A local-first Visual Direction Loop: open a review surface on an agent-produced interface, point at visible targets, compose Annotations, and verify the result by hand. No prose location descriptions, no lost context.

The Annotation is the unit of work. A Builder-Reviewer points at one or more visible targets, drafts a note in a card anchored to the target, attaches reference images, and queues the result. Sending delivers one Visual Intent Envelope carrying each Annotation with its own identity. Each Annotation is verified on its own.

Install

Requires Node 20+. No Rust toolchain, no hosted account.

npm install -g visual-intent-layer
visual-intent setup            # register in this project
visual-intent setup --global   # register for every project on this machine

setup looks for each harness in scope and registers the server with each one natively. A harness counts as present when any of these is true, and the report names which one matched:

Harness	Configuration location (env override)	Command	Extra
pi	`$PI_CODING_AGENT_DIR`, else `~/.pi/agent` or `~/.pi`; project `.pi/`	`pi`	—
Codex	`$CODEX_HOME`, else `~/.codex`; project `.codex/config.toml`	`codex`	`/etc/codex`; `~/.codex/packages/standalone/releases/*/bin/codex`
Claude Code	`$CLAUDE_CONFIG_DIR`, else `~/.claude`; plus `~/.claude.json`	`claude`	—
opencode	`($XDG_CONFIG_HOME or ~/.config)/opencode`; project `opencode.json(c)`	`opencode`, `opencode2`	—

A command only counts when a real executable file resolves, so a directory or a non-executable file with a harness's name is ignored. Writes land in the same locations detection checks, so an override that detection honours is also where setup writes and where status looks. Without a harness filter, every detected in-scope harness is configured. Every run prints its own package version, the detected harnesses, each known-but-absent harness with the evidence that was checked, and an explicit line when nothing was detected — so a pi-only result is never ambiguous between a detection result and a fallback. When nothing is detected, project scope still writes the shared .mcp.json so pi and Claude Code keep working by default, and the report says so.

Harness	Project scope	Global scope
pi	`.mcp.json` (shared) + Skill	`~/.config/mcp/mcp.json` + Skill
Claude Code	`.mcp.json` (shared; first-use approval)	`claude mcp add-json … --scope user` (or the printed manual command)
Codex	`.codex/config.toml` (trusted projects)	`codex mcp add …` when the binary exists, else `~/.codex/config.toml`
opencode	`opencode.json`	`~/.config/opencode/opencode.json`

Every write merges the server entry and preserves existing servers. Setup never overwrites another server's configuration and refuses a file holding invalid JSON or TOML (naming the harness and path) instead of touching it. An existing server entry is compared against the entry setup would write: a matching entry is reported as current and left byte-identical, and a differing entry is reported as outdated — naming the fields that differ — and repaired on re-run, so an entry left by an older release or copied from a snippet heals without hand-editing.

Transport

Setup registers whichever transport will actually work on this machine. When visual-intent resolves on PATH the entry is command: "visual-intent" with args ["mcp"]; otherwise it is the packaged npx -y --package visual-intent-layer@<version> visual-intent-mcp form.

Setup flags

--global — user-global scope instead of the current project.
--harness <name> — restrict the matrix to named harnesses; repeatable.
--status — read-only: report registration state, location, transport, evidence.
--print-only — show every planned write without touching disk.
--no-skill — skip the pi Skill install.

Use

Start a session yourself, or let the agent open one:

visual-intent open --html ./checkout.html     # opens the browser, prints the URL
visual-intent open --app http://localhost:5173
visual-intent serve --port 3742               # local service only

Starting a session launches the default browser on this machine and prints the review URL as well, so the loop is a product experience rather than a URL to copy. Automatic opening is suppressed with --no-open or VISUAL_INTENT_NO_OPEN=1 for headless, remote and scripted use; the URL is still printed. Reopening the same artifact revision reuses the open session and its URL, so a review does not accumulate tabs.

One rail, two tiles

The Review Surface has one chrome region and one state. The rail on the right holds the artifact's identity, its revision, the agent's position and every Annotation — unsent and sent alike, in one list with the state pill carrying the difference, so nothing leaves the view when it is sent. An Annotation that needs a decision comes first; closed ones stay behind one toggle. Verdicts are decided where the Annotation sits: approve, reject with another pass, or mark obsolete. Amending one that was already sent supersedes it and delivers the amendment; it is never rewritten in place.

