[ranxianglei]

Problem (#447)

bili pi never disabled pi's own context compaction, so pi's native summarizer (shouldCompact, enabled: true by default) could fire ahead of / alongside bili's ACP compression and lossily rewrite the context bili is managing — losing the detailed history.

Fix

The pi overlay (<piHome>-bili) is now always built when the real pi home exists, and always carries a settings.json with compaction.enabled=false (merged over the user's real settings.json, which is never mutated). This makes bili's ACP compression the sole compression path under bili pi, mirroring how codex/claude already get their native auto-compact aligned (codex model_auto_compact_token_limit, claude CLAUDE_CODE_AUTO_COMPACT_WINDOW).

Changes:

• refreshOverlayHome / mergeOverlayEntry now accept a set of generated files (models.json + settings.json) so both stay out of the real-home symlink and merge-back sets (the 真实 models.yml 可被塞进死端口代理地址且永不解包:"real file never touched" 前提被嵌套 merge-back 递归与遗留插件写破,裸 omp 永久拨死号 #410 guarantee now holds for both).
• preparePiHttpRewrite returns undefined only for a missing real home; models.json is still rewritten for HTTP/wrapped-HTTPS upstreams when present, and copied verbatim otherwise.
• New writePiCompactionDisabledSettings merges the real settings.json with compaction.enabled=false (other keys — packages, theme, reserveTokens, … — preserved).

Verification

• npm run typecheck ✅
• npm test — 899 pass, 0 fail ✅
• npm run build ✅
• Integration check: realistic pi home (HTTP vllm provider + user settings.json) → overlay settings.json = {packages, theme, compaction:{reserveTokens:16384, enabled:false}}, models.json baseUrl rewritten to /bili/, real settings.json untouched ✅

Launcher-only change (no request-pipeline change), so the codex E2E suite is not exercised by this diff.

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-02_pi-disable-native-compaction (037070d)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-448

Each push to this PR publishes a new version under the pr-448 npm tag.

Option B — Download artifact

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

tar xzf billion-context-pr448.tgz
npm install -g package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 ework agent · qwen3.8-27b

[bot] 🏷 Reviewed PR #448 (fixes #447). Verified and analyzed in detail — assessment below.

Verification (all green):

• npm run typecheck ✅
• npm test — 899 pass, 0 fail ✅
• npm run build ✅

The root cause is real and the fix targets it correctly. For pi, raw-HTTPS providers (the common case) are classified into httpsDomains (cert-MITM), not into httpRewrites/httpsRewrites (discoverRoutes, src/launcher.ts:267-270). So for a typical HTTPS-only pi setup, preparePiHttpRewrite returned undefined, the <piHome>-bili overlay was never built, PI_CODING_AGENT_DIR was never set, and pi ran from its real home with native compaction left at its default (enabled: true). That's exactly the gap #447 describes. Making the overlay always build (when the real home exists) closes it. I reproduced both scenarios manually:

• HTTPS-only providers (no rewrites) → overlay now built, settings.json = {"theme":"dark","compaction":{"enabled":false}}.
• Realistic home (HTTP vllm provider + user settings.json) → settings.json = {packages, theme, compaction:{reserveTokens:16384, enabled:false}}, vllm baseUrl → /bili/…, openai (HTTPS) untouched, real settings.json byte-for-byte unchanged, auth.json symlinked.

The #410 guarantee now holds for both generated files. refreshOverlayHome/mergeOverlayEntry take a set (settings.json + models.json), so both stay out of the real-home symlink set and the merge-back set at any depth. isGeneratedDraft correctly matches .settings.json.<pid>.tmp / .models.json.<pid>.tmp. Clean generalization of the prior single-file logic; the omp/hermes/dsh call sites still pass a single string, which the Array.isArray guard handles.

writePiCompactionDisabledSettings is correct. Reads the real settings.json, shallow-copies the compaction object, sets enabled=false, writes to the overlay — preserving packages/theme/reserveTokens/etc. Real file is never mutated. Missing/malformed/non-object settings.json and non-object compaction are all handled gracefully (falls back to {} then forces enabled=false).

Consistent with codex/claude, with an intentional mechanism difference. Same goal (ACP is the sole compression path). codex's local compaction is benign (same-session truncation the kernel deactivates by message id), so codex gets threshold alignment (model_auto_compact_token_limit). pi's native summarizer is lossy — it rewrites the context bili manages — so pi gets full disable. That's the right call for pi.

Two small notes (non-blocking):

1. The description says models.json is "copied verbatim otherwise," but the code re-serializes it via JSON.stringify(modelsRoot) whenever it's a present, valid object — even when no rewrite matched. For plain-JSON models.json the parsed content is identical (only whitespace/formatting differs), so it's harmless in practice. Either do a true byte-copy in the no-rewrite case or soften the description.
2. The whole fix hinges on pi reading compaction.enabled (boolean) from settings.json to gate its native summarizer. That's pi-internal (separate repo) and isn't referenced anywhere in this codebase, so I'm taking the author's word for it — worth a one-line pointer to the pi source that confirms the exact key/shape before merge.

Bottom line: LGTM to merge. Correct fix, well-tested (the new/updated preparePiHttpRewrite tests cover the no-rewrite overlay build, the non-existent-home undefined, compaction disable, and key preservation), clean verification. Launcher-only, so no codex E2E needed. I'd only want the compaction.enabled key-shape confirmation (note 2) before merge; note 1 is optional polish.