install-and-verify

Reticle: prove the change, do not guess

Reticle embeds a dev-only SDK in the user's running app and exposes it to you as reticle_* MCP tools. You look, act, observe, and assert against the real app. No screenshots, and no browser download for the verify loop: it drives the tab the user already has open. (A driven browser, which reticle_lease and --drive use, does need Chromium; Reticle says so when it needs one.)

Where the rest of the documentation is

Everything not in this file is at https://docs.reticle.sh, and it is built to be fetched rather than browsed. Append .md to any page URL to get its source with no site chrome, so you can pull one page instead of a whole document you mostly do not need:

curl https://docs.reticle.sh/llms.txt                  # every page title and URL, small enough to read whole
curl https://docs.reticle.sh/cli/doctor.md             # one CLI command: flags, real output, exit codes
curl https://docs.reticle.sh/tools/act-and-wait.md     # one tool: arguments and what a verdict means
curl

Read llms.txt first and pick the one page that answers the question. That is almost always cheaper than pulling a large file and hoping the part you need survived. There is a /llms-full.txt with the entire site in one file; use it only to seed a context window deliberately.

Every page arrives with the four rules that matter prepended, whether you asked for them or not, so a single fetch orients you without a second call.

Which path am I on

• NOT_FOUND → SETUP below.
• File exists → VERIFY below. If reticle_session { action: "list" } then returns an empty list, go to references/troubleshooting.md; do not restart setup.

Both paths are about THIS PROJECT. The machine step is separate and comes first: one command that puts the CLI on PATH, registers the MCP server with the coding agents it finds, and pre-approves Reticle's own tools where an agent has a per-server approval rule (Claude Code, for one). A user who would rather choose each step can follow https://docs.reticle.sh/install-manual.md instead.

If you can see reticle_* tools, that already happened and you can ignore it. If you cannot, hand the user that one line to run in a terminal and stop there. init can register the MCP server itself, but doing it from inside a client that has already read its server list means the tools cannot appear until the client restarts, which ends your turn in the middle of setup. The terminal-first order is what removes that step, so do not work around it.

Read this before you touch SETUP

Installed means a verdict was produced. Writing config files is not installed.

Every earlier point looks like success and is not:

• init exited 0. Files were written. Nothing connected.
• The reticle_* tools appeared. Your client can reach a daemon. The app is very likely still uninstrumented.
• A session is listed. The app dialled in. The user has still seen nothing happen.

Do not tell the user Reticle is set up until step 5 has produced a verdict. The single most common outcome in the field is an agent that finishes step 1, reports success, and leaves a user with config files and no instrumented page.

Finish the setup steps

Finish the setup steps; ask the user only when a step needs their decision. Running init, fixing wiring it could not, starting the dev server and opening the browser are setup steps, not decisions.

The repo already answers which framework, package manager, port, editor or MCP client, so work those out rather than asking. Say what you did in one line.

Three places always need the user:

1. No recognisable dev script in package.json. Say so; do not invent one.
2. Your host asks the human to approve a command. That prompt belongs to the host. Never bypass or suppress it, and take a refusal as the answer. init writing a pre-approval rule for the reticle server is not that: it is a scoped, announced config change the human asked for by running the command, and it covers only Reticle's own tools.
3. On a first install only, the client restart that loads the MCP tools (step 2). Even then, keep going until the app is running and connected, and stop for the restart last, with only the drive left.

Setup requires a client restart, which ends your turn. This skill survives that restart. After the restart, re-read this file and resume at the step you were on. Do not start over, and do not skip forward.

SETUP

One command wires the project. A second one proves a flow.

It detects the framework and package manager, wires the build config, installs the SDK, registers the MCP server, starts the dev server, opens the app, and waits for a session to connect from inside it. That connection IS the proof onboarding worked: the SDK is in the page and the tools have something to talk to. It exits non-zero if nothing connected, and prints exactly what is left to do.

