Feat/hardening ci and tap - #3
Merged
Merged
Conversation
Picks up the typesafe `env` export added by a newer Convex version. No behaviour change — the file is generated, and this keeps it matching the installed client.
… poller A DNS outage turned one dead host into a wall of stack traces: the inbound-payment poller ran every fifteen seconds regardless, each pass opening connections that took the Node default of many seconds to give up, each failure printing in full. The real problem — the provider is unreachable — was buried under its own repetition. Two changes. Calls now carry an eight-second deadline, so a hung connection is abandoned rather than left to occupy the poll it belongs to. And repeat failures against the same host collapse: the first prints, then every twentieth, until a success resets the count. The poller itself moves from a fixed interval to self-scheduling with exponential backoff, fifteen seconds out to five minutes. It only backs off when no wallet reached the provider at all — a single wallet failing is a wallet problem, not an outage, and must not slow down everyone else's payments.
Four things a user had to sit through, each fixed at its cause. Aide opened almost every sentence with "Let me", including twice in a row, and the engine's own stall filler said much the same thing — so a pause sounded like Aide starting over. The fillers now split in two: openers only before the first tool call, continuations for a mid-reply lull, worded so they can never be mistaken for the start of a turn. The prompt allows the covering opener once per turn, not once per tool. Aide also claimed to have opened pages it never opened, and asked a blind user whether they could see something. Both are now explicit prompt rules, and tests assert the rules still exist — a deletion inside a long prompt is easy to miss in review. The speech worker's request deadline started at enqueue rather than at synthesis, so queued work burned its budget waiting its turn. Enough short sentences at once and the last few timed out untouched, killing the worker and taking the rest of the reply with it. The clock now belongs to whatever the worker is actually synthesising, and a worker that dies has its pending requests replayed on a fresh process instead of rejected. A request is only text; failing it drops the user to the robotic browser voice mid-sentence for no reason. Finally, AIDE_DIST_DIR: two Next processes in one tree share one build directory and silently overwrite each other, which surfaces much later as a missing-chunk error in whichever server was there first.
Aide moves real money on behalf of users who cannot see the screen to notice when something has gone wrong, so the checks that used to live in one developer's terminal now run on every push and pull request. Five checks, in parallel, each one guarding a failure this codebase has actually had: - Types and Tests are the two gates run by hand before every commit so far. The test job spans both vitest projects, including the Convex functions, which only run inside an edge-runtime VM. - Production build catches routes that compile only under `next build` — the dev server is more forgiving than a deploy is. - Speech worker installs edge-tts and imports both speech entry points. When that path breaks, a blind user hears the robotic browser fallback and gets no explanation anywhere on screen. - No secrets committed refuses a tracked .env, local Convex state (it holds an admin key), or anything shaped like a live credential. All five run against placeholders, never a real provider: what they prove is the wiring, not the keys. Dependabot is monthly and grouped on purpose — a wall of open PRs gets ignored, which is worse than a slightly stale lockfile.
There was no way to tell Aide to stop listening. The mic closes itself after ninety seconds of quiet, but that is Aide's decision, not the user's, and any stray touch reopens it — so a user who wanted a private moment had no way to take one. A mute button would not solve it. A button is only a control if you can find it, and finding it is exactly what these users cannot do. A count of taps needs no target: anywhere on the screen or the trackpad, on the surface their hand is already resting on. Three rather than two, because a hand double-taps by accident and does not triple-tap by accident. Both directions are announced, because the whole state change is otherwise silent and invisible. The closing notice carries the way back inside it — it is the last thing heard before Aide goes quiet, and there is nothing on screen to remind them afterwards. The hold is enforced at startRecognition(), the single function that can open the mic. Its half-dozen callers — a tab regaining focus, a reply finishing, the restart backoff, interrupt() — would each have undone it otherwise, quietly, at a moment the user had no reason to be watching for. The level meter releases its own capture stream too, or the browser's recording dot stays lit and the announcement is a lie. Two smaller things fall out of it: the third tap's trailing click lands on the Aide orb's interrupt handler and would cut off its own announcement, so interrupt() ignores a click that closely follows a toggle; and the gesture is inert in speak-only browsers, where there is no microphone to hold. Aide is told about the gesture in the prompt so it can answer when asked, and told it cannot perform it — the same rule as page navigation. A claim with no action behind it leaves someone believing a microphone is closed when it is open.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.