IBM Session Health Monitor

Agentic coding assistants like IBM Bob can quietly go wrong in the middle of a session, and nothing tells you. A fix gets reapplied to a file that already has that exact fix. The same error comes back after Bob says it's resolved. A session runs 30, 40, 50 tool calls deep because the original request was broader than one clean unit of work. None of this looks like a crash — it looks like progress — so a developer only notices after wasting real time (and real tokens) on a session that was stuck long before they checked in. Bob Session Health Monitor is a lightweight layer built directly on top of Bob's own hook lifecycle (PostToolUse, Stop, UserPromptSubmit) that watches every tool call as it happens, detects three specific failure patterns, and — the moment the next prompt is submitted — silently hands Bob a diagnosis and a concrete next step, before Bob acts on that prompt. It catches three things: the same file being rewritten three or more times in a short window (a stuck fix), twenty-five or more tool calls with no pause (a runaway session), and the same error-shaped response reappearing after a fix was attempted in between (an unresolved loop). Each one maps to a hardcoded, plain-English scenario — what it usually means and what to do about it — so the verdict is instant and fully deterministic, with zero extra model calls and zero added latency. A companion dashboard shows every session's health score and flag history at a glance. The target user is any developer or team relying on an agentic coding assistant for real work, who wants a safety net that catches a stuck session automatically instead of a human having to notice thirty tool calls too late. It requires no workflow change: once the hooks are in place, it runs silently in the background of every normal Bob session. What makes this a genuinely new angle rather than another wrapper around an LLM call: it turns Bob's own hook architecture into a self-monitoring nervous system, using Bob to watch Bob, with no separate judge model in the loop at all. Every verdict is traceable to a specific rule a human can read and audit, rather than an LLM's opinion that might hallucinate or drift between runs. That also made it straightforward to extend the exact same detection engine to watch Claude Code, Codex CLI, and Gemini CLI sessions too — proving the pattern isn't a one-off script but a reusable architecture that treats "the assistant's own hooks" as a general-purpose instrumentation layer, with IBM Bob as the reference implementation everything else stays compatible with. The result is a tool that makes agentic coding sessions self-aware of their own failure modes in real time — not by adding another AI on top, but by paying closer attention to the AI that's already there.