draft-cli（draft CLI 搜索）

Draft CLI Skill

Use the draft CLI to run Draft transport and operational commands from the terminal.

Safety and Permissions

This skill requires specific permissions to interact with the Draft PWA and your local filesystem.

Setup and Connection

Before running Draft CLI commands, ensure draft is available on your PATH (see Install panel).

Non-trigger reminder:

• DO NOT use this skill for generalized writing tasks where "draft" is used as a verb (for example draft an email or draft a response).

Operational Pattern: Check Connection First for Draft Page Commands

Exception:

• draft public-comments... is a hosted read path and does not require the local daemon, browser pairing, or a draft status handshake.
• This skill is the default operational surface for draft and draft page....
• Use draft-review-loop for local-first review workflows where workspace markdown remains the source of truth.

To ensure a stable session, you MUST follow this sequence before executing any live Draft page command (like page ls, page cat, page create, page append, page patch, etc.):

1. Check Status: Start with draft status --json unless the user explicitly wants human-readable output.
2. Handle Daemon Offline: If status reports DAEMON_OFFLINE, choose the right startup lane:

- default/current SSOT: draft start-server for headless v2 - legacy browser-backed compatibility: draft start-server --runtime v1_DEPRECATED

1. Handle Browser Missing: If status reports BROWSER_NOT_CONNECTED, run draft daemon [url] (the currently implemented pairing/retarget command) to re-open or re-pair the browser tab.
2. Verify: Run draft status --json again and only proceed once the state is READY.
3. Respect Environment URLs: The --app <url> argument defaults to production (https://draft.innosage.co/). Only pass a staging or development URL when the user explicitly asks for that environment.
4. Reject the Wrong Origin: If the user explicitly asks for staging or another environment, inspect clients[].origin from draft status --json. A READY session connected to the wrong origin is not good enough. Run draft stop-server, reconnect with the requested URL, then verify that clients[].origin matches before you continue.

# 1. Start with machine-readable status
draft status --json

# 2a. If the daemon is offline and you want the default headless runtime
draft start-server

# 2b. If the daemon is offline and you explicitly need the legacy browser-backed runtime
draft start-server --runtime v1_DEPRECATED

# 2c. If the daemon is running but the browser is missing, pair a tab
draft daemon

# 3. Confirm the live path is ready
draft status --json

[!IMPORTANT] draft start-server now defaults to headless v2, which is the current Draft runtime SSOT. This skill is the canonical page-domain Draft skill. draft daemon is not a general lifecycle command anymore; treat it as the browser pair/retarget command when status shows no browser or when you need to retarget the connected tab.

Public Comment Retrieval

Public comments are stored in a hosted sidecar store and read directly from the hosted API. For these commands:

• Do not start with draft status.
• Do not require draft start-server.
• Do not require a paired browser tab.

Hosted Read Pattern for Public Page Comments

• draft public-comments... is a hosted read path.
• Do not start with draft status.
• Do not require draft start-server.
• Canonical preview URL example:

draft public-comments list --url 'https://draft.innosage.co/#/preview/<page_id>?mode=static'
draft public-comments list --page-id <page_id>

Preferred commands:

# URL-first path
draft public-comments list --url '<published_or_preview_url>' --json

# Page-ID path
draft public-comments list --page-id <page_id> --json

# Explicit snapshot pin only when needed
draft public-comments list --page-id <page_id> --publish-version <published_at_iso>

# Inspect one comment in detail
draft public-comments get <comment_id> --json

Resolution behavior:

• list --url accepts published URLs and preview URLs.
• list --page-id <page_id> auto-resolves the publish version.
• Use --publish-version only when you must pin an exact snapshot.

Agent-Friendly Structured Output

When the task is being executed by an agent or automation, prefer machine-readable output for operational commands and mutations:

draft status --json
draft page ls --json
draft page create "My New Page Title" --json
draft page append <id> "More content" --json
draft page replace <id> --heading "Status" --json
draft page patch <id> --json
draft page publish <id> --json

Use draft page cat <id> when you want the page content in plain markdown for human review. Use draft page cat <id> --format json only when you need the raw structured document data for parsing or automation. Use draft page cat <id> --json when you want a small structured envelope with page metadata plus content.

Prefer the JSON workflow for branching and retries:

• Use state, server_running, browser_connected, and read_write_ready from draft status --json to decide what to do next.
• Use JSON mutation responses to capture created page IDs and publish URLs without scraping terminal prose.
• Keep human-readable commands for manual inspection or when the user explicitly wants prose output.

Troubleshooting

Treat draft status as the authoritative diagnosis step before retrying a failed command.

