[delete-cloud]

Summary

Step ① of the personal-agent RAG roadmap: prove that enabling the kb plugin actually makes RAG run in the agent loop (it never has — kb is wired in core/app.py but absent from the default [agent.plugins].enabled, so retrieval was only ever validated standalone via CLI/evals).

Key fact: KBPlugin is a context-injection plugin (build_context), not a model-callable tool — once enabled it auto-retrieves and injects matched KB chunks into the context on every user message.

What this adds

• tests/integration/test_kb_in_loop.py — one hermetic test that:
  • enables kb via a temp config override (does NOT touch shipped agent.toml / helm configmap),
  • injects a deterministic fake embedding_fn (KBPlugin supports it) and monkeypatches the KB OpenAI clients to hard-fail if a real endpoint is hit,
  • indexes a sentinel fact into a temp LanceDB KB,
  • runs one real PipelineAdapter turn with a mocked LLM provider,
  • asserts the sentinel chunk appears in the context the provider receives ([Context Pack] / ## Repo references).
• No production code changes. Purely additive, no network, no external services.

Verification

• uv run pytest tests/integration/test_kb_in_loop.py tests/coding_agent/plugins/test_kb_plugin.py -q → 14 passed (run outside codex sandbox).

Scope / not in this PR

• ② o6n cutover (cluster-side ops).
• ③ helm kb.enabled knob + [kb] render + SiliconFlow embedding secret + indexing Job for SRE/Obsidian corpora.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Tests
  • Added integration test for KB plugin context injection verification.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: e3999ac2-27a4-45ed-8b81-61c8f13f2f92

📥 Commits

Reviewing files that changed from the base of the PR and between 23bcae4 and fe7d12c.

📒 Files selected for processing (1)

• coding-agent/tests/integration/test_kb_in_loop.py

---

📝 Walkthrough

Walkthrough

A new async integration test file is added that verifies the KB plugin injects RAG context into the agent loop. It configures an agent with the kb plugin, indexes a sentinel document using a fake embedder, monkeypatches embedding clients and the plugin class, mocks the LLM provider, runs one turn, and asserts the prompt contains the injected context.

Changes

KB In-Loop RAG Context Injection Integration Test

Layer / File(s)

Summary

End-to-end KB RAG injection test coding-agent/tests/integration/test_kb_in_loop.py

Adds test_kb_plugin_injects_indexed_context_into_agent_loop: writes agent.toml enabling the kb plugin, creates and indexes a sentinel document into a KB instance with a fake embedding function, monkeypatches KB embedding client getters and the module-level KBPlugin with a TestKBPlugin subclass, replaces the LLM provider with an AsyncMock that streams fixed events and captures messages, then asserts plugin state, StopReason.NO_TOOL_CALLS, final message "ack", one provider call, and that the serialized prompt includes context pack markers, repo references, and the sentinel fact.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Poem

🐇 A sentinel fact tucked in a doc so small,
Fake embeddings dodge the network's call.
The plugin stitches context neat and true,
"ack" comes back — the agent knew!
No real LLM, just mocks in play,
The rabbit checks RAG works today. ✨

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/kb-in-loop-validation

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}