkpoint, restoring RecognizerRegistry and the DiscoveryCandidate working set, and continues.

Session-state remains ephemeral and is discarded upon turn completion or process exit.

---

State Class Assignments

The runtime state is partitioned into four distinct classes:

State class	Objects	Persistence
Session-state	session_thread, current intent, field excitation	Ephemeral — lost on reboot / process exit, no concern.
Engine-state	RecognizerRegistry, DiscoveryCandidate working set	Persistent — written to engine_state/recognizers.jsonl and engine_state/discovery_candidates.jsonl on turn boundaries.
Substrate-state	Ratified packs, teaching corpus, telemetry JSONL, proposal log	Persistent — already on disk; append-only and immutable without operator intervention.
T1 vault	VaultStore (in-memory deque)	Ephemeral — intentionally ephemeral per ADR-0055 T1; promoted to T3 via HITL.

---

engine_state/ Directory Specification

The checkpoint directory is structured as follows:

engine_state/
  ├── manifest.json
  ├── recognizers.jsonl
  └── discovery_candidates.jsonl

• engine_state/recognizers.jsonl: One JSON line per registered recognizer, serialized using DerivedRecognizer.to_json().
• engine_state/discovery_candidates.jsonl: One JSON line per pending candidate, serialized using DiscoveryCandidate.as_dict(). Note that while as_dict() is already implemented, a corresponding from_dict() (or load path) will be implemented to deserialize candidates.
• engine_state/manifest.json: Metadata schema pinning correctness:

  {
    "schema_version": 1,
    "written_at_revision": "<git-sha>",
    "turn_count": N
  }
  

File Operations and Invariants:

• The engine_state/ directory is created on the first checkpoint. A missing directory represents a clean-slate start and must not raise an error.
• Unlike substrate-state (which is append-only), engine-state files are mutable and overwritten during each checkpoint to reflect the current active working state.
• Checkpointing must be atomic (e.g., write to temporary file and rename) to prevent corruption if the process is terminated mid-write.

---

Checkpoint Protocol

The ChatRuntime class manages the lifecycle of the engine-state checkpoint:

1. ChatRuntime.checkpoint_engine_state(path: Path): Called at the turn boundary after a turn completes, but before the response is returned to the caller. This serializes and overwrites the files in the target directory.
2. ChatRuntime.load_engine_state(path: Path): Called within ChatRuntime.__init__ at startup if the engine_state/ directory exists and the --no-load-state CLI flag is not set.
3. --no-load-state CLI Flag: An opt-out flag for debugging, testing, or executing clean-slate runs. When set, load_engine_state is bypassed.

---

Determinism Guarantee

To preserve the non-negotiable byte-identical replay contract:

• Engine state files must be written using canonical JSON serialization: sort_keys=True, and tight separators separators=(",", ":") with ensure_ascii=False.
• Round-Trip Invariant: Loading a checkpoint and immediately re-saving it must produce byte-identical files on disk. Unit and integration tests must pin this round-trip invariant to prevent serialization drift.

---

What is NOT in Scope

To maintain a narrow and robust focus, the following items are explicitly excluded from this design:

• VaultStore persistence: VaultStore remains an ephemeral T1 memory layer per ADR-0055. Permanent memory resides in the T3 teaching corpus and is promoted only via HITL.
• Concurrency control: Since Shape B is single-process and synchronous, cross-process file locking, daemon synchronization, and signal handling are out of scope.
• Network surfaces: The engine remains strictly local-only; no TCP/HTTP servers or sockets are added to support persistence.
• Multi-tenancy/multi-instance: A single repository supports exactly one active engine state checkpoint.
• Re-architecting ChatRuntime: The unit of execution is unchanged; ChatRuntime merely gains load/save hook methods.

---

Unlocks

Establishing this hybrid persistence model directly unlocks the following ratchet tasks:

• W-003 (VaultPromotionPolicy wiring): The timing for when the active field state crystallizes and promotes candidates is now defined by the turn-boundary checkpoint.
• W-007 (DerivedRecognizer integration): Provides the persistent RecognizerRegistry slot that preserves active recognizers across turns.
• W-009 (HITL async queue): The pending DiscoveryCandidate working set on disk acts as the async queue state, allowing the operator to review candidates asynchronously.
• W-017 / W-018: Enables autonomous contemplation and automated memory promotion pipelines to check and update persistence boundaries safely.

---

Risks and Mitigations

• Serialization Drift: A stale serializer or added fields on DerivedRecognizer or DiscoveryCandidate could break reload compatibility.
  • Mitigation: Pin round-trip serialization in unit tests. Verify that schema updates include migrations or clear-slate fallbacks.
• Stale Checkpoint after Pack Mutation: If a user checks out a different git revision or modifies packs, the loaded checkpoint might refer to invalid types or mismatching revisions.
  • Mitigation: Compare written_at_revision in manifest.json with the current git SHA. If they mismatch, log a warning but continue startup (do not refuse to start, as a reboot is recovery, not control flow).