DeepSeek-Reasonix/docs/CLI.md

Reasonix CLI Reference

README  ·  简体中文  ·  Guide

This reference covers interactive sessions, one-shot automation, session resume, permission flags, and the most useful in-session commands. For provider configuration, plugins, and sandbox policy, see the Guide.

Start a session

reasonix
reasonix --model deepseek-pro
reasonix --effort high
reasonix --dir /path/to/project

Ordinary requests always enter the executor. There is no automatic simple / light / full task mode to pick. The dedicated planner runs only for an explicit Plan, an approval boundary, or Goal start.

Running reasonix without a subcommand starts the interactive terminal UI. If the selected connection has no credential, the local connection picker opens instead of sending a request. History and local commands remain available while authentication is incomplete.

Flag	Purpose
--model NAME	Select a configured provider or provider/model reference.
--effort LEVEL	Override reasoning effort for this session.
--max-steps N	Set a one-off maximum tool-call round budget; 0 uses automatic execution.
--dir PATH	Change the workspace root before loading config and tools.
--add-dir PATH	Add another writable tool directory; repeat for multiple directories.
-c, --continue	Resume the most recent session.
-r, --resume [QUERY]	Open the session picker, or resume a matching session.
--copy	Continue in a writable copy of the resumed session.
--allowed-tools RULES	Add session-only permission allow rules. Repeatable; --allowedTools is an alias.
--permission-mode MODE	Start with a specific permission posture.
--dangerously-skip-permissions	Deprecated compatibility flag; migrates conservatively to workspace-write. Use --permission-mode danger-full-access for YOLO.

Flags may appear before or after the prompt where applicable.

Update the native CLI

reasonix upgrade                  # install the latest official release
reasonix upgrade --check          # report the target without installing
reasonix upgrade --force          # reinstall the current official release

The updater selects only strict vX.Y.Z non-prerelease GitHub Releases. During the 1.x compatibility period, old channel arguments and --channel are still accepted, but resolve to the same official release and print a deprecation notice. Legacy [cli].update_channel values are ignored and removed the next time Reasonix saves the configuration. The reasonix update alias behaves the same way.

Configure providers

reasonix setup                    # manage the user-global config
reasonix setup --local            # manage ./reasonix.toml
reasonix setup /path/to/config.toml

In an interactive terminal, reasonix setup is a staged provider manager. It lists configured providers and lets you:

• add OpenAI-compatible or Anthropic-compatible providers;
• edit endpoints and model lists;
• update API keys or test the connection and refresh models;
• choose the default model; and
• remove providers.

Choose Save and exit to review and confirm the pending operations. Canceling discards them. Setup reloads the latest config while saving: unrelated desktop or CLI changes are retained, while an overlapping change is reported as a conflict instead of being overwritten.

Provider definitions contain only the api_key_env variable name. Key values are stored in the shared Reasonix home .env, even with --local. Adding, replacing, or explicitly clearing a key creates a fresh private credential slot and atomically switches only the selected connection to it. Existing fixed variables remain readable and migrate only when that connection is edited.

Inside the TUI, /setup opens the same connection flow and /auth is an alias. The key field is masked; press Ctrl+T to test the draft connection, Enter to save, or Escape to cancel. /? is an alias for /help. Authentication that is not ready never turns ordinary input into a provider request.

reasonix doctor credentials
reasonix doctor credentials --json
reasonix doctor credentials --probe
reasonix doctor credentials --repair --dry-run
reasonix doctor credentials --repair

The default diagnostic is read-only. --probe tests temporary create and atomic rename without replacing .env. Repair is limited to a current-user- owned regular file inside Reasonix home; it does not take ownership, remove deny rules, grant Everyone, follow links/reparse points, or kill a file holder.

