feat(compact): add compactModel config to override compaction model

[laurentftech]

Closes #1445

Summary

• Add compactModel?: string to GlobalConfig — optional override for the model used during compaction
• Expose via /config set compactModel <model-id> with model validation
• Skip forked-agent cache-sharing path automatically when compactModel differs from mainLoopModel (different model = guaranteed cache miss)

Motivation

When running a slow local reasoning model (e.g. Qwen3.6 27B) as the main loop, compaction blocks on that same slow model — even though summarization doesn't require reasoning. This allows routing compaction to a fast API model (e.g. Gemma 4) without affecting the main conversation model.

No cache tradeoff for local-model users: local inference has no prompt cache anyway.

Usage

claude config set compactModel <your-fast-model-id>

Unset to revert to default (uses mainLoopModel).

Files changed

File	Change
src/utils/config.ts	compactModel?: string field + GLOBAL_CONFIG_KEYS registration
src/tools/ConfigTool/supportedSettings.ts	Config UI entry with validateModel
src/services/compact/compact.ts	Read compactModel from config, disable forked path when model differs, pass override to queryModelWithStreaming

🤖 Generated with Claude Code

[jatmn]

Thanks for the contribution. I took a pass through the PR intent, the linked issue, the changed paths, and the related open code-reviewer-agent PR, and found issues that need to be addressed before this is ready.

Findings

• [P2] Remove the unrelated duplicate code-reviewer agent changes
  src/tools/AgentTool/built-in/codeReviewerAgent.ts:72
  This PR is titled and described as adding compactModel for compaction, but it also adds the built-in code-reviewer agent, its tests, README routing entry, and an .openlore/ ignore rule. That agent work is already being reviewed separately in #1381, and this branch contains an older copy of it: for example, #1381 has since added shell-tool denials after review, while this PR still leaves Bash/PowerShell available through wildcard agent tools despite the prompt's read-only guarantee. Please drop the code-reviewer/README/OpenLore-related changes from this compact-model PR and keep that work in the dedicated agent PR.

• [P2] Check tool-search support against the model that will actually compact
  src/services/compact/compact.ts:1270
  The streaming fallback now sends the request with model: compactModel, but isToolSearchEnabled() is still called with context.options.mainLoopModel. When the main model supports tool references and the configured compact model does not, this can enable ToolSearchTool and deferred MCP-tool handling for a request that is then sent to the unsupported compact model. That makes the primary use case fragile: switching compaction to a fast override model can fail or shape the request incorrectly based on the unrelated main-loop model. Please make the tool-search decision use compactModel consistently for the compact request.

• [P2] Add a supported way to clear compactModel
src/tools/ConfigTool/supportedSettings.ts:62
  The linked issue says unset compactModel should restore the existing behavior of using mainLoopModel, and the PR body says "Unset to revert to default". The config tool path added here only supports get vs set; omitting value reads the setting, empty strings fail validateModel, and the only "default" deletion path is hard-coded for remoteControlAtStartup. After a user sets compactModel, they cannot return to the default behavior through the documented config command without manually editing the global config file. Please add an explicit clear/default/unset path for this setting and cover it with a focused test.

[laurentftech]

Thanks for the detailed review. Addressed all three P2 findings in 8499c32:

[P2] Code-reviewer agent removed — dropped codeReviewerAgent.ts, its test, the builtInAgents.ts registration, the README routing entry, and the .openlore/ gitignore rule. That work stays in #1381.

[P2] isToolSearchEnabled now uses compactModel (compact.ts:1271) — the tool-search capability check now uses the same model as the compaction request, so a fast override model that doesn't support tool references won't get ToolSearchTool injected.

[P2] compactModel unset path added — introduced clearable?: boolean on SettingConfig (generic, not hardcoded per-key). When clearable: true and value is "default", ConfigTool deletes the key from global config before reaching validateOnWrite. Running /config compactModel default now restores the main-loop-model fallback. compactModel description updated to document the "default" escape hatch.

[jatmn]

Thanks for the update. I rechecked the changed paths and found issues that still need to be addressed.

Findings

• [P2] Resolve compactModel aliases before sending compaction requests
  src/services/compact/compact.ts:1159
compactModel is validated with the same validateModel path as the main model setting, and that validator explicitly accepts aliases such as sonnet, opus, and haiku as valid. Unlike the main-loop model path, though, this value is read straight from global config and passed to queryModelWithStreaming as model: compactModel, and the API request only strips context tags rather than expanding aliases. That means /config set compactModel sonnet succeeds, but the next compaction sends model: "sonnet" instead of the resolved model ID, so the documented config surface can fail at runtime. Please normalize/parse the stored compact model the same way the main model path does before comparing it to mainLoopModel and before passing it to the API.

• [P3] Return the cleared value in the config-tool result
  src/tools/ConfigTool/ConfigTool.ts:205
  The new clearable-setting branch successfully deletes compactModel, but it returns only value: null. Both the visible ConfigTool UI and mapToolResultToToolResultBlockParam render successful set operations from content.newValue, so /config set compactModel default reports Set compactModel to undefined to the user and to the model even though the key was cleared. Please return the same result shape that the set path expects, and add a focused regression test for the clear/default case so the reset behavior and confirmation stay in sync.

[laurentftech]

Both findings addressed in ed1aad0.

[P2] Alias resolution — compactModel is now resolved via parseUserSpecifiedModel() before any downstream use. The raw config value (e.g. "sonnet") is expanded to the full model ID before being passed to isToolSearchEnabled, getMaxOutputTokensForModel, and queryModelWithStreaming. The cache-sharing comparison (compactModel === context.options.mainLoopModel) now compares two resolved IDs on both sides, since mainLoopModel is already resolved at startup via the same function.

[P3] Clearable result shape — the clearable branch now returns newValue: null instead of value: null. mapToolResultToToolResultBlockParam renders set operations from content.newValue, so /config set compactModel default now reports Set compactModel to null rather than Set compactModel to undefined. Added a regression test covering the set/clear/get/error rendering paths in ConfigTool.test.ts.

[jatmn]

Thanks for the update. I rechecked the changed paths and found one issue that still needs to be addressed.

Findings

• [P3] Coerce compactModel values before validation
  src/tools/ConfigTool/supportedSettings.ts:68
  The ConfigTool input schema still allows value to be a string, boolean, or number, but the new compactModel setting passes v as string directly into validateModel. That cast does not convert the runtime value, and validateModel immediately calls model.trim(), so a malformed tool call such as { "setting": "compactModel", "value": true } throws TypeError: model.trim is not a function instead of returning a normal validation error. The existing model setting avoids this by calling validateModel(String(v)); please apply the same coercion or explicitly reject non-string values before validation, and add a small regression test for the bad-value path.

[jatmn]

closing as abandoned and newer #1629 has overlap