A small island over the artifact holds two icon-only tiles: point at things, and box an area. Operating the artifact is the unarmed resting state rather than a third tile. Pointing is a gesture: clicking targets a thing the artifact owns, and dragging across words targets exactly those words. P, B and V arm and disarm them.

Type and press Enter in the Annotation card to queue it; Cmd/Ctrl+Enter sends the whole queue. Escape unwinds exactly one level — close the card, then clear the selection, then return to operating the artifact — and never discards unsent writing.

The Annotation model

An Annotation is durable and individually identified: targets, note, references, delivery state and resolution. The envelope keeps its support for relationships, but this iteration's surface expresses no relation.
Unsent text survives a surface reload, a service restart and a browser restart. Nothing discards a note silently.
Several targets can be gathered into one Annotation; a drawn Area records the revision and scroll position it was drawn at and reports the elements it encloses.
Reference images are added by picker, paste or drop, are content-addressed by their own bytes, and are refused visibly (and unread) when disallowed or larger than 5MB. Only image types are accepted.
Relational Intent stays in the domain model and the envelope keeps its support for relationships, but this iteration's surface expresses no relation. The capability is deferred deliberately, not removed.
The product never writes style values into the artifact or its source.

Resolution, honesty and the agent

When the artifact changes, each target is located again and reported as Matched, Recovered, Ambiguous (with its candidates, never auto-selected) or Deleted. Whether an Annotation was written before the revision now on screen is a separate, Annotation-level fact shown as such. Provenance Confidence — exact source span, inferred, or unavailable — is a separate axis and never shares the word "exact" with target resolution.

The surface states the agent's position in a sentence, whether a tool call is currently held or direction will be read at the agent's next Check-In, and when the agent last checked in. Where the host can hold the call the agent is awaiting you and sending is urgent; where it cannot, the surface says the agent has stepped away and the Annotations are queued durably. Agent acknowledgement is never presented as implementation or as verification, and the workflow never requires the agent to be waiting.

Check-In is a convention, not a capability

Steering and interruption are seen at a Check-In: the point between an agent's own steps where it reads new direction. There is no push channel and no wake mechanism, because a server cannot put anything into an agent's running turn. The product publishes the convention instead of detecting a host capability it cannot rely on, and records every MCP call for a session so "last checked in" is a fact the surface states rather than an assumption.

An agent that was not holding a call can retrieve what arrived with the check_in tool: newly delivered Annotations with their intent, amendments that superseded something, a pending stop request, and the current state of what it was given before. Sending the queue is Next-Pass Intent, amending something already sent is Steering Intent, and asking an agent to stop is Review Interruption — a request, never a claim that work stopped. The envelope's review-interruption value is reserved and never emitted: an interruption names no target, so it cannot be one.

The nearest canvas-agent prior art refuses check-in outright and schedules later requests instead; polling through the MCP Tasks extension is the one push-free pattern the specification sanctions. Both are recorded in the shipped skill.

A single drawer, hidden while it has nothing to report and badge-counted when it does, carries what needs a decision: everything that will leave the machine, any unresolved or ambiguous Annotation, and the fact that the artifact has moved on. The overflow menu holds the rare actions — reload artifact, copy artifact path, copy evidence for the queue, open the disclosure, end session, and choose the chrome theme — and never a frequent one.

Agent tools (MCP)

The model-visible tool surface is deliberately small:

Tool	What it does
`open_visual_review`	Opens the loop for a saved HTML file or a running localhost app, may declare `capabilities: { embeddedUI, subscriptions }`, and holds the call for the human.
`check_in`	Reads new direction between the agent's own steps, without an `envelopeId`: deliveries with their intent, amendments, a pending stop request, and current state.
`get_intent_status`	Reads one delivered batch by `envelopeId`, including each target's resolution.
`acknowledge_intent`	Confirms receipt of a batch or one Annotation. Never implementation and never verification.

embeddedUI and subscriptions are the only host capabilities the product negotiates; there is no steering flag. A host declares them through the capabilities argument on open_visual_review, not through a side channel. The tool descriptions ship the Check-In convention, so an agent following them knows to call check_in between its own steps.

