The budget was cut to 60s because a five-minute evaluation held the single
llama-server slot, and a voice turn arriving mid-evaluation waited behind it.
That collision is now solved where it belongs: the background client yields the
slot while a turn is in flight.
With the gate in place the short budget only truncates a Thinking model
mid-synthesis, which costs an observation and saves no latency on any real turn.
Kami's call, 2026-08-02.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BFaeSbLMEVG5ey8tejU3y2
llama-server is started without -np, so it serves one request at a time and
everything else queues. Mail extraction is allowed two minutes on a Thinking
1.7B, and the reader hands core up to 25 messages back to back. A turn arriving
mid-extraction therefore waited for whatever was left of that budget: the router
timed out into the classifier cascade and its 36.8% floor, and the phraser, which
has no floor, simply waited. Memory evaluation had the same shape with a five
minute budget.
llm.Gate is the bound. Foreground requests never wait. Background requests run
one at a time and yield while a foreground request is in flight, plus a quiet
window after it that covers the gap between the router call and the phraser call
of one turn. Clients get their priority from llmClientFor or
llmBackgroundClientFor, so which side a caller is on is decided at wiring time.
It gates only what goes through those clients, which the comment on Gate says.
mail intake: the extraction timeout no longer wraps the capture writes. A model
answering at 119 seconds of a 120 second budget left the first CaptureTask one
second and the third none, so candidates the model had already produced were
dropped with a deadline error. The mailbox name is validated before it becomes
provenance, since "email:" is not a source and neither is an arbitrary string
posted at the socket. The enable log prints the normalised candidate bound
rather than the configured one, which said "max 0" and then wrote three.
Found in review of #64.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
Both windows the evaluator keeps over the notes table were row budgets over
every writer. The dedupe read 200 recent notes and kept the eval ones, so after
200 ordinary notes an old observation left the window and the next evaluation
wrote the same sentence again. The snapshot asked for MaxItems notes and then
discarded her own, so once hourly evaluation had run for a few weeks the model
saw almost no real notes. Both reads are now filtered in SQL, by
RecentNotesBySource and RecentNotesExcludingSource.
Two smaller things in the same area. The dedupe key stripped any trailing
bracketed clause, so an observation ending in one hashed differently from its
stored form; it now strips only the recorded action. The evaluation timeout was
five minutes on the one llama-server that also answers voice turns, which made
a collision a five-minute mute assistant, and is now sixty seconds.
Found in review of #55.
Loading a different gguf was a one-line edit to phraser.model_path plus a
restart. It is now an owner-triggered IPC call, off unless configured.
internal/phraser/swap.go holds the safety properties as code:
- Never two models resident. The old llama-server is killed and reaped
before the new one is launched. One 1.7B fits the Vega iGPU; a
blue/green overlap would OOM the box, so it is not offered.
- Atomic from a turn's point of view. Swap drains the in-flight turns
(they finish on the old model), then refuses arrivals with ErrSwapping
until the new server has answered /v1/models. No turn ever sees half a
swap; refused turns fall back to the classifier cascade.
- A failed load rolls back. If the new model does not start or does not
probe, the previous one is reloaded and the call returns RolledBack
with the error. If the rollback also fails the daemon says so and
degrades to the classifier rather than pretending to serve.
Holders of the completion client are re-pointed, not rebuilt: llm.Client
guards its base URL and LLMPhraser.OnSwap re-points it, so the router, the
replier, the mail extractor and the memory evaluator follow the new port
without knowing a swap happened.
Reach is deliberately narrow. phraser.swap_models is an exact-match
allowlist of absolute paths a human wrote, rejected at startup otherwise,
so "swap the model" can never mean "load any file on my disk"; the running
model is always swappable back to. MethodSwapModel is AuthStepUp, the same
rung as mutating the tool allowlist, and /models gates POST through the
same stepUpOK the tools page uses. Nothing calls Swap on a timer and no
act, intent or utterance reaches it.
Vikunja #250
Ships the real, local, testable part of the memory-evaluation plan
(docs/plans/03-memory-evaluation.md): Maven reads back her own recent
memory on a slow ticker, asks the resident model what it notices, and
records the confident answers as notes.
internal/memeval — not internal/memory/eval.go as the plan says, because
internal/store imports internal/memory for the vector backend and an
evaluator has to read store.Fact/Note/Nudge, which would close the
cycle. Evaluate() gathers RecentFacts/RecentNotes/RecentNudges, prompts
under a GBNF grammar bounded to three {observation, confidence,
suggested_action} objects, drops anything under min_confidence,
deduplicates against what earlier runs wrote, and writes the rest as
notes with source infer:memory-eval. /dash already renders notes with
their source, so the output is visible with no UI change.
cmd/mavend/memoryeval.go drives it on its own goroutine and ticker, not
on the 60s tick: an evaluation is a multi-second round-trip on the same
llama-server that answers voice turns, and it runs hourly at most. The
memory_eval config block is absent by default and absence means the
goroutine does not exist. No llama-server phraser also means no loop —
there is no template fallback, because a "memory evaluation" assembled
from templates is a fixed sentence pretending to be an observation.
What it deliberately cannot do, since this is the feature most likely to
turn Maven into a nag:
- It cannot speak. No dispatcher reference, no channel, no nudge. An
observation is a thought she wrote down and he reads on /dash.
Announcing them is a separate decision with its own opt-in.
- It cannot act. suggested_action is recorded as text and interpreted
by nobody — no reminder, routine or fact is created from it.
- It says nothing about an empty store: no memory means no LLM call,
so there are no observations invented out of two facts.
- Its own notes are excluded from the next evaluation's input, and are
written with a nil embedding so they stay out of the recall pool.
The plan's remaining items (dispatching observations, an /eval IPC
method and trace view, RecentEvents) and the fact that output quality is
entirely unmeasured are written up at the bottom of the plan doc.
kami