> ## Documentation Index > Fetch the complete documentation index at: https://arizeai-433a7140.mintlify.site/llms.txt > Use this file to discover all available pages before exploring further. # 09.01.2026: GitHub Issues from PXI, PII Detection, and Client Updates > PXI can file the GitHub issue for the bug it just found, a new evaluator screens conversations for PII, prompt versions land over REST, and the TypeScript client picks up prompt, trace, project, and user helpers. Most of what PXI finds ends up as a GitHub issue eventually. Until now, the "eventually" was on you: copy the trace link, write up the repro, check whether someone already filed it. This release lets PXI do that part. Alongside it, we shipped a PII Detection evaluator that looks in the places leaks actually happen, filled a REST gap that kept sending people back to GraphQL, and added the TypeScript client helpers you asked for most. # File GitHub Issues from PXI September 1, 2026 **Available in arize-phoenix 20.5.0+** When PXI lands on a real defect, ask it to file the issue. It searches the repository for duplicates first, then drafts an issue that links the traces and spans it was looking at. Under the hood this talks to [GitHub's hosted MCP server](https://github.com/github/github-mcp-server), so PXI uses GitHub's own tools rather than something we reimplemented. We spent most of the design time on the token, because an agent that can write to your GitHub needs to be boring about security: * **You file as yourself.** Add a fine-grained personal access token with Issues read/write under **Settings → Assistant → Personal settings → GitHub**. The token stays in your browser. Phoenix never persists it, and it never reaches the transcript, tool spans, or error messages, because it rides on the connection to GitHub rather than through the agent. * **You approve every write.** Before anything posts, PXI shows the exact repository, title, and body, and waits. If nobody is around to approve, the write tools are not even offered to the agent. * **Admins can set a shared fallback.** Store a workspace token as an encrypted secret, or turn the whole feature off under **Settings → Assistant → System settings → GitHub tools**. * **Enterprise and air-gapped deployments work too.** Point `PHOENIX_AGENTS_GITHUB_MCP_URL` at your own `github-mcp-server`, or set `PHOENIX_AGENTS_DISABLE_GITHUB=true` to remove the tools entirely. If GitHub is unreachable, the turn continues without the GitHub tools instead of failing. PXI carries on with what it can still do. Learn about the AI engineering agent built into Phoenix # Detect PII in Conversation Records August 28, 2026 **Available in arize-phoenix-evals 3.6.0+ (Python) and @arizeai/phoenix-evals 2.4.0+ (TypeScript)** Agents leak personal data in places the end user never sees: a tool result that returns a full customer record, a retrieved document with someone's home address in it, a system prompt that quotes a support ticket. The new PII Detection evaluator is built to look there. You decide what slice of the interaction to hand it. Pass only the user and assistant turns if that is what you care about, or the fuller record with system instructions, tool calls, tool results, and retrieved content. Each result carries a `pii_detected` or `no_pii_detected` label, a score of `1.0` or `0.0`, and an explanation that ends with a `FINDINGS` block listing every instance and its category, so a downstream filter can act on "email address" without parsing prose. Direction is `minimize`: a detection counts against you in aggregates, which is the point. ```python theme={"theme":{"light":"github-light-default","dark":"github-dark-default"}} from phoenix.evals import LLM from phoenix.evals.metrics import PiiDetectionEvaluator llm = LLM(provider="openai", model="gpt-4o-mini") evaluator = PiiDetectionEvaluator(llm=llm) scores = evaluator.evaluate( { "conversation": ( "User: Reset my account.\n" "Assistant: I can help. What email is on the account?\n" "User: jane.doe@acme.com" ), } ) print(scores[0].label) ``` ```typescript theme={"theme":{"light":"github-light-default","dark":"github-dark-default"}} import { openai } from "@ai-sdk/openai"; import { createPiiDetectionEvaluator } from "@arizeai/phoenix-evals"; const evaluator = createPiiDetectionEvaluator({ model: openai("gpt-4o-mini"), }); const result = await evaluator.evaluate({ conversation: "User: Reset my account.\nAssistant: What email is on the account?\nUser: jane.doe@acme.com", }); console.log(result.label); ``` Choose what to screen, format the conversation, and read the findings # Create Prompt Versions over REST August 26, 2026 **Available in arize-phoenix 20.5.0+** This one closes a gap that has bothered me for a while. You could create a prompt over REST, but every version after the first meant reaching for GraphQL or one of the SDKs. Now `POST /v1/prompts/{prompt_identifier}/versions` adds a version to an existing prompt, addressed by name or GlobalID, so a deploy script can promote a prompt with the same `curl` it uses for everything else. * **Tag in the same call.** Pass `tags` and the new version is labelled on create. If a tag already sits on another version, it moves and keeps its description and owner instead of being recreated blank. * **Bad parameters fail loudly.