superdoc-custom-actions

name

superdoc-custom-actions

description

Author a custom SuperDoc LLM-tools "action" — a named, deterministic document-edit verb that plugs into the core preset's superdoc_perform_action tool. Use when the user wants to add a new action (e.g. "build an action that adds footnotes / bolds table borders / stamps a banner") or asks how to extend SuperDoc's LLM tools with their own document operation. Works with installed packages only — no SuperDoc source checkout needed or expected.

Build a SuperDoc custom action

A custom action is a named, namespaced verb (superdoc.<verb>, e.g. superdoc.add_footnote) that a customer authors once and exposes to an LLM through superdoc_perform_action, alongside the 40 built-in core actions. An action is deterministic application code, not a prompt.

Your job when this skill is active: turn a plain-language request into a correct, registered, live-verified ActionSpec, plus a test. Never hand back an action you haven't executed against a real document.

---

1. The canonical ActionSpec — two execution tiers

ActionSpec = {
  name:        "superdoc.<verb>"       # namespaced; must not collide with built-ins
  description: string                  # shown to the model: what it does + when to use it
  input:       JSON Schema (object)    # the action's FLAT args; defaults are honored
  # exactly ONE of:
  steps:  [ {action, args}, ... ]      # TIER 1 — declarative (PREFER THIS)
  run:    (doc, args) -> result        # TIER 2 — native function in your process
}

Tier selection rule — always try in this order:

1. steps — can the goal be expressed as a sequence of built-in core actions? Then it MUST be steps: you inherit target resolution, placement handling, receipts, and per-step verification that the built-ins already guarantee. Steps are pure data (shippable as JSON) and behave identically from Node and Python.
2. run — the goal needs an API domain the built-in actions don't cover (footnotes, headers/footers, table borders, bookmarks…) or computed logic. A native function in YOUR language, called with the typed session-bound doc handle.

Tier 1 example — verified working

from superdoc import define_action

stamp = define_action(
    name='superdoc.stamp_confidential',
    description='Insert a confidentiality banner at the top and flag it with a review comment.',
    input_schema={'type': 'object', 'properties': {
        'label': {'type': 'string', 'default': 'CONFIDENTIAL'},
    }, 'required': []},
    steps=[
        {'action': 'insert_paragraphs',
         'args': {'texts': ['{{label}}'], 'placement': {'at': 'document_start'}}},
        {'action': 'add_comments',
         'args': {'selectors': [{'kind': 'textSearch', 'terms': ['{{label}}'], 'occurrence': 1}],
                  'commentText': 'Auto-stamped. Verify distribution before sending.'}},
    ],
)

Templating: "{{label}}" as a WHOLE string substitutes the raw value (arrays and objects survive); "Stamped: {{label}}" interpolates as text. A caller's changeMode flows into every step that doesn't pin its own. Receipts aggregate per step: succeeded / partial (some steps landed — failedStep says which and why) / failed.

Tier 2 example — verified working

Take MODEL-FRIENDLY args (anchor text, ordinals — things an LLM can actually produce) and resolve low-level targets INSIDE the action. Never expose raw TextTarget/ref shapes as action args.

def _add_footnote(doc, args):
    # Resolve anchor text -> a zero-width TextTarget right after the match.
    # NOTE: use includeText + block['text'] — the default textPreview field is
    # truncated to 80 chars, so anchors later in a paragraph are never found.
    # On large documents, paginate with offset/limit (see examples/reference_pack.py)
    # — one unpaginated call may not return every block.
    needle = args['anchorText']
    for block in doc.blocks.list({'includeText': True})['blocks']:
        text = block.get('text') or ''
        pos = text.find(needle)
        if pos >= 0:
            end = pos + len(needle)
            at = {'kind': 'text', 'segments': [{'blockId': block['nodeId'], 'range': {'start': end, 'end': end}}]}
            result = doc.footnotes.insert({'at': at, 'type': args.get('type', 'footnote'), 'content': args['content']})
            after = doc.footnotes.list({})   # re-inspect: prove the footnote exists
            return {'inserted': result, 'footnoteCount': after.get('total', len(after.get('items', [])))}
    raise ValueError(f'anchorText {needle!r} not found in any block')

add_footnote = define_action(
    name='superdoc.add_footnote',
    description='Insert a footnote whose marker lands right after the given anchor text. args: {anchorText, content, type?}.',
    input_schema={'type': 'object', 'properties': {
        'anchorText': {'type': 'string'}, 'content': {'type': 'string'},
        'type': {'type': 'string', 'enum': ['footnote', 'endnote'], 'default': 'footnote'},
    }, 'required': ['anchorText', 'content']},
    run=_add_footnote,
)

