Execute deterministic close/claim without the LLM

[agushudono]

Problem

The management cycle decides exits deterministically in JS (getDeterministicCloseRule + updatePnlAndCheckExits), but the actual close_position/claim_fees call was delegated to the LLM. When the LLM provider was unavailable (in production: "400 Your credit balance is too low"), every management cycle failed at step 0 — so stop-loss, trailing-TP, OOR, and low-yield exits never executed.

Observed impact: a position sat at its stop-loss threshold and bled to -33% for ~6 hours because no close could fire while the provider was down. Fee claims were also blocked.

Fix

runManagementCycle() now executes the deterministic actions directly:

Action	Before	After
CLOSE (stop-loss / trailing-TP / OOR / low-yield)	LLM calls the tool	JS calls executeTool("close_position") directly
CLAIM (fees ≥ minClaimAmount)	LLM calls the tool	JS calls executeTool("claim_fees") directly
INSTRUCTION (free-text condition)	LLM	LLM (unchanged — can't be parsed in JS)

The LLM is now only invoked for INSTRUCTION positions, and that call is wrapped in try/catch so a provider failure no longer aborts the cycle after deterministic closes have already run.

Why this is safe

• Goes through executeTool(...) (not raw closePosition), so all side-effects are preserved: recordClose/recordPerformance (inside closePosition), notifyClose + auto-swap base→SOL (inside executeTool), decision logging, pool-memory annotation.
• Live Telegram toolStart/toolFinish progress still updates per position.
• runSafetyChecks for close_position/claim_fees falls through to pass: true, so nothing is newly blocked.
• Standalone notifications remain suppressed while a live message is active (existing behavior).

Net result: stop-loss / trailing-TP / OOR / low-yield / claim now work regardless of LLM availability. The LLM is value-add for manual instructions, no longer a single point of failure for safety.

🤖 Generated with Claude Code