Troy / Project
Roadmap
Planned Troy work and its state on main: the read pipeline, text recognition engines, a troy command, browser features and the September 2026 proposals.
On this page
This page collects planned work from the project’s own documents and marks where each item stands. Three states are used: in v0.1.5 (released), on main (merged, not released) and planned (not in the code). For what works today, see Troy.
The documents describe intent as well as behaviour, so nothing below is a promise of a date.
Sources:
- docs/PRD.md, the product requirements document (PRD), revised 9 August 2026
- docs/DESIGN.md, section 9 and section 10
- The original spec, 7 to 8 August 2026
- The engine plan, 8 August 2026
- The bugs and enhancements spec, September 2026, status “proposed”
The read pipeline
On main. The PRD marked these requirements “In build”. All five stages from the design document now exist in src/read/ and run in the agent panel and in scripts/read.mjs:
- Settle. Wait for
readyState,document.fonts.readyand two frames with no layout shift, then give up after a timeout and read anyway. - Extract. One evaluation that walks the document, including open shadow roots, and returns text, role, box, a stable selector and a visibility verdict.
- Visible. Count an element as painted only if it has a non-zero box, is not hidden, has non-zero opacity, is not clipped away, and its text colour differs from its background, which catches white-on-white text.
- Cover. Find only the painted regions the Document Object Model (DOM) cannot explain: a
canvas, an image or Scalable Vector Graphics (SVG) element without accessible text, a cross-origin iframe, or a large painted area with no text. - Fuse. Drop invisible blocks, order text into columns and reading order, and merge optical character recognition (OCR) lines back in by position. Every line keeps its box and its source.
See How Troy reads a page for how each works.
Acceptance criteria from the PRD, and where they stand on main:
- A plain article triggers zero OCR calls. Met, and checked by a test on the article fixture.
- A canvas page and an image-with-text page both return their text, attributed to pixels. Not met. The canvas is found and reported as a region, but no OCR engine is connected, so its text is not returned.
- Visually hidden DOM text never appears in the output. Met on the hidden-text fixture.
- Selectors from
troy read --jsonare accepted bytroy clickandtroy fillwithout translation. Not met. There is notroycommand.scripts/read.mjs --jsonreturns standard Cascading Style Sheets (CSS) selectors, the same formclick.mjsandfill.mjstake with--selector, except that selectors crossing a shadow root (>>>) are rejected by both.
OCR engines
The design: Apple Vision on macOS through a Swift helper, Tesseract elsewhere, behind one interface the caller does not choose.
- On main: the one interface (
OcrEngine), a stub that reads nothing, and an Apple Vision engine that compiles its Swift helper on first use. - Planned: passing an engine to the panel and the read script, and Tesseract for Windows and Linux. The original spec also describes a
--deepflag that escalates to a vision model.
A troy command
Planned. The spec and the engine plan describe a package published on npm as @anishfyi/troy with the command troy:
troy read <url>: Markdown by default,--jsonfor structure. The plan lists--session,--deepand--timeout.troy goto,troy fill,troy click,troy state: the action scripts ported to TypeScript.troy launchandtroy session ls|rm: persistent profiles.troy plugin install: install the Claude Code skill.
None of these commands exist. package.json is marked private and nothing is published to npm. On main, node scripts/read.mjs covers part of troy read, but it reads an open tab rather than opening an address.
Milestones
From PRD section 8, with the state on main added. Continuous integration (CI) is the automated test run on every push.
| Milestone | Scope | PRD status | On main |
|---|---|---|---|
| Design locked | Spec approved | Done | Done |
| M1 | Scaffold, Cdp port, CI |
Done | Done |
| B1 | The browser | Done | Done |
| v0.1.0 to v0.1.5 | Releases, cask | Done | Done |
| M2 | troy read, DOM-only pages |
Next | Script only |
| M3 | OCR engines | Planned | Apple Vision, unwired |
| M4 | Cover and fuse, scored | Planned | Built and tested |
| M5 | Action layer in TypeScript | Planned | Planned |
| B2 | Read pipeline, live tab | Planned | In the panel |
| M6 | --deep, docs, plugin, npm |
Planned | Planned |
Browser features
Planned in the PRD:
- Context menus, find in page, reopen closed tab, tab reordering.
- Address bar suggestions from local history and open tabs.
- Toolbar surfaces for extensions and an overflow menu.
- Auto-update.
On main: history recording when the user turns it on, cleared when it is turned off. The PRD had called the non-functional switch the highest-priority correctness gap in the shipped product. Nothing reads the recorded list yet, so suggestions from history remain planned.
Proposed in the September 2026 spec
The bugs and enhancements spec, status “proposed”, lists these enhancements:
- A real
troy read <url>binary that navigates, settles, reads and prints, with--json. - OCR on Windows and Linux, with Tesseract behind the same engine interface.
- Trustworthy endpoint discovery: a single-instance lock, writing the endpoint only after the port answers, and readers that check the process is alive.
- A shadow DOM resolver so
>>>selectors can be acted on. - A
page_scrolltool and viewport control, because OCR sees only the viewport. - A same-origin iframe walk that keeps selectors inside the frame.
- About panel and Help menu on Windows and Linux.
- An extension list in the interface.
- Auto-update.
It also proposes cutting a 0.2.0 release once its first round of fixes is in. That is a proposal in the document, not a scheduled release.
Non-goals
From the PRD: beating Chrome at general browsing, any engine other than Chromium, cloud sync, accounts, telemetry or a store, notarisation by Apple, and the Model Context Protocol (MCP) as a transport. The Chrome DevTools Protocol (CDP) is the protocol.