Configure fee display currency

Use the user-global command to inspect or select the display currency:

reasonix config currency             # show the saved and resolved currency
reasonix config currency auto        # wallet hint, then original price currency
reasonix config currency CNY
reasonix config currency USD

auto remains unresolved in configuration. With one valid wallet currency it can become a runtime session hint; otherwise CLI uses the original currency or sorted currency buckets. Language and host locale never select a price table. The preference is user-global and cannot be overridden by project reasonix.toml; --local is therefore not supported. Custom prices are preserved.

In an interactive session, /currency shows the saved and resolved values, and /currency auto|CNY|USD changes the preference and refreshes the current runtime without discarding the conversation.

Configure automatic compaction

The desktop app and CLI share the user-global automatic compaction threshold. Inspect the effective percentage and its source, set the global default, or add a project override:

reasonix config compact-ratio              # show effective value and source
reasonix config compact-ratio 75           # set the user-global default
reasonix config compact-ratio --local 75   # override in ./reasonix.toml

The editable range is 30–85%, with 80% as the built-in default. Lower values compact earlier, may increase summary calls and cost, and may reduce prompt-prefix cache reuse; higher values retain more context before compaction. Below the threshold, complete tool results may increase ordinary request cost; at pressure they are durably pruned before the cache-aligned summary runs. Project reasonix.toml takes precedence over the user config. Changes apply to new CLI sessions; an already-running session keeps the threshold it loaded at startup.

One-shot and automation

Use -p / --print when a script needs only the final answer:

reasonix -p "summarize this repository"
reasonix -p "summarize this repository" --output-format json
reasonix run "implement the TODOs in main.go"
reasonix run --auto "implement the TODOs in main.go"
echo "explain this code" | reasonix run

reasonix run keeps the normal streamed terminal presentation unless -p or a structured output format is selected. It also accepts --model, --max-steps, --effort, --dir, --add-dir, --continue, --resume QUERY, --copy, --allowed-tools, --permission-mode, and --auto / -y (legacy aliases for --permission-mode workspace-write).

Benchmark arms

--ablate switches whole subsystems off so a benchmark can attribute a change in success rate to one of them. It accepts a comma-separated list of evidence, planner, subagent, retrieval and compaction, plus none (the default, everything on) and all. Sub-agents inherit the parent's arm, and the arm name is written to the --metrics file so a recorded run is self-describing.

reasonix run --ablate evidence,planner --metrics run.json "fix the failing test"

This is a measurement tool, not a tuning knob: switching a subsystem off makes Reasonix worse at the work it was added for.

Trajectory recording

--trajectory PATH appends the run's full event stream — tool dispatches and results with absolute start/end times, reasoning, retries, readiness and recovery decisions — as one timestamped, sequenced JSONL record per event, so a run can be replayed and its time attributed offline (tool execution vs. the model thinking between calls). Records reuse the shared eventwire JSON contract under an event key, wrapped in schema_version, seq, and ts (unix ms). Every completed line survives a killed run. Unlike --events-jsonl, the file contains prompts, tool arguments, and reasoning: treat it with the same care as a session transcript.

reasonix run --metrics run.json --trajectory run.trajectory.jsonl "fix the failing test"

Turn phases

While a turn runs, the host publishes a content-free phase so a frontend can say what the turn is doing. The CLI shows it on the spinner line; the desktop app shows it in the composer.

These phases describe execution timing, not verification evidence. Desktop check-result cards follow actual running verification tools, not phase names.

Phase	Emitted when	capability_phases bucket
working	the turn starts, after each tool batch returns, after model generation	ProviderWaitMs
checking	a tool batch is about to execute	ToolExecMs
verifying	an actual verification tool runs	ToolExecMs

A phase is billed to its bucket when the next phase opens, so the durations in --metrics split a turn into model wait versus tool execution without replaying the run. Spans under a millisecond are dropped, and a turn that ends through an error or a pause rather than an answer does not bill its last span, so the buckets read as a lower bound rather than a full partition of the turn.

