feat(app): a Repeat reads switch in the main window and a Repeat reads card on the Optimizer tab

[HackPoint]

What

Repeat reads are on by default in 1.8.0, and until now the only way to turn them off was an environment variable or a hand-edited JSON file, with nothing in the app showing what they had done. This adds a switch to the main window and a card to the Optimizer tab.

Nothing the MCP server sends changes (no tool output, tools/list or hook), and nothing is added to the hook path. In lumen-core only the unusable-settings-file message changes: it now names the switch.

The switch

• On the Context tab of the main window, on by default. It writes lumen_settings.json (atomically, readable by its owner only, the other keys of a valid file kept). The server reads that file on every tool call, so a change applies to the next Lumen read in every running Claude Code session, with no restart, and survives a restart. The window says so.
• It shows the settings file, not the server. It cannot see a LUMEN_REPEAT_READS in the MCP config Claude Code gives the server (which outranks the file), nor a LUMEN_DB there that points the server at another ledger. Its text says both.
• A file that cannot be used is shown as off, with the reason: the same off the server applies. Switching on rewrites it. While a write is in flight the switch stays focusable (marked aria-disabled) and ignores a second flip, so a keyboard user does not lose their place. A setting that cannot be read shows neither state and offers Retry. A change that cannot be saved puts the switch back and says why.
• lumen doctor and the unusable-file message name the switch as the repair.

The card

On the Optimizer tab, below every total: unchanged answers served, edits served as a delta, tokens not re-sent, and how often the full result was asked for again with fresh=true (n of m, labelled as re-requests). It is counted apart from every total (token totals, by-tool and by-channel breakdowns, net value) and shows no dollar figure and no percentage. The token figure covers only answers whose count was measured: the card says how many when that is fewer than all, and shows a dash when none. It says when the settings file turns repeat reads off, and counts the reads answered in full because Lumen could not tell the conversation, with a link to Setup. An empty ledger says "No repeat read has been answered short yet." with no figure. A backend that does not send the figures gets no card.

lumen-stats: the optimizer report gains repeatReads; no schema change. Tauri: two commands read and write the setting through lumen_core::settings.

Styles

The card shares the Optimizer page's existing card, title, row-list, figure and note rules by adding its selectors to them, instead of repeating their declarations. With its own copies the page's stylesheet was 8.80 kB, over the 8 kB build budget; it is 7.83 kB now (main: 7.13 kB, limit 8 kB). home.css is at 7.93 kB of 8 kB: the next rule on either page has to share, or the budget has to move. The existing blocks render as before; the card takes their values (rows 0.5rem apart, figures 0.84rem). The template has no class without a rule; a guard in lumen-core enforces it.

One wording decision, yours to change

The brief says "tokens saved". The card says "tokens not re-sent", counted against reading the whole file again: with repeat reads off a read returns the outline, usually far smaller, so a saving against the whole file would overstate what repeat reads add. The card and the README name the baseline.

Tests

• Frontend, each state driven by a fixture and asserted on the rendered text: 19 for the switch (every file state, failed read and Retry, failed write, switching on over a bad file, the switch held still with the focus kept while a write is in flight, nothing claimed before the first read), 25 for the card (none yet, off, each figure and its baseline, partial and no measurement, "0 of m", "1 delta" not "1 deltas", the unplaced-reads line and Setup link, no card without the figures), 22 for the service (read and write ordering, errors cleared, answers that are not a setting).
• Rust: 10 in lumen-stats for the new aggregate, one of which seeds all four served routes and checks that none reaches any of the six totals; 17 for the Tauri commands (every file state, a failed write names the file and the operating system's reason, the variable not read, every command the window invokes is registered, and the argument keys the window sends are the parameters each handler takes); 1 in doctor; the settings message test updated; in lumen-core one test that a reader holding the file open still sees the whole old file after a write.

Gate

Local gate green on 31772f8 (446 s): fmt, clippy -D warnings (workspace and Tauri crate), workspace tests with RUSTFLAGS=-D warnings, frontend build under the style budgets and tests with the coverage thresholds, e2e, Tauri crate and hook e2e tests, MSRV 1.94.0, install verify, debug app build. Run with LUMEN_RANKED_TIME_BUDGET_MS=60000: the lumen-core ranked-outline budget test is wall-clock bound and flakes on a loaded machine; nothing here touches it.

Each new test was proved to fail without its fix: the settings write done in place, the handler removed from the list, the operating system's reason dropped from the message, the switch disabled while writing, the plural, the argument sent as { enabled: on }, the handler's parameter renamed, and any one of the four served routes added to any of the six route lists in lumen-stats (24 changes; the test before this seeded two routes and passed 12 of them).

Not run locally: the tauri-driver app e2e, which runs on CI Linux (GUI build + test). No e2e touches the switch, and the argument test compares the two sources rather than running the Tauri runtime.

CRLF clone (core.autocrlf=true): the lumen-core, lumen-stats and lumen-cli tests and the new Tauri tests, none failed.