billion-context-omp 0.2.6 → 0.2.7

package/README.md

@@ -144,7 +144,7 @@ Create `~/.omp/acp-omp.json` (global) and/or `<project>/.omp/acp-omp.json` (proj
 | Key | Default | Description |
 |-----|---------|-------------|
 | `debug` | `false` | Enable verbose **debug-level** events in the log. The always-on log (lifecycle events, errors, warnings) is written regardless; `debug` only adds extra diagnostics. Also enabled by env `ACP_DEBUG=1`. |
-| `transformMode` | `"provider"` (default since v0.2.6) or `"context"` — where the compression surgery intercepts. `provider` transforms the provider wire payload (request-local, structurally immune to feedback re-entry). `context` is the legacy rewrite-in-place mode. |
+| `transformMode` | unset (resolved per API) or explicit `"provider"` / `"context"` — where the compression surgery intercepts. Unset: `provider` where the host applies the wire-payload replacement AND the wire body has a codec path (anthropic-messages, ollama-chat; openai-completions on hosts >= 17.3.8 — upstream PR can1357/oh-my-pi#8717, issue #83); `context` everywhere else (issue #79). Explicit: always honored. `provider` transforms the provider wire payload (request-local, structurally immune to feedback re-entry); `context` rewrites the context event (legacy mode). |
 | `autoUpdate` | `true` | On session start (throttled to one check per 3 minutes), check npm for a newer version and auto-install it. Disable to avoid all startup network calls. |
 | `modelContextLimit` | *(auto)* | Override the context limit (in tokens). Defaults to the model's `contextWindow`. |
 | `toolBashDefaultTimeout` | `60` | Seconds injected into the `bash` tool when the model omits `timeout`. Without this a forgotten timeout can hang for thousands of seconds. `0` restores unbounded behavior. |

package/README.zh-CN.md

@@ -146,7 +146,7 @@ billion-context-omp 开箱即用，无需配置。可选键写入 JSON 配置文
 | `debug` | `false` | 开启**debug 级**详细日志事件。常开日志（生命周期事件、错误、警告）无论如何都会写；`debug` 只增加诊断信息。也可用环境变量 `ACP_DEBUG=1` 开启。 |
 | `autoUpdate` | `true` | 会话启动时（节流为每 3 分钟最多一次）检查 npm 是否有新版本并自动安装。关闭可避免所有启动期网络请求。 |
 | `modelContextLimit` | *(自动)* | 覆盖上下文上限（token 数）。默认取模型的 `contextWindow`。 |
-| `transformMode` | `"provider"`（v0.2.6 起默认）或 `"context"` | 压缩手术拦截位置。`provider` 变换 provider wire 载荷（请求局部，结构性免疫回灌环路）。`context` 是旧版原地改写模式。 |
+| `transformMode` | 未设置（按 API 解析）或显式 `"provider"` / `"context"` | 压缩手术拦截位置。未设置：host 会应用 wire 载荷替换且 wire 体有编解码路径的 API 用 `provider`（anthropic-messages、ollama-chat；openai-completions 需 host >= 17.3.8，上游 PR can1357/oh-my-pi#8717，issue #83），其余用 `context`（issue #79）。显式设置总是生效。`provider` 变换 provider wire 载荷（请求局部，结构性免疫回灌环路）；`context` 改写 context 事件（旧版模式）。 |
 | `toolBashDefaultTimeout` | `60` | 模型省略 `timeout` 时注入 `bash` 工具的秒数。没有它，一次忘记的 timeout 可能挂起数千秒。`0` 恢复无限制。 |
 | `toolOutputMaxBytes` | `200000` | 工具结果文本的硬字节上限（经 `tool_result` 钩子实施）。拦截 omp 自身上限管不住的失控输出。触发时模型会被告知完整输出在哪；调低（如 `8192`）可更省上下文，`0` 禁用。 |
 | `compress.maxContextLimit` | `"75%"` | 触发**强制压缩**提醒的上下文用量阈值（绕过增长门控与节拍）。接受比例（`0.75`）或百分比字符串（`"75%"`）。越低 = 越早/越激进压缩。 |

package/dist/config.d.ts

@@ -28,14 +28,20 @@ export interface CompressConfig {
  */
 export interface AdapterConfig {
     /** Where the compression surgery intercepts (issue #52). "provider"
-     *  (default since v0.2.6) leaves the agent array untouched and transforms
-     *  the WIRE payload at before_provider_request — request-local, no
-     *  re-entry, structurally immune to omp's feedback-view loops (the recap /
-     *  subagent re-feed pathology, issues #22/#52). Unknown provider formats
-     *  pass through untransformed (fail-open). "context" rewrites the context
-     *  event — battle-tested legacy mode, kept for compat.
-     *  Existing setups pin the mode explicitly via `transformMode` in
-     *  ~/.omp/acp-omp.json; unset now resolves to "provider". */
+     *  leaves the agent array untouched and transforms the WIRE payload at
+     *  before_provider_request — request-local, no re-entry, structurally
+     *  immune to omp's feedback-view loops (the recap / subagent re-feed
+     *  pathology, issues #22/#52). Unknown provider formats pass through
+     *  untransformed (fail-open). "context" rewrites the context event —
+     *  battle-tested legacy mode, kept for compat.
+     *  When omitted, the mode is resolved per model API (issue #79):
+     *  "provider" where the host actually applies the wire-payload
+     *  replacement AND the wire body has a codec path — anthropic-messages,
+     *  ollama-chat, and openai-completions on hosts >= 17.3.8 (upstream PR
+     *  can1357/oh-my-pi#8717, issue #83); "context" everywhere else
+     *  (older hosts drop the replacement; bedrock/cursor/responses/google
+     *  bodies have no codec path yet). Explicit pinning in
+     *  ~/.omp/acp-omp.json always wins. */
     transformMode?: "context" | "provider";
     /** When omitted, the adapter reads `ctx.model.contextWindow` live each turn.
      *  Set explicitly for tests/headless runs. */