An approval prompt raised inside a tool batch bills to ToolExecMs: the batch stays open from checking until the next working, and no user-wait phase is emitted. ReviewMs, SubagentWaitMs, UserWaitMs and CompactMs stay zero because nothing opens those phases inside a turn — reviewing is published only at run exit, after the turn's phase clock has already closed.

Output formats

Format	Behavior
text	Human-readable text. With -p, prints only the final answer.
json	Emits one final result object.
stream-json	Emits one shared eventwire JSON object per line, followed by the final result object.

reasonix -p "list the risky changes" --output-format text
reasonix -p "summarize the diff" --output-format json
reasonix run "run the tests" --output-format stream-json

The final structured object has this shape:

{
  "type": "result",
  "subtype": "success",
  "is_error": false,
  "duration_ms": 123,
  "num_turns": 1,
  "result": "...",
  "result_from_reasoning": false,
  "session_id": "...",
  "total_cost": 0,
  "currency": "USD",
  "total_cost_usd": 0,
  "usage": {
    "input_tokens": 0,
    "output_tokens": 0,
    "cache_read_input_tokens": 0,
    "cache_creation_input_tokens": 0
  }
}

result_from_reasoning is present, and true, when the turn finished with an empty visible message and result therefore carries the turn's reasoning instead. A thinking model may answer entirely in the reasoning channel; without the field a caller cannot tell that text apart from a visible answer, and without the fallback result would be "" and -p would print nothing. It is omitted whenever the model emitted visible text, which is the ordinary case.

total_cost is present only when a single selected display amount exists (ISO code in currency). Prefer the structured cost_quote field when present: it carries the original estimate, original_totals, occurrence-time valuations (official_table for dual-region public prices), cost_complete, display_complete, display_status, and billing_mode (payg or subscription_equivalent for pay-as-you-go equivalent estimates such as MiMo Token Plan).

total_cost_usd remains a numeric compatibility alias when total_cost exists and does not imply USD. Mixed original currencies no longer fail the run: cost_complete remains true when usage/pricing facts are known, display_complete is false, and original_costs/original_totals list per-ISO totals so clients never invent a cross-currency sum.

Global display preference is [billing].display_currency (auto|CNY|USD); legacy [desktop].currency still migrates. Provider list prices use each entry's frozen billing_currency and are never rewritten by display switches. Diagnose with reasonix doctor billing.

Execution failures use subtype: "error_during_execution" and is_error: true. Structured modes keep runtime errors in JSON instead of also printing a duplicate human-readable error. Authentication failures also include optional error_code, authentication_status, and recovery_actions fields. The same fields appear on the final run_done record from --events-jsonl. For example, a missing key reports missing_credential and actions such as configure_credentials, select_model, and diagnose_credentials; no model request is made.

The completion validator has been removed. A clean model stop without tool calls ends the turn directly; a response with tools continues through the tool loop, and a truly empty response is retried at the frozen-request boundary. Legacy completion_validation, completion_evaluator_model, and REASONIX_COMPLETION_VALIDATION_MODE settings remain readable but are ignored and are no longer emitted by the config renderer. Explicit budgets, tool-safety and protocol recovery boundaries remain active. Goal completion is a model declaration; no host quality gate or independent Goal evaluator runs. See migration details.

Redacted machine interfaces

Use the dedicated event flag when an automation needs lifecycle telemetry but must not receive prompts, reasoning, tool arguments, tool output, or approval text:

reasonix run --events-jsonl "run the focused tests"

Every line has schema_version, sequence, and kind; the final line is kind: "run_done". --events-jsonl is intentionally separate from the richer --output-format stream-json contract and cannot be combined with --output-format.

