Memory/QMD: reuse default model cache and skip ENOENT warnings by tyler6204 · Pull Request #12114 · openclaw/openclaw

tyler6204 · 2026-02-08T21:30:20Z

Summary

symlink the default QMD models directory into each agent's custom XDG_CACHE_HOME so preinstalled models are reused instead of re-downloaded
keep existing target model directories untouched (no overwrite)
treat missing default model cache (ENOENT) as an expected first-run no-op instead of logging a warning
add/extend tests for symlink creation, existing-target preservation, missing-default no-op behavior, and no-warning assertion for the missing-default case

Testing

pnpm tsgo
pnpm exec vitest run src/memory/qmd-manager.test.ts

Greptile Overview

Greptile Summary

This PR adds a symlinkSharedModels() step to QmdMemoryManager.initialize() so per-agent XDG_CACHE_HOME directories can reuse a shared/default QMD models cache (avoiding repeated model downloads). It also extends qmd-manager tests to cover symlink creation, preserving an existing target directory, and treating missing default models (ENOENT) as a silent no-op.

Key area to double-check is the first-run behavior of symlink creation: the implementation currently relies on the target directory structure existing before fs.symlink() is called, and if it doesn’t, the code will warn and fall back to re-download behavior.

Confidence Score: 3/5

This PR is close, but symlink creation likely fails on a fresh state dir due to missing parent directories.
The new model-cache reuse is a good direction, but symlinkSharedModels() appears to attempt to create the symlink before ensuring the target parent path exists, which would reliably throw ENOENT on first run and negate the intended behavior (falling back to downloads). Tests also don’t fully guard against silently falling back in some scenarios.
src/memory/qmd-manager.ts (symlinkSharedModels)

greptile-apps

_{2 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

greptile-apps · 2026-02-08T21:31:30Z

src/memory/qmd-manager.ts

+      try {
+        await fs.symlink(defaultModelsDir, targetModelsDir, "dir");
+      } catch (symlinkErr: unknown) {
+        const code = (symlinkErr as NodeJS.ErrnoException).code;
+        if (process.platform === "win32" && (code === "EPERM" || code === "ENOTSUP")) {
+          await fs.symlink(defaultModelsDir, targetModelsDir, "junction");
+        } else {


Symlink parent dir missing

symlinkSharedModels() tries to create targetModelsDir directly, but .../xdg-cache/qmd/models’s parent directory (.../xdg-cache/qmd) is not created anywhere. Since initialize() only does mkdir(this.xdgCacheHome) and mkdir(path.dirname(this.indexPath)) (which also resolves to .../xdg-cache/qmd only if index.sqlite is created first), fs.symlink() will throw ENOENT on a fresh state dir and you’ll just warn + fall back to re-downloading. Create path.dirname(targetModelsDir) before calling symlink so the intended sharing works on first run.

Prompt To Fix With AI

This is a comment left during a code review. Path: src/memory/qmd-manager.ts Line: 521:527 Comment: **Symlink parent dir missing** `symlinkSharedModels()` tries to create `targetModelsDir` directly, but `.../xdg-cache/qmd/models`’s parent directory (`.../xdg-cache/qmd`) is not created anywhere. Since `initialize()` only does `mkdir(this.xdgCacheHome)` and `mkdir(path.dirname(this.indexPath))` (which also resolves to `.../xdg-cache/qmd` only if `index.sqlite` is created first), `fs.symlink()` will throw `ENOENT` on a fresh state dir and you’ll just warn + fall back to re-downloading. Create `path.dirname(targetModelsDir)` before calling `symlink` so the intended sharing works on first run. How can I resolve this? If you propose a fix, please make it concise.

greptile-apps · 2026-02-08T21:31:32Z

src/memory/qmd-manager.test.ts

+    it("does not overwrite existing models directory", async () => {
+      // Pre-create the custom models dir with different content.
+      await fs.mkdir(customModelsDir, { recursive: true });
+      await fs.writeFile(path.join(customModelsDir, "custom-model.bin"), "custom");
+


Test can be flaky

In "does not overwrite existing models directory", you pre-create customModelsDir before QmdMemoryManager.create(), but initialize() still runs symlinkSharedModels() and may log a warn if the default models dir exists yet the symlink creation fails for unrelated reasons (e.g., permissions on platforms where directory symlinks aren’t available). The test doesn’t assert on logWarnMock, so it can intermittently pass while the feature is effectively broken. Consider asserting no warn occurred in this test as well, or explicitly simulating the platform behavior you expect.

