shidoshi

An opinionated way to augment Jupyter Lab for iterative work.

shidoshi adds %ask / %%ask magics to Jupyter that let you talk to an LLM from inside a notebook — using the notebook itself, in order, as the conversation history. No separate chat pane, no copy-pasting context: your code cells, their outputs, and your notes are the context.

Install

Requires Python ≥3.13 and JupyterLab. Set an API key before use:

export OPENAI_API_KEY=sk-...       # for the openai provider (default)
export OPENAI_BASE_URL=...         # optional, e.g. to point at a proxy
export OPENROUTER_API_KEY=...      # for the openrouter provider

Installing across Jupyter environments

%load_ext shidoshi runs import shidoshi inside the running kernel process. That means shidoshi has to be installed into whichever Python environment the kernel you're using actually runs in. It ships a small shidoshi command for setup, but the library itself is not a standalone tool — so uvx / uv tool install (which run a tool in an isolated subprocess, separate from any kernel) don't apply here.

• Per-project venv with its own JupyterLab (e.g. a uv-managed project): add shidoshi as a normal dependency of that project.

  uv add shidoshi
  # or: pip install shidoshi
  

• One shared JupyterLab, many kernels (each notebook's kernel points at a different project venv registered via ipykernel install): install shidoshi into each kernel's venv. Installing it only where JupyterLab itself lives will not make it importable from other kernels.

  # inside the venv backing a given kernel
  uv add shidoshi
  # or: pip install shidoshi
  

Skipping %load_ext — the shidoshi kernel

To avoid typing %load_ext shidoshi in every notebook, register a kernel that loads it for you:

shidoshi install-kernel

Pick Python 3 (shidoshi) from the Jupyter kernel list and the magics are already there. Nothing else changes: it is a stock Python kernel running this environment's interpreter — your imports, variables, and debugger all work exactly as before. The generated kernel.json just appends --IPKernelApp.extensions=shidoshi to the normal ipykernel_launcher command, with an absolute path to this environment's Python.

With no location flag, this installs into your per-user kernel directory — the only location every Jupyter on the machine can find, regardless of which environment actually launches it. After installing, the command runs jupyter --paths for you and prints a warning if the kernel it just wrote won't actually be visible to the jupyter on your PATH — worth reading if you pass --sys-prefix or --prefix and the kernel doesn't show up in the launcher.

Useful flags:

Installing is refused if a kernelspec of that name already exists, so it won't quietly replace one you made by hand. Register one per environment with a distinct --name.

Remove it with jupyter kernelspec remove shidoshi.

With uv

Add shidoshi to the project, then register the kernel from inside it. Which location flag you need depends on where JupyterLab itself runs from, because Jupyter only searches its own sys.prefix, your user directory, and the system directory:

uv add shidoshi

# A: JupyterLab in an ephemeral env (uv's default suggestion).
#    Its sys.prefix is a uv cache dir, so --sys-prefix would be invisible.
uv run shidoshi install-kernel --user
uv run --with jupyter jupyter lab

# B: JupyterLab as a project dependency — sys.prefix *is* the project venv.
uv add --dev jupyterlab
uv run shidoshi install-kernel --sys-prefix
uv run jupyter lab

B keeps the kernel scoped to the project and disappears with the venv; A is the one that works with uv run --with jupyter — and matches install-kernel's default, so uv run shidoshi install-kernel (no flag) does the same thing. If a freshly installed kernel doesn't show up in the launcher, run jupyter kernelspec list the same way you start Lab — that prints exactly the directories being searched, or trust install-kernel's own post-install warning, which runs that same check for you.

Either way this replaces the kernel step in uv's Jupyter guide — you don't need uv run ipython kernel install --env VIRTUAL_ENV ... as well, because install-kernel records VIRTUAL_ENV for you when it detects a venv.

That variable matters more than it looks. uv pip install resolves its target from VIRTUAL_ENV (falling back to CONDA_PREFIX, then a base interpreter) — never from the kernel that's running. So if you start Jupyter from a conda-activated shell, a kernel without VIRTUAL_ENV will import from your project venv while !uv pip install quietly installs into your conda base. Pinning it keeps both views on the same environment.

