Repository navigation
Releases
V-Agent 1.1.7
Release fixing API connections and updates, reading every statistic from the engine itself, translating the rest of the interface and correcting the GPU memory reported on Windows.
- GPU memory on Windows (
agent/hardware_stats.py, issue #1):Win32_VideoController.AdapterRAMis a 32-bit value, so every card with 4 GB or more was reported as 4 GB (a 12 GB Radeon RX 6700 XT read 4095 MiB), shrinking model-fit labels and automatic sizing. The size now comes from the driver's 64-bitHardwareInformation.qwMemorySize, or from DXGI when the driver does not write it; WMI is kept only when nothing larger is reported, so integrated GPUs keep their shared memory. Thesourcefield records which one was used. - Statistics from the engine, never estimated (
agent/engine_manager.py,agent/desktop_runtime.py): llama.cpp sends its own timings on every chunk and the prompt read's progress (timings_per_token,return_progress); Strata Rust is read from its/api/stats(live tokens, and each request's exact read and write times); Project Maya from its/metrics. Strata (PolyStrata) now also shows the prompt read's progress. Prompt sizes are no longer estimated from character counts, and streamed chunks are no longer counted as tokens. With an API, the provider's own counts are used (Claude reports the input at the start), the output speed is measured as received, and the wait before the first token is no longer shown as a reading speed. - Exact tool cost: The token cost of the enabled tools (Settings) is counted by the engine or provider (llama.cpp
/apply-template+/tokenize, Maya/tokenize, Strata Rust and Claudecount_tokens, Strata's own tokenizer and template). It is marked as estimated only on OpenAI-compatible APIs, which have no counting endpoint. - "Use model" with an API did nothing: A KV-cache setting left over from a local engine was sent with the API model, the core refused it and the button failed silently. API models now ignore local-only settings, and a failed save is reported.
- Updates with an API model: Engine and app updates no longer ask to unload an API model (only a model loaded in local memory blocks them). Installing an engine right after a restart checks for the release first instead of failing.
- OpenCode plans (
agent/remote_protocols.py, connection form): Separate presets for OpenCode Go / Go Plus (one endpoint; the plan comes from the key) and OpenCode Zen (account balance; its free models only work inside the OpenCode app). The currenthttps://opencode.ai/inferenceendpoints are supported with their per-protocol paths, and OpenCode always usesAuthorization: Bearer. - Translations (
agent/ui/locales.json,agent/ui/standalone/i18n.js): Over 300 interface, status and error messages from the core added in English, Portuguese, French and German, including update notices. Messages with values (paths, counts, models) are translated through templates, and "prefix: detail" and multi-sentence messages part by part. The language audit now also fails on Spanish that is missing from the catalog.
V-Agent 1.1.6
V-Agent 1.1.6
Release making the terminal tools usable by any model without a harness-specific fine-tune: Linux habits are explained instead of failing silently, output is clean, background jobs and servers behave, and nothing changes what a working program does.
Changes & Updates Since v1.1.5
Clean, non-interactive console (agent/shell_hygiene.py, agent/pc_mcp.py, agent/sandbox_exec.py)
- Environment: every command runs without colours, prompts or pagers (
NO_COLOR,GIT_TERMINAL_PROMPT=0,GCM_INTERACTIVE=never,PIP_NO_INPUT…) and with unbuffered Python output, so background jobs show progress while they run.CI,PYTHONUTF8andPAGERare deliberately not set: they change what users' programs do. - Output: escape codes removed, carriage-return progress bars collapsed to their final state, padding stripped, repeated lines summarised, PowerShell CLIXML errors made readable, the console's old code page and UTF-16 decoded line by line, binary output replaced by one line.
- Several lines actually run:
cmd /cran only the first line of a multi-line command and reported success; such commands now run as a batch script with the last exit code.
Linux habits explained
- Notes, even on success: single quotes,
$VAR,;,/dev/null,-xoptions on cmd built-ins,%%iin batchforloops, andfind/timeoutonly when this PC really runs the Windows programs of that name. - Hints on failure:
ls/cat/grep/rm/export…,./script,python3,mkdir -p,NAME=value cmd, prompt signs and markdown fences, PowerShell&&andcurl, missing keyboard input, Git credentials and timeouts.
