Configuration
Settings live in .enigma.json (global ~/.enigma.json, overridden by a repo-local one). Set them with enigma config <key> <value>, the interactive menu (enigma config), or the dashboard. Add -g / -l to target global or local.
| Key | Values | Default | What it does |
|---|---|---|---|
commit-emoji | on / off | on | Leading gitmoji on commit subjects, the agent’s and the gate’s. |
output-style | off / lite / full / ultra | off | Compress the agent’s chat prose (memory-file section; restart the agent). Enforced, not just asked for: the verify Stop hook scans the final reply for filler, a preamble, and status rows that pad a table, and denies the stop on the ones it can catch precisely. |
minimal-code | off / lite / full / ultra | full | Anti-overengineering intensity for code the agent writes. |
permission-bypass | on / off | on | Skip the agent’s approval prompts (a deliberate security downgrade). |
claude-trust | on / off | on | Pre-answer Claude Code’s “Is this a project you trust?” prompt for any path, so it stops asking in every new folder (and in your home directory, where Claude never remembers the answer). Also a deliberate security downgrade: that prompt is what makes you look at unfamiliar code before an agent runs in it. Turning it off keeps the folders you already trusted. |
resource-cap | 1 - 100 | 60 | Share of the machine work an agent runs may take: cores, build concurrency, and a lowered scheduling priority so your own work stays responsive. Bulk file work also drops to idle priority. |
low-memory-cap | 1 - 100 | 80 | On a machine with 16 GB of RAM or less, the memory use above which an agent stops launching heavy work instead of throttling it; the cap above is also halved there. Global-only: it describes the machine, not a repo. |
auto-sync | on / off | on | Refresh deployed skills when launching a tool through enigma. |
statusline | on / off | on | Show the enigma badge, model, context and cost in Claude Code’s status bar, plus live gate progress during a run. |
statusline-refresh | <0-60> | 10 | Seconds between status bar refreshes. Every tick starts a new process, so raise it where that is slow; 0 removes the timer and refreshes only when the conversation moves (the gate line then stops animating during a run). |
remote-skills | on / off | on | Fetch skills newer than the bundled set from GitHub. |
skill-update-policy | overwrite / keep | overwrite | On update, replace a skill you edited locally, or keep your edits. |
usage-api | on / off | off | Fetch the real usage %/reset from Anthropic via your local Claude Code login. |
parallel-subagents | on / off | off | Allow the agent to fan out sub-agents in parallel. |
subagent-limit | 0 - 20 / off | 4 | How many subagents one Claude Code session can run at once. A launch past the limit is refused and the agent is told to wait or do the work itself; a subagent can never launch its own. 0 turns subagents off. Set it per project with -l. |
relay | on / off | on | Hand work across a cleared context: a handoff saved with /handoff reaches the next session in the project once (Claude Code, Codex, Kimi Code, OpenCode). |
relay-at | 50k - 1000k / off | 300k | Context size at which Claude Code saves a handoff, clears and continues on its own. off keeps only the manual /handoff. |
compact-window | 100k - 1000k / auto | 400k | Context size at which Claude Code compacts a conversation and its subagents. Every call re-reads the whole context, so a smaller window spends less per call. auto leaves Claude Code’s own value. |
auto-lint | on / off | off | Lint edited files via a post-write hook. |
guardrails | on / off | on | Enforce project conventions (UUID keys, Prisma as default ORM, …) via a post-edit hook that feeds violations back to the agent, plus a turn-end sweep of the lines the change added for the rules that only apply to new code. Set guardrails: false in a repo’s .enigma.json to stand that sweep down there. enigma guardrails stats reports which rules keep being broken. |
trim | on / off | on | Remove the blank line agents leave at the end of a file, after each edit and on the staged files at commit time. Only a file with content followed by blank lines is touched. |
verify | on / off | on | Check a “this is finished” claim against what the turn actually produced, and deny the stop when the evidence contradicts it. The same hook sweeps the added lines for guardrail violations, claim or no claim; checks the reply itself against the output-style level; and - where the quality gate is on - denies a stop that reports the gate as skipped or leaves committed work no gate run has seen. Global; to skip one project put verify: false in its .enigma.json. |
verify-command | command | (none) | Command the completion gate runs before a “done” claim is allowed through (e.g. npm test). Global-only, and carried by neither a config export nor an import. |
compress | on / off | off | Register the context-compression MCP server in your agents. |
code-graph | on / off | on | Index your codebase as a graph your agents query instead of re-exploring it: find the code a task concerns, trace a change’s blast radius, read a file’s API surface. Also rides along in a Claude Code session - orientation at start, locators per prompt, blast radius after an edit. |
browser | on / off | on | Register the Chrome DevTools MCP server in your agents, so they can open the app in a real (isolated) Chrome to click through it and read the console, network, load-speed trace and Lighthouse audit before calling a UI or speed change done. Isolated profile, never your logins; needs Chrome and Node; toggling applies immediately. |
dashboard | off / on-demand / always | off | Local management dashboard mode. |
dashboard-port | <0-65535> | 0 | Preferred dashboard port (0 = auto: 80, then 24282). Restart the dashboard to apply. |