The following read-only commands expose persisted state without transcript, label, command, output, path, PID, or host-name content. Here, read-only means the commands do not mutate transcript, runtime, recovery, or query state. The first redacted-machine invocation may initialize a private identity key in the Reasonix user-state directory:

reasonix session list --json [--dir SESSION_DIR | --project-root PATH]
reasonix session show <machine-session-id> --json [--dir SESSION_DIR | --project-root PATH]
reasonix session status <machine-session-id> --json [--dir SESSION_DIR | --project-root PATH]
reasonix session recovery [<machine-session-id>] --json [--dir SESSION_DIR | --project-root PATH]
reasonix task list --json [--dir SESSION_DIR | --project-root PATH] [--session MACHINE_SESSION_ID]
reasonix task show <task-id> --json [--dir SESSION_DIR | --project-root PATH] [--session MACHINE_SESSION_ID]
reasonix task monitor list --json [--dir PROJECT_DIR]
reasonix task monitor status <task-id> --json [--dir PROJECT_DIR]
reasonix task monitor events <task-id> --json|--jsonl [--dir PROJECT_DIR] [--after N] [--follow]
reasonix hook list --json [--project-root PATH] [--home-dir PATH]
reasonix hook status --json [--project-root PATH] [--home-dir PATH]

For session and task, --dir explicitly selects the session storage directory, while --project-root resolves the selected project's session store. The two options cannot be combined. Without either option, Reasonix selects the current project's session store. For hook, --dir is an alias for --project-root. hook list reports active or invalid; invalid means the configured event cannot execute because its event, command/context source, or tool-event matcher is unusable. Matchers on non-tool events are ignored.

Machine session IDs are keyed opaque hashes, not transcript file names. They remain stable for the same session and Reasonix user-state directory, while a different installation key produces unrelated IDs and prevents offline guesses from timestamps or model labels. Preserve the private identity key when moving the Reasonix state directory if automation depends on existing machine IDs. Task finished_at is empty while a task is running, and artifact_complete=true is emitted only for a terminal task whose persisted artifact exists. A running record without a live session lease is reported as interrupted; opening that session also repairs the persisted lifecycle state.

Schema compatibility rules for version 1:

• consumers must ignore unknown fields;
• fields are not removed or retyped within the same schema version;
• empty collections are encoded as [];
• argument errors exit with status 2, state/query errors with status 1;
• machine-command errors are JSON objects with a stable error.code.

Resume sessions

reasonix --continue
reasonix --resume
reasonix --resume provider-config
reasonix --resume <session-id>
reasonix --resume provider-config --copy

• --continue resumes the newest saved session immediately.
• Bare --resume opens the searchable picker in an interactive terminal.
• --resume QUERY accepts an exact session ID or path, or a unique title or preview substring. Missing and ambiguous matches fail with a descriptive error.
• --resume=true and --resume=false remain accepted for compatibility.
• --copy leaves the original transcript untouched and continues in a new writable session. Use it when another Reasonix process owns the original.

For one-shot runs, reasonix run --resume QUERY "task" accepts a session file path, a session ID, or an opaque machine session ID from --events-jsonl / reasonix session show --json. Session leases prevent the desktop app and CLI from writing the same transcript concurrently.

Permissions

reasonix --permission-mode read-only
reasonix --permission-mode workspace-write
reasonix --permission-mode danger-full-access
reasonix -p "run the focused tests" --allowed-tools "Bash(go test ./...)"

Preset	Behavior
read-only	Read the workspace; writes and external side effects require a scoped authorization.
workspace-write	Write inside the workspace and private session temporary directory. This is the default.
danger-full-access	Run as the current OS user without Reasonix filesystem or network sandboxing. Explicit host deny rules still apply before launch.

Inline scripts, pipes, substitutions, and shell -c forms follow the same preset and sandbox boundary as other commands. Syntax alone never creates an approval request.

--allowed-tools is a session permission override, not a provider tool-schema filter. Rules may be comma- or space-separated, and the flag is repeatable. Configured deny rules always win over command-line allow rules.

