feat: /acp compact — dedicated compression model (session shared-prefix + models.json)

[ranxianglei]

Summary

Implements #240: a /acp compact command that offloads compress summary-writing to a separate call, in two modes:

• session — the session's own model, in a separate call that reuses the session prompt prefix (system prompt + active tools + the exact transformed messages Pi just sent). Because the prefix matches what the main model already sent, the provider prompt cache keeps the input cheap, and the compression reasoning stays out of the main model's context.
• <id> — a cheaper models.json model (reuses its baseUrl/apiKey; a different cache namespace, so no prefix sharing, but a lower per-token price).

Unset → the main model writes summaries (unchanged default). Any call failure (network/timeout/API/empty) → fallback to the main model's summary for that range — the session is never interrupted.

Changes

• src/compress-model.ts: CompressionModelClient (listModels / resolveModel / summarize / summarizeContext), SESSION_MODEL_REF.
• src/compress-tool.ts: branch on session vs models.json; build the shared-prefix Context; per-range instruction appended to the captured prefix.
• src/index.ts: capture the ACP-transformed messages in the context transform; pass pi to the tool.
• src/runtime.ts: per-session setLastSentMessages / getLastSentMessages.
• src/commands.ts: /acp compact [session|<id>|reset] + status display.
• src/user-config.ts / src/config.ts: compressionModelId (+ saveCompressionModelId writer).
• Tests: 22 new (compress-model + acp-compact, incl. 4 session tests). 450/450 pass.
• Docs: README + CONFIGURATION.

Related

• Follow-up (the primary lever — Prompt Cache stability / batch rollover) is being tracked as a separate issue.

[github-actions]

📦 Built Extension Artifact

Branch: feat/acp-compact-compression-model (c5e7de8)

Option A — Install from npm PR tag (recommended)

pi install npm:billion-context-pi@pr-241

Each push to this PR publishes a new version under the pr-241 npm tag.

Option B — Download artifact

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

tar xzf billion-context-pi-pr241.tgz
pi install ./package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 ework agent · vllm-qwen/qwen3.8-27b

[bot] Reviewing this PR now. Branch feat/acp-compact-compression-model checked out locally (2 commits: 631d51e + c5e7de8). Running typecheck, tests, and build, and reading through the diff — I'll report findings here shortly.

[ranxianglei]

🤖 ework agent · vllm-qwen/qwen3.8-27b

[bot] Review complete — LGTM, all CI checks green. Merging is left to you (human-only per AGENTS.md).

CI (run on branch feat/acp-compact-compression-model, commits 631d51e + c5e7de8)

Check	Result
npm run typecheck	✅ pass
npm test	✅ 450/450 pass (22 new: 12 in tests/compress-model.test.ts, 10 in tests/acp-compact.test.ts)
npm run build	✅ dist/index.js 535.59 KB, self-contained — zero runtime refs to pi-ai (type-only imports erased)

Verified integration points

• Pins intact: acp-kernel 0.0.46, pi-coding-agent 0.83.0 unchanged; new devDep @earendil-works/pi-ai pinned exactly to 0.83.0 (type-only, so devDep placement is correct). The 1.3K-line lockfile diff is just pi-ai's transitive deps (AWS/Anthropic SDKs) — expected.
• Shared-prefix capture (src/index.ts:367): runtime.setLastSentMessages(sid, rebuilt) stores exactly the array returned as { messages: rebuilt } — the final ACP-transformed messages (tags + injected nudge). Correct source.
• Prefix fidelity: getSystemPromptText(ctx) uses ctx.getSystemPrompt() (same prompt Pi sends), convertToLlm(AgentMessage[]): Message[] exists in pi-coding-agent 0.83.0, and collectActiveTools's {name, description, parameters} mapping matches pi-ai's Tool/Context shapes exactly. So the session-mode request should be byte-identical to the main model's prefix → cache hit.
• Fallbacks: per-range try/catch → main-model summary retained on any failure; unresolvable ref, missing session model, and empty response all degrade gracefully with a ⚠️ note in the panel. Session never blocked.
• Tests: the 4 session tests use a real mock SSE server and assert the shared prefix ("lorem ipsum") is actually present in the outbound request body, plus the unreachable-model fallback path. Good coverage of the load-bearing behavior.
• Docs: README + CONFIGURATION cover both modes, the fallback guarantee, resolution rules, and the global-file scope.

Non-blocking observations

1. Fallback quality cliff: when a compression model is set, the tool description tells the main model to pass a minimal placeholder ("compressed"). If the compression call fails for that range, the block's stored summary is that placeholder — much weaker than the old "main model writes full summaries" behavior, and the only recovery is decompress + re-compress. The session is never interrupted (as designed), but consider either (a) stating this tradeoff explicitly in the tool description/docs, or (b) asking the main model for a 1–2 sentence real summary instead of a bare placeholder so the fallback is still useful.
2. Project-level override edge case: /acp compact writes the global ~/.pi/acp.json, but loadUserConfig lets a project .pi/acp.json override per-field. If a project file sets compressionModelId, the next reloadConfig (forced by lastUserConfigKey = undefined) re-reads and the project value wins over the command's in-memory update. CONFIGURATION documents the scope, but the command's "applies across sessions" message could mislead in that corner.
3. Memory: lastSentMessages holds a full copy of the transformed messages per session until clearSessionTracking — fine, just noting it's a duplicate of data already in memory.
4. Minor: commandArgString(args: string) is typed string but handles arrays (test convenience) — the type is a slight lie; harmless.
5. truncateContent(content, 120000) caps range content only in models.json mode; session mode sends the full (cached) context with no cap. Inconsistent but each is reasonable for its mode.

None of these block the merge. Nice work — the session-mode design (isolation without a cheaper model, via cache) is the interesting part.