JevSeek v0.2.3 — skip unnecessary argument generation

JevSeek v0.2.3 — schema-determined argument bypass

What's changed

• Skip unnecessary generation: after Jev selects a tool, code supplies its arguments directly when the entire argument object is fixed by a conservative schema subset. This removes a Flash/Bonsai generation call without adding another Jev call.
• Supported fixed cases include closed empty objects, required constants/singleton enums/nulls and closed required nested objects. Optional, free-form, reference or unsupported schemas retain the original generator and original context. No requested parameter is silently replaced with a default.
• Usage/activity records distinguish a skipped generator call from actual provider usage. JEV_ARGUMENT_OFFLOAD=off restores the original always-generate behavior.
• The complete public research archive is included in the source download: successes, negative results, limits, cancelled work and incomplete prototypes. Unvalidated neural argument filling and other non-winning paths remain research-only, not enabled in the app.
• Build metadata records SHA-256 hashes of bundled application source files. Offline EXE checks now verify both the fixed-argument bypass and unchanged free-form fallback, with model requests blocked.

Update

Close the old app and replace it with JevSeek.exe. Existing keys, preferences, saved conversations and local Bonsai assets remain in the normal personal data directory. Review uncertain/interrupted tool effects before resuming; an update does not replay them automatically.

Windows x64, unsigned. Check SHA256SUMS.txt. The app bundles Python and Fixed Version WebView2; Python/Node are not required to launch it, but tools needed by your workspace projects are not bundled. Do not disable SmartScreen/Defender. The Microsoft runtime terms and privacy notice still apply.

Scope and verification

This is a narrow, deterministic cost optimization—not general intelligence, automatic neural argument offloading, or a claimed whole-agent saving percentage. The previously published34%/85%estimated API savings concern a separate bounded decision endpoint; they are not v0.2.3coding-session savings.

The release gate covers offline Python/frontend tests, the production build, byte/secret/source audits, and an actual EXE isolated from the checkout with Windows-only PATH: native file/shell tools, temporary vault roundtrip/removal, bundled browser/native bridge, and MCP helper. No paid model calls, local model inference or benchmark VM launches are required. Test reports remain local; hashes identify the published assets. Windows11 is the tested host, not certification of every Windows version or a fresh VM.

Research ledger · Backend contract · Security

JevSeek v0.2.2 — context fix and archive recall

JevSeek v0.2.2 — context overflow fix and intent-relevant archive recall

Fixed

• Fixes the repeatable Context accounting reserve exhausted failure: archive-count metadata is now included before fitting the context budget.
• Safely removes optional duplicate/history rows from the working projection when needed, without deleting their original saved records. Latest execution/status and user instructions remain protected.
• The actual reported session was projected read-only to reproduce the old bookkeeping error and confirm the corrected projection fits; its saved log was unchanged and no actions were replayed.

Added: searchable original memory

• Private, session-local SQLite FTS5 archive of original user messages, assistant conversation and tool results, including paged full saved outputs.
• A small intent-relevant history tier alongside recent results. Jev can select recall to search for missing evidence or page a precise event sequence. Activity shows Searching saved history.
• Original provenance and status retained. Old source is not assumed current, assistant text is not execution evidence, and recall never changes workspace files or clears interrupted-effect uncertainty.

Research published alongside the implementation

Intent-relevant memory investigation covers MemGPT, Lost in the Middle and LongMemEval; the overflow diagnosis; a reproducible synthetic retrieval study; token/timing measurements; and limitations.

On the intentionally old-evidence fixture, recent-8 found 0/10 old facts, lexical top-4 found 9/10, and Jev chose the correct record or abstained on 12/12 questions from a lexical/diversity candidate pool. These are different metrics on a small constructed example, not proof of general accuracy or infinite context. The study's extra semantic reranking call remains research; production uses local retrieval and Jev's existing routing to request recall. No hidden extra memory-provider calls were added.

Update and resume

Close the old app and replace it with the attached JevSeek.exe. Reopen the saved conversation and choose Resume for a blocked run, or send a follow-up. Existing keys and history remain in place. Review any interrupted effects before acknowledging them; updates do not automatically replay work. Prior Bonsai and v0.2.1 follow-up fixes remain included.

Windows x64, unsigned. Verify SHA256SUMS.txt. Python/Node are bundled for the app, not for arbitrary workspace projects. Microsoft WebView2 runtime terms/privacy still apply; do not disable Windows security protections.

Honest boundaries

This is a finite working window over a growing disk archive, not a literally infinite context window. All accumulated user instructions remain pinned; if mandatory intent/schema/metadata itself cannot fit, the app can still pause safely rather than silently discard constraints. Search may miss synonyms or relevant pages; archives consume storage and logs still load into RAM. A damaged optional cache does not block ordinary context construction and can be removed with the app closed to rebuild it.

Verification for this update: scoped local retrieval/paging measurements, a 12-call Jev-only synthetic research run, read-only overflow reproduction, production build/static syntax and source/package privacy audits. No broad test suite, native UI smoke, generation-model task or Terminal-Bench sweep was run for this update. No VM was restarted. Private user traces and cancelled benchmark work are not in the release.