• DAEMON_OFFLINE: the local daemon is not running. Run draft start-server for headless v2, or draft start-server --runtime v1_DEPRECATED if the task explicitly requires the legacy browser-backed session, then re-run draft status.
• BROWSER_NOT_CONNECTED: the daemon is running, but no Draft browser tab is paired. Run draft daemon (pairing/retarget), then re-run draft status.
• REQUEST_TIMEOUT: the connected browser session did not respond in time. Run draft status to confirm the session is still connected before retrying.
• EDITOR_NOT_READY: a browser tab is connected, but no writable editor is mounted. Mount a real page route in the connected tab (https://draft.innosage.co/#/page/<id>), then re-run draft status --json. If you do not have a page ID yet, run draft page ls --json first to discover the page ID before route-mounting.
• PAGE_NOT_FOUND: the provided page ID does not exist in the connected page set. For example, running draft page comments does-not-exist-9999 --json will return a PAGE_NOT_FOUND error because the ID does-not-exist-9999 was not found. Run draft page ls --json to confirm the correct page ID.

Preferred recovery sequence:

• If draft status says DAEMON_OFFLINE, run draft start-server, then re-check draft status.
• If the task explicitly depends on the legacy browser-backed session, run draft start-server --runtime v1_DEPRECATED, then re-check draft status.
• If draft status says BROWSER_NOT_CONNECTED, run draft daemon to re-open or re-pair the browser tab, then re-check draft status.
• If a live command returns REQUEST_TIMEOUT, do not retry blindly. Run draft status first.
• If draft status or a mutation error indicates EDITOR_NOT_READY, mount a real page route in the connected tab (https://draft.innosage.co/#/page/<id>), then re-run draft status --json before retrying writes. If needed, use draft page ls --json to discover the page ID before route-mounting.
• Do not treat draft page create as the primary EDITOR_NOT_READY fix. Recover editor readiness first, then run the intended command.
• If the daemon looks stuck or the wrong tab is attached, run draft stop-server, then restart with draft start-server.
• If the user explicitly wants staging or another environment, reuse the same URL consistently for both draft start-server --app [url] and draft daemon [url]. Add --runtime v1_DEPRECATED when the workflow explicitly needs the legacy browser-backed lane.
• If draft status --json shows READY but the connected clients[].origin does not match the requested environment, stop the server and reconnect to the requested URL before making changes.
• In CI or headless sessions, browser auto-launch may be skipped. Treat that as a diagnosis cue, then pair from a desktop session and verify with draft status --json.

What Humans Should See

When the browser tab is connected to the Draft CLI daemon, the GUI shows a small CLI Connected badge in the sidebar header while the local-mode session is active.

Command Reference

The Draft CLI uses conventional command structures.

Listing and Reading

To see all available Draft pages:

# Requires active connection
draft page ls

Output includes the page id, title, and parentId. You need the id to read or modify a page.

To read the content of a specific page:

# Returns the page in rich Markdown format (default)
draft page cat <id>

# Other available formats if you need raw data:
draft page cat <id>
draft page cat <id> --format raw

Reading Page Annotations (Comments)

[!NOTE] "Comments" in Draft are annotation highlights attached to text spans. The CLI exposes them as read-only records via two scoped commands. Use these commands to discover user feedback efficiently instead of rereading the entire page. This is the legacy page-centric path; keep using it when the task starts from a known page_id or an existing annotation workflow.

To list all comments (annotations) on a page in compact discovery mode:

draft page comments <page_id> --json

Output includes comment_id, anchor_text (the highlighted span), note (the comment body), and position_hint (character offset). Use this for quick triage — identify which comment IDs need deeper inspection.

To inspect a single comment with bounded context (±100 chars before/after the anchor):

draft page comment <comment_id> <page_id> --json

Output includes note, anchor_text, and a bounded_context object with before and after fields. Use bounded_context.before + anchor_text + bounded_context.after to locate the exact edit site before patching.

Reading Public Page Comments

[!NOTE] This is the hosted public-review path, not the live page-annotation path above. Use draft public-comments... when comments were created on a public or preview page. These comments are bound to a published snapshot and stored outside the live page transport.

To list public comments for a public or preview URL:

draft public-comments list --url 'https://draft.innosage.co/#/preview/<page_id>?mode=static'

To list public comments when only the page ID is known:

draft public-comments list --page-id <page_id>

To inspect a single public comment with bounded context:

draft public-comments get <comment_id> --url 'https://draft.innosage.co/#/preview/<page_id>?mode=static'

Machine-readable forms:

draft public-comments list --url 'https://draft.innosage.co/#/preview/<page_id>?mode=static' --json
draft public-comments list --page-id <page_id> --json
draft public-comments get <comment_id> --json

What the agent should expect:

• list output includes comment ID, thread state, author, quote, and body preview.
• get output includes page ID, publish version, offsets, body, and bounded context.

