fieldrich-vscode-extension

VS Code extension for FieldRich AI OS chat, repository questions, expert consultation, and software development tasks.

Images (1.7.2): drag images into chat, paste with Ctrl+V, or use + to attach PNG, JPEG or WebP files. Click a photo to view it at full size, zoom, and download it in the website or extension. Both interfaces restore saved chat photos and support removal and image-only messages. Chat experts and Agent mode use the local Qwen vision model. Image analysis requires the server image integration; see image setup and verification.

Open the FieldRich sidebar and select Agent to ask it to investigate, implement, fix, or test code in the first open workspace folder. It reads actual files, searches source, makes focused edits, runs commands, and feeds the observed results back to the AI OS until it finishes or reports a blocker. Agent mode does not require an index. Normal chat, repository chat, and Consult keep their existing API contracts.

Select Commands → Automatic (terminal access) for unattended command execution. The default, Ask before running, prompts before each shell command. Automatic execution runs with your user permissions; commands are not sandboxed and can access more than the workspace. The setting is saved at User scope, and repository settings cannot enable it. Trusted VS Code workspaces are required. For multiple workspace folders, the first folder is the execution root; open the desired folder first.

The model selector also applies to Agent mode. The selected model must be available on the configured AI OS. The default is qwen3.6:35b. Use Stop to cancel the outstanding AI request and the active command. Completed edits remain in place for review; Stop does not roll them back. Use managed process tools for temporary servers and long-running checks; readiness, logs and cleanup are tracked. Interactive commands are not supported.

Steer a running agent: keep typing in the message box while Agent is working, then press Enter or the send arrow. Your update shows Queued and then Applied to this run. For example: “Focus on the backend first” or “Use SQLite instead of in-memory storage.” The agent reconsiders any outstanding model decision with your latest instructions before its next action. An action already executing finishes normally; use Stop to cancel it. Shift+Enter adds a new line. Session/model changes remain locked while a run is active.

Live updates remain in context through compaction, are considered during final review, and are restored when you send continue in the same session/workspace. Later directions take precedence where they change earlier requirements. Rejected updates display an error and return to an empty composer for editing. Updates are limited to 4,000 characters each, 16,000 characters and 32 messages across a task and its continuations. When this limit is reached, start a new request with the current requirements.

Agent and chat transcripts are separate views of the selected session. Reload restores your selected session, Agent/Chat mode, model and unsent message. Agent conversations are retained in workspace state without the previous 30-run cutoff; the latest three summaries provide follow-up context. They are stored locally in this VS Code profile, while normal chat history remains on the AI OS. Full run logs and original file contents are stored under the extension's global storage, in agent-runs/<run-id>/. Show log opens the FieldRich Agent output channel and shows each run's recovery directory. files.json contains original contents for files changed by file tools; original: null means the file was newly created. Review these records before manually restoring individual files. Terminal commands can change additional files and are not covered by these file-tool checkpoints; review Source Control too.

On first access after activation, saved run results for the current workspace restore missing history, including older runs dropped by the previous cutoff. Runs also write reload checkpoints with the request, plan, accepted updates and recent activity. If reload interrupts a run, its checkpoint appears as interrupted; send continue to resume after inspecting current files. A reload does not automatically restart commands. Recovery is limited to files successfully saved before interruption. Agent history loads before remote chat/repository requests, so a failure in those requests does not hide it.

Deleting a session removes its agent transcript from workspace history and records the deletion so reload recovery does not reimport it. Recovery files remain in extension storage until manually removed.

File tools reject paths outside the workspace, symlinks, hard-link edits, protected credentials, binary/large files, and overwriting a file with unsaved editor changes. Existing files must be read before editing, and concurrent changes invalidate the edit. Root AGENTS.md is loaded automatically; reading a nested file supplies the applicable parent instructions, and new or changed instructions gate subsequent writes. Git ignore rules are used for discovery when available. Secrets and generated/dependency directories are excluded from discovery. Commands have separate terminal permissions and are not subject to file-tool restrictions.

Runs have an initial implementation budget (fieldrich.agent.maxSteps, default 80, maximum 200) and a separate review-repair allowance (fieldrich.agent.reviewRepairSteps, default 80, maximum 200). Pending review repairs receive blocks of up to 20 additional decisions while successful edits or checks show progress, bounded by the sum of those settings. Existing settings such as 40 initial steps are preserved and reported at startup. Set the repair allowance to 0 to keep the initial budget as the entire implementation/repair limit. There are three repair rounds and a final fourth review, each review limited to eight read-only decisions. Stop cancels any phase.

Earlier observations are compacted while preserving the original request, plan, repository instructions, pending review findings, source changes and check results. The context window (fieldrich.agent.contextTokens, default 32768), ten-minute inference deadline and maximum ten-minute command timeout remain configurable/bounded as before. A changed project requires observed successful verification after its latest edit. Install/check commands with outer pipelines, || fallbacks, unconditional separators or nested shells are rejected before execution; run the actual checker directly and use cwd for paths with spaces. Inspection commands can still use pipelines, but they do not count as verification. This checks outer command syntax, not the correctness of arbitrary scripts.

A command rejected before execution is recorded in the action log and does not become a failed build/test. Requesting a check still requires an actual successful check before completion; a rejection alone cannot satisfy verification. Correcting a refused workspace-root command to a direct build in frontend no longer leaves a false root-level failure. Checks that actually run and fail remain pending until a check of the same type passes in the same directory. Verification state also reports whether a successful check covers the latest edit.