Two related notes:

• !uv add was always safe — it finds the project by walking up for pyproject.toml, so it ignores VIRTUAL_ENV and targets the project venv either way.
• %pip install needs uv venv --seed; uv venvs have no pip in them by default. Prefer !uv add.

Pass --env to set anything else the kernel should launch with (repeatable), including an override for VIRTUAL_ENV:

uv run shidoshi install-kernel --sys-prefix --env OPENAI_BASE_URL=http://127.0.0.1:18080/v1

Because the spec pins an absolute interpreter path, it can only ever start the environment shidoshi is installed in. That is the advantage over the ipython_config.py route below: ~/.ipython is shared by every Python environment under your $HOME, so putting the extension there makes every kernel on the machine try to import shidoshi, including ones that don't have it.

c.InteractiveShellApp.extensions = ["shidoshi"]

Quickstart

%load_ext shidoshi

(skip this line if you're on the Python 3 (shidoshi) kernel)

%%ask
What does the `history.build_history` function in this file do?

The response streams into the cell's output as Markdown.

Images in web mode

Run shidoshi web, open a notebook, then drop PNG, JPEG, GIF or WebP files between cells. Each file becomes a rendered, editable note. The insertion marker shows where the notes will go; Add images inserts before the selected cell, or at the end if no cell is selected. Uploads support files up to 20 MB and 25 megapixels.

Images are saved beside the notebook with unique filenames and referenced by ordinary relative Markdown links. Keep these files alongside the .ipynb when sharing or moving it. Branching into another directory copies referenced images; deleting a note keeps its file, so undo and other references continue to work.

Dropping an image does not call the LLM. Ask and Agent include images only from eligible cells preceding the prompt being run, following the same pin/exclusion/aside rules as other notebook context. Images below the prompt are not included, even when pinned. Use a model that supports image inputs. Markdown image references in prompts, raster Jupyter cell attachments, and preceding Python plot outputs are supported too.

Agent refreshes notebook image attachments for each turn. Moving an image below the prompt or excluding it removes its bytes from subsequent requests; the agent's existing text conversation memory, including earlier descriptions, remains. Missing or unreadable images produce a visible error before a provider request is sent.

Magics reference

• %ask <prompt> — line magic for a one-line prompt.
  • Prefix with model| or provider:model| to override the configured default model for just this call, e.g. %ask openrouter:openai/gpt-4o|summarize this.
  • Add --no-search to make web search unavailable for this call, or --search to override a disabled profile. With model | prompt, put flags before |; prompt text after it is preserved verbatim.
  • Add --debug before | to also show the full request/response payload.
• %%ask [model] — cell magic; the whole cell body is the prompt (multi-line is fine, and it can reference images via Markdown ![]()/<img> syntax or bare local file paths — they're inlined as base64). An optional model name on the magic line overrides the default for this call. Also supports --search, --no-search, and --debug.
• %%skip — runs the cell normally, but the cell is left out of the context sent to the model entirely. Use it for scratch or exploratory cells you don't want the model to see.
• %%pin — runs the cell normally; its content and output are always included in context and are exempt from the auto-trim behavior below. Use it to protect a fact, constant, or definition you don't want dropped over a long session.
• %%agent [model] — a multi-step, tool-using agent instead of a one-shot answer. See below.
• %agent_resume approve|reject|edit — answers a %%agent turn that paused for approval (only relevant if agent.approvals is configured — see Configuration).
• %shidoshi config ... — inspect and edit shidoshi's own config from inside a notebook cell. See Configuration.

%%agent — the agentic magic

Where %%ask answers once from what the notebook shows, %%agent can take several steps and look at what the kernel actually holds — including variables no cell output ever displayed. It runs deepagents in-process, as a backend parallel to %%ask's; %%ask is unchanged.

%%agent
Which of my dataframes has missing values, and where?

The answer renders as Markdown with the agent's steps in a collapsed 🧠 Agent steps panel above it.

Tool rungs

--tools chooses what the agent is allowed to do. The default is actor: model-written code runs in your live namespace, with no confirmation step. That is the reason %%agent exists rather than being a slower %%ask, so it does not sit behind a flag — but know that it is what a bare %%agent cell does.

--no-tools is the way back down — short for --tools none, for when you want the conversation without the hands. --tools proposer is the middle ground: the agent hands you a cell and your Run button is the approval step, with no separate permission prompt to click.