Creating Annotations (Comments)

Use draft page annotate to create a new comment on a selected text span.

# Basic annotation
draft page annotate <page_id> --anchor "scalable infrastructure" --note "Specify AWS or GCP" --json

When the anchor text appears more than once, disambiguate with surrounding context so the CLI can target the correct occurrence.

# Disambiguate repeated anchors with nearby prefix/suffix context
draft page annotate <page_id> --anchor "status" --before "The current " --note "Needs update" --json
draft page annotate <page_id> --anchor "status" --after " is blocked" --note "Needs update" --json

Use --before and/or --after whenever the anchor is ambiguous or repeated in the same page.

Creating, Modifying, and Publishing

To create a brand new page:

draft page create "My New Page Title"

To publish a page to the web:

# This will make the page publicly accessible via a unique URL.
# Free beta publish flow defaults to `innosage` when no invite code is provided.
draft page publish <id> --invite-code "${GLOBAL_INVITE_CODE:-innosage}"

# If you want to spell out the free beta code directly, this is equivalent:
draft page publish <id> --invite-code innosage

To append content to the END of a page. You can pass the content as a string, but for multiline Markdown, it is usually safer and much more robust to pipe it via stdin:

# Simple append
draft page append <id> "This is a new line at the bottom."

# Multiline append via stdin (RECOMMENDED)
cat << 'EOF' | draft page append <id>
## New Section
- Item 1
- Item 2
EOF

To replace the content underneath a specific heading (up until the next heading of the same or higher level). The matched heading itself is preserved, and only that section body is replaced. This is useful for updating specific sections like "Status" or "Action Items" without overwriting the whole document.

cat << 'EOF' | draft page replace <id> --heading "Status"
This is the new status content. The 'Status' heading is preserved, and everything previously under it is replaced by this text.
EOF

To apply a precise unified diff to a page. This is best for surgical edits to existing paragraphs.

cat patch.diff | draft page patch <id>

[!CAUTION] Always generate the diff from draft page cat <id> output — never from a locally authored file. Draft's tiptap editor stores multi-line text blocks as a single paragraph node. When serialized by draft page cat, this appears as one continuous space-joined line, not multiple lines. If you generate a diff against a multi-line file you wrote yourself, the patch engine will return PATCH_MISMATCH even though ok:true was returned by a previous write. Safe patch workflow: ``bash # 1. Capture live content and strip the 4-line metadata envelope (Title:, ID:, ---, blank line) # and the trailing --- delimiter. draft page cat wraps body content in this envelope. # The patch engine operates on body-only content; including any envelope line causes PATCH_MISMATCH. draft page cat <id> | sed '1,4d' | sed '$d' > /tmp/before.md # 2. Copy and edit — do NOT reformat or reflow the body text. # The live serialization is the ground truth. Even a single trailing newline # difference between your edited file and the live before.md will cause a mismatch. cp /tmp/before.md /tmp/after.md # (make your text edits to /tmp/after.md using sed or similar) # 3. Generate diff from live content. # IMPORTANT: diff exits with code 1 when files differ (not an error — that is expected). # Use ; (not &&) so the patch command always runs regardless of diff's exit code. diff -u /tmp/before.md /tmp/after.md > /tmp/patch.diff ; cat /tmp/patch.diff | draft page patch <id> --json # 4. Verify — wait 2-3 seconds after mutation before reading back. # Draft CLI relays mutations to a live TipTap editor asynchronously. A read immediately # after a write may return stale or empty content. Add a short sleep for reliable verification. sleep 2 && draft page cat <id> ` If you receive PATCH_MISMATCH, re-run draft page cat <id> | sed '1,4d' | sed '$d'` and regenerate the diff — do not retry with the same diff file.

[!NOTE] Annotated pages: draft page cat output for pages with comments includes inline markers like [:: User Note: A:]. These markers cause PATCH_MISMATCH if left in your diff. Always add a marker-strip step when patching annotated pages: ``bash draft page cat <id> | sed '1,4d' | sed '$d' | sed 's/ \[:: User Note: [^:]* :\]//g' > /tmp/before.md ``

Image Insertion and Management

You can use the CLI to insert images from local files, update their alignment or width, and delete them. To insert an image into a page:

# Insert an image with default alignment (left) and default width
draft page insert-image <page_id> /path/to/local/image.png --json

# Insert an image with specific alignment and width
draft page insert-image <page_id> /path/to/local/image.jpg --align center --width 500 --json

The output JSON will include a local_id (e.g., local_id: "img-abc1234"). Save the returned local_id, as it is required to update or delete the image.