Sync vs async — match the host app's doc handle (this trips agents). The example above is synchronous. But if the app dispatches through dispatch_async / drives an AsyncSuperDocClient (a very common real case — check the app first), your run MUST be async def and MUST await every doc.* call. The kit awaits the value your run returns, but it does NOT make the inner doc.* calls awaitable for you — a sync body on an async handle gets un-awaited coroutines and fails with 'coroutine' object is not subscriptable. Async variant of the body above:

async def _add_footnote(doc, args):
    needle = args['anchorText']
    for block in (await doc.blocks.list({'includeText': True}))['blocks']:
        pos = (block.get('text') or '').find(needle)
        if pos >= 0:
            end = pos + len(needle)
            at = {'kind': 'text', 'segments': [{'blockId': block['nodeId'], 'range': {'start': end, 'end': end}}]}
            result = await doc.footnotes.insert({'at': at, 'type': args.get('type', 'footnote'), 'content': args['content']})
            after = await doc.footnotes.list({})
            return {'inserted': result, 'footnoteCount': after.get('total')}
    raise ValueError(f'anchorText {needle!r} not found in any block')

run receipts are synthesized for you: preRevision/postRevision, and on failure partialMutation + a recovery hint. Raise on failure — the exception becomes a truthful failed receipt with shape {status:'failed', errors:[{code, message}], partialMutation, recovery} — the message lives at errors[0].message (NOT a top-level error); assert on that in your tests. The examples/ directory next to this file contains complete, live-verified packs (footnotes + page headers, table borders) authored exactly this way (in the synchronous form — apply the async rule above if the host is async).

---

2. Register + advertise

Simplest path — hand your actions to the toolkit (no preset to manage):

from superdoc import create_agent_toolkit

kit = create_agent_toolkit({'provider': 'openai', 'actions': [stamp, add_footnote]})
# kit['tools'] enum + kit['system_prompt'] now include your actions;
# kit['dispatch'] / kit['dispatch_async'] route them. No register_preset, no id.

create_agent_toolkit({actions}) builds the extended preset internally over core and returns tools + system prompt + dispatch that already agree; provider shaping (anthropic/openai/vercel/generic) is automatic. (Node: createAgentToolkit({ provider, actions }).) Pass base to extend a preset other than core, or includeCoreActions to keep only a subset of the built-ins alongside yours. Schema enum, tool description, system prompt, and dispatch move together — coherence is the kit's job, not yours.

Dispatch args are FLAT — {action, ...yourArgs}, never a nested args object:

# async host (this uses dispatch_async); a sync host uses kit['dispatch']
receipt = await kit['dispatch_async'](doc, 'superdoc_perform_action',
    {'action': 'superdoc.add_footnote', 'anchorText': 'the report', 'content': 'See appendix.'})
# NOT {'action': ..., 'args': {...}} — a nested `args` raises
# "Unknown argument(s) for superdoc_perform_action: args".

Integrating into the host app — find its toolkit seam. Real apps differ; read the app before writing files:

• The common case: the app already calls create_agent_toolkit(...) / createAgentToolkit(...) (often in one startup function). Just add your actions to THAT call — one import + one arg. No register_preset, no PRESET constant, no preset id threaded through dispatch.
• If the app has an action-discovery convention (an actions/ folder it auto-loads, an ACTIONS list it imports), follow it and feed that list into the toolkit call. Do not add scaffolding, CLI flags, or discovery frameworks the app doesn't have.

Advanced — a named, registered preset. Only if the app uses the standalone functions (choose_tools / get_system_prompt / dispatch_superdoc_tool) or genuinely needs a reusable registered preset:

from superdoc import extend_preset, register_preset

register_preset(extend_preset('core', id='custom_superdoc_preset', actions=[stamp, add_footnote]))
# then pass preset='custom_superdoc_preset' to choose_tools / get_system_prompt / dispatch_superdoc_tool

Reuse the app's existing extended-preset id if it has one; otherwise the conventional default is custom_superdoc_preset — not an invented codename.

---

3. Discovery — you have no source access, and you don't need it

Built-in action names + args (decides steps-vs-run):

from superdoc import choose_tools
perform = next(t['function'] for t in choose_tools({'provider': 'openai', 'preset': 'core'})['tools']
               if t['function']['name'] == 'superdoc_perform_action')
print(perform['parameters']['properties']['action']['enum'])   # the 40 built-ins
print(sorted(perform['parameters']['properties']))              # every advertised arg
print(perform['description'])                                   # grouped hints per action

Native doc.* surface (for the run tier) — the client self-describes; use that FIRST, it beats trial-and-error:

from superdoc import SuperDocClient
with SuperDocClient() as client:
    contract = client.describe()                       # ALL operations (~439) with ids like 'doc.tables.setBorders'
    print(sorted({op['id'].split('.')[1] for op in contract['operations']}))   # the namespaces
    spec = client.describe_command({'operationId': 'tables.setBorders'})       # full input contract for one op
    print(spec['operation'])

Then confirm against a live session on a COPY of a sample doc:

    doc = client.open({'doc': 'work-copy.docx'})
    print([m for m in dir(doc.tables) if not m.startswith('_')])   # methods (snake_case + camelCase both work)
    print(doc.blocks.list()['blocks'][:3])                          # rows: ordinal, nodeId, nodeType, textPreview, ref
    doc.close({'discard': True})

Note: dir(doc) does NOT list the namespaces (they're dynamic attributes on the Python handle) — client.describe() is the namespace directory; dir(doc.<ns>) works once you know the namespace.