--no-tools does not mean offline: web search is provider-hosted rather than a kernel tool. Combine --no-tools --no-search when you want the model to use only the notebook conversation and its existing knowledge.

Set a different default in config with agent.permission_rung — permission_rung = "proposer" restores the gated behavior for every cell. Go a step further and require an explicit approval before run_code itself even runs — see Configuration.

These rungs also govern the opt-in project filesystem and host shell; see Letting %%agent work with real files.

Memory

Each notebook gets one conversation thread, seeded once from the cells above the first %%agent call. Later cells continue that conversation rather than rebuilding context each time.

The thread lasts as long as the kernel and no longer. This is deliberate: what the agent remembers is largely kernel state — variables it listed, values it read — and a restart destroys exactly that. A conversation that outlived the kernel would keep tool results reporting variables that no longer exist, worded as fact. So a restart clears the thread and the notebook is read again as it currently stands, which is what rerunning it from the top means anyway.

Use --fresh to opt out of the thread entirely and get %%ask-style behavior — context rebuilt from cells, nothing remembered — or --thread NAME to keep a side conversation separate.

Inspecting a turn

--debug works as it does for %%ask, adapted to a graph that runs more than once. Above the answer you get the request panel — provider and model, the composed system prompt, every message the model will see, and the tools this rung binds — followed by a live panel of raw LangGraph events as the turn runs. Messages already in the thread are labelled from thread, because on a continued conversation only your new prompt is passed in; the rest comes from the checkpoint.

The stream panel is LangGraph's fullest per-step view — task, task_result and checkpoint events with step numbers, the middleware nodes that a plain update stream never names, and a per-task error when one fails. Failed steps are flagged in the summary line so you don't have to expand them to find the one that broke. A running token count sits above it, for both providers.

Token deltas are collapsed into a per-node count, so what you read is the node transitions — model → tools → model — rather than several hundred one-token lines.

An unrecognised flag is an error rather than a no-op, so a typo'd --tools tells you instead of quietly running with the default.

Getting help

%ask --help and %agent --help print the flags, and for %%agent the rung table. Use the single-% line form: IPython rejects a cell magic whose body is empty before the magic itself runs, so %%agent --help on its own can never reach us.

%agent [flags] [model |] your request is also a one-line shorthand for %%agent, matching %ask. With an explicit model | request, flags belong before |; text after it is always treated as the request.

Letting %%agent work with real files

By default, Deep Agents' files live only in the in-memory conversation state. They are useful as temporary notes, but they are not your project files. You can opt into a rooted filesystem backend when you want %%agent to inspect a project beside the notebook, propose changes to it, or run approved shell commands there.

This is an explicit per-configuration capability. Existing notebooks keep the state backend and receive no host-filesystem access until you select filesystem.

Start read-only

Add this to project, notebook, or global configuration:

[agent.harness.backend]
type = "filesystem"
root_dir = "."

Save the notebook, then start with an observer turn:

%%agent --tools observer
Summarize the Python package beside this notebook. Read files only.

The notebook must be running through a real JupyterLab or Jupyter Notebook server. Shidoshi resolves the workspace from the server's notebook path; an unsaved notebook or a standalone/headless kernel has no trustworthy path and therefore fails closed. The configured root_dir must already exist, must be relative to the notebook's directory, and cannot contain ...

For example, given /work/analysis.ipynb:

The paths seen by the model are virtual POSIX paths. /src/app.py means src/app.py below the resolved workspace, not the host machine's /src. Containment is enforced again at the backend boundary after resolving symlinks; changing path spelling cannot escape the workspace.

Tools and permission rungs

The filesystem capability follows the same --tools rung selected for the rest of %%agent:

This table is additive to the kernel tools described above. In particular, actor can still run Python in the live kernel unless you separately configure an approval for run_code. Provider-hosted web search is also separate and is controlled by --search / --no-search.

The active notebook, Git metadata, .env files, and Shidoshi's own runtime storage are always protected. Add project-specific virtual glob patterns with deny_paths:

[agent.harness.backend]
type = "filesystem"
root_dir = "project"
deny_paths = ["/private/**", "/data/raw/**", "/*.pem"]

Denied paths cannot be read, searched, written, or reached through a symlink. Matching is deliberately case-insensitive so a protected /.env cannot be addressed as /.ENV on a case-insensitive volume. Denies are filesystem-tool boundaries, not shell sandbox rules; an approved shell command is unrestricted host execution.

