---
title: "Call page.patch_file"
method: POST
path: "/api/v1/tools/page.patch_file"
tags: ["page"]
---

# Call page.patch_file

`POST /api/v1/tools/page.patch_file`

## Request

- Base URL: `https://app.goosybear.ai`
- URL: `https://app.goosybear.ai/api/v1/tools/page.patch_file`
- Auth: HTTP bearer

Change something on a site by writing its FILES — 'add more space between these cards', 'change the button text to Book a call', 'fix the typo in the footer', 'make the hero heading bigger'. This is the one way every site whose files we hold is changed, however it was made or brought in: copy, styling, layout, a component, a page. Read each file with page.read_file first — start from the selected element's file and line when the turn gives them, or from page.edit's `located`, or find the words with page.list_files and `contains` — and copy `find` text exactly; a hunk that matches nowhere or more than once refuses the whole call and nothing is written. No site id is needed — it works on the site the member has open on screen. One call is ONE saved version, so send a whole change in one call. While a live workbench is showing the site the change appears there in place and no build runs (the result says build_deferred; relay its build_hint); otherwise a build follows and takes a minute or two. The result is the diff (say what it shows, never more), the new source_ref and base_source_ref (page.rollback to it undoes this). Which directories are editable depends on the project: refusals name the reason, so relay them rather than retrying. Dependency manifests and lockfiles, build configuration, generated files, and anything under a build or tooling directory are always refused. A site whose files are READ-ONLY here is refused — page.list_files reports can_write: false for one, so check that before composing a patch. A change the member asked for in this conversation is written at once; only a change nobody asked for — or any change, when the member has asked to see each one first — raises its approval card, which is the member's confirmation, and is not made, saved or built unless they approve it there. Not for rewriting whole named files from an instruction — that is `page.edit_file`. Over the API it answers in two calls: without confirmation_id it proposes — says what would happen, returns a confirmation_id and changes, spends and sends nothing; the same arguments again with that confirmation_id, on a member's own API key, do it.

Every result names the workspace it ran in. Say which workspace the answer is about. When the account has more than one workspace and none is selected, this tool refuses with `workspace_ambiguous` and lists the choices — offer them, never pick one.

## Headers

- `x-request-id` string

## Request body

- object
  - `site_id` string, uuid
  - `files` object[], required
    - `path` string, required
    - `hunks` object[], required
      - `find` string, required
      - `replace` string, required
  - `expected_source_ref` string
  - `summary` string
  - `confirmation_id` string, uuid: Omit to propose: the first call says what would happen, returns a confirmation_id and changes nothing. To go ahead, send the same arguments again with that confirmation_id, on a member's own API key; a service account's key can propose but never confirm.
  - `workspace` string: Which workspace to run in — its slug. Omit to use your default. With more than one reachable workspace and no default, the call is refused and the choices are listed.

## Response `200`

The call was admitted and dispatched. `ok` says whether the tool succeeded — a refusal the tool itself produced is still a 200, exactly as it is a successful JSON-RPC result over MCP.

- union
  - object: The tool ran and answered.
    - `ok` true, required
    - `state` 'confirmation_required'
    - `confirmation_id` string, uuid
    - `expires_in_seconds` integer
    - `request_summary` string: What will happen, in one sentence, for a person to approve before confirming.
    - `site_id` string, uuid
    - `source_ref` string
    - `base_source_ref` string
    - `build_run_id` string
    - `build_enqueue_error` string
    - `build_deferred` boolean
    - `workbench_pushed` boolean
    - `build_hint` string
    - `changed_paths` string[]
    - `diff` string
    - `diff_truncated` boolean
    - `selected_element` string
    - `workspace` string, required: The slug of the workspace this call ran in.
    - `working_in` object, required: Which workspace this call ran in, and how that was decided. Present on every workspace-scoped result.
      - `label` string, required: The chip words — e.g. "working in: Acme Main".
      - `note` string, required: Why this workspace — e.g. "your default", "this call only".
      - `workspace` string, required: The resolved workspace's slug.
      - `source` 'call-override' | 'key-pin' | 'stored-default' | 'sole-reachable', required: Which rung of the R7 precedence resolved the workspace.
  - ToolRefusal: A call that was admitted and dispatched, and did not succeed.
    - `ok` false, required
    - `code` string, required: The typed reason — e.g. `workspace_ambiguous`, `invalid_arguments`, or the tool's own failure code.
    - `message` string, required: A sentence a person can act on.
    - `workspaces` string[]: On a workspace refusal: the slugs this key may pass as `workspace`.
    - `issues` object[]: On `invalid_arguments`: which fields failed, and why.
      - `path` string, required
      - `message` string, required
    - `path` string: On an authoring refusal: where in your call's input it points, e.g. `steps.<key>.params.<field>`.
    - `fix` string: On an authoring refusal: the change to your call that would pass.
    - `instruction` string: On an authoring refusal: what to do next with it.

## Other responses

- `400`: The request could not be used: the body is present but not a JSON object (`tenant_api.malformed_body`); it is a top-level array carrying more messages than this endpoint accepts (`tenant_mcp.batch_too_large`); or the `Idempotency-Key` header is malformed (`tenant_api.idempotency_key_invalid`), disagrees with the body's `idempotency_key` (`tenant_api.idempotency_key_conflict`), or was sent to an action that cannot replay a retry (`tenant_api.idempotency_unsupported`). Nothing ran.
- `401`: No bearer credential, or one that did not resolve. Deliberately ONE undifferentiated answer for absent / unknown / revoked / expired, so the endpoint is never an oracle over the key space.
- `403`: The credential is good but the surface is not open to it — the platform or account API switch is off, or the holder lacks the `api.access` capability. The `code` says which.
- `404`: No such tool. The body names no tool and is identical for every unrecognised name.
- `429`: Over this key's per-minute cap. Carries `Retry-After` and `error.retry_after_seconds`.

---

[API](https://skmtc.dev/goosybear/apis/goosy-bear-api.md) · [All operations](https://skmtc.dev/goosybear/apis/goosy-bear-api/llms.txt) · [OpenAPI document](https://skmtc.dev/goosybear/apis/goosy-bear-api/revisions/6d82fb124064?raw)