Edits and dependency installs mark managed services in the affected directory stale. Old readiness output cannot verify new source: stop/restart the service and run assertions against the current implementation, or stop it and run in-process tests. A frontend build cannot clear a pending backend runtime check or a failed backend test. Failed checks remain pending until a passing check of the same type (build or behavior) runs in the same directory. The model must still select meaningful tests; the tool layer cannot prove their coverage.

After an interrupted run, send continue or resume in the same session and workspace. The agent retains the original goal, plan and pending findings, reads current files, and performs new checks. Older v1.2 findings are recovered from the extension's saved JSONL run log when available. Model quality and environment access still limit which tasks can be completed.

Malformed reviewer responses, network failures and temporary server errors receive up to two retries within the same review, preserving its request and evidence. If no valid verdict is available, the result explicitly says Independent review unavailable and retains the implementation report and genuine pending findings. It does not turn reviewer errors into code defects or request source repairs for them. Completion still requires a valid approval. Sending continue starts a fresh run with current-file checks and another review.

Build and verify:

npm ci
npm run typecheck
npm run test:agent
npm run package

An optional live check creates a disposable calculator project and a temporary backend session, asks the actual model to fix the code and run its test, then deletes the temporary session. It permits only the fixture's test command and retains the disposable project for inspection:

npm run test:agent:live -- http://YOUR-AI-OS/fieldrich-api qwen3.6:35b

Create a VSIX using the existing VSCE workflow (install/deploy.bat on Windows), or package it with VSCE and --no-dependencies after building. The extension and React sidebar are bundled. Version 1.3.0 uses the same AI OS protocol as 1.2.0; upgrading a working 1.2 installation requires only the new VSIX and Developer: Reload Window. If upgrading from before 1.2, the backend must provide /api/ai/agent/decision. This release requires no additional backend or web-chat deployment.

Additional real-model regressions exercise Python/C projects and the original React + Node.js JWT request from an empty folder:

npm run test:agent:languages -- API_URL qwen3.6:35b
npm run test:agent:fullstack -- API_URL qwen3.6:35b
npm run test:agent:workflow -- API_URL qwen3.6:35b

Architecture and activation instructions are in docs/agent-implementation.md. Current release checks and limits are in docs/agent-v1.5.1-validation.md; earlier reports are retained for 1.5, 1.4, 1.3 and 1.2.

Version 1.6.0 adds free internet search for agents and shared backend model modes, and includes the connection reliability improvements from 1.5.2. Install fieldrich-vscode-extension-1.6.0.vsix and run Developer: Reload Window. The backend search integration must also be installed and the API restarted; see web search setup.

Agent inference now uses native HTTP with the full ten-minute deadline and a fresh socket per request, avoiding fetch's separate five-minute header timeout and stale keep-alive connections. Network interruptions and temporary HTTP 429, 500, 502, 503 or 504 responses receive at most two retries with the same task and evidence, without consuming action steps or rerunning tools. Model schema errors still use protocol correction. Stop and live steering interrupt retry waits. Connection failures include the method, endpoint, underlying error code and relevant troubleshooting advice. Settings → Test connection has a fifteen-second deadline, checks JSON, and leaves the saved/active URL unchanged when testing an unsaved address.

The COMSOL server's actual backend directory is /opt/FieldRich-AI-OS. Its logs exposed RuntimeError: cannot release un-acquired lock during session activation: FastAPI can enter and exit a synchronous yield dependency on different worker threads, while RLock requires the same thread. docs/server-session-lock.patch replaces it with Lock, preserving serialization. The patch and regression tests are saved on that server; the prior dependency file is backed up at /home/advent/repo-backups/20261022-fetch-failures/dependencies.py.before. Run sudo systemctl restart launch_api.service on the server to activate that backend fix. Restart requires interactive administrator authentication. The extension transport update itself uses the existing protocol v2.

Run npm run test:connection -- http://100.123.32.10/fieldrich-api qwen3.6:35b to check session access and a real model decision without executing tools or creating sessions. The server's nginx API route already sets proxy_read_timeout 86400s; no nginx change is needed for the observed configuration. The lock regression reproduced the logged exception before the fix; 20 targeted backend checks pass after it. All 50 extension tests pass against a separate source copy on the server with Node 22. A live check through nginx returned HTTP 200 for sessions and a valid decision from qwen3.6:35b in 30.9 seconds. The Windows sandbox permits typechecking but blocks the local test/build tooling. Actual Windows VS Code interaction and a five-minute delayed response are not covered by those automated checks.

The current extension ID is fieldrichsoftware.fieldrich-vscode-extension. Some older installs use fieldrich.fieldrich-vscode-extension. install/deploy.bat removes the older ID after successfully installing the current extension to prevent duplicate toolbar buttons and views. When installing a VSIX manually, uninstall or disable the older ID before reloading. The extension reads available saved run records from the older identity's adjacent storage directory for the same workspace, validates them, and preserves the original files. Do not delete the older extension's storage before migration. Select the relevant session and Agent mode to see recovered conversations. Records from a different machine/profile/storage location are not automatically imported. Show log displays the active version, extension ID and storage location, plus recovery diagnostics.