pi-exe-dev: compaction fails with 404 — aliased "@llm" model IDs reach the gateway without rewriting

[mydnicq]

Environment

• pi 0.85.1, bundled pi-exe-dev extension, exe.dev VM
• Custom LLM integration: ollama-cloud models via the llm gateway (https://llm.int.exe.xyz, provider exe-dev-ollama-cloud,
  e.g. glm-5.3-flash, deepseek-v4-flash:0731)

Summary

/compact (and auto-compaction / branch summaries) always fails with:

  Error: Compaction failed: Turn prefix summarization failed: 404:
  {"message":"model \"deepseek-v4-flash:0731@llm\" not found","type":"not_found_error",...}

Same failure for any model (e.g. glm-5.3-flash@llm), while normal agent turns work fine.

Root cause

1. The extension registers integration models under aliased IDs (@) and maps them back to native
   IDs in a before_provider_request hook.
2. pi only dispatches before_provider_request for agent-loop requests. pi's compaction builds summarization request
   options without onPayload (completeSummarization / createSummarizationOptions in core/compaction/), so the hook never
   fires there.
3. The aliased ID therefore reaches the gateway unrewritten → 404. Verified directly against the ollama-cloud route: POST
   https://llm.int.exe.xyz/ollama-cloud/v1/chat/completions with model: "glm-5.3-flash" → 200; with model:
   "glm-5.3-flash@llm" → 404 not_found_error.

Suggested fix (in the extension)

Register integration models under their native IDs and only alias when a native ID actually collides within a provider
(same model served by multiple integrations). No payload rewrite is then needed on any request path, and compaction works
unchanged for custom integrations like ollama-cloud, whose gateway only knows native IDs.

  // providerInfosFromIntegrationCatalogs
  const nativeIDCounts = new Map<string, number>();
  for (const c of entry.candidates)
    nativeIDCounts.set(c.nativeID, (nativeIDCounts.get(c.nativeID) ?? 0) + 1);
  for (const candidate of entry.candidates) {
    const alias = (nativeIDCounts.get(candidate.nativeID) ?? 0) > 1;
    const modelConfig = integrationModelConfig(candidate, alias); // id = alias ? `${native}@${name}` : nativeID
    models.push(modelConfig);
    if (alias) modelAliases.set(modelConfig.id, candidate.nativeID);
    ...
  }