Prompt To Fix With AI

This is a comment left during a code review. Path: src/memory/qmd-manager.test.ts Line: 652:656 Comment: **Test can be flaky** In "does not overwrite existing models directory", you pre-create `customModelsDir` before `QmdMemoryManager.create()`, but `initialize()` still runs `symlinkSharedModels()` and may log a warn if the default models dir exists yet the symlink creation fails for unrelated reasons (e.g., permissions on platforms where directory symlinks aren’t available). The test doesn’t assert on `logWarnMock`, so it can intermittently pass while the feature is effectively broken. Consider asserting no warn occurred in this test as well, or explicitly simulating the platform behavior you expect. How can I resolve this? If you propose a fix, please make it concise.

…ler6204

…ler6204)

…tyler6204

…tyler6204)

…yler6204

…yler6204)

…tyler6204

…tyler6204)

QmdMemoryManager overrides XDG_CACHE_HOME to isolate the qmd index per-agent, but this also moves where qmd looks for its ML models (~2.1GB). Since models are installed at the default location (~/.cache/qmd/models/), every qmd invocation would attempt to re-download them from HuggingFace and time out. Fix: on initialization, symlink ~/.cache/qmd/models/ into the custom XDG_CACHE_HOME path so the index stays isolated per-agent while the shared models are reused. The symlink is only created when the default models directory exists and the target path does not already exist. Includes tests for the three key scenarios: symlink creation, existing directory preservation, and graceful skip when no default models exist.

…tyler6204

…tyler6204)

)

@tyler6204

…12114) (thanks @tyler6204)

@tyler6204

…law#12114) * Memory/QMD: symlink default model cache into custom XDG_CACHE_HOME QmdMemoryManager overrides XDG_CACHE_HOME to isolate the qmd index per-agent, but this also moves where qmd looks for its ML models (~2.1GB). Since models are installed at the default location (~/.cache/qmd/models/), every qmd invocation would attempt to re-download them from HuggingFace and time out. Fix: on initialization, symlink ~/.cache/qmd/models/ into the custom XDG_CACHE_HOME path so the index stays isolated per-agent while the shared models are reused. The symlink is only created when the default models directory exists and the target path does not already exist. Includes tests for the three key scenarios: symlink creation, existing directory preservation, and graceful skip when no default models exist. * Memory/QMD: skip model symlink warning on ENOENT * test: stabilize warning-filter visibility assertion (openclaw#12114) (thanks @tyler6204) * fix: add changelog entry for QMD cache reuse (openclaw#12114) (thanks @tyler6204) * fix: handle plain context-overflow strings in compaction detection (openclaw#12114) (thanks @tyler6204)

@tyler6204

…law#12114) * Memory/QMD: symlink default model cache into custom XDG_CACHE_HOME QmdMemoryManager overrides XDG_CACHE_HOME to isolate the qmd index per-agent, but this also moves where qmd looks for its ML models (~2.1GB). Since models are installed at the default location (~/.cache/qmd/models/), every qmd invocation would attempt to re-download them from HuggingFace and time out. Fix: on initialization, symlink ~/.cache/qmd/models/ into the custom XDG_CACHE_HOME path so the index stays isolated per-agent while the shared models are reused. The symlink is only created when the default models directory exists and the target path does not already exist. Includes tests for the three key scenarios: symlink creation, existing directory preservation, and graceful skip when no default models exist. * Memory/QMD: skip model symlink warning on ENOENT * test: stabilize warning-filter visibility assertion (openclaw#12114) (thanks @tyler6204) * fix: add changelog entry for QMD cache reuse (openclaw#12114) (thanks @tyler6204) * fix: handle plain context-overflow strings in compaction detection (openclaw#12114) (thanks @tyler6204)

Integrated upstream improvements: - CRITICAL: Fix bundled hooks broken since 2026.2.2 (openclaw#9295) - Grok web search provider (xAI) with inline citations - Telegram video note support with tests and docs - QMD model cache sharing optimization (openclaw#12114) - Context overflow false positive fix (openclaw#2078) - Model failover 400 status handling (openclaw#1879) - Dynamic config loading per-message (openclaw#11372) - Gateway post-compaction amnesia fix (openclaw#12283) - Skills watcher: ignore Python venvs and caches - Telegram send recovery from stale thread IDs - Cron job parameter recovery (openclaw#12124) - Auto-reply weekday timestamps (openclaw#12438) - Utility consolidation refactoring (PNG, JSON, errors) - Cross-platform test normalization (openclaw#12212) - macOS Nix defaults support (openclaw#12205) Preserved DEV enhancements: - Docker multi-stage build with enhanced tooling (gh, gog, obsidian-cli, uv, nano-pdf, mcporter, qmd) - Comprehensive .env.example documentation (371 lines) - Multi-environment docker-compose support (DEV/PRD) - GOG/Tailscale integration - Fork-sync and openclaw-docs skills - UI config editor (Svelte) - Fork workflow documentation Merge strategy: Cherry-picked 22 upstream commits, preserved DEV Docker architecture. Docker files unchanged: Dockerfile, docker-compose.yml, docker-setup.sh, .env.example Co-Authored-By: Claude Sonnet 4.5 <[email protected]>