JevSeek v0.2.1 — follow-up execution hotfix

JevSeek v0.2.1 — follow-up and premature-completion hotfix

Fixed

• Preserves the previous assistant response as explicitly labeled conversation, so “proceed” can refer to offered work. Assistant claims are never treated as execution evidence.
• Reviews a proposed done against the current request and recorded effects. When work remains, Jev selects a relevant tool instead of immediately emitting an inactivity summary.
• Tells the response generator not to invent an inability to use tools just because none ran. Ordinary conversational questions can still finish without tools.
• Shows completion review in live activity and records its API usage.

Included in this release

This packages the existing local v0.2.0 Bonsai implementation plus the hotfix, hence v0.2.1 rather than v0.1.1. Settings → Generation model offers the official Prism or community CRACK Bonsai model, with automatic downloads, progress, integrity checks and DeepSeek switchback. Bonsai replaces DeepSeek generation; Jev still requires a key and receives routing context. Model files download separately (~7.2 GB each). Existing sessions and keys are preserved. JevSeek's MIT license notice is included in the EXE.

Updating

Close JevSeek, replace the old executable with JevSeek.exe, and reopen it. Use the new EXE—not an older shortcut pointing elsewhere. Your saved history and credentials remain in their existing Windows locations. To continue the affected conversation, send a new follow-up message; simply reopening history deliberately does not execute work.

Windows x64, unsigned. Python/Node are not needed to run the app; project-specific developer tools are separate. Verify the download against SHA256SUMS.txt. Do not disable Windows security protections. Microsoft WebView2 terms/privacy apply. See the repository's release and security documentation.

Scope and verification

The failure was diagnosed from actual session events: immediate done selections with no tool calls, combined with missing prior assistant context. This release was compiled and its source and decompressed package audited for private data. No new test suite, live-model check or benchmark was run for this hotfix, at the owner's request. Earlier Bonsai/runtime validation predates this change. Completion review is model-based, adds Jev latency/API cost, and is not a guarantee that every model decision will be correct. Historical research results describe the earlier routing policy.

The cancelled Terminal-Bench run, VM, private logs, API keys and local sessions are not included or restarted.

JevSeek v0.1.0 — Windows Desktop

JevSeek v0.1.0

The first standalone Windows release of JevSeek: a desktop coding agent where Jev selects the next action, DeepSeek supplies its arguments, and native tools execute it. Actual results inform the next decision.

Download and run

1. Download JevSeek.exe below (about 393 MiB).
2. Launch it, review the included Microsoft runtime terms, and enter your own DeepSeek and TypeSafe/Jev API keys.
3. Choose a trusted workspace and send a task.

Python, the React frontend/fonts, and Microsoft Fixed Version WebView2 are bundled. No separate Python, Node or WebView2 installation is needed to run the app. First launch extracts the runtime; allow roughly 1.5 GB of temporary disk space. Project-specific developer tools and external MCP servers are not bundled.

Included

• Chat, local conversation history, live tool activity and read-only session artifacts.
• Native file read/write/edit and PowerShell execution; optional trusted MCP connections.
• Persistent sessions, factual context compaction and explicit interrupted-tool review.
• Light/dark/system themes and keyboard navigation.
• In-app API-key setup using Windows Credential Manager; saved keys are not exported to tool subprocess environments.

Important

• Windows x64 only. Unsigned build: Windows may warn. Verify the download against SHA256SUMS.txt; do not disable Defender or SmartScreen. Windows 11 was tested; a fresh Windows VM has not been certified. Normal Windows PowerShell and .NET Framework 4.8 components are required.
• Not a sandbox. Tools run with your account's privileges. Use trusted projects/servers and do not run as administrator.
• API-backed inference: your keys and provider credits are required. Selected task context goes to providers/enabled MCP services.
• Microsoft WebView2 includes Defender SmartScreen, which collects and sends end-user information to Microsoft under its Privacy Statement and Edge Privacy Whitepaper. The app presents the runtime terms. The bundled Fixed Runtime does not auto-update; future JevSeek releases must supply runtime security updates.
• Personal state lives under %LOCALAPPDATA%\JevSeek. No developer keys, sessions, machine MCP configuration or private workspace are included.

Verification

This exact executable passed isolated checks with Python/Node removed from PATH: native vault save/read/remove, SDK file tools, PowerShell, bundled WebView2 UI/bridge and the packaged MCP helper. These were offline integration checks, not a new live-provider benchmark. Its decompressed contents were scanned for known local secrets and private paths with no findings.

Assets

• JevSeek.exe — standalone Windows application.
• JevSeek-source.zip — audited source-only export, including research documentation and reproduction code.
• SHA256SUMS.txt — SHA-256 checksums for both files.

See the README, research breakdown, security documentation, and third-party notices. An overall license for JevSeek's own source has not been selected; dependencies retain their respective licenses.