Agents & MCP

Self-diagnosis recipes#

Concrete things to ask once the MCP server is connected. Each of these is a prompt you can paste, and a pattern worth turning into a skill.

Before starting work#

"Have I tried this before?"#

Using clawmetry, search my recent sessions for anything touching <the thing>. If you find one, read its trace and tell me how it ended before you start.

This is the single highest-value one. Most repeated agent spend is a second attempt at a problem whose first attempt is sitting in the store, unread.

"What did the last attempt cost?"#

Using clawmetry, get the cost summary for the last 7 days and list my five most expensive sessions. If any of them are about this task, tell me what approach they used.

Cost-aware agents make different choices. An agent that knows the last attempt was $12 will propose reading the schema before regenerating the migration.

While working#

"Am I looping?"#

Using clawmetry, list my events for this session filtered to tool_call. If the same tool and arguments appear more than three times, stop and tell me instead of trying again.

Three identical calls is the same threshold ClawMetry's own stuck_loop detector uses. An agent that checks this on itself catches the loop before the detector has to.

"Is this error familiar?"#

Using clawmetry, list events of type error from the last 14 days and see whether any match the error I just hit. If one does, read that session's trace and tell me what fixed it — or what did not.

The agent has your entire error history. It is the fastest incident search you have, because it is scoped to your machine and your work.

"What did I change since it last worked?"#

Using clawmetry, find the most recent session where <test or command> succeeded, then the most recent where it failed. Diff what the agent did differently.

After a failure#

Post-mortem on one session#

Using clawmetry, get the full trace for session <id>. Walk it and tell me: the first step that went wrong, what the agent believed at that point, and what evidence it had for that belief.

The why endpoint on the dashboard does this visually; asking the agent to do it produces something you can paste into an issue.

"Which tool keeps failing me?"#

Using clawmetry, list tool_result events from the last week and group them by tool. Tell me which tools have the highest failure rate and show me one example of each failure.

A tool with a high failure rate is usually a configuration problem — a wrong path, a missing credential, a permission — not a model problem. This finds it.

Weekly, as a habit#

The spend review#

Using clawmetry, get my cost summary for the last 7 days and the 7 before that. Tell me what changed, which runtime drove it, and whether any single session is responsible for more than 20% of the total.

The waste review#

Using clawmetry, list my sessions from the last week. For each one over $2, read enough of its trace to say whether it produced a durable result — a file change, a merged fix — or whether it ended without one. Total the second group.

That second number is the one that changes behaviour. It is also exactly what ClawMetry's no_progress detector estimates, so the two should agree; if they do not, the agent's reading is worth trusting more, because it can tell "wrote a file" from "wrote a file that was then reverted".

Making the agent do this unprompted#

Recipes only help if someone remembers to run them. Two ways to make it automatic:

Put it in the agent's instructions. A line in CLAUDE.md or the equivalent:

Before starting a non-trivial task, use the clawmetry MCP server to check
whether a previous session already attempted it, and say what you found.
If you make the same tool call with the same arguments three times, stop.

Make it a skill. → Skills and prompts

A note on trust#

The agent is reading its own past output. That is genuinely useful and it is also a feedback loop: an agent that misdiagnosed something last week will read that misdiagnosis back as evidence this week.

The events are facts — a tool was called, it returned this error, it cost this much. The agent's narration in a past transcript is not. Prompts that ask for the tool stream ("list events of type error") tend to be more reliable than ones that ask for a summary of what the agent said it was doing.