New row in the optional-env-vars table documenting LANGGRAPH_RECURSION_LIMIT.

docs/runbook/deploy-new-version.md

@@ -189,6 +189,7 @@ These are **never required** for the system to run. They're available for tuning
 | `USE_QDRANT_NATIVE_MMR` | Switches MMR diversification from the local text-Jaccard implementation to Qdrant's server-side MMR (`models.Mmr`, requires Qdrant server v1.10+). Accepts `1`, `true`, `yes`, `on`. | unset → local Jaccard MMR (the default — see ADR-007) | A/B comparison, experimentation, or if the corpus shifts toward paraphrased content where embedding-cosine similarity beats token-Jaccard. Safe to toggle: falls back to local MMR on any error. |
 | `LANGCHAIN_TRACING_V2` + `LANGCHAIN_API_KEY` + `LANGCHAIN_PROJECT` | Enables LangSmith Tracing — every turn becomes a Trace under a Thread (grouped by `conversation_id`). | unset → no LangSmith export | Always recommended in production for debuggability. See §6.14. |
 | `HF_TOKEN` | Hugging Face Hub token, used by the `huggingface_hub` library if you need authenticated model downloads. | unset → public models work fine; private models fail | Only if you need to pull a private/gated model. |
+| `LANGGRAPH_RECURSION_LIMIT` | Caps the number of agent-node executions per user turn. Hitting the cap triggers the same friendly "service unavailable" message as a quota error and writes a `🚨 INCIDENT type=recursion_limit` log entry. | unset → `8` (a healthy turn never exceeds ~5 steps) | Only raise (e.g. to `15`) as a temporary mitigation while diagnosing an unexpected agent loop. The fix is always in the code, not this knob. See [`fault-handling-and-common-issues.md`](fault-handling-and-common-issues.md) for the `recursion_limit` symptom entry. |
 
 > Reminder: API keys (`OPENAI_API_KEY`, `QDRANT_API_KEY`, `QDRANT_URL`) are **required**, not optional — they belong in Space *secrets*, not regular variables. See [`rotate-credentials.md`](rotate-credentials.md).