UsageRun recipes
Run recipes
Run recipes: a Cyberine usage recipe.
- Command map for an orchestrating session — cyberine guide
- --help on any subcommand is safe (kz4pg)
- Plugin CLI cheatsheet — plugin/references/cli-cheatsheet.md
- First-time setup — cyberine setup
- Project-level setup — cyberine init
- Smart bare cyberine
- 0. Global cyberine command (cross-directory run monitor, à la claude agents)
- Failover preflight — cost before a run spends anything (cyberine route, cyberine run)
- Wait for / digest a background run — cyberine wait, cyberine digest
- Recurring scheduled dispatches — cyberine cron
- Batch-integrate finished runs — cyberine land
- See what's runnable + dispatch by id — cyberine next
- JSON reads of the board — status --json and next --json (ka7fx)
- List + switch models — cyberine models, cyberine model
- Brainstorm tasks interactively — cyberine chat
- Task workflow inside cyberine chat — shared manifest with the nf-tasks skill
- 1. Multi-adapter demo (same task -> every available adapter, compare)
- 2. One-off task from the CLI
- 3. Switch the OpenCode model / reasoning effort (config only)
- 4. Switch the OpenRouter loop model (config only)
- 5. Full gateway path (Anthropic / OpenAI / any provider via LiteLLM)
- Global credentials (cyberine from any directory)
- API-key providers
- OAuth providers
- MiniMax mmx — text worker + image/speech tools
- Vision / multimodal tasks
- DeepSeek request isolation — user_id
- The opencode2 worker (OpenCode 2.0 preview)
- The muse worker (Muse Code, Meta)
- Muse via the raw Meta Model API (loop-muse-build)
- Spawned CLI workers and PWD
- Retiring a worker — RETIRED_WORKERS
- The web UI — cyberine web (browser fleet monitor + chat)
- 6. Workflows + live fleet view (the orchestration layer)
- Meter tracks belong to their own meter (kw1e9)
- The sparkline colours by value (kefc7)
- The accent budget — two claims per screen (kpbk7)
- Divider and border weight (k6shr)
- The bold budget — one bold per region (km2w9)
- The border vocabulary is corner-less, by decision (kwk4f)
- One dot separator, no raw colour literal, no doubled hue (k1rse)
- The wide hub: a 140-column breakpoint that splits list and preview (kxxph)
- The streaming label pulses, and its copy escalates (kk92a)
- Meters render at sub-cell resolution (k9y4k)
- Code fences use four tones, not eight (kptd7)
- A judged winner looks like one (ksdrd)
- The fuzzy picker highlights what it matched (kenqx)
- The stall indicator stays inside the palette (kfapa)
- One state grammar across every pane (kctsa)
- The chat REPL is a bottom-anchored column (kntve)
- Diffs are filled, not coloured (k7pc0)
- Wide terminals — the hub measures rows, not the screen (k09p3)
- Keybindings — remap fleet + chat actions (~/.cyberine/keybindings.json, T535)
- Fleet list — TASK-first columns, filter, scrolling
- Fleet list — narrow terminals
- Fleet list — quick reply (Space)
- Agents view — one unified list of chats (c)
- Fleet list — collapsible folders ([+]/[-]), quick-sum, hide filter
- Run detail — task panel, prompt, verdict, diff, race timing
- Fleet list — RATE column
- TUI — flicker-free rendering (synchronized output)
- TUI/chat — local notifications (T534)
- TUI/chat — update notifier (T533)
- TUI — worker toast notifications
- TUI end-to-end cases — drive chat and the fleet hub through a real terminal (npm run tui:e2e)
- Workflow Monitor — live phase/agent drill-down (w)
- Stopping a run or worker
- Foreground run wall-clock timeout — cyberine run --timeout
- Live progress output — streaming per-worker status
- Auto-routed implementer fleet on the foreground run
- cyberine run refuses a repository with no commit yet (k4zpd)
- Foreground cyberine run with a build fleet runs through the engine (isolated worktrees)
- Judge backend fallback and unjudged runs
- Stale / zombie runs
- CLI reaper — cyberine reap [--dry-run]
- Always-on operations — graceful drain, health, queue intake, maintenance
- Standby worker pool (opt-in)
- Steering a running loop worker (between tool rounds)
- 7. Implementer fleet: do a real task (multiple providers compete, judge picks the best diff)
- Which verdict wins a step (k0fxw)
- Judge phase is fail-bounded (T551)
- AdversarialJudge rationale carries the skeptics' reasons (kxqep)
- Auto-evolve: Cyberine authors and implements its own task (mine -> implement -> gate)
- LSP/compiler diagnostic feedback loop (T436)
- Verify-error steering — test-runner output as navigation index (T447)
- Long-term memory — episodic records, retrieval + injection (waves 1-4)
