feat(cli): prefix-cache-preserving compaction (Reasonix-style)

[miroyong]

Summary

Ports Reasonix's prefix-cache strategy to the Continue CLI's auto-compaction so providers with automatic prompt caching (DeepSeek, OpenAI) can serve cache hits across turns.

Problem: today compactChatHistory collapses the history into [system, summary], and getHistoryForLLM trims everything before the compaction index. Any turn after that starts with a different prompt prefix than the previous turn, cold-starting the provider's prompt cache. Additionally, the system message is reconstructed on every streaming iteration — re-reading AGENTS.md/CLAUDE.md and re-running git status — so any change invalidates the whole prefix.

Changes

• compaction.ts
  • compactChatHistory pins the cache-stable prefix verbatim (system message, first user turn when ≤1.5K tokens, and any prior digests) and keeps a token-budgeted recent tail (12K tokens, ≥2 messages), splicing the new digest in the middle instead of replacing everything.
  • No-op path: when the conversation already fits within pinned prefix + tail, the summarizer is not invoked and history is returned unchanged.
  • getHistoryForLLM returns the full history — after compaction the stored history is the compacted layout (pinned prefix + digest + tail), so trimming before the compaction index would drop the cache-stable prefix.
  • The digest prompt is now a structured briefing (## Standing facts & constraints, ## Goal, ## Decisions & rationale, ## Files & code, ## Commands & outcomes, ## Errors & fixes, ## Pending & next step), mirroring Reasonix's summarySystemPrompt.
• services/SystemMessageService.ts: memoizes the constructed system message per (mode, additionalRules, format, headless), keeping the prefix byte-identical across turns.
• services/ChatHistoryService.ts: getHistoryForLLM returns the full history (the compacted layout is already prefix-preserving).
• Tests updated/added: compaction.test.ts, compaction.infiniteLoop.test.ts (new tokenizer mock surface), ChatHistoryService.test.ts, SystemMessageService.test.ts.

Test plan

• npx vitest run src/compaction.test.ts src/compaction.infiniteLoop.test.ts src/compaction.pruneLastMessage.test.ts — 38/38 pass
• npx vitest run src/services/ChatHistoryService.test.ts src/services/SystemMessageService.test.ts src/systemMessage.test.ts src/session.test.ts src/messageConversion.test.ts src/services/ServiceContainer.test.ts — 99/99 pass
• npx tsc --noEmit — clean
• npx eslint on changed files — clean

Note: the full CLI suite hangs in the TUIChat UI tests on this Android/Termux environment (unrelated, pre-existing).

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[miroyong]

Summary

This PR ports Reasonix-style prefix-cache-preserving compaction to the CLI. The core idea: after a compaction, the prompt prefix stays byte-identical across turns (system message + first user turn verbatim + prior digests), so providers with automatic prompt caching (DeepSeek, OpenAI) serve cache hits on subsequent requests.

Changes

compaction.ts

• Structured briefing prompt (inspired by Reasonix's summarySystemPrompt): output under fixed headings — ## Standing facts & constraints, ## Goal, ## Decisions & rationale, ## Files & code, ## Commands & outcomes, ## Errors & fixes, ## Pending & next step — instead of free-form prose. Rules require terse fragments, exact identifiers/paths/numbers, and no invented content.
• getPinnedPrefixLength: computes the leading messages that a compaction must preserve verbatim — system message, the first user turn (when ≤ 1.5k tokens, so it stays a brief and a fold never summarizes the user's stated facts away), and any prior compaction digests (so a later fold never re-summarizes an earlier digest into nothing).
• Bounded recent tail: RECENT_TAIL_TOKEN_BUDGET = 12_000 / MIN_RECENT_TAIL_MESSAGES = 2 — the verbatim tail after a fold is budgeted so large tool outputs cannot keep the history above the auto-compaction trigger and re-fire compaction every turn.
• No-op fix: when the whole history fits within the budget, compaction returns compactionIndex = -1 without invoking the summarizer (the previous loop condition never included the last message, so small conversations were compacted unnecessarily).

ChatHistoryService.ts

• getHistoryForLLM now returns the full history (previously it dropped everything before compactionIndex). This is what actually preserves the prefix: the LLM receives system + pinned first turn + digests + tail, and compactChatHistory filters internally for what gets summarized.

SystemMessageService.ts

• Caches the rendered system message by (mode, rules, format, headless) — relevant for cacheability since the system message is part of the prefix.

Tests

• compaction.test.ts rewritten for the new contract (pinned prefix, tail budget, no-op path, digest chaining).
• compaction.infiniteLoop.test.ts updated for the new tokenizer mock and exercises the pruning loop with a large history.
• ChatHistoryService.test.ts / SystemMessageService.test.ts updated/added for the new behavior and the system-message cache.

All CLI tests pass on the CI matrix (Node 18/20/22/24 × ubuntu/macos/windows).

[miroyong]

I have read the CLA Document and I hereby sign the CLA