A command center for sharper Pi sessions. Pi-Rogue adds the missing control layer around coding agents: route work explicitly, ask for strategic review at the right time, preserve exact context without flooding prompts, and keep long work anchored to explicit goals.
pi install npm:@fiale-plus/pi-rogueOne public package. One install. Explicit commands only.
Supported host: Pi >=0.80.6 <0.81.0 on Node.js >=22.19.0.
Pi-Rogue is built for agent sessions that should compound instead of drift:
- Spend intelligence where it matters — use router/advisor signals to escalate hard turns and stay lightweight for routine work.
- Keep evidence retrievable — large tool outputs become compact handles, not prompt sludge.
- Make autonomy visible — goals, loops, and autoresearch are explicit, inspectable, and stoppable.
Start with /pi-rogue for the cockpit, then jump into the subsystem you need.
| Subsystem | Command | What it gives you |
|---|---|---|
| Pi-Rogue | /pi-rogue |
The cockpit: status, health checks, and pointers to every subsystem. Start here. |
| Advisor | /pi-rogue-advisor |
Low-friction strategic review. A tiny local trained gate learns from examples when to escalate, then calls a strong advisor model only when the turn deserves it. |
| Router | /pi-rogue-router |
Capability-aware routing telemetry and opt-in model switching. It is the multi-model layer: use whatever model fits the current turn or sequence instead of pretending one model is always best. |
| Goal / loop / autoresearch | /pi-rogue-orchestration (or /goal, /loop, /autoresearch) |
Visible session orchestration: define success, run periodic work, or start solo/parallel research flows without hidden budgets or background mystery. |
| Context broker | /pi-rogue-context |
Bounded context memory for tool outputs, diffs, snapshots, subagent results, advisor briefs, and memory notes. Prompts stay small; exact evidence stays one lookup away. |
/pi-rogue status # cockpit: health and command pointers
/pi-rogue-advisor <question> # ask for strategic guidance
/pi-rogue-router status # inspect route telemetry and mode
/pi-rogue-orchestration goal set <goal> # anchor long-running work
/goal set <goal> # shortcut for goal set
/pi-rogue-orchestration loop 5m <task> # run an explicit periodic loop
/loop 5m <task> # shortcut for loop
/pi-rogue-context brief # see compact stored context handles
Router defaults to observe-only recommendations. auto_model is explicit and limited to future model switches; it does not spawn agents or mutate tools.
@fiale-plus/pi-rogue is the single consolidated public artefact. It bundles:
packages/advisor/— advisor routing, binary gate, review/check-in behavior.packages/router/— local trajectory telemetry, model cards, training/evaluation workflows.packages/context-broker/— bounded broker runtime andcontext_lookuptool integration.packages/orchestration/— goal, loop, autoresearch, and lab primitives.packages/core/— shared contracts/helpers.
Former workspace-only lab helpers were archived; the supported production path is the consolidated bundle and shared core/context runtime.
| Surface | Common commands |
|---|---|
| Pi-Rogue | /pi-rogue status, /pi-rogue help, /pi-rogue doctor |
| Advisor | /pi-rogue-advisor status, settings/config, on/off, mode, model, review light|strict|off, <question> |
| Router | /pi-rogue-router status, mode observe, mode auto_model, profile <name>, models, configure |
| Orchestration | /pi-rogue-orchestration goal set/show/clear/list, /goal set/show/clear/list, /loop status/off/<interval> <instruction>, /autoresearch status/clear/<instruction>, /pi-rogue-orchestration lab status/clear/<instruction> |
| Context | /pi-rogue-context status, brief, lookup <handle|text>, pin <handle>, export <handle>, prune |
- Canonical package README — install scope and bundled command surface.
- Advisor README and binary gate runbook — escalation policy, trained gate, and evaluation workflow.
- Router README, routing dataset workflow, and routing labels — route telemetry and learning loop.
- Context broker README and context footprint proposal — prompt-safe artifact storage and lookup.
- Advisor Board agent/skill taxonomy — boundary between advisory agents, executable skills, and role Markdown.
- Advisor Board replay PoC — deterministic ledger/replay gate before live board behavior.
- Orchestration README — goal, loop, autoresearch, and lab behavior.
- Release guide — canonical
pi-rogue-<semver>release process.
npm install
npm run check
npm testLegacy .autoresearch scratch data is archived at ~/.pi/archived-autoresearch/pi-rogue/.