Approved file changes

Workspace mutation supports two policies:

[agent.harness.backend.filesystem]
write = "ask"  # ask | deny

• ask exposes write/edit tools only at the actor rung. Every proposed mutation pauses before touching disk.
• deny removes write/edit tools entirely.

Autonomous filesystem writes are intentionally unsupported. With ask, the notebook displays an approval card containing the virtual target and proposed diff. You can approve once, reject with guidance, or edit the arguments. A filesystem edit creates a new diff and requires a second, explicit approval; the old card is retired and cannot approve the revised proposal.

Shidoshi snapshots the target when it shows the preview. If the path, file type, existence, or contents change before approval, the action becomes stale and a fresh preview must be approved. This prevents an approval from applying to a different file state than the one you reviewed.

The widget is the normal interface in JupyterLab and Notebook 7. The equivalent command fallback is:

%agent_resume approve --action act-0123
%agent_resume reject use docs/example.toml instead --action act-4567
%agent_resume edit {"file_path":"/docs/example.toml","content":"..."} --action act-0123
%agent_resume continue
%agent_resume abandon

For one pending action, --action may be omitted. When several tool calls are pending together, decide each stable action ID and then run continue; the decisions are resumed in their original order. abandon cancels the whole pending batch without advancing the agent. Another turn on that thread cannot silently replace an unresolved batch—resolve it or abandon it first. A different named thread remains independent.

Use --thread NAME on %agent_resume when the paused action belongs to a named side thread. Approval must resume the same live thread and filesystem root that created the proposal; moving the notebook or changing the root while paused fails closed.

Approved host shell

Shell execution is a separate, actor-only opt-in:

[agent.harness.backend.shell]
enabled = true
execute = "ask"
default_timeout_seconds = 300
max_timeout_seconds = 600
max_output_kb = 100
pass_env = ["PATH"]

execute = "ask" is the only supported execution policy: every command pauses before it starts. The card shows the exact command, parsed compound segments, risk indicators, real working directory, timeouts, output cap, and environment policy. Editing a shell proposal is itself an approving decision; for a multi-action batch it runs only after continue.

This shell is not a sandbox. It runs on the Jupyter host with the workspace as its working directory and can use absolute paths, follow symlinks, access the network, or modify files outside the workspace with the kernel process's permissions. Filesystem deny_paths do not constrain it. Approve only commands you would run yourself in a terminal.

The shell does not inherit the kernel environment. Shidoshi constructs a new environment from the names in pass_env; missing names are simply absent. Credential-shaped names containing KEY, TOKEN, SECRET, PASSWORD, or CREDENTIAL are rejected by configuration validation. Values are never shown in configuration diagnostics or the approval card.

Timeouts are enforced outside model control. Both timeout settings accept 1–3600 seconds, default_timeout_seconds cannot exceed max_timeout_seconds, and an individual command cannot request more than the configured maximum. max_output_kb accepts 1–1024 KiB and truncates captured output after the limit. The defaults are five minutes, ten minutes, and 100 KiB respectively.

Scratch files and large tool results

Deep Agents may offload large intermediate tool results into private scratch storage so the model context stays manageable. Scratch is not part of the workspace: the model cannot list, search, edit, or create deliverables there. It can only read an exact artifact path returned by the harness.

The default location is Shidoshi's per-user state directory, outside the notebook project:

[agent.harness.backend.scratch]
location = "user_state"       # user_state | notebook_local
retention_days = 7            # 1..365
max_size_mb = 256             # 1..1024

Scratch is isolated by notebook and agent thread, created with user-only permissions, capped by max_size_mb, and garbage-collected after the retention window. Explicitly abandoning a pending thread attempts immediate cleanup. The scratch host path is not exposed to the model.