** Invocation parameters whose family does not match `model_provider` come back as a `422` rather than being stored and failing later in the playground. ```bash theme={"theme":{"light":"github-light-default","dark":"github-dark-default"}} curl -X POST "$PHOENIX_ENDPOINT/v1/prompts/summarizer/versions" \ -H "Authorization: Bearer $PHOENIX_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "version": { "description": "Lower temperature", "model_provider": "OPENAI", "model_name": "gpt-4o", "template_type": "CHAT", "template_format": "MUSTACHE", "template": { "type": "chat", "messages": [{"role": "user", "content": "Summarize: {{document}}"}] }, "invocation_parameters": {"type": "openai", "openai": {"temperature": 0.2}} }, "tags": [{"name": "production", "description": "Current production prompt"}] }' ``` Version, tag, and deploy prompts # TypeScript Client Additions August 28 – September 1, 2026 **Available in @arizeai/phoenix-client 7.7.1+** (`deletePrompt` since 7.6.0, the other helpers since 7.7.0) Four helpers that already existed in the REST API or the Python client and kept coming up as missing here. Now they exist. * **`deletePrompt`**: delete by `name` or `promptId`. Deletion cascades to every version along with its tags and labels, so double-check the name before you call it. Needs Phoenix server 13.20.0+. * **`transferTraces`**: move traces between projects by GlobalID or OpenTelemetry trace ID. Traces are re-parented, not copied, so nothing is duplicated and nothing is left behind. Needs Phoenix server 20.4.0+. * **`setProjectRetentionPolicy`**: assign an existing retention policy to a project, or pass `policyId: null` to fall back to the default. * **`getCurrentUser`**: find out who the current credentials belong to. With authentication disabled you get an anonymous user with `auth_method: "ANONYMOUS"`, which is handy for scripts that run in both modes. * **Session token counts**: `getSession` and `listSessions` now carry cumulative `tokenCountPrompt`, `tokenCountCompletion`, and `tokenCountTotal`, so you can rank sessions by spend without fetching their spans. ```typescript theme={"theme":{"light":"github-light-default","dark":"github-dark-default"}} import { deletePrompt } from "@arizeai/phoenix-client/prompts"; import { setProjectRetentionPolicy } from "@arizeai/phoenix-client/projects"; import { transferTraces } from "@arizeai/phoenix-client/traces"; import { getCurrentUser } from "@arizeai/phoenix-client/users"; const user = await getCurrentUser(); console.log(user.auth_method); const { transferredTraceCount } = await transferTraces({ traceIdentifiers: ["VHJhY2U6Mg=="], destinationProjectIdentifier: "production", }); console.log(transferredTraceCount); await setProjectRetentionPolicy({ projectName: "support-bot", policyId: "UHJvamVjdFRyYWNlUmV0ZW50aW9uUG9saWN5OjI=", }); await deletePrompt({ prompt: { name: "old-summarizer" } }); ``` The optional `openai` peer dependency now accepts `^6.10.0 || ^7.0.0`. If you are already on the OpenAI SDK v7, the install stops complaining. Browse the full client API # Approve PXI Browser Scripts as a Whole September 1, 2026 **Available in arize-phoenix 20.5.0+** One decision per script instead of a card per operation. In manual edit-permission mode, PXI now shows you a single description of everything the script is about to change. You accept or reject once, and that answer covers every state-changing action in the run. A script you reject cannot change state at all. The per-operation cards from last release were correct but exhausting for anything longer than two steps, and this is the fix. # Claude Fable 5.1 in the Playground September 1, 2026 **Available in arize-phoenix 20.5.0+** `claude-fable-5-1` is available in the playground through Anthropic and AWS Bedrock, with adaptive thinking controls and token pricing wired in for cost tracking. It is also the model we now recommend for PXI sessions. # Additional Improvements August 26 – September 1, 2026 **Available in arize-phoenix 20.5.0+ and arize-phoenix-evals 3.6.0+** * **Tool settings stick in the playground.** Tool choice and the strict flag survive runs against Anthropic and Bedrock models instead of resetting between runs. * **Annotation filters match the right span.** Annotation filters in the span and trace DSL now correlate against the span the filter is reading, which fixes some surprising results on traces with annotations on several spans. * **Faster paging through project traces.** `listProjectTraces` uses keyset pagination instead of offsets, so page two hundred costs about the same as page two. * **Async evals no longer stall.** The evals rate limiter used to block the event loop while it waited. It does not anymore. * **The code sandbox survives a crowded start.** Several server processes fetching the shared WASM runtime at once used to race each other and fail with a missing file. Downloads are now atomic, so every process sees a complete binary or none. * **Fresh token prices**, including LiteLLM reasoning token rates, so built-in cost tracking reflects current list prices. * **Login handoff lands on a neutral screen** instead of the error boundary. * **Long annotation names** no longer overflow the experiment comparison layout.