按源仓库内容呈现，保留标题、案例、代码、表格、链接以及原文引用的演示图片。
The app is broken and the code looks fine
Reading the source again will not tell you why. The evidence is in the running app, and most of it never reaches the screen: the request that failed, the store that never moved, the handler that was never bound, the element covered by something invisible.
Reticle reads all of it from inside the page. Not installed? RETICLE_INSTALL_SOURCE=npx_skill npx @reticlehq/server@latest init, then the `install-and-verify` skill.
Do not start by guessing. Start by reproducing.
reticle_look({ action: "page", sessionId, mode: "interactive" }) // controls only, with refs
reticle_act_and_wait({ sessionId, ref, action: "click", until: { kind: "element", query: { testid: "..." } } })The verdict already narrows it: no / contradicted means a channel saw something incompatible with the UI: you have the bug. unknown means Reticle drove it and could not tell, which is a different investigation from "it failed".
Read the act result before you call anything else. Its summary block already carries the whole causal window of that one click: net {total, errors, headline}, consoleErrors, stateDiffs [{path, from, to}], storageDiffs [{key, from, to}], route, signals, layoutShift, longTasks: real before→after diffs, not counts. The console, network, storage and state calls below are for going DEEPER into something the summary already pointed at. Calling all three straight after an act pays three model round trips (and, in an approval-gated client, three human clicks) for evidence you were already holding.
summary.stateUnwatched: true means the state channel is dark: no subscribable store is registered, so an empty stateDiffs means unwatched, not unchanged, and no state-based conclusion is available until someone registers the store. It is the one field here that is about your instrumentation rather than about the app.
When the summary does point somewhere, read that channel scoped to what you just did: pass since from the act result, or you are reading a buffer that predates the click:
reticle_observe({ action: "console", sessionId, since })
reticle_observe({ action: "network", sessionId, since })
reticle_look({ action: "state", sessionId, store })The diagnosis table
| What you see | What it is | Next |
|---|---|---|
| No request fired at all | the handler is not bound, or a client cache served a stale value | reticle_look { action: "element" } the element for its file:line; check the cache if it is TanStack Query |
Request fired, 4xx/5xx | a real backend failure the UI swallowed | the response body is in the network entry |
2xx but nothing changed | the app never read the response: verdict says outcome_unread | usually a real bug in the success path |
| Request fine, DOM fine, store stale | UI-vs-state desync, invisible to any screenshot | reticle_look { action: "state" } is the only witness; this is the highest-value read here |
| Click did nothing, no error | the control is dead, occluded, or disabled | reticle_look { action: "element" } returns occluded, box (0×0), styles.cursor, opacity |
| Element "not found" that you can see | virtualised list has not mounted it | reticle_scroll_to. A query that misses is not evidence of absence |
| Looks logged in, behaves logged out | the token never persisted | reticle_storage |
| Console error that predates your click | a pre-existing fault, not your bug | call it out before continuing |
Get the file, not a theory
reticle_look({ action: "element", sessionId, ref })Returns the component and source: { file, line } for the element you are looking at, plus whether it is occluded, sized, disabled or off-theme. That is the pointer to open. End the investigation with a location, not a hypothesis.
When the flow is long
Do not ping-pong act → snapshot → act → snapshot to find the failing step. Drive the journey with reticle_act { steps: [...] } and assert once; the first step whose consequence fails is the one to look at, and you paid a fraction of the calls to find it.
Before you say it is fixed
Re-run the reproduction as a verdict (reticle_act_and_wait with the consequence named up front) rather than eyeballing the page. A fix confirmed by looking at it is the failure mode that created this ticket.
And never weaken the check to make it pass.
Troubleshooting Reticle itself (as opposed to the app): curl https://docs.reticle.sh/troubleshooting.md. Everything else: curl https://docs.reticle.sh/llms.txt. If Reticle could not see something you needed, reticle_session { action: "feedback" } with kind: "gap".