In non-interactive runs (reasonix run / -p) there is no prompt to answer. read-only therefore fails closed for writes and side effects unless a narrow authorization was supplied at startup. workspace-write runs normal builds, tests, pipes, and inline scripts inside the OS sandbox. danger-full-access must be explicit and still cannot bypass configured deny rules.

Additional directories

reasonix --add-dir ../shared
reasonix -p "update both projects" \
  --add-dir ../frontend \
  --add-dir ../backend

Relative paths resolve from the workspace root and must already exist as directories. Reasonix resolves symlinks, removes duplicates, and extends the file-writer and sandboxed Bash write boundaries for the session. These additions are runtime-only and are not written to configuration.

Interactive controls

The /model, /provider, and /resume commands use searchable pickers. Approval prompts use the same row-selection behavior while retaining their single-key shortcuts.

Key	Action
Up / Down, Ctrl+P / Ctrl+N	Move through picker or approval rows.
j / k	Move while the search is empty; after search input starts, enter j / k as query text.
Type	Filter a searchable picker.
Enter	Select the highlighted row.
Esc	Cancel the current picker or approval.
y / a / n, number keys	Allow once, allow the displayed scope for this session, or deny.
Shift+Tab	Cycle Read only → Workspace write → YOLO → Plan.
Ctrl+Y	Toggle YOLO; the runtime permission preset is danger-full-access.

The responsive footer keeps interaction state on the left and, when space allows, places model and effort on the right. Its second row shows available repository and session telemetry such as cache hit rate, context use, compaction headroom, background jobs, and balance. ready means the composer is idle; that slot changes when a picker, approval, image paste, shell mode, or other interaction needs attention. Narrow terminals move or compact complete groups instead of cutting labels in half. Visible labels follow /language.

Use /theme auto|light|dark to select the terminal background mode, or choose a named accent from /theme. Both composer borders, the insertion cursor, selection, scrollbar, and footer use the active CLI theme. See Keyboard shortcuts for transcript navigation, multiline input, rewind, and clipboard controls.

Clipboard actions are deliberately split by content type. Local transcript and composer selections use the native system clipboard and report success only after that write completes; SSH falls back to an explicitly labelled OSC 52 request. Text paste remains the terminal's bracketed-paste action (Cmd+V on macOS and the terminal's configured shortcut elsewhere). While Reasonix owns the mouse in a local session, right-click with no selection reads clipboard text through the same paste path; right-click with a selection copies it. Over SSH, use the terminal paste shortcut because the remote process cannot read the local clipboard; /mouse restores the terminal's native right-click menu. Image paste is application-owned: use Ctrl+V on macOS/Linux, Alt+V on Windows, or /paste-image; the footer shows Pasting image… until the attachment token is ready. Where the terminal forwards that shortcut instead of pasting itself, a clipboard holding no image falls back to a text paste, so the key never swallows plain text.

In-session commands

Type /help in an interactive session for the complete command list. Slash completion, help, dispatch, and aliases are generated from the same registry, so the displayed list matches the commands the TUI accepts.