Let errors teach you — but confirm params with describe_command, never with a successful call. Most inputs are validated against the real contract and error messages enumerate the valid unions. Two traps, both observed live:

• Some ops silently drop unknown params: headerFooters.parts.create returns success: true for {'content': ...} while writing nothing — the success receipt lies about the param having done anything.
• Union error messages can mislead: blocks.list({'in': ...}) prints in.kind must be one of: "story","story","story" and storyType must equal "body" even though richer story kinds are valid. Only describe_command({'operationId': 'blocks.list'}) shows the real union.

Budget for iteration: some ops reveal ONE missing required field per probe (tables.setBorders took six rounds that way — describe_command gives the whole contract in one).

Stories (header/footer content). Body is only one story of a document. Header/footer text is written with ordinary content ops pointed INTO a story: doc.create.paragraph({..., 'in': {'kind':'story','storyType':'headerFooterPart','refId': ...}}), or a slot story with onWrite: 'materializeIfInherited' (auto-creates and wires the part). Wiring evidence comes from doc.headerFooters.get(...) (refId/isExplicit); header/footer TEXT has no in-session read-back (blocks.list returns body blocks regardless of in, doc.get omits header parts) — save to a file and inspect the zip (word/header1.xml) for text-level evidence.

Finding table/row/cell handles: doc.blocks.list() rows with nodeType == 'table' carry the table nodeId; doc.tables.get_cells(...) / doc.tables.get_properties(...) take it from there (probe their input shape with a small call and read the validation error).

---

4. Verification — two layers; degrade honestly, never silently

Layer 1 — contract validation. ALWAYS, no document needed. Every operation your action calls must be confirmed against client.describe_command({'operationId': ...}): the op exists and your arg shapes match its contract. Never confirm a param with a successful call — some ops silently drop unknown params and report success anyway. This layer is seconds of work and catches most authoring mistakes.

Layer 2 — live semantic verification. When a document is available. A run on a COPY of a real document showing a succeeded receipt AND independent evidence the document actually changed (re-inspect: doc.blocks.list() shows the banner at ordinal 0, doc.footnotes.list({}) count went up, borders read back different; for header/footer text, saved-zip XML).

Finding a document for layer 2, in order:

1. Glob the workspace for *.docx, skipping outputs (out*.docx), Word lock files (~$*), and .venv/node_modules/state dirs. Prefer conventional homes first: sample_docs/, fixtures/, tests/data/.
2. Check suitability, not just existence: open a COPY, one countsOnly inspect — does it contain what the action targets (images, footnotes, tables)? If not, can you PLANT the conditions on the copy first (built-in actions or raw ops, e.g. doc.footnotes.insert before testing a renumber action)?
3. No usable document? Open a blank one — built in. client.open({}) (no doc at all) starts a session from the SDK's embedded blank document, in both languages; doc.save({'out': ...}) materializes it as a real file. PLANT the conditions your action needs on it (built-in actions or raw ops, e.g. create_table, doc.footnotes.insert), then verify against it.
4. Conditions you cannot synthesize (images, embedded objects, messy real-world structure) and a human is present: ask the user for a representative document. This is the one legitimate reason to stop and ask.
5. Nobody to ask (headless/CI) and synthesis impossible: deliver anyway, as a DRAFT — do not block, do not fabricate.

A generated-fixture verification is real layer-2 evidence for the mechanics, but say so in your summary ("verified against a generated fixture") — it proves the action works, not that it handles the customer's real documents.

Draft labeling — when layer 2 did not run: the action file's docstring AND your final message must say plainly: "DRAFT — contract-validated, not yet executed against a document." Still write the pytest (see below) so the customer's first real run IS the verification. Never present an unexecuted action as verified; a false "tested ✓" is worse than no action at all.

A test is required in BOTH modes — in tests/, following the existing suite's pattern; if the app has NO test suite yet, create tests/ plus a conftest that wires the environment the same way the app does (e.g. read SUPERDOC_CLI_BIN from the app's .env). Tests open a fresh doc (or tmp copy), dispatch through the toolkit (or a registered preset), and assert receipt fields plus re-inspected evidence. Failure honesty matters — if your action can partially apply, test that path too.

