Harsh Mahajan
4fc65470e5
Web: - Shell, turn engine and conversation view rebuilt around a single assistant store - Evidence, approvals, chat history, sharing, doors (Explain/Troubleshoot/Add to chat), settings and credits - Approval card survives a reload; Escape yields to open menus Backend: - Typed AI error envelope, resumable turns and capabilities endpoint - Org admin controls, chat history records, credits and approval policy - Query-context evidence and explicit chat mode - AI migrations registered after main's, as schema versions 129-133 Claude-Session: https://claude.ai/code/session_017PUuNehbpAvsekHyhEBmFt
Replay LLM proxy
Test infrastructure for the stack-ee runs. It is an OpenAI-compatible endpoint that
o2-ai uses as its LLM gateway, so a real OpenObserve and a real o2-ai answer a turn
without a model or a key. It holds no customer data and no keys, and it never records:
every recording is written by hand.
Run
REPLAY_PROXY_PORT=5199 node tests/ui-testing/playwright-tests/assistant/stack/replay-proxy/server.mjs
Point o2-ai at it through its gateway settings. The key only has to be non-empty:
| o2-ai variable | Value |
|---|---|
O2_AI_GATEWAY_ENABLED |
true |
O2_AI_GATEWAY_URL |
http://127.0.0.1:5199 (o2-ai appends /v1) |
O2_AI_MODEL |
replay |
O2_AI_API_KEY |
any placeholder, for example replay-not-a-key |
The proxy listens on 127.0.0.1 unless REPLAY_PROXY_HOST says otherwise.
| Route | Answers |
|---|---|
POST /v1/chat/completions |
one recorded step, streamed (chat.completion.chunk frames, then [DONE]) or as one chat.completion |
GET /v1/models |
the single replay model |
GET /__replay/served |
{ scenario, step, stream } per answer, for the run report (no message text) |
Scenario selection
A flow puts a scenario tag in its prompt, for example why 5xx? [replay:complete]. The
proxy takes the first [replay:<name>] in the latest user message, or complete when
there is none, so bracketed page context or link text after the prompt never picks a
scenario. An unknown name answers 404 with the scenario's name, so a missing recording
fails loudly. The mock harness does not read this tag: mock flows pick a scenario with
assistantPage.send(text, scenario) / nextScenario, which stack modes skip.
The step is the number of assistant messages after that user message: step 0 answers the prompt, step 1 answers after the first tool round, and so on. The last step repeats if the agent asks for more rounds than the recording has.
Recording format (recordings/<scenario>.json)
{
"version": 1,
"scenario": "complete",
"steps": [
{ "tool_calls": [{ "name": "<an MCP tool o2-ai exposes>", "arguments": {} }] },
{ "content": "## Markdown answer" }
]
}
scenariomatches[a-z0-9_]{1,40}and the file name.stepsholds 1 to 20 entries. Each one needscontent,tool_calls, or both.- A recording that breaks these rules stops the proxy at start-up.
Stack flows assert structure only (states, data-test ids, network calls, telemetry),
so the text in a recording can change without breaking a flow.
Checks
node --test tests/ui-testing/playwright-tests/assistant/stack/replay-proxy/proxy-check.mjs