Two defects in the run identity and the headline metric, both found by
running the new agent adapter and reading its output rather than trusting
it.
`RunIdentity::detect` finds the model through the environment, which is
how the daemon is configured. The `agent` subcommand takes `--model`
directly, so every run started that way reported `no model` while holding
a loaded one. An identity that is wrong exactly when a model was used is
worse than no identity: it would have labelled every benchmark row as
modelless. It now stamps the model it actually loaded.
The metered loop passed `None` for a generation step's token count, so
`summary.tokens` was 0, `microjoules_per_token` was `None`, and the JSON
reported `output_tokens: 0` — which reads as "generated nothing". Every
run down this path silently disabled joules-per-token, the metric this
project exists to report.
The counts were never missing, only discarded. `l3_generate_counted`
already returns them, and its own doc says it exists so that
µJ/accepted-token is "measured rather than left at zero or invented" —
the metered path just called `l3_generate` instead. It now takes a
generator returning `(text, tokens_out)` and records the count on the
step.
First measured figures from the agent path, on a 2.6B local model:
192 output tokens, 6.49 kJ, 111 s — 33.8 J/token at 1.73 tokens/sec. The
joules figure is now real. The token rate is roughly twenty times slower
than this model should run on this hardware, which is a separate thread
to pull.
Co-Authored-By: Claude Opus 5 <[email protected]>
6362476bc9