Then prove a flow. That is the FIRST RUN, and it is a separate call:

Drive the journey that matters and put the verdict on its LAST step: until names the end state before the action fires. What you drive is saved as a flow, so later runs replay it with no model. On a linked project (reticle connect; every plan, Free included, has monthly Harness credits), reticle_verify { action: "explore", persona: "<who does what>" } has the Reticle Harness drive the whole journey for you instead. Before driving anything, replay what is already saved: reticle_verify { action: "flows" } costs no model at all.

What YOU decide, and pass in

The command reads the repository. It cannot read the request, and these live only there.

Add --license <key> if the user gave you one: it writes RETICLE_LICENSE_KEY to .env and keeps .env out of git.

Framework, package manager, port, editor and MCP client are answerable from the repo you are sitting in, so work them out rather than asking.

Then read what it gives you back

A non-zero exit is a to-do list, not a failed install. The command names the cause and prints the REMAINING steps from wherever it stopped; it will not tell you to redo a phase that already worked. Do those and re-run, which is safe.

It is not finished until a verdict exists. Writing files is not an install, and neither is a connected session.

If that command could not run it

Do not choose this path. It is not the thorough version of the one above; it is what you fall back to when the command physically could not do the work. Use it only when init exited without ever printing starting: or ▸ WATCH (an older CLI that stops after writing files), or when it stopped in the same place twice after you did what it asked. A ⚠ in the report is not a reason: re-run the command, which is idempotent and names what is still outstanding.

License key

Enterprise features are unlocked by a signed key, verified locally with Ed25519 and no network call. If the user hands you one, during setup or at any later point:

1. Append RETICLE_LICENSE_KEY=<key> to the project's .env, creating the file if it is not there. The CLI folds a project-local .env into the environment on every invocation, so nothing else has to change.
2. Make sure .env is in .gitignore. A license key in git is a leaked credential, so check before you write and add the line if it is missing.
3. Confirm with npx @reticlehq/server license, which prints active, eval, missing or expired.

Never echo the key back in your reply, and never put it in a commit, a code comment, or a feedback report. The rest: curl https://docs.reticle.sh/enterprise.md.

VERIFY

Verdicts come from reticle_act_and_wait, reticle_assert, and reticle_act { steps } when a step declares expect. Everything else (a bare act, look, navigate, observe) moves or reads the app and proves nothing. A drive that ends without a verdict has no result, however many tools it used.

Prefer reticle_act_and_wait({ ref, action, until }). It names the expected consequence before the action, which is the difference between a check and a rationalisation.

A verdict of verified: "unknown" is not a pass. It means Reticle drove the app and could not tell what happened. Report it as unknown.

Never weaken a check to make it pass.

Take the cheapest path that answers the question

Stop at the first row that fits. Do not hand-drive a flow you could replay.

reticle_flow_replay is not on the advertised tool list: it is reached through reticle_run exactly as written. That is the supported call shape, and it is why you have to be told it exists.

reticle_verify {action:"change"} answers unknown when no saved flow covers the files you changed. Nothing ran, so nothing was proved: that is the honest answer, never a pass, and it is the signal to record one (step 5 above).

The full loop (look, act, observe, assert), how to read a verdict, batching, and regression suites: references/verify.md.

Feedback is first-class

Reticle is built for agents. You are its user, and your experience is the only signal that decides what gets built or fixed. It disappears when your context does.

The moment Reticle breaks, lies, cannot see something you needed, costs too many calls, or gives you trouble during setup:

kind is one of bug, gap, ambiguity, feature_request, improvement. One call, then carry on with the task.

If the tools are unreachable (setup unfinished, daemon down), file the same report from the shell:

Nothing is too minor. Report defects in Reticle; a bug you find in the app under test is Reticle working and belongs in your answer to the user. Never put app source, secrets, or user data in a report.

References

• references/setup.md: per-framework SDK wiring, store registration, desktop apps.
• references/verify.md: the look, act, observe, assert loop and reading verdicts.
• references/troubleshooting.md: no session, port mismatch, non-localhost hosts, daemon problems.