Background jobs and servers
- No silent blocking: a trailing
&ornohupis refused at once (it blocked the call until its time limit) with the way to run it in the background. job_wait until: returns as soon as a server prints its ready line, or when the job goes quiet; jobs keep their latest output in a rolling buffer and each check returns only what is new; the last lines are no longer lost when a job ends.- Long output: beginning and end shown; the whole output (up to 4 MB) saved to
outputs/forfs_read.
Tool description the model actually receives (agent/mcp_prompt.py)
- It now states the cmd quoting, variables and separators, the configured time limit and how to run servers. The description in
pc_mcp.pywas always replaced by an older one.
Engine cleanup (agent/engine_updates.py)
- Old engine folders briefly locked by antivirus right after an update are retried instead of left until the next start.
Installer & verification
- Clean-worktree NSIS installer
V-Agent_1.1.6_x64-setup.exe(260,472,086bytes, SHA-256fac8e46431c915341e5e3e6991b8c5a4339c33e77075f105feaf4fc3abe1d402), verified free of personal data or development artifacts, with Ed25519-signedv-agent-update.jsonandv-agent-update.json.sig. - Tested: full Python
unittestsuite (389tests passed, including30console cases with real commands and background jobs), ruff (0findings), MCP recovery (30/30), profile × access × permission matrix (19/19), the packaged core's shell tool (multi-line scripts, cleaned output, trailing&refused, notes), and a small real model onllama.cppcompleting console tasks that previously failed or stalled for over 10 minutes.
V-Agent 1.1.5
V-Agent 1.1.5
Release aligning provider error handling with the open-source agent harnesses (OpenCode, Goose, Roo Code, Cline, Aider), so a busy provider no longer ends an autonomous run. Includes the 1.1.4 C1 fixes.
Changes & Updates Since v1.1.4
Retries as the open-source harnesses do them
- Busy or throttled providers (
agent/agent_resilience.py,agent/agent.py): HTTP 408/409/425/429, 5xx, 529 "Overloaded" and dropped streams are retried 10 times with exponential waits of 2, 4, 8, 16, then 30 s and ±20 % jitter (about 3.5 minutes), never shorter than the provider'sRetry-After/retry-after-ms(capped at 15 min). Previously 5 attempts over about 2 minutes, ignoringRetry-After. - No network path: DNS or route failures keep retrying for up to an hour (OpenCode's unreachable window); timed-out attempts are retried at most 3 times.
- Never retried: exhausted balance or quota (402,
insufficient_quota/ balance on 429), authentication and other rejected requests stop at once with a clear message. A 429 worded like a context overflow is treated as a throttle. - Empty replies: a reply with no text, no reasoning and no tool call is repeated (twice at most), as Cline and Roo Code do.
Provider errors and output limit (agent/engine_manager.py)
- Status kept:
ProviderHTTPErrorcarries the HTTP status and the requested wait; new plain-language hints for 402 (no balance left), 503 and 529 (provider overloaded). - Output limit as OpenCode computes it: the catalog's output maximum, 32,000 when the provider does not declare one, never above 256,000, fitted to the room the prompt leaves in the window (15 % estimate error, 1 % headroom, minimum 1,024).
Interface
- Translated provider errors (
agent/ui/standalone/i18n.js): "Provider error (N): …. Detail: …" is translated part by part intoen,pt,frandde; retry cards also cover no-connection and empty-reply cases (1,945phrases). - From 1.1.4 C1: Render HTML works on pages with a
<canvas>again (3D/voxel scene diagnostics reach the model); the composer has a double edge with a clear gap.
Installer & verification
- Clean-worktree NSIS installer
V-Agent_1.1.5_x64-setup.exe(260,456,886bytes, SHA-2567d99c03797b02dcefa7dce0b57f7dbf97798a1b8da0f78f5875e8cd55266246d), verified free of personal data or development artifacts, with Ed25519-signedv-agent-update.jsonandv-agent-update.json.sig. - Tested: full Python
unittestsuite (362tests passed: error classes, schedule, jitter,Retry-Afterfrom a real 529 response, output limits), all8Node.cjsUI suites, ruff (0findings),tests/smoke/agent_autonomy_ui.py(14/14, adding an empty reply and an exhausted balance),tests/smoke/ui_language_audit.py(no Spanish left in English), andscripts/clean_release_build.py(Clean release verification passed).
