Replaces the per-plugin Ollama HTTP calls with internal/llm, which picks a
backend from the environment (vLLM or Ollama) behind one Chat interface, plus
internal/plugin/llm_client.go as the plugin-facing wrapper.
Startup now logs llm_backend/llm_endpoint/llm_model instead of the two
OLLAMA_* vars, which no longer describe where inference actually goes.
These files were already running in prod from the vLLM migration but had never
been committed; this is that live state, byte-for-byte.
The run summary had sixty seconds, which is a generation budget, and
this call almost never gets to just generate. Runs end hours apart,
Ollama drops an idle model after about five minutes, so the steady state
is weights on disk and a cold load before the first token. The old
budget expired during that load every time and filed the empty beat that
means "no summary, ever" — there is no retry, the beat itself is the row
that stops the sweep re-picking the run. Five minutes now, sized for
load-then-generate, so a timeout means what the comment always claimed
it meant: the box is down.
That can't sit on a two-minute ticker, so it doesn't. The sweep starts
beside the ticker behind a single-flight flag; ticks that land during a
load skip instead of queueing. Nothing reorders — the summary is written
to the local buffer with the next seq and still ships behind the run's
own log.
Claude-Session: https://claude.ai/code/session_012bxpQQJDjC1mTtLN3VVtBQ
Two halves of the same gap. A dispatch announcing a clear, a retreat or a death
said how an expedition ended and gave a reader no way to reach what happened in
it — so those three facts now carry the run they concluded, and Pete's permalink
can open the log behind the headline.
Finding that run is done by asking what is actually true at the moment a
dispatch is filed rather than by threading a run id through five call sites that
have already let go of it. Both guards on that lookup are load-bearing: a run
with no beats behind it would mint a link to a 404, and without the recency
window a campaign death at the Empty Throne would attach itself to whatever
dungeon that player last walked.
The other half is the summary. Every line of the liveblog is assembled by Pete
out of a beat's own nouns, which is right for a log — it has to be exactly what
happened, in order. But a report is read afterwards and the question it answers
is not "what happened", it is "what was that run", and that is a judgement. So
this is the one piece of prose on the channel, and it earns the model far better
than a dispatch headline does.
It runs on the roster ticker, not at the moment the run ends. A run ending is
already a chokepoint with a dispatch being authored against it, and a second
generation there would stall the command that killed the boss. One per tick, and
a run that can't be summarised is closed out with an empty beat rather than
retried forever — the row is what stops the sweep picking it up again, and a
report with no summary is still the log and the numbers.
The prompt states the one permitted name twice and forbids arithmetic: a model
asked to write warmly about a party will invent a second member of it, and a
total it works out itself will contradict the exact totals printed beside it.
Both were seen against the real box before this was tightened.
Claude-Session: https://claude.ai/code/session_012bxpQQJDjC1mTtLN3VVtBQ
prosolis