feat(native): keep conversation history within a token budget

[bgrana75]

Keep conversation history within a token budget

Builds on #76 (CI); the last commit is this PR's. Finding 5 of the review.

ChatViewModel sent the last 6 messages verbatim (VERBATIM_MESSAGE_COUNT). With detailed answers (~500 tokens each) that reached ~1,700 tokens. The history sits after the retrieved sources in the prompt (RagPure.assembleChatMessages), and the sources change every question, so all of it was read again for every question. In the Expo app the same history took a follow-up's first word from 1.6 s (new chat) to 20–48 s (Pixel 6a, 1.5B, Oct 3).

HistoryBudget (engine, answer/) ports src/routing/historyBudget.ts from #72:

• newest turns first, within 400 tokens;
• questions kept whole, earlier answers cut to their opening sentences (80 tokens), with " …";
• the latest exchange always kept, even over budget.

It's used in two places in ChatViewModel: the answer's history, and the summary's input (the older turns, which could be thousands of tokens).

Golden test. src/routing/historyBudget.ts and its vitest tests come from #72 unchanged, since fullnative-dev doesn't have them. scripts/export-native-golden-history.mjs runs the real TS and writes golden/history-budget.json:

• 438 leadSentences cases: 60 corpus passages and 13 edge cases (abbreviations, decimals, quotes, newlines, Portuguese, one huge sentence), each at caps of 0–200 tokens;
• 34 budgetHistory cases: 17 conversations at the default and at a tight budget, 8 of which drop turns.

HistoryBudgetGoldenTest checks the Kotlin output matches exactly. The Kotlin uses Compress.splitSentences and Compress.approxTokens, which port the same functions (the splitSentences in src/rag/compress.ts and src/routing/context.ts are character-for-character identical, as ResearchContext already relies on).

Not run locally. No Android toolchain here, so CI (#76) is the first place the Kotlin compiles and the golden test runs; the TS tests pass locally (7/7). Not measured on a phone. No CodeRabbit CLI review was run before pushing (cr isn't installed here).

Touches ChatViewModel.kt at the history and summary lines only, not at send()'s start (#78) or preload().

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

• Configuration used: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: 421c7433-8ccc-472b-932d-1865a3d579cf

---

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches 💡 1

🛠️ Fix failing CI checks 💡

• Commit to this branch
• Create a new PR

---

---

🧪 Generate unit tests (beta)

• Create a new PR

---

---

---

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

---

_{Comment @coderabbitai help to get the list of available commands.}

[rferrari]

Approve. The golden file regenerates byte for byte from the TS, and HistoryBudgetGoldenTest
passes on the merged branch. One difference from #72: the budget applies to the last 6
messages rather than the whole history. That's stricter, and fine with the 400-token
budget. Bounding the summary's input too is right; a long detailed chat was the slow case.

[bgrana75]

@coderabbitai review

[coderabbitai]

⚠️ Action not completed

Review rate limited.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[bgrana75]

@coderabbitai review

[coderabbitai]

⚠️ Action not completed

Head commit changed.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[bgrana75]

@coderabbitai review

[coderabbitai]

⚠️ Action not completed

Pull request is closed.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.