V-Agent 1.1.4
V-Agent 1.1.4
Fix release: API models no longer run out of output inside their reasoning, a reasoning-only round is followed by action instead of an empty reply, and the remaining untranslated interface text is translated.
Changes & Updates Since v1.1.3
Output maximum for API models
- "No limit" means the model's own maximum (
agent/engine_manager.py,agent/remote_protocols.py): With no output limit set, V-Agent used to omitmax_tokens, so providers applied their own small default and long reasoning used it all, ending the reply empty. Requests now carry the model's declared output maximum (also read frommax_completion_tokens,top_providerandlimits.output), or 65,536 when the provider does not declare one, kept within the window left by the prompt. - Safe fallback: A provider that refuses the value gets the same request again without it, remembered per model, so this never breaks a reply. A limit set by the user is sent unchanged. How much the model thinks is still set only by the thinking effort.
Reasoning-only rounds
- Recovery (
agent/agent.py): A round that ends with reasoning but no reply and no tool call is asked to act (up to twice) instead of ending the turn. - Clear message (
agent/ui/app/composer.js): If nothing comes, a translated message says the model used up its output while reasoning (or finished reasoning without an answer) instead of "the model finished without returning text".
Translations
agent/ui/app/chat.js,agent/ui/app/composer.js,agent/ui/app/shell.js,agent/ui/v-agent.html: "Reasoning…", the live-arguments label, truncated-result notes and the screen-reader names of the page regions are now translated intoen,pt,frandde(1,928phrases).
Installer & verification
- Clean-worktree NSIS installer
V-Agent_1.1.4_x64-setup.exe(260,448,639bytes, SHA-2563730477801edbb6c4ae0ad8bfda320d0032f267561780175aa9bc6844be9d3d6), verified free of personal data or development artifacts, with Ed25519-signedv-agent-update.jsonandv-agent-update.json.sig. - Tested: full Python
unittestsuite (358tests passed, including the output maximum and its fallback intests/test_remote_providers.py), all8Node.cjsUI suites, ruff (0findings),tests/smoke/agent_autonomy_ui.py(12/12, adding the reasoning-only round and the output maximum sent to the provider), the newtests/smoke/ui_language_audit.py(no Spanish left on screen through every live state of a reply inen,pt,frandde), andscripts/clean_release_build.py(Clean release verification passed).
V-Agent 1.1.3
V-Agent 1.1.3
Release making long autonomous runs survive provider errors and full context windows, letting the agent ask the user mid-task, and strengthening the code checker.
Changes & Updates Since v1.1.2
Ask the user mid-task
- New
agent__ask_usertool (agent/ask_user.py,agent/agent.py,agent/ui/app/shell.js,agent/ui/app/chat.js): Available to every engine — local models and API providers — when the request comes from the interface. The model asks one short question with up to 4 options; the user picks one, types an answer, or lets the agent decide. - Bounded: At most 3 questions per turn. After 15 minutes without an answer the agent continues on its own judgment and states its assumptions. Requests from external API clients never receive the tool, so nothing waits for an answer nobody can give.
- Languages: Tool instructions are in English; the dialog and cards are translated into
en,pt,frandde.
Long autonomous runs that keep going
- Automatic retries (
agent/agent_resilience.py,agent/agent.py): Transient engine/provider failures in the middle of a turn (HTTP 408/429/5xx/529, dropped connections, replies cut before their end mark) repeat the round up to 3 times (2, 6 and 15 s), shown as an "Automatic retry" card. Authentication, bad-request and missing-model errors still stop the turn. - Context kept within the window: Past 60 % of the model's window, old tool batches are archived to the chat's
context/folder and old outputs are shortened. When the engine still refuses a prompt as too long, the history is shortened further and the round repeated. Shown as a "Context trimmed" card.
Stronger code_check
- Python (
agent/code_insight.py,agent/verify.py):pyflakesstatic analysis reports undefined names, locals used before assignment and unused imports; imported modules missing from the user's Python are flagged. - HTML, CSS, JavaScript: Missing local scripts, stylesheets and frames are errors; duplicate ids, script lookups of ids the page lacks, missing
url(...)files and relativeimport/requiretargets that do not exist are reported. - On write (
agent/pc_mcp.py): The same analysis runs when a file is saved, so the model sees the problem immediately.
