Session artifacts: directory expansion and office files

Problem

Models often register a generated folder as one workspace artifact (kind: file, workspacePath pointing at the directory). Clients then open or download it through GET /file or GET /file/bytes, which require a regular file and return HTTP 400.

Office outputs (Word / Excel / PowerPoint) were also second-class: they were not treated as artifact-like files, and the preview path tried to read them as text.

Decision

Directories are never artifacts. If a registration points at a directory, the store expands it to one artifact per recordable file inside that directory. The original directory path is not stored.

Office documents are first-class file artifacts (kind: document) and are opened as downloadable binaries, not text previews.

The chat turn-output list shows at most 3 artifact cards, with the same expand/collapse control already used for edited files.

Behavior

• Expansion happens in the session artifact store so every ingest path (record_artifact, hooks, client POST) behaves the same.
• Walk is recursive, skips hidden names, Excel lock files (~$*), symlinks, and well-known junk directories. Cap is 100 files and 4 directory levels; either limit is disclosed to the model and store warnings instead of silently dropping files.
• Chat grouping uses a recorded directory path as a prefix only for artifacts from that same record_artifact call (or artifacts with no tool call id). Later files written under the same folder stay on their own turn.
• Each child uses its filename as the title, inherits tool/source metadata, and infers kind from its own extension.
• A leftover directory path (empty, unlistable, or not expanded) is rejected; it is never marked available.
• Word / Excel / PowerPoint / OpenDocument extensions map to kind: document. The client shows a type label and a download action instead of a CodeMirror text preview.

Out of scope

• Directory artifacts or a folder browser kind.
• Auto-zip at record time.
• Auto-registering every write_file / shell output without record_artifact.

Adjustments after #9385 triage

Upstream analysis (QwenLM/qwen-code@main) confirmed the original decision. Keep store-side expansion, kind: document, and the 3-card collapse. Fine-tune the landing, do not change the product rule.

Coordinate with #9142

#9142 (fix/artifact-workspace-path-contract) stats the locator in record_artifact and rejects directories (TARGET_IS_DIRECTORY). That stops the 400, but it is not this issue: the user still does not get one card per generated file.

Keep #9142's path canonicalization (workspace-root-relative workspacePath, worktree prefix, no invented w/agent/ prefixes). Replace its directory rejection with expansion:

• If the locator is a regular file, verify + canonicalize as in #9142.
• If it is a directory, succeed and let the store expand. The tool result should say the directory was expanded, not Recorded artifact for the folder itself.
• If it is empty, missing, escaped, or not a regular file, fail as in #9142.

.xlsx in #9142's write_file whitelist must be document, not file. .csv stays file (text-previewable).

Client

Upstream Web Shell already has Download (canDownloadArtifact, readWorkspaceFileAsBlob) and GET /stat. Reuse those. Do not add a second download stack.

• kind: document and other non-text kinds (pdf, image, video, audio) never fall through to the 'source' / CodeMirror preview.
• If GET /stat returns type: directory (legacy dirty records), disable Open and do not call /file or /file/bytes.
• Add document to ARTIFACT_FORMAT_ICONS and kind labels.

write_file reminder

On current main the whitelist is ARTIFACT_KIND_BY_EXTENSION, not a plain extension set. Office/OpenDocument entries belong there as document. Do not call record_artifact again for a path write_file already recorded.