- Rules — project + global conventions (like Claude Code rules)
- Ecosystem instruction files (AGENTS.md / CLAUDE.md)
- LSP worker context (lsp_query tool, T312)
- Persona — chat planner voice (CYBERINE_PERSONA)
- Output styles — /output-style persona overlay (T539)
- Confluence ring — welcome-screen brand mark
- Skills — Claude-Code-compatible SKILL.md
- cyberine mcp install — register the cyberine MCP server with a harness
- Backlog tasks — what's pending beyond run-memory
- Claude Code memory adapter (CYBERINE_CC_MEMORY) — opt-in
- Shared memory backend — ContextQ (or any similar server)
- ContextBrain — Code-Context 3-Act adapter (LocalBrain / ContextQBrain, T441/T454)
- Project groups — link dirs to share runs / chats / memory
- Expose Cyberine as an MCP server — npm run mcp
- Consume external MCP servers as tools — McpToolAdapter
- Oversized tool results — spill-to-disk + retrieval (spill_retrieve)
- History compaction spills oversized tool results instead of truncating to 80 chars
- Malformed tool calls — validate-then-repair on the loop tool-input path (kmk7x)
- Edit pacing — a loop implementer writes before its round budget runs out
- Repeat-tool nudge — advisory reminder when a worker repeats the exact same call
- Durable worker transcripts — cyberine transcript
- Capability-aware routing — cyberine route
- Smart (LLM-as-router) for low-confidence tasks (T61)
- Router eval — compare routing strategies on the corpus (T66)
- Model strength table and fleet.json override (kyhae)
- Fleet bench — solve rate and dollars per task (kvxah)
- LLM candidate re-ranker — Act 1 SCOPE filter (T446)
- Per-task model variant selection (T60)
- Machine-readable route recommendation — cyberine route --json (T62)
- Routing-explain — why a worker was picked or skipped (kc0nz)
- OpenTelemetry GenAI export — CYBERINE_OTEL_ENDPOINT (kyv0t)
- Swarm mode — cyberine run --mode swarm (plan -> parallel-chunk implement -> merge)
- Lane grouping — when lanes help vs flat fan-out
- Operating modes — --mode compete|swarm (+ CYBERINE_MODE env)
- Research mode — cyberine research (plan -> search -> verify -> synthesize)
- Website rebuild target (Next.js + Fumadocs) — site/
- Documentation site: VitePress retired (T670)
- Fleet policy file — cyberine fleet (T422, T425, T426)
- Unified health check — cyberine doctor (T540)
- Adapter liveness — is it FUNDED, not just keyed (--liveness, kwrp9)
- Selftest sweep with a real SKIP channel — npm run sweep (T707)
- Security-gate drift check — npm run selftest:gate-list-drift (T707, retargeted 2026-09-07)
- Sweep stamp and pre-push hook — .githooks/pre-push (2026-09-07, gate removed 2026-09-30)
- Sweep flake ledger — npm run sweep:flakes (kdpxk, 2026-09-28)
- Dependency audit gate — npm run selftest:dep-audit (T673, multi-tree since T707)
- Credit ledger — track provider balances, gate low-credit workers
- TUI stdout shim — synchronized output, and how to smoke it
- Task footprint estimator — history-driven budget planning (kx9rt, calibrated in k5m89)
- Hard spend cap — opt-in, race-free ceiling across ALL providers
- Quota cooldown — automatic circuit-breaker after 402/429
- Blocker visibility — waitingFor on stalled workers
- Provider circuit breaker and bounded retry — one dead provider degrades the fleet, not the run
- Fallback model chain — per-worker model degradation
- CLI flag-compat diagnosis + structured-output knobs (T656)
- Claude cold-spawn session-persistence suppression (T627)
- Claude prompt answering — --permission-prompts none (k7hha)
- A non-success result subtype is a failed run (kq5rd)
- Where a worker's CLAUDE_CODE_* settings came from — toolingSnapshot.workerEnv (kg4pt)
- Which Claude account a worker bills — oauthToken (k8rhz)
- Background fleet worker — claude --bg (k56eq)
- Idle-stream watchdog — abort a stalled stream, fleet worker + chat (T562, T570)
- Worker-internal subagents — delegate_subtask (T313)
- Tool-call pipeline — batch multi-tool sequences in one model turn (T319)
- Result projection — configurable worker output fidelity (T328)
- Stateful session reuse — warm CLI-agent bindings + session tree ingestion (T333)
- Pluggable execution backends — run workers in Docker or on a remote host (T332)
Configuration file
Cyberine supports layered JSON config files as an alternative to .env files. Precedence (highest first):
Command map for an orchestrating session — cyberine guide
cyberine guide prints an intent-to-command cheatsheet for an ORCHESTRATING agent session (any harness with a shell, not only Claude Code) — the recipe to run first in a fresh session before...