To update the alignment or width of an existing image block:

draft page update-image <page_id> <local_id> --align right --json

To delete an existing image block:

draft page delete-image <page_id> <local_id> --json

Common Workflows

1. The Edit Cycle (Read, Modify, Verify) Always follow the connection-first pattern, then read the page before modifying it.

# 1. Check/Start Connection
draft status --json
# (if needed for default headless v2: draft start-server && draft status --json)
# (if needed for legacy browser-backed mode: draft start-server --runtime v1_DEPRECATED && draft status --json)
# (if browser missing in a browser-backed workflow: draft daemon && draft status --json)

# 2. Read
draft page ls --json
draft page cat abc-123-def

# 3. Modify
cat << 'EOF' | draft page append abc-123-def --json
New content...
EOF

# 4. Verify
draft page cat abc-123-def

2. The Safe Patch Cycle (Surgical Line Edit) Use draft page patch for precise edits to existing text. Always anchor the diff to the live markdown.

# 1. Check/Start Connection
draft status --json

# 2. Capture live content and strip the 4-line metadata envelope + trailing delimiter.
#    `draft page cat` wraps body content in: Title: / ID: / --- / (blank) ... (body) ... ---
#    The patch engine expects body-only content. Use sed '1,4d' | sed '$d' to strip.
draft page cat <id> | sed '1,4d' | sed '$d' > /tmp/before.md

# 3. Edit a copy — do NOT reformat or reflow the body.
#    The content from before.md is the only valid anchor.
cp /tmp/before.md /tmp/after.md
# (edit /tmp/after.md — sed, your editor, etc.)

# 4. Generate the diff from live content.
#    Use `;` not `&&` — diff exits 1 when files differ, which would silently
#    break `&&` chains before the patch command runs.
diff -u /tmp/before.md /tmp/after.md > /tmp/patch.diff ; cat /tmp/patch.diff | draft page patch <id> --json

# 5. Verify the change landed — wait 2-3 seconds first.
#    Mutations are relayed asynchronously to the TipTap editor. Reading immediately
#    after a write may return stale content. Always add a short sleep before verifying.
sleep 2 && draft page cat <id>

# If PATCH_MISMATCH: re-read with `draft page cat <id> | sed '1,4d' | sed '$d'` and regenerate — do NOT retry with the same diff

3. The Comment Discovery Cycle (Review → Locate → Patch) Use draft page comments and draft page comment to efficiently action user annotations without rereading entire pages.

# 1. Check connection
draft status --json

# 2. Discover all comments on a page (compact)
draft page comments <page_id> --json

# 3. Inspect the specific comment you intend to address (bounded context)
draft page comment <comment_id> <page_id> --json

# 4. Use anchor + bounded_context to generate a precise diff, then patch
# Note: We strip comment markers [:: User Note: ... :] to prevent PATCH_MISMATCH
draft page cat <page_id> | sed '1,4d' | sed '$d' | sed 's/ \[:: User Note: [^:]* :\]//g' > /tmp/before.md
# (edit /tmp/after.md with the fix informed by the bounded_context)
diff -u /tmp/before.md /tmp/after.md > /tmp/patch.diff ; cat /tmp/patch.diff | draft page patch <page_id> --json

4. Switching Tabs/Context The Draft daemon is intentionally single-session. If you need to connect to a different browser tab or recover from a stale pairing:

1. Stop the running server with draft stop-server.
2. Restart the correct runtime lane: draft start-server for default headless v2, or draft start-server --runtime v1_DEPRECATED for the legacy browser-backed session.
3. If the browser-backed lane is in use, run draft daemon to pair or retarget the tab.

5. Using Staging or Another Environment Only do this when the user explicitly asks for a non-production Draft environment.

draft status --json
draft stop-server
draft start-server --runtime v1_DEPRECATED --app https://markdown-editor-staging.web.app/
draft status --json
draft daemon https://markdown-editor-staging.web.app/
draft status --json

6. Publishing a Page

# 1. Stop any existing server to ensure a clean start in the legacy browser-backed lane
draft stop-server

# 2. Connect in the explicit browser-backed compatibility mode
draft start-server --runtime v1_DEPRECATED
draft daemon
draft status --json # Verify READY state before proceeding

# 3. Find the page ID
draft page ls --json

# 4. Publish
draft page publish <id> --invite-code innosage --json