Dependencies
pyflakes 3.4.0(MIT) bundled.
Installer & verification
- Clean-worktree NSIS installer
V-Agent_1.1.3_x64-setup.exe(260,444,915bytes, SHA-25698c0ea8078919c8e7ec7ca7f510a850d8f15d9d8aa70d248c3f8ca2a7427053e), verified free of personal data or development artifacts, with Ed25519-signedv-agent-update.jsonandv-agent-update.json.sig. - Tested: full Python
unittestsuite (357tests passed, including newtests/test_agent_resilience.pyandtests/test_code_insight.py), all8Node.cjsUI suites, ruff (0findings),tests/smoke/agent_autonomy_ui.py(real interface against a scripted local provider: question with options, "let the agent decide", a mid-turn 503, a context overflow —10/10), MCP recovery (30/30), profile × access × permission matrix (19/19), a real local model onllama.cpp(20/20, including an automatic context trim on an 8K window, and a realagent__ask_userround trip), the packaged core'scode_check, andscripts/clean_release_build.py(Clean release verification passed).
V-Agent 1.1.2
V-Agent 1.1.2
Fix release: buttons that did nothing when clicked now act, or say why they cannot.
Changes & Updates Since v1.1.1
"Use model" and "Disconnect API" with providers
- The visible button acts directly (
agent/ui/app/model-controls.js): "Use model" / "Stop using" now load or release the model themselves instead of pressing a hidden button whose enabled state was computed elsewhere and could stay disabled, swallowing the click without a word. - A stuck local load no longer blocks API models (
agent/ui/app/composer.js,agent/ui/app/controls.js): Allama.cpp/Strataload still marked as "preparing" in the core (stuck or left over, for up to 15 minutes) no longer blocks provider models: "Use model", "Disconnect API" and sending messages work. A leftover "preparing" state is checked with the core and cleared.
No more silent clicks
- Busy-time explanations (
agent/ui/app/shell.js,agent/ui/app/settings.js,agent/ui/app/events.js,agent/ui/app/load-settings.js): During a reply or an engine operation, file access, approvals, "Add files", "Refresh", "Reconnect MCP", "Add model" and the draft model's "Choose…" now say why they cannot act instead of ignoring the click. - Model list refresh (
agent/ui/app/models.js): After a download or import, the model list refreshes even while a reply is running. - Translations (
agent/ui/locales.json): The new messages are translated intoen,pt,frandde(1,905phrases).
Release build
scripts/clean_release_build.py:vcvars64.batis now called with intact quotes whencl.exeis not already onPATH.
Installer & verification
- Clean-worktree NSIS installer
V-Agent_1.1.2_x64-setup.exe(260,301,717bytes, SHA-2562867bd44121e2aab6ed1b360450f962e65f196072845553c0af67aec36d63de3), verified free of personal data or development artifacts, with Ed25519-signedv-agent-update.jsonandv-agent-update.json.sig. - Tested: new
tests/smoke/api_use_model_ui.py(real interface against a local fake provider; fails on 1.1.1 in all five reported cases, passes on 1.1.2), all8Node.cjsUI suites, full Pythonunittestsuite (338tests passed), ruff (0findings) andscripts/clean_release_build.py(Clean release verification passed).
V-Agent 1.1.4 C1
V-Agent 1.1.4 C1
Minor update for V-Agent 1.1.4, delivered through the signed components channel (no installer; applied on the next restart).
- 3D pages (
render_check.py): Render HTML works again on pages with a<canvas>. Since 1.0.8 its summary failed on the canvas size, so Three.js / voxel scene diagnostics (empty instance groups, instances outside the camera, invalid matrices, moving parts intersecting blocks) never reached the model. Verified on the packaged 1.1.4 core: fails without C1, works with it. - Composer (
ui-patch.css): Double edge with a clear gap — the surface's own border plus an outer line 5 px outside it that follows the chosen radius and takes a hint of the accent color while typing. The deco style keeps its own framed edge.
Signed with the maintainer's Ed25519 key (v-agent-components.json and v-agent-components.json.sig).
V-Agent 1.1.1
V-Agent 1.1.1
Release introducing optional Project Maya integration, custom pre-compiled build selection across all four local engines, per-model engine load options, an intermediate installer progress window on app updates, clean engine uninstall, and llama.cpp update reliability fixes.
