Summary

In Ruflo 3.51.1, a successful hooks_intelligence_trajectory-end reports that SONA learned a pattern and persists it to .swarm/sona-patterns.json, but the normal hooks_route path does not consult that persisted SONA optimizer. In addition, trajectory feedback is only promoted to the semantic pattern store when the trajectory has no steps. The common start → step → end path therefore records SONA state without making the feedback searchable through hooks_intelligence_pattern-search.

This was reproduced against the installed 3.51.1 package before applying a local compatibility shim. The shim is not part of this report; this issue describes the vendor behavior and the integration test expected upstream.

Environment

• Ruflo / bundled @claude-flow/cli: 3.51.1
• Node.js: 26.10.0
• OS: macOS on Apple Silicon
• Embeddings status: ONNX, semanticGrounded: true

Reproduction

1. Call hooks_intelligence_trajectory-start with a task and agent.
2. Record at least one step with hooks_intelligence_trajectory-step.
3. Call hooks_intelligence_trajectory-end with success: true and non-empty feedback.
4. In a fresh process, inspect SONA state / query the same task through the public routing path, and search feedback with hooks_intelligence_pattern-search.

Actual behavior

• trajectory-end returned sonaUpdate: true and a SONA pattern key; the pattern was written to .swarm/sona-patterns.json and reloaded by a fresh SONAOptimizer process.
• hooks_route returned its AgentDB or local semantic/keyword route instead of a SONA-pattern route. The hooksRoute implementation tries bridgeRouteTask and then routeTaskLocal; it does not call SONAOptimizer.getRoutingSuggestion().
• hooks_intelligence_pattern-search did not find the trajectory feedback. In hooksTrajectoryEnd, feedback promotion is gated on !hasSteps, so feedback from a trajectory that recorded steps is not indexed in the pattern namespace.
• A single success increases SONA confidence from 0.50 to 0.55, below getRoutingSuggestion()'s 0.60 threshold. The response can therefore say “learned” even though one successful outcome is not yet eligible to affect routing; please clarify/document this evidence threshold or make the response explicit.

Relevant implementation areas in the published package:

• @claude-flow/cli/dist/src/mcp-tools/hooks-tools.js: hooksTrajectoryEnd, hooksRoute, routeTaskLocal, hooksPatternSearch
• @claude-flow/cli/dist/src/memory/sona-optimizer.js: processTrajectoryOutcome, getRoutingSuggestion, saveToDisk, loadFromDisk

Expected behavior

• A qualifying persisted SONA routing pattern is considered by the normal hooks_route inference path, with an explicit confidence/valid-agent gate and an identifiable result source; otherwise, trajectory-end should not imply that routing was improved.
• Non-empty feedback is promoted to a durable, searchable pattern whether or not trajectory steps were recorded, with trajectory metadata and outcome confidence preserved.
• Add a fresh-process regression test proving both that an eligible persisted SONA pattern affects hooks_route and that feedback from a stepful trajectory is found by semantic pattern search.
• Keep process-local counters distinct from persisted counters in telemetry; hooks_intelligence_unified-stats already documents their different scopes.

Related issues

• Route Q-learner never persists CLI feedback (autoSaveInterval); trained LoRA/SONA not consumed at inference — routing appears non-self-improving #2222 covers a separate route-Q-learning persistence defect and notes that learned SONA/LoRA signals were not consumed at inference. Its reported Q-learning save fix does not close the SONA optimizer integration gap observed in 3.51.1.
• trajectory-end: feedback is never distilled into a searchable pattern; step-less trajectories always report patternsExtracted: 0 #2351 addressed indexing feedback for step-less trajectories; the stepful-feedback case above remains.

Acceptance criteria

• Successful, supported SONA patterns above the documented threshold can affect hooks_route after a process restart.
• trajectory-end with non-empty feedback and one or more recorded steps makes that feedback retrievable through hooks_intelligence_pattern-search.
• Response fields and metrics distinguish “recorded/persisted” from “eligible and used for inference.”
• Integration tests cover a fresh process, threshold behavior, and both stepful and step-less trajectories.