Bug: summaryModel config override is ignored when resolving LCM summarizer

[uf-hy]

When running lossless-claw via OpenClaw, setting a summarizer model override in config does not currently affect which model is used for LCM compaction summaries.

Repro

1. Configure a summary model override, e.g.:

• plugins.entries.lossless-claw.config.summaryModel = "openai-resp/gpt-5.1-codex-max" (runtime config)
  or
• legacyParams.config.summaryModel = "openai-resp/gpt-5.1-codex-max"

2. Trigger compaction (manual /compact or automatic compaction).
3. Observe the summarizer still resolves to legacyParams.model (often the session/chat model).

Expected

If summaryModel is provided via config, it should take precedence over legacy provider/model hints.

Root cause

createLcmSummarizeFromLegacyParams currently only considers legacyParams.provider + legacyParams.model when calling deps.resolveModel(...), and ignores summaryModel from config.

Notes

When an explicit summaryModel override is provided, passing providerHint to resolveModel can also clobber a cross-provider model ref. In that case, providerHint should be omitted.

Related: #65

[HenryLoenwind]

Sorry to chime in here, but you've just looked at the code and might be able to answer this easily: When the readme says "(from openclaw)" as the default, what does it mean? The hardcoded Claude model openclaw uses? The default agent's model? The primary (fallback?) agent the session that's being summarised belongs to? The model that is associated with that session? The override model of the agent run that caused a summarisation to be needed? Whatever openclaw would use if it were to compact a session?

[uf-hy]

"(from openclaw)" means the model currently associated with the session being compacted (option 4 in your list).

When OpenClaw calls the LCM plugin, it passes the session's active model via legacyParams.model, and that becomes the default summarizer if you haven't set an explicit summaryModel override.

This is why the override matters: say you're running a session with Claude 4.6 Opus, but you'd rather use something cheaper/faster for compaction summaries (like Claude 4.5 Haiku and more). Now you can set summaryModel to do exactly that.

Before this fix, the config was ignored, so compaction always used whatever model the session was running — which could get expensive or slow for large compactions.

Hope that helps!

[HenryLoenwind]

Thank!, Then I'm fine with the default. I'm a bit short on models that are cheaper than my main one but good enough that I would trust them to compact memories.

But, TIL (because a compaction just went horribly wrong, one more reason to get lossless-claw) that openclaw has a compactionModel config key in the agent structure. Why not read that? The plugin should have access to the whole config, shouldn't it?

[uf-hy]

that's actually a really good point. I genuinely didn't know OpenClaw had "agents.defaults.compaction.model" until you mentioned it, so thanks for that...

I'm planning to add it to the model resolution chain. The idea is: if nothing is explicitly configured, it would fall back to "compaction.model" from the OpenClaw agent config (which makes sense — if someone already set it for built-in compaction, why not reuse it for LCM too).

That said, since "summaryModel" in the plugin config was introduced first, it'll take priority over "compaction.model" for backward compatibility — existing users who already configured "summaryModel" won't see any behavior change.

I'll update the PR to include this. Thanks for flagging it~

[uf-hy]

Resolved by #68.