REFERENCE

Slash commands

Every Theoria-CLI slash command, with its forms and examples.

Type / in Theoria to see the commands, or /help for this list in the terminal and /help <command> for one command. In the forms below, <value> is required, [value] is optional, and a|b means one of them.

You can also ask in plain language, for example "shorten the proof of sum_squares", and the agent runs the matching command.

Lean and proofs#

/lean#

Choose, install and check the Lean and Mathlib this project uses.

  • /lean
  • /lean config
  • /lean status
  • /lean install
  • /lean remove
  • /lean search <status|install|remove>
  • /lean extras <status|install|on|off|remove> [hammer]
Examples
/lean config
/lean install
/lean search install
/lean extras install hammer

Without arguments it opens a menu. A folder that is a Lean project of its own keeps its own toolchain and dependencies. See Lean environments.

/formalize#

Turn an informal statement into a Lean statement, with sorry as its proof, checked by Lean.

  • /formalize <statement | file with one> [--name <theorem_name>]
Examples
/formalize the sum of the first n odd numbers is n^2 --name sum_odd
/formalize notes/conjecture.md

/prove#

Prove a theorem. Lean decides whether the proof is complete.

  • /prove <file.lean> [theorem]
  • /prove <theorem name>
  • /prove <description of the result>
Examples
/prove formal/Basics.lean sum_odd
/prove sum_odd
/prove the theorem about odd sums in Basics.lean

/compress#

Shorten verified proofs and lower their heartbeats, keeping every statement.

  • /compress <file.lean> [declaration] [rounds] [skill=<path>|none] [--emphasis loc|clarity|balanced]
  • /compress <theorem name> [rounds]
Examples
/compress formal/Basics.lean
/compress formal/Basics.lean sum_odd 3
/compress sum_odd --emphasis clarity

The score combines proof-token reduction, heartbeat reduction and measured compatibility; a candidate is kept only if Lean accepts it with the statement unchanged. Its prompt can be edited: see Workflow prompts.

/workflow-cancel#

Cancel the running /formalize, /prove or /compress, including its Lean check.

  • /workflow-cancel
Examples
/workflow-cancel

In the terminal interface, Esc does the same.

Research#

/next-steps#

Suggest ranked next research steps for papers or a line of research.

  • /next-steps <arxiv-id | url | file | topic> [more papers…] [--n <1-10>] [--focus theory|numerics|formalization]
Examples
/next-steps 2401.01234
/next-steps paper.pdf 2312.04567 --n 5
/next-steps "sphere packing in dimension 8" --focus formalization

The agent gathers the papers and related or citing work, gauges recent momentum, distills the open questions, and proposes ranked extensions with a first step, difficulty and risk.

--n sets how many ideas (default 5–8); --focus biases them toward one kind of work. Nothing is recorded unless you ask: record an idea with `/claim new hypothesis` or start an `/attack` on it.

/attack#

Attack a research objective with a persistent goal.

  • /attack status
  • /attack new <objective>
  • /attack fork [objective]
  • /attack resume [run-id]
  • /attack pause
  • /attack complete
  • /attack update focus <text> | note <text> | ready [summary]
Examples
/attack new prove the Collatz bound for n < 2^20
/attack update focus reduce to the odd case
/attack resume run-003

Each attack is a bounded research run. While a session is in one, its goal and the relevant claims and artifacts are added to the agent's instructions. /runs lists the project's runs; /help attack shows every form.

/attack replaces /run. Existing runs keep their IDs and can be resumed with /attack resume <run-id>. In Telegram, use /attack_new, /attack_fork, /attack_resume, /attack_pause and /attack_complete.

/claim#

One claim: create, change, link evidence, verify. /claims [status] lists them.

  • /claim <id>
  • /claim new [hypothesis|supported] <statement>
  • /claim set <id> <hypothesis|supported|refuted> [why]
  • /claim link <id> <supports|refutes|verifies|derived-from> <artifact|path>[:<theorem>]
  • /claim verify <id>
  • /claim test <id> <data> <test>
Examples
/claim new hypothesis every Keller map in dimension 2 is invertible
/claim link claim-004 verifies formal/Keller.lean:keller_dim2
/claim verify claim-004
/claims verified

A claim becomes verified only by Theoria: formally, when a linked Lean theorem is proved without sorry and with standard axioms, or statistically, from a linked data file.

