fix(panel): correct session-accounting label, mark tokensCompressed cumulative

[ranxianglei]

Fixes the panel copy items from ranxianglei/billion-context-pi#267.

Session-accounting label: "never shrinks" was false

The Context (session accounting, host footer scale) line claimed the number "never shrinks; includes compressed originals". Measured in a real session, the host-reported value really does drop after a compress (summaries replace the pruned originals in the window) — e.g. a T2 distillation took it from 55.5k to 48.4k.

Relabeled to includes compressed originals; shrinks slower than the sent view: it still carries the compressed originals, so it lags the sent view, but it is not append-only. Doc comments on StatusPanelInput.tokenCount and buildStatusPanel updated to match.

Blocks figure marked cumulative

Blocks: N active / M total (X tokens compressed) — state.stats.tokensCompressed is the cumulative historical total, which read as a contradiction next to the current Session-only number. Now labeled tokens compressed, cumulative.

Tests

tests/panel.test.ts assertions updated for both labels. Full suite: 568 pass.

Note for the bcp side: after this lands and is published to npm, billion-context-pi should bump its pinned acp-kernel to the new version (see ranxianglei/billion-context-pi#268).

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-01_panel-copy-fixes (2bbb1a4)

Option A — Install from npm PR tag (recommended)

In your adapter project:

npm install acp-kernel@pr-182

Each push to this PR publishes a new version under the pr-182 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf acp-kernel-pr182.tgz
npm install ./package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 ework agent · qwen3.8-27b

[bot] 🏷 Reviewed PR #182 (branch 2026-09-01_panel-copy-fixes, 2 commits).

Panel copy fix (81f61ed) — the change this issue describes: ✅ verified

• src/panel/panel.ts:89 session-accounting line → includes compressed originals; shrinks slower than the sent view (consistent with the measured T2 drop, 55.5k→48.4k).
• src/panel/panel.ts:150 and :159 Blocks line → tokens compressed, cumulative (both the >0-active and 0-active branches).
• Doc comments on StatusPanelInput.tokenCount (panel.ts:14-16) and buildStatusPanel (panel.ts:51-55) updated to match.
• tests/panel.test.ts:36 and :58 assertions updated.
• No leftover "never shrinks" in panel.ts. (The "append-only" mentions in src/search/* are a different concept — the host's raw session log — and are correctly left alone.)
• Pre-flight all green: typecheck clean, npm test 568 pass, build clean, check-pr.sh OK (branch name + no version bump).

Sync fix (0240d90) — not described in this issue: ⚠️ needs a decision
This commit is separate from the panel work and isn't mentioned in the issue body. What it does:

• Adds CompressionBlock.expanded?: boolean (src/types.ts:39) — marked "host-set".
• syncBlocks (src/sync.ts:72) now keeps an expanded block deactivated (and omits it from the deactivated report) instead of re-activating it.
• 2 tests in tests/sync-config.test.ts cover it.

Logic is sound and backward-compatible (non-expanded blocks keep the legacy resurrection), and expanded round-trips through persist. Two things to confirm before this lands:

1. Scope — the issue is titled fix(panel) and only describes the copy fix. Should the sync fix be described here, or split into its own PR?
2. It's inert without host cooperation — no kernel code path sets expanded (deactivateBlock sets active:false only; decompress/collectBlockContent don't touch it), and it isn't documented in README/DESIGN. So it only takes effect if a host manually sets block.expanded = true after a user decompresses. Is billion-context-pi doing that? If yes, a one-line note in the API docs would help; otherwise consider an expanded option on deactivateBlock.

Formatting — npm run format:check flags ~92 files, but that's pre-existing/repo-wide (master already fails on panel.ts + panel.test.ts) and CI does not run format:check. Non-blocking.

I won't merge (human-only). Once the sync-fix scope is confirmed and the description covers it, this is ready for you to merge.

[ranxianglei]

Review 结论:✅ 通过,可合并

• 只动 panel.ts/sync.ts/types.ts,与 0.0.55 的 prompt 改动(merge write-side quote fidelity into tuned compress prompts (verbatim quotes require refs; task state as history; no \uXXXX copying) #205,compression-rules.ts)无文件冲突,无行为耦合
• 两处文案修正语义准确:"never shrinks" 确实不成立(summaries 替换被 prune 的原文后会降),改为 "shrinks slower than the sent view" 既纠正事实又保留了双账本差异的提示;tokensCompressed 标 cumulative 消除与当前 active blocks 的表面矛盾
• CI 绿

发布顺序(bcp #268 依赖):本 PR 合并 → npm 发版(npm view acp-kernel version 确认,注意 registry 元数据可能延迟几分钟)→ bcp #268(已 rebase 到 v0.1.58,详见其 PR 评论)升级内核引用后合并。

[ranxianglei]

🤖 ework agent · qwen3.8-27b

[bot] 🏷 收到,谢谢审核。我的独立复核与你的结论一致:

• 面板文案修正准确、完整:src/panel/panel.ts:89(会话账本行)、:150/:159(Blocks 行)+ doc 注释 + tests/panel.test.ts:36/:58 断言;typecheck / test(568 pass)/ build 均绿。
• sync 的 expanded 改动向后兼容(非 expanded 块保留原复活行为),你确认并入本 PR,无异议。

一点提醒(不阻塞):expanded 目前没有任何内核代码路径会置位(deactivateBlock 只设 active:false,decompress/collectBlockContent 不改写),需 bcp 侧在用户解压后手动置 block.expanded = true 才真正生效。若 bcp #268 顺带补了这步会更完整——需要的话我可以去 bcp 核对一下。

按规则我不能执行 merge(仅人工操作)。请自行合并本 PR(branch 2026-09-01_panel-copy-fixes,对应 issue #182)。

发布顺序我这边理解一致:合并本 PR → CI 自动打 tag 并发 npm → npm view acp-kernel version 确认新版上线(registry 元数据可能延迟几分钟)→ bcp #268(已 rebase 到 v0.1.58)升级内核引用后合并。

合并后若要我帮忙核对发布版本、或去 bcp 确认 expanded 的设置,说一声即可。

[ranxianglei]

冲突了

[ranxianglei]

🤖 ework agent · qwen3.8-27b

[bot] 🏷 冲突已解决并推送,分支 2026-09-01_panel-copy-fixes 已更新到 2bbb1a4(fast-forward,无 force-push)。

冲突根因:master 已从 v0.0.48 前进到 v0.0.55(其间合了 #183–#206),与本 PR 在 panel.ts/sync.ts/types.ts/sync-config.test.ts 四处有交集。实际只有 src/panel/panel.ts 一处真冲突,其余三处自动合并。

冲突点:buildStatusPanel 的 doc 注释块——本 PR 改了第 1 条("never shrinks" → "shrinks slower than the sent view"),而 master 的 countTokens 重构(#203/#205)独立把第 2、3 条的标尺从 "chars/4 est." 改成了 "countTokens scale"。两边改在同一块里就撞了。

解法:两者合并保留——第 1 条用本 PR 的修正措辞,第 2、3 条用 master 的 countTokens 标尺。语义一致,无信息丢失。

验证(合并后整树重跑):typecheck 干净、npm test 582/582 pass、build 干净;相对 master 的净 diff 恰好只有本 PR 的 5 个文件(panel.ts/sync.ts/types.ts/panel.test.ts/sync-config.test.ts),无合并杂音;version 继承 master 的 0.0.55(本分支未动过 version)。

现在该分支与 master 无冲突、可合并。请你自行 merge(我按规则不代合)。合并后 CI 自动发 npm,再走 bcp #268 升级引用即可。