Run Summariser

Who the node is. The id is the key the DOT pins.

id

run-summariser

name

Run Summariser

type

agent

phases

Testing

What it does, and the prose the agent is handed when the graph runs.

action28 words

Aggregate the run's trace records into one report of per-step metrics: how many times each step ran, how many of those runs failed, and how long each took.

spec179 words

Read every trace record handed to you on traces for the run in hand and write one report on report. Give every step that appears in the records its own row: the number of records, the number whose outcome is a failure, the first line of the most recent error, and its timing, read as the interval between each of its records and the record before it in the run; mark the first record of a run as unmeasured rather than giving it a number. Under the rows, give the run's totals: records, failures, distinct steps, and the time from the first record to the last. When a record is malformed, count it under a row named unparseable and go on. When rejected arrives from the gate naming what the previous report lacked, rebuild the report from the same records and address what was named; never fill a gap with a guess. Every figure in the report has to be recomputable from the records, and a step that failed every time stays in the report with its failures counted.in full above

model

claude-haiku-4-5

agent

Summariser

skill

skills/run-summariser.md

tools

none

mcp

none

params

timing: interval-from-previous-record, keep_unparseable_row: true

What arrives, what leaves, which nodes it expects to hear from, and what may not.

inputs

traces : json, rejected : status

outputs

report : report

dependencies

step-tracer, anomaly-gate

cannot

no type is refused

will_not

invent a timing for a step that has no record, drop a step from the report because it failed

The keys the static analysis reads. Nothing here instructs the agent.

risk_markers

none

notes59 words

The report is where tracing becomes metrics, and it is deliberately arithmetic over the records and nothing else: no reading of payloads, because the records do not carry them, and no judgement about whether a figure is acceptable, which is the gate's job. rejected is optional because the first report of a run has nothing to be rebuilt from.

The card's own version, and who wrote it.

version

1.0.0

author

autogen

provenance

not stated