/artifact#

Record or show an artifact (a file the project's results rest on). /artifacts lists them.

  • /artifact <id>
  • /artifact add <path> [kind] [description]
Examples
/artifact add data/runs.csv dataset "benchmark timings"
/artifacts

/plan#

The proof or research plan of this run or project.

  • /plan [show]
  • /plan check
Examples
/plan
/plan check

check asks Lean whether the plan is a valid proof skeleton: the objective is open only through its lemmas.

/autoresearch#

Unattended, bounded research campaigns that improve a benchmark result.

  • /autoresearch start <benchmark> [objective="…"] [corpus=<manifest>] [skill=<path>,…|none] [budget=<USD>] [minutes=<n>] [workers=<n>]
  • /autoresearch status [campaign-id]
  • /autoresearch pause [campaign-id]
  • /autoresearch resume [campaign-id]
  • /autoresearch cancel [campaign-id]
Examples
/autoresearch start putnam2025 corpus=corpus.json budget=100 minutes=180
/autoresearch status
/autoresearch pause campaign-002

The start confirmation shows exactly what is authorized; omitted limits are unlimited. See Autoresearch. macOS only for now.

/latex#

The project's paper: create, update, compile.

  • /latex
  • /latex status
  • /latex pdf
  • /latex paper create [brief]
  • /latex paper update [request]
Examples
/latex paper create a short note on the dimension-32 counterexample
/latex paper update add the Lean statement of the main theorem
/latex pdf

/subagent#

Independent child tasks with their own limits.

  • /subagent start <task>
  • /subagent list
  • /subagent status|result|wait|cancel <id>
Examples
/subagent start survey Mathlib for results on Sidon sets
/subagent result task-002

macOS only for now.

Benchmarks#

/benchmark#

Install benchmark suites, run trials judged by Comparator, and compare runs.

  • /benchmark list
  • /benchmark install <id>
  • /benchmark run <id> [train|val|test] [N | task,task] [model=<provider>/<id>] [cost=<USD>] [repeat=<k>] [concurrency=<c>] [timeout=<min>]
  • /benchmark run <id> mode=compress [corpus=<manifest>] [skill=<path>|none] [rounds=<n>] [model=<provider>/<id>]
  • /benchmark status [run]
  • /benchmark compare <run> <run>
  • /benchmark resume [run]
  • /benchmark cancel [run]
Examples
/benchmark install minif2f
/benchmark run minif2f val 10 cost=2
/benchmark run smoke train 2 model=amazon-bedrock/us.openai.gpt-6.1-sol-ultrafast cost=1
/benchmark compare run-004 run-007

A run costs money and asks first; the test split asks again. Benchmark environments are shared, not copied: see Lean environments. macOS only for now.

model=<provider>/<id> selects a model for this proving or compression run. Use a full ID from /model, including any slashes inside the model ID. Terminal and agent launches validate it against the session's registry, including custom models and aliases. Omit it to use the current session model.

The model is shown in the confirmation, saved with the run, and reused for all trials, retries and resumes. An override does not switch the session model. Telegram's /benchmark_run accepts the same option.

The agent can submit one benchmark-run action per model through the theoria tool, with model= in each call's args. Ask it to launch the same tasks with explicit model IDs; each run has its own confirmation and budget. /help benchmark shows the command forms.

Project and interface#

/project#

Projects: list and switch, status, type, MCP servers.

  • /project
  • /project list
  • /project status [refresh]
  • /project type [type]
  • /project new [name]
  • /project servers
Examples
/project status
/project type theoretical-physics
/project list

/new-project [name] is a shortcut for /project new.

/connect-bot#

Connect Theoria to Telegram: pair your phone by QR code.

  • /connect-bot
  • /connect-bot status
  • /connect-bot pair
  • /connect-bot remove
Examples
/connect-bot

Then theoria bot start runs the bot. macOS only for now.

/tui#

Interface settings: footer, editor, telemetry, thinking peek, icons.

  • /tui
Examples
/tui

More settings, including the startup animation, are in Configuration.

/help#

Theoria's commands, or one command in detail.

  • /help [command]
Examples
/help
/help attack
/help benchmark

Pi's own commands#

Theoria runs on Pi, so Pi's commands work too: /model, /login, /settings, /resume, /new, /compact, /mcp and others.

Last updated 2026-11-10View as Markdown