If you need all runtime data beside the notebook, opt in explicitly:

[agent.harness.backend.scratch]
location = "notebook_local"
notebook_dirname = ".shidoshi"
retention_days = 7
max_size_mb = 256

Shidoshi protects that directory from filesystem tools but does not edit your .gitignore. It prints a warning when the directory is not already ignored; add it yourself if the project should not track runtime artifacts:

.shidoshi/

Complete configuration

This example shows every filesystem-backend setting and its default:

[agent.harness.backend]
type = "state"                # state | filesystem
root_dir = "."                # existing path below the notebook directory
deny_paths = []                # virtual glob patterns, additive to fixed denies

[agent.harness.backend.scratch]
location = "user_state"       # user_state | notebook_local
notebook_dirname = ".shidoshi"
retention_days = 7             # 1..365
max_size_mb = 256              # 1..1024

[agent.harness.backend.filesystem]
write = "ask"                 # ask | deny

[agent.harness.backend.shell]
enabled = false
execute = "ask"
default_timeout_seconds = 300  # 1..3600
max_timeout_seconds = 600      # 1..3600; must be >= default
max_output_kb = 100            # 1..1024
pass_env = ["PATH"]

Settings are resolved per turn. The backend type, canonical root, rung, tool inventory, denies, write/shell policy, scratch policy, and limits form part of the compiled-agent cache and thread identity. If one changes during a live thread, Shidoshi refuses to continue with mismatched assumptions; use %%agent --fresh to start with the new configuration or revert the change. --debug shows the resolved host root and effective policy. Normal notebook output shows only a concise scope strip and virtual paths.

Troubleshooting filesystem access

Use shidoshi config show --sources to confirm the effective merged values, shidoshi config validate for schema errors, and shidoshi config schema --html for the generated field-by-field reference.

How context is built

Every prior cell in the notebook — up to the one you're currently running, and accounting for kernel restarts — is turned into conversation history automatically:

• Markdown cells become background text/image context (treated as notes or reference material, not instructions).
• Regular code cells appear as fenced code plus their text/image outputs.
• Prior %ask / %%ask cells become real user/assistant turns. Their responses are reused from a per-cell cache rather than re-sent, so replaying history doesn't resend answers the model already produced.
• %%skip cells are dropped entirely.
• %%pin cells are always kept.

Automatic context-length handling

If a request is rejected for exceeding the model's context window, shidoshi automatically retries, dropping the oldest trimmable history units first (markdown cells, then plain code cells, then whole ask+response pairs — %%pin cells are never dropped), up to 20 times. A banner reports how many cells were dropped so you know context shrank.

Providers

• openai (default) — uses the OpenAI Responses API.
• openrouter — uses OpenRouter's chat-completions API; select it with the provider:model prefix, e.g. openrouter:anthropic/claude-3.5-sonnet.

OpenAI %ask and %%agent always use Responses, even when web search is off. An OpenAI-compatible endpoint therefore needs to implement /responses; Shidoshi rejects providers.openai.api_mode = "chat_completions" instead of silently changing protocols. OpenRouter continues to use Chat Completions.

Web search

Provider-hosted web search is available by default to %ask, %%ask, and %%agent. “Available” does not mean every prompt is searched: the model decides whether current information is needed. A search can add latency and cost, and a model-generated query is sent to the selected provider/search engine.

Turn search on or off

For one turn:

%ask --no-search Explain this function from the notebook
%ask --search What changed in the latest Python release?

%%agent --no-tools --no-search
Review the notebook without inspecting the kernel or accessing the web.

--no-search removes the hosted search tool; it does not make the model call local or offline. The prompt and notebook context are still sent to the chosen LLM provider under its normal data policy.

The explicit model | prompt form keeps everything after | as prompt text:

%ask --no-search openai:gpt-5.5 | Explain what --search means
%agent --search openrouter:z-ai/glm-5.3 | Find the current release notes

For the current kernel session:

%shidoshi config session set tools.web_search.enabled false
%shidoshi config session set tools.web_search.enabled true
%shidoshi config session unset tools.web_search.enabled

Persistently, in a project, notebook, profile, or global config:

[tools.web_search]
enabled = false

Search changes are safe within an existing %%agent thread. Shidoshi rebuilds the model binding and cache policy while retaining conversation memory. A turn paused for approval resumes with the search policy it started with.

