Quick path, after jev setup:
jev harness install --target command-code --apply # jev MCP server + /jev-advice skill
jev hook config command-code # optional pre-execution shell guard (merge into settings.json)The Command Code integration installs a focused, explicitly invoked /jev-advice skill and the jev MCP entry. It guides the agent to read approved saved logs through jev_read_evidence before those logs enter model context. Installation and off reads make no Jev provider requests. Shadow scoring and qualified selection require separate operator opt-in; no startup or post-tool hook runs inference automatically.
Use an installed Python with jev-decision[mcp]; -I requires the package to be installed into that exact interpreter. Start with user scope and a zero budget. This keeps machine-specific interpreter/state paths in your own profile. Replace the three absolute paths with your own:
$jevPython = 'C:/tools/jev/Scripts/python.exe'
$jevState = 'C:/Users/you/AppData/Local/JevDecision'
$projectRoot = 'C:/work/example'
& $jevPython -I -m jev_decision.cli --runtime-home $jevState setup --non-interactive --credential-source env --key-env TYPESAFE_API_KEY --workspace $projectRoot --daily-budget 0 --selection-mode off --harness command-code --scope user
& $jevPython -I -m jev_decision.cli --runtime-home $jevState harness install --target command-code --scope user --dry-run
& $jevPython -I -m jev_decision.cli --runtime-home $jevState harness install --target command-code --scope user --applyOn macOS/Linux, use /absolute/venv/bin/python and ordinary shell invocation without PowerShell's &. For a reviewed project integration, use --scope project --project-root $projectRoot on installation and restoration. Reconcile generated absolute paths before sharing project configuration with teammates.
Detection recognizes command-code on every platform, cmdc on native Windows and cmd on POSIX/WSL. Windows cmd.exe is never detected as Command Code. An executable on PATH establishes detection only; connection and actual invocation remain separate. Command Code executable names
| Scope | MCP entry | Managed skill |
|---|---|---|
| User | ~/.commandcode/mcp.json → mcpServers.jev |
~/.commandcode/skills/jev-advice/SKILL.md |
| Project | <project>/.mcp.json → mcpServers.jev |
<project>/.commandcode/skills/jev-advice/SKILL.md |
Command Code's private local scope can override project and user MCP entries. Project .mcp.json may also be consumed by another client, so review the preview before applying. The installer preserves unrelated entries and refuses to adopt an unmanaged jev entry. These locations and precedence are documented in Command Code MCP.
If the shared project's jev entry is already managed for Claude Code, Command Code installation/restoration reports shared_client_ownership_conflict; the reverse order is also protected. Use user scope for independent client configurations, or have the operator reconcile a shared entry. An external edit to an owned entry remains a conflict rather than being silently restored.
The generated entry binds an absolute interpreter and JEV_HOME. An environment credential uses ${TYPESAFE_API_KEY:-} (or your selected variable name), never its value. The empty fallback allows credential-free off reads. Supply the real variable in the client launch environment only when enabling provider use, or choose the supported OS vault in jev setup. Use a dedicated credential name; JEV_HOME, JEV_ENDPOINT_URL, JEV_OFFLINE_MODE, JEV_HOOK and JEV_HOOK_THRESHOLD are reserved runtime variables. The installed skill's CLI command also binds the absolute runtime home.
To restore the user integration above, preview first, then apply:
& $jevPython -I -m jev_decision.cli --runtime-home $jevState harness restore --target command-code --scope user --dry-run
& $jevPython -I -m jev_decision.cli --runtime-home $jevState harness restore --target command-code --scope user --applyRestoration uses saved ownership records, retains unrelated edits, and reports conflicts instead of overwriting changed managed content. It does not erase the runtime, credentials, budget ledger, or captured evidence. For a project integration, replace --scope user with --scope project --project-root $projectRoot in both restore commands.
Reload the client, inspect /mcp for the jev connection, and inspect /skills for jev-advice. Complete any normal workspace trust or tool approval prompts. The skill sets disable-model-invocation: true; invoke it explicitly instead of expecting automatic discovery from the model's skill catalog. Command Code skills
If a built-in command or another skill has the same short name, use /skill:jev-advice to select the skill explicitly. Project skills take precedence over user skills; check /skills for the active source before using it. Skill selection priority
For an existing saved log, type a path as ordinary text, without attaching the entire file:
/jev-advice C:/work/example/.jev-captures/run-001/stderr.log Explain the first failed test. Use off mode; do not rerun the producer.
The skill directs the agent to call mcp__jev__jev_read_evidence with a bounded page. The equivalent direct CLI call is:
& $jevPython -I -m jev_decision.cli --runtime-home $jevState evidence --file 'C:/work/example/.jev-captures/run-001/stderr.log' --goal 'Explain the first failed test' --mode off --max-lines 200 --jsonFor a command that has not run, use the installed capture subcommand through the ordinary shell tool, under the same permissions as the producer. No repository checkout is needed. It executes a native argv command and saves stdout, stderr, exit status, sizes and hashes. This synthetic example demonstrates a failed producer while returning only a capture reference:
& $jevPython -I -m jev_decision.cli capture --directory 'C:/work/example/.jev-captures/run-001' -- $jevPython -X utf8 -c 'import sys; print("collected 3 tests"); print("FAILED test_export", file=sys.stderr); sys.exit(7)'Use a new capture directory for each run. Retain the helper's exit status 7 and capture.json; the helper does not reinterpret success. Read both saved streams when relevant. Never pipe the original log through the agent and then claim later scoring saved its context tokens. For output above the evidence reader's file limit, retain the original and produce bounded, provenance-preserving chunks before reading; a truncated excerpt is not a complete run.
