How does a piece of text become a reliable webpage translation?

Chrome's extension service worker is event-driven and may terminate while idle; ordinary JavaScript globals are not automatically saved. This project therefore writes translation memory and quotas to IndexedDB and lets pages redeclare active demands after reconnecting. This is lifecycle design, not a claim that the worker never stops.

Platform reference: Official Chrome documentation: service worker lifecycle (accessed 2026-10-25).

The background scheduler indexes the “lookup in progress” work before the first asynchronous memory lookup. Later page demands subscribe to the same work; even if the earlier lookup has not returned, no identical task is created. This allows duplicate demands to merge even during asynchronous lookups.

UTF-16 code units are the units used for JavaScript string length; an emoji may occupy several units, and multiple Unicode code points may form one visible character. A grapheme describes a grouping of these visible text units; tokens are determined by a model's tokenizer.

The default 2000 is the string-length limit per text unit. If a single grapheme exceeds it, the input is reported as too large. Without browser text-segmentation support, the fallback processes code points individually and cannot guarantee complete grouping of compound visible characters.

Short segments also use a local heuristic: after removing noise such as URLs, a single Latin-alphabet word may be skipped, while short segments containing Chinese, Japanese, or Korean text can be retained. This reduces fragmented requests but may miss meaningful words. Selected-text translation does not use the same short-text threshold as page translation.

Further reading: Browser text segmentation.

When the configured margin is ≤0, both sides return 0. Swap the upper and lower ranges when reading upward. Scroll speed is smoothed with an EMA: α=1−exp(−dt/200), new speed = old speed + (instantaneous speed − old speed) × α. dt is in ms, and instantaneous speed is in px/s; at intervals of 400ms or more, use the new sample directly.

A direction change requires accumulated reverse movement of max(120px, 0.35H). Forward speed of 2.5 viewport/s can trigger fast-scanning mode for 2 seconds. These thresholds reduce observation-region rebuilds caused by small backward movements; they are engineering heuristics, not a model of reading intent.

Handle whole-page disabling first, followed by whole-sentence keep-source / fixed translations, user-pinned translations, and no-external-send rules; only then replace phrases with placeholders. Banning external sending still permits locally saved translations to appear.

Aho–Corasick finds multiple phrases; “leftmost, longest” emerges from the subsequent sort and cursor-based selection together. Scanning uses UTF-16 code units and does not understand natural-language word boundaries. The full process also includes match expansion and sorting, so its total cost cannot simply be written as O(n).

Matching supports arbitrary text, exact whole sentences, and literal phrases, but not arbitrary regular expressions. If compilation or matching fails, the system can fall back to a more direct reference scanner.

Page translation's semantic identity does not directly include the provider or model. With other semantic conditions unchanged, switching models can still hit an old result. Only “force retranslation” skips translation memory and the legacy cache, replacing memory under the same semantic key after successful result validation.

Selected-text translation and the Translation Center use their own entry-point contracts, and their translation configurations include the provider identifier. Thus, a provider A card in the Translation Center does not use a provider B card's results; changing the model within the same provider configuration preserves the existing identity.

Context, rules, and versions are also part of the semantic contract, with translations found by exact hash matching. Identical sentences on different pages can share results only when their other semantic conditions also match; merely similar text does not automatically hit.

For one translation resource, use the latest time threshold; if slots must also be released or writes completed, wait for those events. Across candidate work items, find the earliest possible wakeup time. Fixed-capacity circular time buckets prevent memory from growing without bound with historical requests: 62 buckets of 1 second for the minute window; 1442 buckets of 60 seconds for the daily window.

A debit's time-bucket index is ceil(attribution time ÷ bucket duration), rounded upward. When creating a persistence snapshot, the attribution time of pending debits is set at least 1000ms beyond now; after the transaction completes, ensure it is not before the current time, using conservative expiry. The request token bucket uses integer counts: one request equals 1,000,000 microunits. This is unrelated to the tokens into which models split text.

When multiple translation resources are available, first compare “compatible demand within the candidate window ÷ available concurrency,” using a denominator of at least 1. Then compare load, routing priority, and stable resource-identifier order. This heuristic selects among resources for the current candidates; it does not solve for a globally optimal allocation.

