digline

digline-mcp

The MCP server for digline: it lets a coding agent read a digline result and measure a new one, and it does not let the agent promote a baseline.

Not because promotion is refused. Because there is no such tool.

The eight tools

tool	what it does
list_runs	every stored run of a suite, newest first, with the baseline marked
get_run	one stored run — the verdicts, never the payload
get_baseline	the approved reference, same shape
log	which model answered, read down the stored runs; never a gate
compare	a run against the baseline: did it get worse?
diff	two runs, neither of them a reference: should I switch?
explain	one run read back as typed facts, with the exit code
run	execute the suite, with the call count acknowledged first

promote is absent by construction, and so are migrate, view and report. A baseline is an approved reference, and promote writes into .digline/<tenant>/baselines/, which is committed — anything landing there arrives in somebody's diff and must arrive because they put it there. A refusal would be a conversation an agent could argue with; an absence is not.

Every tool, the reading ones included, loads the suite it is named. For a suite.py that executes the project's code, in the server's environment.

The reasoning is ADR 0011, and the judgement layer it turns into surface is AGENTS.md.

Install and configure

uv pip install digline-mcp

One server serves one repository: the perimeter is the repo, and multi-project is N named servers rather than a registry.

{
  "mcpServers": {
    "digline-northwind": {
      "command": "digline-mcp",
      "args": ["--root", "/Users/you/src/northwind"]
    }
  }
}

--tenant and --env are accepted and verify; they never override. The suite decides, as it does everywhere else in digline.

What run costs, before it costs it

run takes a mandatory acknowledge_calls that must equal the suite's planned calls to the target. Called without it, the tool refuses and hands back the number:

run(suite="suite.py")
  → refused: this suite plans 100 calls to the target.
    20 cases × 5 samples = 100 calls to the target; each answer is judged
    3 times by llm_rubric.
    Call again with acknowledge_calls=100.

run(suite="suite.py", acknowledge_calls=100)   → executes

The first call is the probe. An agent cannot spend a hundred model calls without having stated the number — AGENTS.md §7 as a contract rather than as advice.

What never crosses

An MCP response goes into a model's context, and from there into transcripts and caches nobody controls. So the run tools return a verdict-only projection: the name, the identity, the status, the score, the threshold, the tolerance, the measured interval, and the metadata a suite disclosed.

Never the judge's reason — the judge quotes the output, so the reason is the output. Never the sentence explaining a suspension, never undisclosed metadata, never a case's inputs, and never a prompt or its digest unless the suite declares Disclosure(artifacts=True). A declared file outside the perimeter is refused when the run is made; one under .digline or .git is refused by the command that would let it cross. Neither is dropped in silence.