Changes & Updates Since v1.1.0
Optional Project Maya Integration (mw00/project-maya)
- Local Maya engine lifecycle (
agent/maya_engine.py,agent/engine_manager.py,agent/thinking_policy.py): Added optional support forProject Mayawith Windows CUDA 13.2+ / Visual Studio 2026 and AMD ROCm prerequisite diagnostics, official GitHub tag auto-updating, and per-family configuration (GLM-5.3-Flash,GLM-5.3 744B,DeepSeek-V4.1-Flash-1017,MiniMax-M3.1-Flash,Qwen-3.8-Flash-Next). - Never forced: Project Maya remains completely optional and is never pushed or required in first-run onboarding, engine recommendations, or update notifications when not installed.
Custom Pre-Compiled Builds for All 4 Local Engines
- Unified custom build support (
agent/engine_manager.py,agent/maya_engine.py,agent/runtime_paths.py,agent/model_import.py,agent/kv_settings.py,agent/ui/app/engine-controls.js): Users who already have custom builds forStrata · PolyStrata,Strata Rust,llama.cpp, orProject Mayacan pick their executable or checkout folder directly without forced downloads or recompilation. - Persistence & compatibility: Custom paths and detected build metadata survive switches between official and custom variants, companion DLL directories (
exe.parentandlib/) are prepended toPATHon launch, andkv_settings/model_importrespect active custom Strata executables.
Per-Model Engine Load Options
- Dynamic engine flag catalog (
agent/engine_options.py,agent/ui/app/load-settings.js): Added a--help-filtered catalog of engine-specific load flags forllama.cpp,Strata, andStrata Rustwith conflict validation (e.g., quantized KV cache requiring Flash Attention) and per-model persistence.
Intermediate Installer Progress Window
- Zero-blank-screen update handoff (
agent/app_updates.py,agent/ui/app/tour.js,agent/ui/styles/panels.css,agent/desktop_runtime.py): Clicking Install and restart now displays an in-app transition overlay followed by a compact external WinForms progress window (UPDATE_HELPER, defaulting to English and localized acrossen,es,pt,fr, andde) until V-Agent relaunches.
Updater Reliability, Engine Uninstall & Storage Cleanup
llama.cppupdate install fix (agent/engine_updates.py): Fixed updates stalling at 100% downloaded without installing by retrying transient WindowsWinError 32file locks and resetting progress/busy state cleanly.- Engine uninstall & storage pruning (
agent/engine_updates.py,agent/maya_engine.py,agent/components.py,agent/app_updates.py): Added clean removal of managed engines (and safe unlinking of external Project Maya folders without deleting user files), plus automatic cleanup of superseded minor component revisions and already-installed or broken installers.
5-Language UI & Clean Fast-Extraction Packaging
- Translations (
agent/ui/locales.json,agent/ui/standalone/i18n.js): Expanded coverage to1,899translated phrases acrosses,en,pt,fr, andde. - Installer & verification: Clean-worktree NSIS
zlibfast-extraction installer (V-Agent_1.1.1_x64-setup.exe,260,516,288bytes, SHA-256adc62a28a930da6f42efb4f5e848a991735767b0f4f033b2a76df8dd0a58e733) verified free of personal data or development artifacts, with Ed25519-signedv-agent-update.jsonandv-agent-update.json.sig. - Tested: Verified
scripts/build-ui.py(1,899phrases), all8Node.cjsUI test suites (test_palette.cjs,test_permission_ui.cjs,test_response_metrics.cjs,test_ui_i18n.cjs,test_ui_motion.cjs,test_ui_rich_stream.cjs,test_update_notice.cjs,test_v2_ui_features.cjs), full Pythonunittestsuite (338tests passed), andscripts/clean_release_build.py(Clean release verification passed).
V-Agent 1.1.0
V-Agent 1.1.0
Release focused on cleaner tool-execution UI, automatic minor updates without manual intervention, optional local engine setup on fresh installs, hardened remote API compatibility with per-connection UI controls, and clearer MCP parameter diagnostics.
