Add AI agent conversations landed by the AI Sessionizer

[wu-sheng]

Add AI agent conversations landed by the AI Sessionizer

• If this is non-trivial feature, paste the links/URLs to the design doc. The protocol design is in ai-agent: add conversation query protocol for the AI Sessionizer skywalking-query-protocol#166, ai-agent: serve the conversation view over a streamed HTTP route, not GraphQL skywalking-query-protocol#167 and ai-agent: the view route says format, version and errors the HTTP way skywalking-query-protocol#168 (the list and export queries, and the streamed view route); the file formats and the asz.view document are the Sessionizer's under https://skywalking.apache.org/docs/skywalking-ai-sessionizer/next/en/formats/asz-view/; the operator doc of this feature is docs/en/setup/backend/ai-agent-conversation.md.

• Update the documentation to include this new feature.

• Tests(including UT, IT, E2E) are added to verify the new feature.

• If it's UI related, attach the screenshots below.

• If this pull request closes/resolves/fixes an existing issue, replace the issue number. Closes #.

• Update the CHANGES log.

---

The AI Sessionizer collects an agent runtime's transcripts into two file formats, Session Data (.sd) and Session Flow (.sf, an append-only chain of rounds), and pushes every file as one OTLP log record. This PR makes the OAP store them and serve a conversation whole.

Ingest. A new AI_AGENT layer routes the records to the bundled LAL rule lal/ai-agent.yaml. Its output builder, ConversationFile, verifies each file's sha256 and line count against the record attributes and stores it verbatim, or drops it and counts the reason in ai_agent_conversation_files_rejected. Nothing is decoded at ingest, so a cluster needs no shared state.

Storage. Two record models, ai_agent_session_data and ai_agent_session_flow, super datasets on every storage, in a BanyanDB group of their own, recordsAIAgent, with hot, warm and cold stages like the log group. A row belongs to its sender: its id is the service, the instance and the file digest, so a retry lands on the same row and another sender's copy is another row.

Query. The list page and the raw-file export are GraphQL queries, listConversations and getConversationRawFiles. The conversation itself is GET /ai-agent/conversations/{conversation}/v1/view on the core HTTP server beside /graphql: one asz.view 1.0 document, equal key for key to the one asz conversation prints for the same files, streamed as it is rendered, Content-Type: application/vnd.skywalking.asz.view+json; version=1.0 or the +yaml twin, gzip on Accept-Encoding, errors as application/problem+json, under its own timeout. It is not a GraphQL query because a long conversation renders to tens of megabytes: measured, a 136 MB session folds to a 70 MB document in about eleven seconds, past the HTTP server's ten second default. Rounds and files are read window by window (roundReadWindow, fileReadWindow), across every sender when none is named, and over every retained stage on BanyanDB.

Tests. Unit tests compare the built document with the Sessionizer's own output for the same files, cover the file builder, the formats, the raw-JSON preservation, and the route on an in-process server (JSON, YAML, gzip, chunking, UTF-8 at a chunk boundary, the error statuses). The e2e, one shared case on BanyanDB, Elasticsearch, MySQL and PostgreSQL, builds sessions with the Sessionizer's scenario tool, lands one in three stages, pushes everything to the OAP, and reads it back through swctl ai-agent list, files and view (apache/skywalking-cli#234, pinned as SW_CTL_COMMIT), comparing every document with the one the Sessionizer's viewer serves, over HTTP/1.1 and HTTP/2, as JSON and YAML, gzipped or not, plus the export's digests, a rejected file, and the three-round chain. All four passed locally.

Unrelated e2e fix. The Agent PHP e2e image (test/e2e-v2/cases/php/Dockerfile.php) moves from php:8.1-fpm-bullseye to php:8.1-fpm-bookworm: the bullseye apt mirror answers 404 for nginx on the CI runners, so the case failed on every run. The rustup download now follows the build host's architecture, so the image also builds on arm64 machines. The PHP extension directory is unchanged (PHP 8.1). Verified locally: the image builds natively and under --platform linux/amd64, and the PHP e2e passes 30/30.

Known limit. On JDBC, a retry of the same file by the same sender hits the same primary key, which the shared record writer's plain INSERT rejects; BanyanDB and Elasticsearch treat it as a duplicate the reader ignores or an overwrite. Content under a key is immutable by construction, so an insert-or-ignore in the common JDBC writer would be safe; that is a change outside this feature.