feat: cooperative proxy mode — forward native tools to billion-context proxy (内外呼应)

[ranxianglei]

What

The "inside" half of the cooperative plugin protocol (billion-context issue dog/billion-context#1, gitea; protocol half = ranxianglei/billion-context#161). When pi's model baseUrl routes through a billion-context proxy (the /bili/ zero-config prefix), the extension automatically switches to cooperative mode — 有外用外:

• Tools stay native, execution forwards: compress / decompress / search_context / acp_status keep their native pi registration (native UX, permissions, audit), but execute POSTs to the proxy's /__bili/plugin/tool, which runs them under its session lock against the same view the model saw. Result text returns as the native tool result.
• Real session identity: every provider request carries x-bili-plugin: pi + x-bili-plugin-conversation: <pi session id> (via the SDK's before_provider_headers hook). The proxy keys state by pi's real session id — no more content-fingerprint collisions — and suppresses its own wire tool injection (no double compression).
• In-process pipeline skipped: context returns pi's messages untouched (the proxy runs processTurn — tags, folding, nudges — at the wire level); before_agent_start drops the local philosophy prompt (proxy injects it) while keeping the delegate prompt (pi-side feature).

Detection

Stateless per-call check: bili path segment in ctx.model.baseUrl → proxy origin. Model switches mid-session are picked up on the next call; nothing is cached. ACP_COOPERATIVE_PROXY=0 disables entirely.

Compatibility

Without a /bili/ baseUrl every path is byte-identical to before (early-return branches only). MITM-mode (transparent proxy, no prefix) is NOT detected — documented; use the /bili/ prefix baseURL for cooperative mode.

Verification

• npm run typecheck ✅
• New suite tests/cooperative.test.ts 7/7 (URL matrix, header injection + absence, context identity passthrough, prompt ownership, live HTTP forward with conversation-id/args/result assertions, error propagation, kill switch)
• npm test: 300 tests — 299 pass + 1 pre-existing failure on pristine master (e2e-compress-config.test.ts "2w limit fires nudge" — verified failing on a clean master checkout before these changes; unrelated)
• npm run build ✅

Files

• src/cooperative.ts (new) — detection + forwardToolToProxy / tryForwardTool
• src/index.ts — wireProviderHeaders + cooperative early-returns in context/system-prompt wiring
• 4 tool files — tryForwardTool branch at top of execute
• tests/cooperative.test.ts, README.md section, CONFIGURATION.md ACP_COOPERATIVE_PROXY
• devlog/2026-08-16_cooperative-proxy-mode/ — REQ + DESIGN + WORKLOG

Ordering note

No code-level coupling, but the endpoints this forwards to land with ranxianglei/billion-context#161 — merge that first for the combination to work end-to-end.

Refs: dog/billion-context#1 (gitea)

[ranxianglei]

已更新：

1. 合并 master（symlink-hardening fix(decompress): harden toFile against symlink path-escape #140 后续，BEHIND 已消）
2. 修复 pre-existing master CI 红灯：e2e-compress-config 的 2w-limit 断言（feat(config): three-level compress cascade (global > provider > model) #145 引入）被 fix: recommend only viable compressible ranges (>=200 tokens) on all surfaces #148 的 viable-ranges 行为打破 —— kernel 现在只把 tokens*4 >= minCompressRange(5000) 的合并区间计入有效 T1 内容，12×2000-char fixture 全部低于门槛，EMERGENCY nudge 被正确压制。fixture 加大到 3000 chars/msg（内核行为是有意的防碎片设计，不该改内核）。本地 303/303，typecheck/build 绿。