bug: placeholder text '[Directory abstract is not ready]' indexed and returned in search results with high scores

[Aaron052399]

Summary

Leaf directories that lack .abstract.md files return a placeholder string [Directory abstract is not ready] from the content API. This placeholder text is vectorized and returned as high-scoring search results, degrading retrieval quality.

Environment

• OpenViking: v0.3.23
• OS: macOS (arm64)
• Embedding: doubao-embedding-vision-251215 (volcengine)
• VLM: mimo-v2.5 (openai-compatible)
• Config: ov.conf with custom VLM and embedding providers

Steps to Reproduce

1. Write memories via viking_remember() MCP tool, which calls session.commit_async() → MemoryUpdater → viking_fs.write_file().
2. After extraction, leaf directories (e.g. viking://user/default/memories/entities/开发工具/) contain memory files but no .abstract.md.
3. Query the abstract API for these directories:

   GET /api/v1/content/abstract?uri=viking://user/default/memories/entities/开发工具/
   

4. Response:

   {"status":"ok","result":"# viking://user/default/memories/entities/开发工具 [Directory abstract is not ready]"}

5. Run ov search "开发工具" -n 5 — the placeholder text is returned as the top result with score ~0.5.

Expected Behavior

Either:

• Option A: The MemoryUpdater write path triggers SemanticQueue for parent directories, generating proper .abstract.md files for leaf directories (similar to how add_resource() works via content_write._enqueue_semantic_refresh()).
• Option B: Placeholders like [Directory abstract is not ready] are excluded from vector indexing and/or filtered from search results.

Actual Behavior

The placeholder text [Directory abstract is not ready] is:

1. Returned by the abstract API as a 200 response (not an error)
2. Indexed into the vector store with a generic embedding
3. Returned in search results with inflated scores (0.4–0.5) because the text is semantically broad
4. Injected into agent context via the memory prefetch pipeline, wasting tokens on every conversation turn

Analysis

The root cause appears to be a code path gap:

• viking_remember() → session.commit_async() → MemoryUpdater._apply_upsert() → viking_fs.write_file() → direct AGFS write without enqueuing to SemanticQueue.
• MemoryUpdater does enqueue to EmbeddingQueue (line ~729 in memory_updater.py), but skips SemanticQueue.
• In contrast, content_write._enqueue_semantic_refresh() properly enqueues to both queues for resource/skill writes.

This means memory extraction writes never trigger directory abstract generation for their parent directories. The .abstract.md files for intermediate directories (e.g. memories/, memories/patterns/) are generated by other write paths, but leaf directories under entities/, preferences/, agent/*/memories/ are never processed.

Impact

With 140 vectors across ~40 sessions, placeholder results dominate search output. The zero-result rate was measured at 36.9% on v0.3.22, partly because genuinely relevant results are ranked below placeholder noise.

Workaround

Manual reindex with semantic_and_vectors mode generates proper abstracts:

# Per-directory
curl -X POST http://127.0.0.1:1933/api/v1/content/reindex   -H 'Content-Type: application/json'   -d '{"uri":"viking://user/default/memories/entities/开发工具","mode":"semantic_and_vectors","wait":true}'

But this must be repeated for every leaf directory and does not persist across new memory writes.

Additional Context

• ov status (v0.3.23 verbose): Embedding queue processed 183, Semantic queue processed 4 — consistent with the gap described above.
• Logs confirm SemanticProcessor ran for viking://user/default/memories and viking://user/default/memories/patterns, but never for leaf directories like entities/开发工具/, preferences/wangnandou/, agent/hermes/memories/tools/.

[Aaron052399]

Additional Impact: Memory Prefetch Pollution

Beyond search quality, this issue directly affects the agent's reasoning context.

Every conversation turn, the OV memory prefetch pipeline injects the top search results into the agent's system prompt. When placeholder results dominate, the injected context looks like this:

## OpenViking Context
- [0.44] # viking://user/default/memories/preferences/wangnandou [Directory abstract is not ready]
- [0.42] # viking://user/default/memories/preferences/default [Directory abstract is not ready]
- [0.38] # viking://user/default/memories/entities/UI组件 [Directory abstract is not ready]

This is injected on every turn (~300 tokens per turn wasted on noise). The agent receives zero useful context from these results, and if it trusts the injected memory, it may make decisions based on meaningless placeholder text.

With top_k=3 in the memory prefetch and placeholder scores consistently at 0.4–0.5, genuinely useful memories (which typically score 0.3–0.52) are frequently displaced by placeholders.

This makes the issue not just a search UX problem, but a correctness problem for any agent that relies on OV memory injection.

[Aaron052399]

Additional Evidence: Web Studio Confirms Rich Content Exists

The OV Web Studio UI shows that leaf directories like viking://user/default/memories/preferences/wangnandou/ contain 14+ substantive preference files (shell_terminal_style, computer-use操作策略, UI设计风格偏好, etc.), yet:

1. The directory has no .abstract.md file
2. The API returns [Directory abstract is not ready] as the abstract
3. This placeholder text is what gets indexed and returned in search results

The content clearly exists and is meaningful — the system just never generated a summary for it.

Bonus: Duplicate Memory Injection

The memory prefetch also injects the same file twice from different agent paths:

- [0.28] viking://agent/claude-code_claude-code-guide/memories/soul.md
- [0.28] viking://agent/claude-code/memories/soul.md

These are the same content indexed under different agent scopes, causing redundant context injection.