Environment and data

Variable	Meaning
`VISUAL_INTENT_DATA_DIR`	Lifecycle data directory (default `~/.visual-intent-layer/data`).
`VISUAL_INTENT_PORT`	Default service port (default `3742`).
`VISUAL_INTENT_NO_OPEN`	Set to `1` to suppress automatic browser opening.
`VISUAL_INTENT_WAIT_MS`	How long the agent-facing entry tool holds the call, in milliseconds.

Annotations, sessions, attachments, snapshot bytes and Check-In contact all live under the data directory, so a service restart loses nothing. The packaged mcp.json pins the current release version for the npx transport form.

Envelope schema

The portable contract is schema/envelope-v0.2.schema.json (experimental, versioned). One envelope carries one or more Annotations, each with its own identity, targets, note, relationships, references and attachments. TypeScript types are generated from it (npm run build:types).

The 0.1 schema is kept only for reading state written by older releases; new envelopes are 0.2. review-interruption remains a reserved value in the delivery.intent enum and is never emitted, because an interruption names no target and so cannot be an envelope.

Security and privacy

Reviewing is local-first. The service binds to loopback only, requires an unguessable per-session capability, and confines file access to the artifact directory. A saved HTML artifact keeps its own relative and root-relative assets and may load the remote stylesheet, font and image origins it declares — and those origins are disclosed in the surface before the artifact contacts them. Runtime data requests stay blocked by the artifact's content policy. A running local development server may be reverse-proxied through the review service's own origin so its DOM is selectable; only loopback development origins are proxied and authenticated production applications are refused.

See SECURITY.md for the full policy.

Develop

npm install
npm test            # full suite, including the browser-driven primary seam
npm run typecheck
npm run build       # compile the service and bundle the shell + artifact layer
npm run benchmark   # target-resolution mutation benchmark
npm run instrument  # latency / reliability / token-efficiency signals
npm run eval:invocation   # documented routing policy, not live agent judgment

The primary test seam is tests/browser-loop.test.ts: the Visual Direction Loop is driven end to end in a real browser engine against the locally running service — the served asset graph must load, the artifact must render with its styles, and a human path (point, annotate, queue, send, reload, compare, amend, stop, verify, restart) is performed through the DOM the product actually serves. A surface that cannot load its own scripts fails CI rather than shipping.

The design gallery is at /gallery on a running service. It is not part of the product's navigation. tests/gallery-snapshot.test.ts pins every design token exactly, fails on a screenshot difference, checks every semantic surface/ink pair against the contrast floor in both themes, and fails when a baseline is missing rather than silently writing one. Regenerate baselines with UPDATE_GALLERY=1 once the change is intended.

Layout

design.md — normative Review Surface design contract (tokens, roles, states)
CONTEXT.md — the domain language, including the words to avoid
schema/ — versioned Visual Intent Envelope contracts (0.2 current, 0.1 legacy)
src/annotation/ — the Annotation model, durable store, attachments, migration
src/artifact/ — identity, content-addressed revisions, fidelity rewriting, snapshots
src/resolution/ — target resolution and its reduced vocabulary
src/mcp/ — MCP server, the four tools, stdio transport
src/service/ — loopback HTTP service, sessions, Check-In records, security boundary, browser opening
src/ui/ — product-owned shell, artifact interaction layer, icon set, design gallery
src/adapters/ — React/Vite Source Provenance adapter
src/host/ — the surviving host-capability declaration (embedded UI, subscriptions)
src/benchmark/ — mutation benchmark matrix
src/instrumentation/ — product-boundary measurements
src/eval/ — documented invocation-routing eval
fixtures/ — the gallery artifact the Lever and tests drive
tests/ — the browser loop, the Lever contract, and the gallery baselines
scripts/ — UI bundling and envelope type generation
skills/ — thin optional Skill for agents

Documentation

docs/adr/ — architecture decisions, including ADR-0018 on Check-In replacing a steering capability
docs/background/ — research and the product strategy brief
docs/pi-validation.md — the live-pi checklist
docs/agents/ — the issue tracker, triage labels and domain-doc conventions
SECURITY.md — the threat model, supported versions and disclosure process

Maintaining this package

Releases go out through the Publish workflow (npm trusted publishing). /maintain-visual-intent-layer lays out the paths — dogfood locally, cut a pre-release to next, promote a validated pre-release, or fix a bad release — and executes the one you pick. Pre-releases publish to next; latest only moves when a validated pre-release is promoted.