HEXRA by ValerieAI

HEXRA 3.0 adds project intelligence, visible task plans, coordinated file edits, and verification repairs.

Version 3.0.1 adds Valerie's gold logo to the Marketplace listing and extension details. The existing toolbar icon is unchanged.

New in 3.0

• Build / Plan only: plan mode exposes inspection tools and cannot edit files or run commands. Execution Off / Ask / Auto remains a separate control.
• Project intelligence: map source files and entrypoints, read real VS Code diagnostics, and trace document symbols, definitions, and references. Symbol support depends on the language extensions installed in VS Code.
• Task board: live progress for multi-step tasks, saved with conversation history. Long tasks condense older observations for the provider while keeping the full local transcript.
• Coordinated edits: review up to 12 files together, validate the whole change before applying, and undo them with one checkpoint. Stale diffs and unsaved buffers are refused. Undo preserves subsequent user edits.
• Verification repair: a failed automatic check is fed back to Valerie for up to two repair passes. The completion card reports actual checks and incomplete work accurately.
• Quick actions: understand a project, fix workspace errors, review the active file, or generate focused tests. These prepare an editable request before sending.
• Resilient streams: CRLF support, complete-event validation, and no tool execution from interrupted responses. Both the panel and terminal client use bounded conversation context.

The default limit is 24 model/tool rounds per panel task, configurable from 4 to 40 with valerie.maxToolRounds. The existing server allowance still applies. Send Continue to resume unfinished work. These tools require the matching Valerie backend upgrade; installing only the extension against an older backend leaves new tools unavailable.

Install vscode-extension/hexra-3.0.1-valerieAI.vsix using Extensions → … → Install from VSIX, then reload VS Code. The publisher ID is valerieAI and the extension identity is valerieAI.valerie-ask; the displayed title remains HEXRA by ValerieAI. The publisher ID must match the ID registered in Visual Studio Marketplace, not just its display name.

If the older valerieos.valerie-ask extension is installed, disable it before using this package so two copies do not register the same commands or claim Cowork tasks. Sign in again: changing the publisher creates a different extension identity, so its private sign-in and history storage are separate. This release does not migrate or delete the older extension's stored data. Checkpoints from before 3.0 lack content hashes and cannot be automatically undone by the panel; their backups remain on disk for manual recovery.

Mobile Cowork

HEXRA retains authenticated mobile Cowork: while VS Code is open, signed in, and has a project folder, Valerie Mobile can hand one task to the extension and receive its bounded completion summary. HEXRA still enforces the selected Off / Ask / Auto execution permission and its protected-file and sandbox boundaries.

Minimal coding workspace (2.7.1)

Run HEXRA: Open Coding Workspace, click the gold mark in the editor toolbar, or press Ctrl+Alt+Shift+A. HEXRA opens as a dedicated right-side editor panel with Valerie as the coding agent.

The composer uses one compact permission rail:

PLAN · daily usage ring · Execution: Off / Ask / Auto

The gold ring fills clockwise from the server's real shared daily token usage. It refreshes after every billed turn; typing does not affect it. Admin accounts show a complete gold ring because their allowance is unlimited.

• Off removes command execution from Valerie's tools and hard-blocks any attempted command.
• Ask shows the exact command and waits for Run or Deny.
• Auto proactively runs every shell command needed while diagnosing, fixing, building, and testing, without another confirmation. Persistent source edits still use HEXRA's edit tools because each command workspace is disposable.

Execution permission never disables the sandbox. Ordinary commands run in a hidden, disposable Docker container with no network, no host credentials, no Linux capabilities, a read-only container filesystem, bounded CPU, memory, process count, output, and time. HEXRA stages a filtered workspace copy and does not mount the real project writable. .env, .env.*, .gitignore, .sql, .sqlite, .sqlite3, and .db files are excluded before the container starts.

HEXRA never pulls a container image silently. It automatically selects an official Node, Python, Go, Rust, or Debian image that must already be installed. Advanced users can select another preinstalled image with valerie.sandboxImage. If Docker or the image is unavailable, the command is refused and the final report says verification was unavailable.

After an editing turn, HEXRA performs a dependency-free syntax/build check when one can be identified. Ask mode requests approval for that check; Auto runs it; Off reports that it was not run. The final card shows files changed, token use, execution mode, and the checks that actually passed, failed, were denied, or could not start.

Existing production features remain available:

• Browser sign-in to ValerieAI and the account's shared daily token allowance.
• Streaming responses, file attachments, active-editor context, and @file.
• Native side-by-side edit diffs with Apply or Reject.
• Conversation history, new chat, undo checkpoints, and admin self-update.
• /graphify . through the locally signed-in Codex CLI. Off blocks it, Ask asks, and Auto starts it under Codex's own workspace sandbox.

Account and provider routing

Sign in with HEXRA: Sign in to ValerieAI or open the panel and approve the browser authorization. Plus, Pro, and MAX accounts use their configured public OpenRouter coding model and shared ValerieAI daily allowance. Admin uses the private server-selected provider. Provider keys never enter VS Code. The editor credential is stored in VS Code SecretStorage and can be revoked server-side.

Terminal profile

The legacy HEXRA Agent terminal profile remains available for compatibility. The new Off/Ask/Auto sandbox guarantee applies to commands issued from the HEXRA panel; terminal commands remain an explicit interactive terminal workflow.