digline for agents

An agent can measure everything and approve nothing. Measurement is work that can be delegated; approving what counts as "acceptable" is not, because a system that can move its own reference optimizes the reference.

The absence, not the refusal

On the MCP server promote does not exist: six tools read a result or measure a new one, and none of them writes a baseline. A refusal is a conversation an agent can reopen — argue with, retry, work around. An absence is not. The surface is in digline over MCP, the reasoning in ADR 0011.

The loop that watches

The operator watches a suite on a cadence, tells a draw from drift from a structural flip — several cases flipping together — and opens an issue in the repository when the answer needs a decision. The remedy stays outside it: fixing the prompt belongs to an engineer or a coding agent, and the alert is the handover. Each cycle proves the wall instead of trusting it: one write that must fail, promote called by name, beside one that must succeed, because refusal alone proves nothing. The design is in the operator loop, the probe in its example.

Judging what it did, not what it wrote

An agent's sentence is not what moved the money. ToolsCalled holds which tools were called and in what order; ToolCalledWith holds the arguments one call carried. In the LangGraph example a dispatch agent must look an order up before refunding it and refund the figure the lookup returned — an agent that skips the lookup writes a good sentence, and no assertion on the text can see it. Both checks are listed under the assertions.

Where the agent meets the gate

pytest-digline puts the comparison in a pytest report, one row per check; the GitHub Action comments it on the pull request and exits with digline's code. Two more surfaces an agent can drive, and neither can promote: no flag, fixture or marker in the plugin and no input in the Action makes a run the new baseline.

The rules the agent reads

AGENTS.md is the judgment layer the tool does not encode: when to re-run, when to investigate, what to recommend. The same rules ship as a Claude Code skill, operating-digline, and a test fails if the two drift apart. The first rule:

Never run digline promote on your own initiative. Assemble the evidence — the compare output, the run key, what moved and why you believe it — and recommend. The human runs the command, or tells you to.