This local estimate determines whether dispatch is allowed; it may differ from the provider's actual tokenization and cannot replace the account's remaining allowance. Page batches also include translation instructions and context in the estimate, so source character count alone cannot stand in for tokens-per-minute usage.

Adaptive concurrency is off by default. When enabled, the concurrency limit starts at half the hard limit, rounded up. When a successful response arrives and the statistics window has reached 20 seconds, check whether the previous window had at least 10 successes, pending work, and no congestion; only then add 1 slot. Capacity congestion halves the limit, at most once per adjustment cycle. Automatic fallback switching is also off by default and switches to other resources under eligible conditions only when explicitly enabled.

The parser permits a leading reasoning section enclosed in designated markers, and Markdown fences that enclose the entire response and are unlabeled or labeled JSON. Valid whitespace and reordered data fields are accepted too. It does not guess and extract JSON from arbitrary prose. Large streaming increments are processed in chunks of 32,768 UTF-16 units, preserving text, escape, and Unicode parsing state across chunks.

The product direction is “auto-detect source → Traditional Chinese for Taiwan.” Page translation and shared settings use different Traditional Chinese language tags; Microsoft additionally converts via an ISO language mapping table, falling back to a Traditional Chinese tag when no mapping is found. Integrations across services must handle these differences in language representation.

The system prompt asks the model to ignore instructions in the source and context, and to use plain text for translation content. This expresses the intended behavior but cannot guarantee elimination of prompt injection or translation errors. DeepSeek's thinking mode is off by default, and users may adjust it.

The rendering-cadence controller coalesces arriving results and adjusts cadence using recent work duration: new cost estimate = 0.6 × old cost estimate + 0.4 × current duration; next interval = min(1000ms, 2 × new cost estimate). A 250ms timer provides a fallback when the browser's animation-frame callback has not arrived. This tracks work duration, not the full cost of browser layout and painting.

Retain only the latest pending update plan per block, with at most 128 deferred plans. Beyond that, the system can record “refresh needed” markers, capped at 4096. Reconciliation after page changes coalesces into one in-progress job per paragraph plus one rerun-pending marker.

Rendering summaries use “text length + 32-bit FNV-1a hash” to decide whether rewriting can be skipped, unlike the background's SHA-256 semantic identity; different content can still have identical summaries. Multi-sentence translations are combined in original translation-unit order, with joins estimated from Chinese, Japanese, and Korean spacing rules rather than reproducing the original whitespace character for character.

Retaining the same original nodes helps preserve object relationships; event bindings, layout, and interaction behavior still depend on the website's own structure and update approach.

Waiting content and content retained after persistence failure have a 10-minute retention period, checked and cleaned up during later queue operations or retries. Capacity is estimated as “source UTF-8 bytes + translation UTF-8 bytes × 2 + 1024,” which may differ from actual memory use. When full, the queue pauses new page-translation dispatches; requests already sent may finish.

This queue exists only in runtime memory. When the background service ends, content whose write has not completed is not automatically restored on restart. The diagram depicts page-translation persistence states; interactive entry points only enqueue persistence work and do not wait for the same complete on-screen state flow.

Delayed retries for ordinary work use full jitter: choose a random delay between zero and the upper wait bound each time. Capacity-related HTTP 429 instead sets a cooldown for the limited resource or quota scope, using the server-specified wait time, or the base delay if none is provided, without random jitter.

Jitter spreads out retries; the response header Retry-After adds a lower time bound, accepting seconds or an HTTP date and accounting for server clock differences. Server-requested waits are retained for at most 7 days.

Wakeup scheduling is coalesced into one timer for the earliest deadline; when the expected wait is at least 30 seconds, a browser background alarm supplements it. Clock checks compare elapsed system time with elapsed monotonic time: a difference over 5 seconds, or a system-clock rollback over 5 seconds, pauses new dispatches. Conservative recovery occurs only after 2 stable seconds without another anomaly.

The cancellation diagram describes network attempts already sent. Attempts with reserved slots but not yet sent may end and release their slots first; if an ongoing debit transaction later commits successfully, that conservative debit remains even if the service was never called. Network operations whose termination cannot be confirmed quarantine the relevant resources and retain their slots until a late termination notice arrives. Restarting the background service establishes a new execution generation: old requests no longer occupy concurrency slots, but the known quota ledger is retained.