On Windows, capture rejects .cmd/.bat wrappers, including wrappers resolved through PATH, because the operating system can introduce implicit command-shell parsing. Use the underlying executable, such as node.exe --test or node.exe /absolute/tool-entry.js. Running a batch file through an explicitly supplied interpreter requires the same operator authorization as that shell command. Python subprocess security considerations
Evidence pages require UTF-8. Capture preserves bytes even when a producer emits a different encoding. A CP1252/UTF-16/UTF-32 log gets an encoding-specific diagnostic. Configure the producer for UTF-8, or produce a separate UTF-8 derivative using its documented encoding; retain the original and hash, and use the derivative's own hash when reading it. Never replace undecodable bytes or claim a converted hash identifies the original artifact.
For a Command Code agent using Engraphis alongside Jev, recall authorized context through Engraphis first and keep its workspace/session routing. The memory-system recipe supplies bounded relationship, relevance and verification-gap advice through jev_decide; durable memory operations still use Engraphis' governed tools. Installing Jev preserves existing Engraphis MCP entries.
off is the initial policy. To run an explicitly authorized shadow trial, first choose a nonzero daily cap and save the mode. For example, the operator can repeat setup with --daily-budget 0.25 --selection-mode shadow after approving that cap and configuring credentials. A later evidence call may request --mode shadow; it retains the entire sanitized page while collecting scoring metadata. Per-call options cannot upgrade a saved off policy.
select requires a reviewed profile configured by the operator and qualification evidence. Pass an independently established workload identity with the real harness, harness_version, primary_model, and primary_provider; use MCP's workload object or CLI --workload /absolute/workload.json. Do not fill these fields by copying a profile. No Command Code selection profile ships, and no saved policy is upgraded by this skill.
Keep source_sha256 and page metadata. Retrieve an omitted or later range with --mode off --start-line N --max-lines M --expected-source-sha256 HASH. Read errors, contradictions, exit status, and any required unread tail before concluding. A redacted or selected page is advisory evidence, never authorization or a replacement for executed verification. See evidence behavior for fallback and recovery details.
Command Code runs PreToolUse hooks from ~/.commandcode/settings.json or .commandcode/settings.json. jev hook config command-code prints matcher ^(shell|powershell)$, covering shell_command and the native Windows powershell tool through their case-insensitive display names. Merge the fragment into the existing hooks object and reload.
Command Code hooks can allow or deny but cannot ask. So in bypass and dont-ask sessions, where nobody reviews commands, the guard denies a command Jev flags as destructive or sensitive and gives the agent the reason. In other modes it stays silent, so your approval prompt keeps the decision. Add --when always to the hook command to deny flagged commands in every mode. The guard never approves anything, skips plain read-only commands without a request, and on any local failure produces no decision. JEV_HOOK=off disables it. Details and the other harnesses are in HOOKS.md.
The current mod contract exposes afterToolCall, which can replace the result before it is committed to model context. That is a plausible future adapter, but a generic tool result does not establish the immutable source, approved upload scope, matching workload identity, and qualified omission policy this runtime requires. Mods are unsandboxed trusted code and their API is experimental. This integration instead exposes an explicit saved-artifact workflow through existing tools; it adds no permission grants, automatic retries, or event hooks.
On 2026-09-30 the published command-code 1.72.4 package (dist/cli.mjs) was inspected read-only for the hook runner. A hookSpecificOutput.permissionDecision of deny blocks the tool. Matchers compile with new RegExp(pattern, "i") against display names. Plan mode skips tool hooks. shell_command input is passed through unchanged as command, args and cwd. MCP locations, ${NAME:-} expansion and disable-model-invocation are unchanged from 1.66.0.
The locally installed npm package reported command-code 1.66.0 in its package.json. Read-only inspection of its bundled skill/MCP/mod references and dist/cli.mjs confirmed the manual-only skill field and ${NAME:-} stdio environment expansion. The package's skill catalog and MCP paths match the current official documentation above.
package.jsonSHA-256:59ff1b0414e415611a6318f1f38c80aabeeb70aa50feb781d02a503531ae3f8a.dist/cli.mjsSHA-256:adf05d4e64358d631e4627e7cfaee23ac7903690ca44fe5b0e715f8b814b95ca.
Automated checks use temporary profiles, synthetic capture subprocesses, and the local JSON CLI. They establish generated configuration, preservation/restore behavior, and an offline capture-to-evidence path. They do not establish a Command Code UI invocation, provider authentication, live mod behavior, or token/time savings. Record those separately for the actual client version and workload before advertising them.