feat(autocomplete): personalise suggestions via accepted-completion history

[oxoxDev]

Summary

• Adds a persistent history of accepted autocomplete completions stored in the local KV store
• Feeds recent accepted completions back as dynamic style examples during inference, creating a personalisation feedback loop
• Exposes two new RPC methods (autocomplete_history, autocomplete_clear_history) with full controller schema registration
• Adds a "Personalization History" section to the autocomplete settings panel with list view and clear button
• No new dependencies, no schema migrations — built entirely on existing MemoryClient KV infrastructure

Problem

• Autocomplete starts cold every session with no memory of what the user previously accepted (issue Improve Autocomplete and Screenshot Features #71)
• The inline_complete() call already supports style_examples for in-context learning but they were only static (user-configured) — no automatic personalisation existed
• Users had no visibility into what completions were accepted or how the system was learning

Solution

• New history.rs module persists each accepted completion (context, suggestion, app_name, timestamp_ms) under KV namespace "autocomplete" with zero-padded timestamp keys for correct ordering, auto-trimmed to 50 entries
• On every Tab-accept or RPC-accept, a fire-and-forget tokio::spawn saves the completion — zero latency impact on the acceptance path
• Before each inference cycle in refresh(), the 6 most recent accepted completions are loaded from KV and merged with user's static style examples (capped at 8 total), then passed to inline_complete() — the model sees what this user liked before
• Design tradeoff: MemoryClient::new_local() is called inline per operation (not stored on AutocompleteEngine) to avoid changing the Lazy singleton constructor — matches the established pattern in memory/ops.rs

Testing

• cargo check --manifest-path Cargo.toml --bin openhuman
• Pre-push hooks passed: prettier --check, eslint, tsc --noEmit, cargo fmt --check
• yarn -s compile
• cargo check --manifest-path app/src-tauri/Cargo.toml
• Manual: accept 3-5 completions via Tab → entries appear in Settings → Autocomplete → Personalization History
• Manual: observe suggestion quality shift after ~6 accepts
• Manual: click "Clear History" → list empties, suggestions revert to cold-start

Impact

• Desktop only (macOS) — autocomplete is platform-gated; no mobile/web/CLI runtime impact
• Performance: KV reads are local SQLite, expected <5ms per inference cycle; saves are async fire-and-forget
• Storage: max 50 KV entries in ~/.openhuman/workspace SQLite — negligible disk footprint
• Security: all data stays local, no network calls, no PII sent externally

Breaking Changes

• None

Related

• Issue(s): Improve Autocomplete and Screenshot Features #71 (first half — autocomplete personalisation)
• Follow-up PR(s)/TODOs: screenshot feature improvements (second half of Improve Autocomplete and Screenshot Features #71), unit tests for history.rs, E2E test for history RPC roundtrip

🤖 Generated with Claude Code