@tyler6204

…law#12114) * Memory/QMD: symlink default model cache into custom XDG_CACHE_HOME QmdMemoryManager overrides XDG_CACHE_HOME to isolate the qmd index per-agent, but this also moves where qmd looks for its ML models (~2.1GB). Since models are installed at the default location (~/.cache/qmd/models/), every qmd invocation would attempt to re-download them from HuggingFace and time out. Fix: on initialization, symlink ~/.cache/qmd/models/ into the custom XDG_CACHE_HOME path so the index stays isolated per-agent while the shared models are reused. The symlink is only created when the default models directory exists and the target path does not already exist. Includes tests for the three key scenarios: symlink creation, existing directory preservation, and graceful skip when no default models exist. * Memory/QMD: skip model symlink warning on ENOENT * test: stabilize warning-filter visibility assertion (openclaw#12114) (thanks @tyler6204) * fix: add changelog entry for QMD cache reuse (openclaw#12114) (thanks @tyler6204) * fix: handle plain context-overflow strings in compaction detection (openclaw#12114) (thanks @tyler6204)

@tyler6204

…law#12114) * Memory/QMD: symlink default model cache into custom XDG_CACHE_HOME QmdMemoryManager overrides XDG_CACHE_HOME to isolate the qmd index per-agent, but this also moves where qmd looks for its ML models (~2.1GB). Since models are installed at the default location (~/.cache/qmd/models/), every qmd invocation would attempt to re-download them from HuggingFace and time out. Fix: on initialization, symlink ~/.cache/qmd/models/ into the custom XDG_CACHE_HOME path so the index stays isolated per-agent while the shared models are reused. The symlink is only created when the default models directory exists and the target path does not already exist. Includes tests for the three key scenarios: symlink creation, existing directory preservation, and graceful skip when no default models exist. * Memory/QMD: skip model symlink warning on ENOENT * test: stabilize warning-filter visibility assertion (openclaw#12114) (thanks @tyler6204) * fix: add changelog entry for QMD cache reuse (openclaw#12114) (thanks @tyler6204) * fix: handle plain context-overflow strings in compaction detection (openclaw#12114) (thanks @tyler6204)

- Symlink default model cache into custom XDG_CACHE_HOME per agent - Skip ENOENT silently when default models directory is missing - Clean up compaction test imports - Add changelog entry for QMD cache reuse

@tyler6204

…law#12114) * Memory/QMD: symlink default model cache into custom XDG_CACHE_HOME QmdMemoryManager overrides XDG_CACHE_HOME to isolate the qmd index per-agent, but this also moves where qmd looks for its ML models (~2.1GB). Since models are installed at the default location (~/.cache/qmd/models/), every qmd invocation would attempt to re-download them from HuggingFace and time out. Fix: on initialization, symlink ~/.cache/qmd/models/ into the custom XDG_CACHE_HOME path so the index stays isolated per-agent while the shared models are reused. The symlink is only created when the default models directory exists and the target path does not already exist. Includes tests for the three key scenarios: symlink creation, existing directory preservation, and graceful skip when no default models exist. * Memory/QMD: skip model symlink warning on ENOENT * test: stabilize warning-filter visibility assertion (openclaw#12114) (thanks @tyler6204) * fix: add changelog entry for QMD cache reuse (openclaw#12114) (thanks @tyler6204) * fix: handle plain context-overflow strings in compaction detection (openclaw#12114) (thanks @tyler6204)

@tyler6204

…law#12114) * Memory/QMD: symlink default model cache into custom XDG_CACHE_HOME QmdMemoryManager overrides XDG_CACHE_HOME to isolate the qmd index per-agent, but this also moves where qmd looks for its ML models (~2.1GB). Since models are installed at the default location (~/.cache/qmd/models/), every qmd invocation would attempt to re-download them from HuggingFace and time out. Fix: on initialization, symlink ~/.cache/qmd/models/ into the custom XDG_CACHE_HOME path so the index stays isolated per-agent while the shared models are reused. The symlink is only created when the default models directory exists and the target path does not already exist. Includes tests for the three key scenarios: symlink creation, existing directory preservation, and graceful skip when no default models exist. * Memory/QMD: skip model symlink warning on ENOENT * test: stabilize warning-filter visibility assertion (openclaw#12114) (thanks @tyler6204) * fix: add changelog entry for QMD cache reuse (openclaw#12114) (thanks @tyler6204) * fix: handle plain context-overflow strings in compaction detection (openclaw#12114) (thanks @tyler6204)