Command	Purpose
/continue-checks [guidance]	Resume the immediately preceding paused task-completion check while preserving its verified tool evidence. The command is one-shot and refuses stale cards after another user turn.
/model	Search configured models and switch the active model.
/provider	Choose a provider, then choose one of its configured models.
/resume	Search recent sessions and switch to one.
/takeover	Take over the last refused session (or a listed entry) from the resident serve: this CLI becomes the writer and remote viewers become read-only spectators until they reclaim. After a desktop reclaim it re-takes the remembered session directly; a session no runtime holds any more is simply resumed.
/status	Show model, effort, cache, Git, background jobs, and balance details.
/theme [auto|light|dark|style]	View or change the CLI background mode and accent palette.
/currency [auto|CNY|USD]	View or change the user-global fee display currency and refresh the runtime.
/paste-image	Read a clipboard image and insert an editable attachment token.
/mouse	Toggle in-app mouse selection, scrollbar, and wheel handling; SSH sessions start with capture off so the terminal's native selection works.
/effort	View or change reasoning effort.
/output-style	Select an answer style.
/verbose	Toggle expanded reasoning display.
/sandbox	Inspect sandbox status.
/goal [objective]	Start a continuous goal, or inspect its runtime statistics.
/goal status	Show the active goal plus turns, requests, tokens, work time, and the last continuation reason.
/goal pause	Pause the running goal (keeps todos, Delivery checkpoint, and runtime history).
/goal resume	Resume a manually paused or genuinely blocked goal without changing a numeric quota.
/goal clear	End goal mode permanently.
/docs [question]	Show the embedded corpus identity, or search it locally and ask the configured AI to answer from version-matched evidence.
/reasonix:docs [question]	Preferred built-in fallback when an existing custom command or compatible plugin/skill alias owns /docs; if this spelling is also owned, the menu selects the next free reasonix:-qualified name without displacing it.
/mcp, /skills, /hooks	Inspect and manage extensions.
/remember <note>	Append a standing note to the project instruction document; # <note> is a shortcut.
/memory [subcommand]	Inspect instructions, memory provenance, recall, revisions, and recovery.
/rewind	Restore conversation and/or code to an earlier turn.
/tree, /branch, /switch	Inspect or navigate conversation branches.
/reload	Reload the agent runtime (extensions, tools, skills, commands, hooks, providers) while keeping the session. Queued once while a turn runs, then fail-atomic: a failed rebuild keeps the current runtime.

Switching model or effort rebuilds the runtime while preserving the active conversation, session-scoped permission overrides, additional directory access, and session ownership. /reload uses the same fail-atomic rebuild. Execution modes no longer exist: planning, verification, and review strength follow task risk per turn.

/preset, /work-mode, and /profile remain hidden compatibility commands. Recognized legacy values are accepted, report that the setting is retired, and leave the session on standard execution; unknown values still return an error.

Session catalog diagnostics

The desktop session catalog is a disposable SQLite query projection; transcript JSONL and sidecars remain authoritative. Inspect it read-only or replace only the projection:

reasonix doctor sessions [--json]
reasonix sessions reindex [--json]
reasonix sessions reindex --dir /path/to/sessions --dir /another/path

Without --dir, reindex includes global sessions and all projects saved by the desktop app. See Session Catalog and Desktop Startup for failure, migration, and data-safety guarantees.

History search uses a separate disposable projection:

reasonix doctor catalogs [--json]
reasonix catalogs reindex history [--dir PATH ...] [--json]

See History Search Catalog. Usage statistics use a separate disposable rollup projection: reasonix catalogs reindex usage [--json] See Usage Catalog.

Inspect or rebuild the disposable task projection independently:

reasonix doctor catalogs [--json]
reasonix catalogs reindex tasks [--project PATH ...] [--json]

See Task Catalog for the authoritative FileStore boundary, cross-project routing, and rebuild behavior.

Memory diagnostics and recovery

Bare /memory shows all active project/global facts without hiding same-name entries. Facts include their stable ID, revision, scope, type, freshness, and description. Slash completion offers the available subcommands, active IDs and names, and owned archive paths.

Command	Purpose
/memory instructions	Show resolved instruction precedence, directories, imports, and diagnostics.
/memory recall	Explain the latest automatic recall query, hits, scores, reasons, freshness, and budget.
/memory revisions <id-or-name>	Show the active revision and immutable history.
/memory restore <id-or-name> <revision>	Restore old content as a new monotonic revision.
/memory archived	List archived facts and their owned paths.
/memory recover <archive-path>	Recover an archive as a new revision without overwriting active data.

These commands run against the active session controller. When the session lives on a remote host (reasonix remote connect / a desktop remote web window), they use the remote memory catalog and never fall back to local desktop memory. See Context Engine v2 for authority, automatic recall, write confirmation, and migration behavior.