MTPLX 2.5.4

Agent sessions got the attention this cycle — especially Pi. If you drive MTPLX from Pi, OpenCode, or any tool-calling client, warm turns should now stay warm.

Faster warm turns in agent sessions

• Tool turns re-used less cache than they should have. Every tool round carries a short transient hint that shifted the cached prefix by ~200 tokens, so each turn re-processed more prompt than it needed to. The engine now records the stable boundary and restores from it directly.
• A postcommit that was about to finish is worth waiting for. When your next message arrives while the engine is a few hundred milliseconds from finishing the previous turn's cache commit, it now waits briefly (bounded, 0.6s) instead of throwing that work away and re-processing the difference. In our agent harness this turned a 1,449-token re-prefill into 436 tokens and cut that turn's time-to-first-token from 2.7s to 1.1s.
• Background cache maintenance no longer sneaks into your turn. The engine processes a request as more than one internal job, and background SSD cache work could slip into the tiny gap between them — its GPU work then drained ahead of your prompt, showing up as a ~0.8s stall and a scary-looking dip on the tokens-per-second gauge (the stream itself was fine; the average lied). Idle work now waits out those gaps, and the SSD encode/writer additionally yield to any queued or running request. If you've seen unexplained pauses at the start of a turn in agent sessions: this was it.
• The SSD tier no longer hydrates candidates that can't win. If a fresher in-RAM snapshot already matches more of your prompt, the multi-gigabyte disk read for an older SSD candidate is skipped entirely.

The session cache tells you what it's doing (#229, #230)

• The daemon now prints the resolved cache budget at startup — total, per-session cap, and whether sizing is automatic — plus the exact environment variables to override it. (This line was promised in the 2.4.2 notes; it was being logged at a level nobody sees. Sorry.)
• If a long conversation outgrows the per-session cap, you get one clear warning with the numbers and the setting that raises the ceiling — instead of silent cold prefills after a restart.
• MTPLX_SESSION_BANK_MAX_BYTES=8GB now parses ("8G", "8GB", "8GiB" all work). Unparseable values warn instead of silently using the default.
• The app no longer drops explicit cache sizes you set in Settings when the policy is "target default".

Long-context decode on 32k+ agent sessions (#228)

The app was forcing a paged-attention route at 32k context with launch-day thresholds that were never re-measured. Reporters measured it 4-7x slower at 43k. The app now defers to the engine's measured thresholds (64k).

Smaller things

• Vision sessions: the near-prefix cache restore lane is now explicitly capped at the first image token, so it can never resurrect cache computed from a different image's pixels. Same-image warm reuse (the content-keyed path) is unchanged.
• mtplx serve --no-auth — explicit auth off-switch for localhost binds (#235). Non-localhost binds still require a key.
• Chat completion responses can now include a llama.cpp-style timings object (#237 — thanks to the contributor) for clients that read prompt/decode throughput from the response body.
• The expected-value adaptive depth policy's cost constants now reflect measured reality on current kernels, so depth-3 drafting engages when it should (it was firing on 13% of eligible rounds despite 65% acceptance).

For the curious

The gate for this release ran a Pi-shaped agent transcript (16k context, 7 tool turns) alternating baseline and candidate under fan-verified thermal control. Full receipts live in the repo's measurement logs.