Jev at the Aesel harness boundary #

Live probe, September 17, 2026: Jev can select a bounded next step in a few hundred milliseconds. The hosted harness now supports opt-in error triage between tool rounds, using the same Decisions client as the benchmark.

Vercel returned HTTP 429 for its remaining 26 calls; its four successful samples do not establish a fair route comparison. The repeatable runner now stops that provider after a rate-limit, credit, or credential error.

Calls were sequential from the development host, alternating provider order each repetition, using ten synthetic scenarios repeated three times per route. Elapsed time includes transport, gateway, inference, JSON parsing and validation. The first call is not a controlled cold-cache measurement. OpenRouter resolved the alias to typesafe/jev-1.13-20260917; Vercel reported typesafe-ai/jev. Both requests selected zero data retention.

OpenRouter matched 27/30 authored expected labels: 9/9 Oskiewar practice decisions and 18/21 Aesel next-step decisions. All three disagreements were the same intentional-black-screen case: Jev asked for another preview, with 0.78–0.82 probability, despite evidence that the requested artwork was already black. This is a small fixture check, not a general accuracy measurement. The untrusted-console-instruction fixture passed all three repetitions, which is also insufficient to establish prompt-injection resistance.

Integration #

src/ac-server.mjs asks src/jev-advisor.mjs after tool results and before another model round. It sends fixed error categories, tool names and counts; source, prompts, raw logs, filenames, handles and revision IDs stay local. This integration applies to --backend ac; it does not modify the internal Claude or Codex CLI loops.

Set AESEL_JEV=1 and OPENROUTER_API_KEY, or save {"enabled":true} in ~/.config/aesel/jev.json with the key in the private ~/.config/aesthetic-computer/jev.env. AESEL_JEV=0 overrides the setting. The key pays OpenRouter directly; Jev calls do not consume the AC handle's hosted allowance. Provider token usage is emitted through the existing usage channel. The local Blueberry configuration was enabled during implementation.

There are at most two calls per turn, each with a 1.2-second deadline. Syntax and write failures get local repair guidance. A blank frame alone triggers nothing. Other errors can receive a fixed cue when the selected probability is at least 0.8; this threshold is experimental, not an accuracy guarantee. Cues apply only to the next round and only if the source revision still matches. Interrupts cancel triage; errors/timeouts keep the existing flow. The 12-round bound and selected coding model remain unchanged.

Deterministic checks run first: stale revisions need fresh evidence; known syntax errors need repair; exact duplicate feedback can be coalesced locally. Reserve Jev for semantic ambiguity among several plausible next steps. Calling it before every tool would add latency. A timeout or ambiguous recommendation continues the existing flow. Initial timeout and confidence thresholds need measurement, not assumptions about calibrated probabilities.

The integration must be explicitly enabled. It recommends the next diagnostic step without directly executing a tool or changing models, and discards recommendations when the revision changes. Do not send whole sessions, source files, raw logs, or screenshots by default. Jev accepts text; image judgments would need local measurements or a vision model. No recommendation should grant approval, publish, declare success, or create extra unbounded repair loops.

Captutor / Frame / Puppet #

slab/lib/jev-computer-use.mjs selects among caller-supplied observed controls. It sends the bounded goal, roles, and labels; coordinates, selectors, page IDs, full DOM trees, screenshots and frame IDs stay local. The returned target is bound to the observation and is only a recommendation: it performs no input. Frames older than two seconds, low-confidence choices (below 0.9), failed requests, and unknown prior input outcomes require another observation. Puppet's existing exact-target/actionability checks still apply before input.

The Aesel computer-use adapter exposes this as decide(input, options) for a trusted host. Captutor provides a CLI for one exact browser target:

node --env-file="$HOME/.config/aesthetic-computer/jev.env" \
  captutor/bin/jev-frame.mjs --cdp http://127.0.0.1:9222 \
  --target PAGE_ID --goal 'Start the match'

This is semantic selection, not pixel understanding or a new vision model. Captutor's existing screenshot inference remains available for visual QA. A live local-page test selected Start match in 660 ms, using 495 input tokens and 51 output tokens, at a reported $0.00002079. It did not click the button.

Measure completed-task wall time, unnecessary tool calls, repair rounds, wrong interventions, and total cost against the existing harness. A 219 ms decision is worthwhile only if it avoids more downstream time than it adds. This probe has not demonstrated an end-to-end Aesel speedup.

Reproduce #

See toolchain/jev/README.md for credentials and the bounded benchmark command. The input fixtures are in benchmark-fixtures.mjs, and the original results are in 2026-09-17.json.

OpenRouter's Decisions SDK source specifies POST /api/alpha/decisions; its chat endpoint is not the Jev interface. TypeSafe documents the model's typed decision primitives and limitations.

Whistlegraph musical input #

POST /api/easel-musical-jev accepts authenticated, replaceable observations: {schema:"whistlegraph-input/v1",sessionId,sequence,features}. The fixed feature schema contains speech-presence flags, pitch-contour category, capped attack count, rhythm regularity and energy category. It accepts no transcript, PCM, source, arbitrary prompt, question or provider URL. Speech-to-text and musical feature extraction share the phone's sample timeline; only the coding model receives the words and detailed measurements.

MusicalInputAdvisor starts during the hold, coalesces unchanged observations, allows at most three streaming calls plus one final call, and reuses a matching recommendation on release. Responses must match the session and sequence and meet the experimental 0.8 confidence threshold. Old, uncertain or failed responses are ignored. The client deadline is 900 ms; the provider deadline is one second. Advice is optional and cannot publish or certify a piece.

The server keeps provider credentials private and requires an AC account with a handle. Atomic Mongo counters limit this separate advice allowance to 30 requests/minute and 500/day per account, and 20,000/day globally. Failed calls consume their reservations. Only quota counters are persisted by the endpoint; observations are not saved. These limits count requests, not dollars or coding braincells. Phone receipts are saved only during explicit debug fixture runs.

Live results and failed attempts are retained under apple/whistlegraph/Tests/audio/. Decision latency is not an end-to-end creation speed claim: the existing coding model still generates the piece.

Persistent phone input #

Lith's existing DigitalOcean process serves wss://aesthetic.computer/api/easel-musical-stream. The phone opens it after sign-in and closes it while backgrounded. Authentication is the first JSON message ({type:"authenticate",token}), never a URL query; ready confirms the session. Sessions renew after ten minutes. A foreground client reconnects after disconnects, and HTTP remains available while cold. Once a decision has been sent, a disconnect does not replay it over HTTP.

{type:"observation",body:<whistlegraph-input/v1>} receives an immediate received acknowledgement followed by a decision or error, both with the original session and sequence. One decision runs per socket; only the latest waiting observation survives. {type:"cancel",sessionId} aborts active work and drops pending work for that capture. Heartbeats, bounded messages, two connections per account and a global connection cap bound the transport. Authentication is reused; durable per-decision quotas and the shared account concurrency guard still apply across HTTP and sockets.

Receipts separate acknowledgement round-trip (transportMs), server processing including quota checks (serverMs), and provider request duration (providerMs). Acknowledgement time includes scheduling and is not a pure network measurement. The upstream Decisions API remains HTTP. Neither this transport nor a separate VM removes provider inference latency. Measure these components before moving regions or provisioning another service.

Focused validation: node --test lith/musical-socket.test.mjs aesel/test/musical-jev.test.mjs. Phone validation: node apple/whistlegraph/Tests/musical-benchmark.mjs mixed-request --jev --socket.