fix(compress): accept JSON-string content in parseCompressInput

[ranxianglei]

背景

非严格工具 provider(vLLM openai-completions,supportsStrictTools:false)会把嵌套数组参数字符串化:模型调用 compress 时传的 content 是 JSON 字符串而非数组。

此前 parseCompressInput 里 Array.isArray(obj.content) 为 false → 走 toRange(obj)(顶层无 startId/endId/summary)→ null → 静默解析出 0 个 range,压缩从不执行(只有一条 warn 日志,无任何报错)。

改动

src/compress-tool.ts 的 parseCompressInput 在数组判断前,若 content 是字符串则先 JSON.parse(非法 JSON → warn + 0 ranges,不崩溃)。与 billion-context-pi PR #164(normalizeRanges)对称。

tests/basic.test.ts 增加 JSON 字符串形式的回归测试。

验证

• npm run typecheck(tsc --noEmit --project tsconfig.build.json):clean
• npm test:464 pass / 0 fail

不合并,等 review。

[ranxianglei]

LGTM — verified against all five call sites (stream.ts, stream-openai.ts, stream-responses.ts, compress-loop-responses.ts, loop/core.ts) and the new regression tests pass (25/25 in tests/basic.test.ts). Mirrors the sibling fix in billion-context-pi.

One non-blocking note: if content is a JSON string of a single object (e.g. '{"startId":"m1","endId":"m2","summary":"x"}'), JSON.parse succeeds but the result isn't an array, so the fallback reads startId/endId/summary from the top-level obj (line 70) — which only carries content — and you get 0 ranges. Same result as before the patch, so no regression; just flagging that the lenient path only covers array strings. If you want the single-object hallucination to work too, one line: after the parse, if (content && typeof content === "object" && !Array.isArray(content)) content = [content];. The schema declares content as an array, so I'd merge as-is.