personas/personas/_shared/community-skills/feynman-cli/references/repl-and-slash.md

REPL & Slash Commands

REPL flow

feynman (no args) launches the REPL. Conversational: type prompt, Enter, read response. The REPL retains session memory, indexes artifacts to outputs/ (when running in a project), and exposes slash commands.

REPL templating:

• @<path> — inject file contents into the prompt (e.g. @papers/foo.pdf)
• !<shell> — run a shell command and inject its stdout (e.g. !ls outputs/)

Quit: Ctrl-D or /exit (when supported by your version).

Slash commands — research workflows

Command	What it does
/deepresearch <topic>	Multi-agent investigation → research brief in outputs/
/lit <topic>	Structured literature review (consensus / disagreement / open Qs)
/review <artifact>	Peer review with severity-graded feedback + inline annotations
/audit <paper-or-arxiv-id>	Compare a paper's claims against its public code
/replicate <paper>	Plan or execute a replication for a paper / claim / benchmark
/compare <topic>	Build agreement / disagreement matrix across sources
/draft <topic>	Paper-style draft from research findings
/autoresearch <idea>	Autonomous loop iteratively optimizing toward a goal
/watch <topic> [cadence]	Recurring research monitoring on a topic

Slash commands — project / session / control

Command	What it does
/help	Grouped command list
/init	Bootstrap AGENTS.md + session-log folders for a new project
/log	Durable session log: work / findings / open Qs / next steps
/jobs	Active background work — running, scheduled, watches
/outputs	Browse research artifacts
/search <query>	Search prior session transcripts (semantic + keyword)
/preview [path]	Render artifact as HTML or PDF
/feynman-model	Model picker (default model + per-subagent overrides)

Slash commands — pi-subagents (when installed)

pi-subagents adds direct agent control:

Command	What it does
/run <agent> <task>	Run one agent with a task
/chain <a> "task" -> <b> "task"	Sequential chain
/parallel <a> "task" -> <b> "task"	Parallel run
/subagents-status	Async status overlay
/agents	Agents Manager overlay (edit / override)

Per-step task forms:

• /chain scout "scan code" -> planner "make plan" — each step has its own task
• /chain scout -- scan code -> planner -- analyze auth — -- delimiter
• /chain scout "analyze auth" -> planner -> implementer — bare steps inherit {previous} (chain) or first task (parallel)
• Shared task: /parallel scout reviewer -- audit src/ — same task to both

Quotes ("…" or '…') and -- are interchangeable.

Sample REPL session

$ feynman
Feynman 0.2.39 — github-copilot/gpt-5.3-codex (medium thinking)
> /deepresearch Current approaches to mechanistic interpretability in LLMs
[plan] Reading 12 source candidates…
[researcher] Fetching arXiv:2401.…
[writer] Drafting brief
✓ outputs/mechanistic-interp-brief.md (2,341 lines, 47 sources)
> /preview outputs/mechanistic-interp-brief.md
✓ Opened in browser
> /log
✓ Wrote outputs/session-log-2026-05-01.md

Tips

• Long-running workflows (/deepresearch, /lit) run for minutes. Use /jobs from another window to monitor.
• /watch <topic> weekly schedules via pi-schedule-prompt. List with /jobs. Cancel by editing the schedule (no built-in unwatch verb).
• /preview requires pandoc (feynman setup preview once).
• /feynman-model lets you set different models per role — e.g. cheap Haiku for researcher (high token volume), Opus for reviewer.
• @<path> works with directories — Feynman recursively reads supported formats (pi-docparser handles PDFs/Office/etc).
• Slash commands shadow built-ins of the same name when defined as a custom command. Project-scope shadows user-scope.