Interface & Tool Trace
- Collapsible tool trace (
agent/ui/app/chat.js,panels.css): Tool executions inside assistant responses are grouped under<details class="tool-trace-group">(collapsed by default) with a compact summary bar showing running step / error counts and individually expandable<details class="tool">entries for inspecting arguments and outputs on demand. - Complete 5-language translations (
agent/ui/locales.json): Added full translations (es,en,pt,fr,de) across1,478phrases including the onboarding tour, collapsible tool trace, optional engine step, and every new API connection/header control, placeholder, and validation message.
Automatic Minor Updates & Optional Local Engines
- Hands-free minor updates (
agent/desktop_runtime.py,agent/components.py,agent/ui/app/core-connection.js): Startup component checks run shortly after launch (1.5 s) with progressive backoff retries (20 s–300 s), fetch up to100releases per page with strict signed-manifest checks, hot-reload UI patches into the active window, and firev-agent-components-updateevents automatically without requiring manual update clicks. - Optional local engine onboarding (
agent/ui/app/tour.js,shell.js,engine_updates.py,engine_manager.py): First-run onboarding step 2 allows choosingllama.cpp,Strata, orStrata Rust, switching directly to an external API, or skipping engine installation cleanly; update banners are suppressed for engines that are not installed, and missing-engine errors distinguishStrata Rustfromllama.cpp.
Hardened Remote API Protocols & UI Controls
- Protocol & header hardening (
agent/remote_protocols.py,agent/engine_manager.py,agent/agent.py,agent/ui/app/engine-controls.js):- Sends global
User-Agent: V-Agent/1.1.0by default (with per-connection override presets/custom value), stripsX-Stainless-*SDK headers, and rejects CRLF/non-Latin-1 header values. - Supports provider presets and auth modes (
auto,openai,anthropicx-api-key+anthropic-version+anthropic-beta,azureapi-key+api-version+ classic/openai/deployments/<dep>404 fallback,opencodeper-conversation persistent UUID v4x-opencode-sessionacross turns/restarts/compaction, andopenrouterHTTP-Referer/X-Title/delta.reasoning/ SSE comments). - Enforces SSE anti-buffering headers (
Accept: text/event-stream,Cache-Control: no-cache,Accept-Encoding: identity) on streams with transparentgzip/deflatefallback decompression. - Exposes provider preset, auth mode, User-Agent, session header mode, Anthropic beta/version, Azure API version,
HTTP-Referer,X-Title, and custom headers cleanly per connection in the UI menu.
- Sends global
MCP Diagnostics
pcMCP parameter errors & recovery (agent/pc_mcp.py,tests/smoke/mcp_recovery.py): Reports missing requiredKeyErrorparameter names clearly to the model so calls can recover cleanly on retry.
Installer & Verification
- Clean-worktree NSIS
zlibinstaller (V-Agent_1.1.0_x64-setup.exe,260,348,539bytes, SHA-256ae455714d889e86e858e5f2c96078b1e686ace00d55bb7ccd8fc7323ec84961a) verified free of personal data or development artifacts, with Ed25519-signedv-agent-update.jsonandv-agent-update.json.sig. - Tested: Verified
scripts/build-ui.py(1,478translated phrases),static_analysis.py(no stray control chars across271text files;ruff/bandit/ESLintskipped as not installed), all8Node.cjsUI test suites (includingtest_ui_i18n.cjsandtest_v2_ui_features.cjsafter the final UI translation commit),api_edge_cases.py(19checks,0failed),xss_model_card.py,language_checks.py(45checks,0failed), andClean release verification passed(clean_release_build.py). On Pythonunittest(308tests),307passed and1test (test_pc_hf_web_admin_computer_tools_end_to_end) hit a local port19380collision because an existing desktop instance ofV-Agentwas running on the build host.
V-Agent 1.0.9 C4 (Minor Update)
V-Agent 1.0.9 C4 (Minor Update)
- Safe
outputs/chunk reads across all profiles: capsfs_readchunks onoutputs/dumps at 3,400 characters in bothoffset/max_charsandlinea/ventanamodes so Light mode (4,000result_chars) never overflows into nestedoutputs/dumps. - Clearer
pc__fs_readguidance: confirms chunkedoutputs/reading directly in the tool description. - Includes all C1–C3 fixes: autonomous turn and draft recovery across all models and profiles, unshortened MCP instructions/token budgeting, fast Light-mode draft views, compact draft banners, browser/web fixes, and localized release-note notifications.