throughline

name	throughline
description	Publish your implementation plan as a live HTML URL the user can read on their phone and hand to another agent. Use whenever you produce a multi-step plan, spec, or roadmap the user should review or that another agent will execute/critique — instead of (or in addition to) dumping a long markdown plan into chat. Keep the plan updated as you complete tasks.

Throughline — publish plans to a live URL

You write a structured plan once; Throughline renders it as a readable HTML page (phases, task status, decisions, before/after diffs, live progress) at a shareable URL, plus machine-readable .md/.json at the same URL for the next agent, plus one-click "Open in Proof" for human+agent comments.

When to use

• You've drafted an implementation plan / spec / migration roadmap with phases and tasks.
• The user wants to review on their phone, or steer you before you build.
• You'll hand the plan to another agent to execute or critique.

Prefer this over a wall of markdown in chat for anything longer than ~15 lines.

Setup

Base URL defaults to the deployed https://throughline.henryzh.dev. Override with $THROUGHLINE_URL (e.g. http://localhost:8787 in dev, or the *.workers.dev URL).

1) Create a plan

Build a JSON object (schema below) and POST it. Capture the returned editToken — you need it to update the plan later.

curl -s -X POST "$THROUGHLINE_URL/api/plans" \
  -H 'content-type: application/json' \
  -d @plan.json

Response: { id, url, ownerUrl, editToken, formats:{html,markdown,json} }.

Give the user ownerUrl (it carries an #t= fragment that unlocks edit/Proof actions in the page; the fragment never reaches the server). Or just url for read-only sharing.

Or use the CLI: node bin/throughline.mjs push plan.json (add --proof to also create the Proof doc).

2) Update as you work

As you finish tasks, flip their status by id (you get ids back in the .json). Always reference the plan by editToken.

curl -s -X PATCH "$THROUGHLINE_URL/api/plans/$ID" \
  -H "authorization: Bearer $EDIT_TOKEN" -H 'content-type: application/json' \
  -d '{"tasks":[{"id":"abc123","status":"done"}]}'

The user's open page updates live (SSE). Patch supports: status, summary, notes, tasks:[{id,status,title,note,owner}], and append-only addPhase / addDecision / addDiff.

3) Hand off to another agent

Give the other agent the .md or .json URL — it's the same plan, machine-readable:

curl -s "$THROUGHLINE_URL/p/$ID.md"     # for an executing/critiquing agent

4) Push to Proof (collaboration loop)

For comments / suggestions / provenance with a human, create the Proof doc:

curl -s -X POST "$THROUGHLINE_URL/api/plans/$ID/proof" -H "authorization: Bearer $EDIT_TOKEN"

Returns { shareUrl } — a Proof document where humans and agents comment and suggest edits.

Plan JSON schema

{
  "title": "string (required)",
  "summary": "markdown string",
  "status": "draft | active | blocked | done",
  "author": "ai:<model> | human:<name>",
  "repo": "owner/name",   // optional — powers the "Execute / Critique in Claude Code" deep-link buttons
  "cwd": "/abs/path",     // optional — absolute working dir; used instead of repo when set
  "decisions": [ { "title": "...", "choice": "...", "rationale": "markdown", "locked": true } ],
  "phases": [ {
    "title": "...",
    "goal": "one-line markdown intent",
    "tasks": [ { "title": "...", "status": "todo|doing|done|blocked", "note": "markdown", "owner": "ai:.. | human:.." } ]
  } ],
  "diffs": [ { "title": "...", "before": "code", "after": "code", "lang": "ts" } ],
  "notes": "freeform markdown reasoning",
  "links": [ { "label": "...", "href": "https://..." } ]
}

Guidance for good plans:

• Phases = milestones; tasks = checkable units. Set realistic status up front (some done, the current one doing).
• Put genuinely settled choices in decisions with locked:true; put open questions with locked:false so the user can weigh in.
• Use diffs to show the key before/after — it's the thing HTML does that markdown can't.
• Mark provenance with owner (ai: vs human:) so the page shows who's doing what.
• Set repo (owner/name) or cwd (absolute path) so the page's Execute / Critique in Claude Code buttons open a session in the right project. The deep link only pre-fills a prompt (pointing back at this plan's .md) — it never runs anything until the user reviews it, and it opens on the same machine as their browser.