dashboard-bind | loopback / lan / custom | loopback | Which interface the dashboard listens on. Anything but loopback is reachable from the network and refuses to start without a token. Restart to apply. |
dashboard-bind-address | <ip> | (unset) | Interface to bind when dashboard-bind is custom (e.g. a Tailscale IP). |
api-port | <1-65535> | 8000 | Default port for the local OpenAI-compatible API server (enigma api). A –port flag overrides it per run. |
api-rotation | off / round-robin / least-used / fill-first / random | off | How enigma api spreads requests that name no account across your accounts. Any strategy but off also retries a request that hits a usage limit on the next account. –rotation overrides it per run. |
api-account-pool | add / remove <account> | (empty) | Accounts the rotation may use, in order. Empty = every account of the backend. |
api-client-context | on / off | on | Let a request pick its own account, profile or pack. Off = enigma always decides and refuses a request that names one. |
api-account / api-profile / api-pack | <name> / none | (unset) | Context a request runs under when it names none and rotation is off. Unset = the active account. |
proxy | on / off | off | Route enigma claude through a local measuring proxy. |
prompt-secret-guard | on / off | off | Scan outgoing prompts for secrets (via the proxy). |
prompt-secret-mode | redact / reject | redact | What the guard does on a detected secret. |
usage-stats | on / off | off | Read your Claude Code transcripts for real token usage and cost. |
recall | on / off | off | Build a searchable local memory of your coding sessions from transcripts; exposed to agents over MCP. |
recall-llm | on / off | on | Have an LLM curate recall in the background: keep the observations worth remembering, drop the trivial ones. Needs recall on; with no login it falls back to the deterministic heuristic. |
recall-provider | claude-local / anthropic / openai | claude-local | Which model curates recall. claude-local reuses your Claude Code login; the others need an API key. |
recall-model | model id | (none) | Model for recall curation; blank uses the provider’s default. |
recall-api-base | URL | (none) | Base URL for the anthropic or openai-compatible provider; blank uses its default. |
recall-api-key | key | (none) | API key for the anthropic or openai provider, stored encrypted. ENIGMA_RECALL_API_KEY overrides it. |
gate | on / off | on | The experimental AI quality gate (the /gate command and enigma gate). Agents auto-drive it once a code task is committed, on whatever branch it is on - no need to ask or run enigma gate init. Turn it off per project with gate: false in that repo’s .enigma.json (or /gate off from your agent), everywhere with enigma config gate off -g. See Quality gate. |
gate-protected-branches | <branch names> | (none) | Branches the gate refuses to validate, matched exactly (add/remove one at a time). Empty means every branch is fair game, the default branch included - and there the pipeline opens no PR and pushes straight to the branch, so list it here when it must only be reached through review. |
gate-trivial-lines | 0 - 1000 | 20 | Changed lines at or under which a commit skips the gate entirely: no review round and no PR, so a one-line style fix is pushed as is. Changes to auth, dependencies, CI, migrations, env files or binaries always go through, whatever their size. 0 sends every change through the pipeline. |
gate-severity | error / warning / info | warning | Lowest finding severity that stops a gate run to ask you what to do. Findings below it are still found and reported, the run just does not wait on them - and every pause you answer with a fix costs the pipeline a full re-review of the diff, so this is the knob for a run that keeps finding one more nice-to-have instead of finishing. |
ci-watch | on / off | on | Watch the GitHub Actions run your push triggers and tell the agent when it breaks, with the failing log attached, so you do not have to relay it. Needs the gh CLI. A background process does the waiting, a green build says nothing, and the agent is never blocked. See CI failure notifier. |
claude-attribution | on / off | off | Let Claude Code credit itself on commits and PRs (Co-Authored-By, PR footer, session link). |
claude-survey | on / off | off | Let Claude Code show its “How is Claude doing?” session survey. |
kimi-trust | on / off | on | Pre-answer Kimi Code’s “Trust this folder?” for the folder you launch in, so its project MCP servers load. Folders already trusted stay trusted when you turn it off. |
search-guard | on / off | on | Refuse a find across the whole disk (from /, your home or a drive root) - the searches nobody notices are still running. |
self-audit | on / off | on | When the agent reports work finished, send it back to review its own change as a hostile reviewer (at most 3 rounds per prompt). Needs verify. |
logo-color-policy | ask / adapt-background / adapt-logo | ask | How a logo/background contrast clash is fixed. Logos are always sourced, never invented; restart the agent to apply. |
gh-telemetry | enabled / disabled | disabled | GitHub CLI usage analytics. |
token-price | <usd> | 0 | $ per 1M input tokens for dashboard money estimates (0 = per-source defaults). |
token-speed | <tok/s> | 0 | Prefill speed for dashboard time estimates (0 = default). |
enigma config compress on
enigma config minimal-code ultra
enigma config dashboard always/improve
A slash command deployed to every agent (it runs inside the agent, not the CLI). Two modes: implement edits a focused area, advisor is read-only and writes self-contained plans for another agent to execute.
/improve ui # implement: ui | security | performance | seo | refactor
/improve audit # advisor (read-only): audit | quick | deep | branch | next | plan ...