Skip to contents

One object tying together the recorded design, personas, declared tool specifications, orchestration metadata, served model identifiers, generation parameters, and package versions. Its manifest_hash changes when a hashed component changes. It does not hash transcripts, replies, or unrecorded ambient state such as values captured by a tool closure.

Usage

agent_manifest(run)

# S3 method for class 'agent_manifest'
print(x, ...)

Arguments

run

An object accepted by as_agent_run() (a chat agent, a conversation, a preset, a pipeline, an experiment, an agent_fanout_result, or an agent_run).

x

An agent_manifest.

...

Unused.

Value

An object of class agent_manifest: a list with manifest_hash, kind, run_id, design, personas, tools, workflow, models, params, pkg_versions, created_at, n_calls, tokens_sent, tokens_received, models_basis, and n_unknown_tokens. Print shows the short hash, kind, models, and headline counts.

Details

The manifest also reports, outside the hash, how many calls the run made and the tokens they reported. n_unknown_tokens counts the calls, failed ones included, whose sent, received, or total count is missing, as LLMR::llm_usage() does. The model list is read from the call records when there are any (models_basis = "calls") and from the participating agents otherwise ("participants"); since the model list is hashed, the two bases can give different hashes for otherwise identical runs. When the call records cannot be read, the manifest warns, reports the call count and token sums as NA, and takes the model list from the participants. An experiment's design records the number of cells that contributed events (n_converted); that count describes the outcome rather than the specification and is not hashed.