What appears in notebook output

• %ask/%%ask show a collapsed tool-activity panel when search activity or citations are returned.
• %%agent shows hosted search calls in the Agent steps panel.
• Source URLs are clickable HTTP(S) links. Provider-supplied titles, queries, and URLs are escaped before rendering.
• A provider can confirm that search occurred without returning citation URLs; Shidoshi labels that state explicitly rather than inventing sources.
• No search panel means search was not observed. It may have been available but unused, or disabled for that turn; use --debug to inspect the effective plan.

include_sources = true asks OpenAI for the complete consulted source list. OpenRouter has no equivalent request switch; routed models may emit inline citations regardless of this setting. No provider guarantees citation URLs.

Common policy

Portable settings apply to either provider:

[tools.web_search]
enabled = true
search_context_size = "medium"
allowed_domains = ["python.org"]
# excluded_domains = ["example.com"]
include_sources = true

[tools.web_search.location]
# country = "IN"       # two-letter ISO country code
# region = "Maharashtra"
# city = "Mumbai"
# timezone = "Asia/Kolkata"

Domains must be bare hostnames: omit schemes, paths, ports, credentials, and wildcards. Up to 100 allowed and 100 excluded domains may be configured.

OpenAI controls

[tools.web_search.openai]
external_web_access = true
return_token_budget = "default"

• external_web_access = false keeps the search tool available but limits it to OpenAI's cached/indexed content. It is not the same as enabled = false.
• return_token_budget is "default" or "unlimited"; it is intended for GPT-5+ reasoning search and may increase latency and cost.

OpenRouter controls

[tools.web_search.openrouter]
engine = "exa"
mode = "instant"
max_results = 5
max_uses = 3
max_total_results = 15
# max_characters = 5000

Supported engines are auto, native, exa, firecrawl, parallel, and perplexity. Modes are engine-specific:

Some native/auto capabilities depend on the routed model. Shidoshi reports a warning when it cannot prove that a configured filter or limit will be honored. Known impossible combinations fail before a request is sent.

Trying newly released provider fields

Provider APIs evolve faster than Shidoshi's named schema. Search-specific escape hatches let you try a new field without waiting for a release:

[tools.web_search.openai.extra_tool_fields]
future_option = true

[tools.web_search.openrouter.extra_parameters]
future_option = true

Shidoshi validates that these are finite, JSON-compatible values, but cannot validate their provider semantics. They cannot override a field Shidoshi already manages. They are included in debug output, warnings, cache policy, and model fingerprints. If a provider later removes or rejects a field, its API error is surfaced; there is no reliable automatic capability negotiation.

These differ from providers.openai.extra_body and providers.openrouter.extra_body, which are for new request-level fields rather than fields inside the web-search tool.

web_fetch is a separate tool and remains off by default. Enable it with tools.web_fetch.enabled = true.

Debug mode

Add --debug to %ask, %%ask, or %%agent to render the effective search plan, serialized provider tool, composed prompt, messages, and raw stream events. This is the fastest way to distinguish “available but unused” from “disabled” and to inspect provider extension fields.

Search troubleshooting

Configuration

shidoshi is configured through layered TOML files — global, project, and per-notebook — merged in that order, with a per-invocation override always winning. Nothing lower is ever silently overridden without a trace: shidoshi config explain <key> tells you exactly which file set the value you're seeing.

Where config lives

A project or notebook config isn't found by convention alone — write one:

shidoshi config init --project            # a short starter file
shidoshi config init --project --full     # every available setting, with defaults filled in

(or %shidoshi config init --project from a notebook cell — every shidoshi config ... command below also works as %shidoshi config ....)

The essentials

config_version = 1

[defaults]
profile = "default"

[profiles.default]
model = "openai:gpt-5.5"

[profiles.default.generation]
reasoning = { effort = "low" }

[agent]
reasoning_effort = "none"  # fast default; raise for harder agent tasks
permission_rung = "actor"   # none | observer | proposer | actor

[tools.web_search]
# Available by default. Search may add latency/cost and sends a generated query
# to the configured provider. Add --no-search to a magic for one no-search turn.
enabled = true

The old flat keys (default_model, agent_model, agent_tools, reasoning_effort, ask_color, skip_color) still work — shidoshi migrates them automatically — but the nested form above is what shidoshi config init writes now, and it's the only shape that can express everything else on this page (profiles, tool policy, context policy, approvals).