Two receipt/fixture facts your tests must account for:

• Status vocabulary differs by layer: custom-action receipts report succeeded/partial/failed, but dispatching a BUILT-IN action directly (e.g. while planting conditions) returns status: "ok". Accept both where appropriate.
• A truly blank document interleaves empty paragraphs around inserted texts (['', 'AAA', '', 'BBB']), so tail/count assertions against the bare embedded blank doc are unstable — plant body content first, then assert.

Never claim success from status: 'succeeded' alone — receipts synthesized for run-tier actions confirm your code returned, not that the effect you intended landed. Re-inspect.

Clean up when you finish. Verification leaves scratch behind — delete it so the only new files are the action module and its test. Remove: any .docx you opened/saved/generated for verification, session-state dirs the host created (e.g. .superdoc-state/, .superdoc-cli-state/), a stray out.docx, and __pycache__/ from test runs. Do NOT delete the app's own sample documents, seed assets, or anything that existed before you started. When in doubt, list what you created and remove exactly that.

"Before" evidence can legitimately be ABSENT: style-inherited formatting reads back with no explicit property at all (e.g. a TableGrid-styled table has no borders key until direct borders are set). Compare "absent → explicit value", don't assume a numeric before-state exists.

Read-back may NORMALIZE identifiers (e.g. a bookmark inserted with one blockId can read back with a normalized id). Verify by stable evidence — names, offsets, counts, saved-XML content — not by identifier equality.

---

5. Naming, schema, and description discipline

• Names: superdoc.<verb> — dots are valid in the merged enum. (Only standalone mode forbids dots, because provider tool names must match ^[A-Za-z0-9_-]{1,64}$.)
• Never collide with built-in names — the kit rejects it.
• description is model-facing: one sentence on WHAT + one on WHEN, and name the args inline (models call better with args: {at, content} spelled out).
• Every arg gets a JSON-Schema property; put default on optionals — defaults are applied on all tiers. Mark true requirements in required.

6. Pitfalls (each of these bit a real implementation)

• Reusing a built-in arg name: custom actions merge into the ONE superdoc_perform_action schema, so a shared arg name maps to a single schema. Reusing a built-in name (anchorText, text, caseSensitive, …) is fine only if your schema matches the built-in's exactly except for description — a differing type, enum (including one side having one), default, limit, or pattern is a real conflict and the kit rejects it (the merged surface can advertise only one schema per name). A description-only difference is allowed, but the built-in's schema wins the merge so your description is not advertised — use a distinct arg name if you need it. Cross- check with §3's sorted(perform['parameters']['properties']).
• Async host apps (very common — inspect the app first): if it dispatches via dispatch_async or an AsyncSuperDocClient, your run function MUST be async def and await every doc.* call. A synchronous body on an async handle hands back un-awaited coroutines and fails with 'coroutine' object is not subscriptable. (See §1's async variant.) A synchronous app is the mirror: a plain def run.
• Do not insert content via raw ops in run actions — e.g. doc.create.paragraph({placement: 'before', ...}) silently IGNORES placement and appends at document end. The built-in insert_paragraphs action resolves placement correctly — compose it via steps instead (or mix: a steps action for the insert, run for the rest).
• Comment anchoring: doc.comments.create with a {text: ...} target can return success without a visible comment. Use the built-in add_comments action (selector-resolved, verified) via steps.
• Two failure shapes: argument-validation errors THROW; runtime failures return {'status': 'failed', ...} receipts. Wrap dispatch in try/except and handle both.
• Refs expire on mutation. Any ref you fetched before an edit is invalid after it — re-list/re-search instead of caching handles across mutations.
• Python handles return plain dicts and accept both snake_case and camelCase method names; results use camelCase keys (textPreview, nodeId).
• Some open string inputs are NOT enum-validated by the host (e.g. tables.setBorders persists any lineStyle string verbatim — zigzag round-trips). Constrain such inputs with an enum in YOUR action's input schema so the model can't invent values.
• insert_paragraphs rejects empty strings in texts — you cannot compose a blank spacer paragraph declaratively; a "" entry fails the whole action. Design steps-tier layouts without empty-line spacers.
• Sandboxed agent environments: the SDK's host bootstrap may chmod its bundled binary and write state under ~/.superdoc-cli — both can be blocked by an agent sandbox. If client.open fails on permissions, set SUPERDOC_CLI_BIN to the installed binary and SUPERDOC_CLI_STATE_DIR to a writable path inside your workspace.

7. Worked examples

The examples/ directory next to this file contains complete, live-verified action packs (declarative steps, footnotes/page headers, table borders) with their receipt-synthesis and verification patterns — in Python and JavaScript (policy_pack.py / policy_pack.mjs are twins; the kit API maps 1:1 across languages). Pattern-match them. There is no bundled API reference by design — the live contract from client.describe() / client.describe_command(...) (§3) is always current; a bundled copy would drift.