@tyler6204

* Memory/QMD: symlink default model cache into custom XDG_CACHE_HOME QmdMemoryManager overrides XDG_CACHE_HOME to isolate the qmd index per-agent, but this also moves where qmd looks for its ML models (~2.1GB). Since models are installed at the default location (~/.cache/qmd/models/), every qmd invocation would attempt to re-download them from HuggingFace and time out. Fix: on initialization, symlink ~/.cache/qmd/models/ into the custom XDG_CACHE_HOME path so the index stays isolated per-agent while the shared models are reused. The symlink is only created when the default models directory exists and the target path does not already exist. Includes tests for the three key scenarios: symlink creation, existing directory preservation, and graceful skip when no default models exist. * Memory/QMD: skip model symlink warning on ENOENT * test: stabilize warning-filter visibility assertion (#12114) (thanks @tyler6204) * fix: add changelog entry for QMD cache reuse (#12114) (thanks @tyler6204) * fix: handle plain context-overflow strings in compaction detection (#12114) (thanks @tyler6204)

openclaw-barnacle bot added the maintainer Maintainer-authored PR label Feb 8, 2026

greptile-apps bot reviewed Feb 8, 2026

View reviewed changes

tyler6204 self-assigned this Feb 8, 2026

tyler6204 added a commit that referenced this pull request Feb 8, 2026

fix: polish qmd cache prep items (#12114) (thanks @tyler6204)

58c220f

tyler6204 added a commit that referenced this pull request Feb 8, 2026

fix: stabilize warning filter assertion in tests (#12114) (thanks @ty…

a9e48cd

…ler6204)

tyler6204 force-pushed the fix/qmd-model-cache branch from de85d2f to a9e48cd Compare February 8, 2026 22:06

tyler6204 added a commit that referenced this pull request Feb 9, 2026

fix: add changelog entry for qmd model cache reuse (#12114) (thanks @…

f09d3fd

…tyler6204)

tyler6204 added a commit that referenced this pull request Feb 9, 2026

test: make warning-filter assertion deterministic (#12114) (thanks @t…

65241af

…yler6204)

tyler6204 force-pushed the fix/qmd-model-cache branch from a9e48cd to 65241af Compare February 9, 2026 06:58

tyler6204 added a commit that referenced this pull request Feb 9, 2026

test: stabilize warning-filter visibility assertion (#12114) (thanks @…

af6f1ac

…tyler6204)

tyler6204 force-pushed the fix/qmd-model-cache branch from 65241af to af6f1ac Compare February 9, 2026 07:11

tyler6204 force-pushed the fix/qmd-model-cache branch from af6f1ac to 23dad73 Compare February 9, 2026 07:21

openclaw-barnacle bot added the agents Agent runtime and tooling label Feb 9, 2026

tyler6204 added 5 commits February 8, 2026 23:24

Memory/QMD: skip model symlink warning on ENOENT

1018717

test: stabilize warning-filter visibility assertion (#12114) (thanks @…

fbb3115

…tyler6204)

fix: add changelog entry for QMD cache reuse (#12114) (thanks @tyler6204

6e53341

)

fix: handle plain context-overflow strings in compaction detection (#…

efa35a6

…12114) (thanks @tyler6204)

tyler6204 force-pushed the fix/qmd-model-cache branch from 23dad73 to efa35a6 Compare February 9, 2026 07:41

tyler6204 merged commit e4651d6 into main Feb 9, 2026
21 of 24 checks passed

tyler6204 deleted the fix/qmd-model-cache branch February 9, 2026 07:43

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Memory/QMD: reuse default model cache and skip ENOENT warnings#12114

Memory/QMD: reuse default model cache and skip ENOENT warnings#12114
tyler6204 merged 5 commits intomainfrom
fix/qmd-model-cache

tyler6204 commented Feb 8, 2026 •

edited by greptile-apps bot

Loading

Uh oh!

greptile-apps bot left a comment

Uh oh!

greptile-apps bot Feb 8, 2026

Uh oh!

greptile-apps bot Feb 8, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Comments

Uh oh!

Conversation

tyler6204 commented Feb 8, 2026 • edited by greptile-apps bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Testing

Greptile Overview

Greptile Summary

Confidence Score: 3/5

Uh oh!

greptile-apps bot left a comment

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Feb 8, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Feb 8, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Comments

tyler6204 commented Feb 8, 2026 •

edited by greptile-apps bot

Loading