Inspecting and editing config

shidoshi config path                # which files are actually in effect, in precedence order
shidoshi config show                # the fully merged, effective config
shidoshi config show --sources      # ...annotated with which file set each value
shidoshi config explain agent.permission_rung   # one setting's value and where it came from
shidoshi config validate            # parse/schema-check every layer, and flag pasted-in secrets
shidoshi config doctor              # validate, plus orphaned sidecars and missing credential env vars
shidoshi config doctor --online     # opt in to current provider metadata checks
shidoshi config set agent.permission_rung observer --project
shidoshi config unset agent.permission_rung --project
shidoshi config schema              # the complete field-by-field surface, as JSON Schema
shidoshi config schema --html       # ...or a single offline, browsable reference page

Each of these also runs as %shidoshi config ... inside a notebook cell, with two differences: edit prints the resolved path instead of spawning $EDITOR (a blocking subprocess doesn't belong in a kernel cell), and a bare call with no --global/--project/--notebook flag automatically includes the current notebook's own sidecar.

Kernel-local overrides are explicit and inspectable:

%shidoshi config session set generation.temperature 0.2
%shidoshi config session show
%shidoshi config session unset generation.temperature
%shidoshi config session clear

For one call only, repeat --set key=value on %ask/%%ask/%agent/ %%agent, or use --profile NAME. Search also has the clearer --search and --no-search aliases. The whole override is parsed and validated before a request, checkpoint, or pending approval is touched.

Provider policy is translated separately for raw OpenAI Responses, ChatOpenAI, raw OpenRouter, and ChatOpenRouter. In particular, OpenRouter's default data_collection = "deny" and require_parameters = true are sent on the request; custom OpenAI-compatible endpoints must implement the Responses API for the OpenAI adapter. They can choose instruction_role explicitly. Configuration display and diagnostics recursively redact credentials, authorization headers, secret-like metadata, and URL query values.

For newly released provider request parameters that Shidoshi does not yet name, use the provider-scoped kwargs-style escape hatch:

[providers.openai.extra_body]
future_option = true

[providers.openrouter.extra_body]
future_option = true

The same table can be set under a profile's providers section. These values reach both the raw and LangChain-backed request paths. Shidoshi rejects keys that collide with request fields it owns (model, tools, reasoning, storage, routing/privacy policy, and limits), recursively redacts the passthrough in diagnostics, and includes it in model-build fingerprints. extra_body is for JSON request-body parameters; arbitrary Python constructor objects remain Python extension points because TOML cannot represent or validate them safely.

What's configurable

Model and generation settings (per named profile — select one with defaults.profile, or ask.profile/agent.profile to use a different one for each magic), provider transport, which tools %ask/%%agent may use (tools.web_search, tools.web_fetch, tools.create_cell, kernel-tool output limits and secret redaction), notebook context policy (what gets included, image handling, trim limits), local response caching, and %%agent's permission model — including, if you want it, a real human-in-the-loop approval gate:

context.include_markdown, include_code, include_outputs, and include_images control notebook history. The nested image policy (allow_local_files, allow_remote_urls, and max_file_bytes) applies to images in both the current prompt and notebook history.

[agent.approvals.run_code]
enabled = true
allowed_decisions = ["approve", "edit", "reject"]

With this set, an actor-rung run_code call pauses instead of running immediately. Answer it with %agent_resume approve, %agent_resume reject [reason], or %agent_resume edit <json args>.

Real project files and approved host-shell execution are configured under agent.harness.backend. They are opt-in and have their own rooted path, protected-path, scratch, timeout, environment, and approval policies. See Letting %%agent work with real files for the complete guide and copy-paste configuration.

shidoshi config schema --html is the complete reference — every setting, its type, and its default, generated straight from the schema shidoshi itself validates against, so it can never drift out of date. Fields that cannot yet be enforced safely (for example durable agent persistence or host-backed skill/memory paths) fail validation with a precise reason; shidoshi does not accept them as inert promises.

Development

uv sync
uv run pytest tests/unit tests/btp -v

Integration tests under tests/integration/ require a live OPENAI_API_KEY (or a proxy via OPENAI_BASE_URL) and are run with:

uv run pytest tests/integration/ -v -m integration

License

Apache License 2.0 — see LICENSE.