[ysyneu]

Problem

• Reading an artifact took three commands plus hand-parsing a token. The steps were artifact-gallery-get → artifact-sign → cutting t out of download_url with sed → artifact-stream --t. An agent has to get all of them right.
• The generated artifact-stream lost data. It went through the typed JSON path:
  • a JSON file printed OK: GET /safari/artifact/stream and its bytes were dropped;
  • files over 10MB were truncated;
  • any transfer longer than the 30s client timeout failed.
• Other byte transfers had the same 30s cap. This applied to session export and to file, skill and CSV uploads, however healthy the transfer.

Change

• New safari artifact-download <art_…|pf_…|link>.
  • Takes an artifact ID, a presented-file ID, or a console link that contains either.
  • Resolves the ID, signs, and streams the file to disk under the server's file name (base name only, so the file cannot land outside the current directory), or to --output (- means stdout).
  • Prints the saved path, byte count and content type; --json prints the same as an object.
  • Removes a partial file on failure.
• safari artifact-stream --t stays at its path-name (the path-is-king invariant). It is now curated on the SDK's streaming Artifacts.Download and saves files the same way.
• Timeouts by kind of call.
  • JSON calls keep the SDK's 30s overall timeout.
  • Byte transfers have no total limit and are cancelled after 60s with no bytes moved. They report transfer stalled: no bytes moved for 1m0s… instead of context canceled. The idle clock also covers waiting for the response to start. Covered: artifact download/stream, session export, safari send --file, skill upload, mapping CSV upload.
  • In broker mode, the SDK copies the broker client and drops its overall timeout for transfers, so the same rules apply there.
• go.mod points at the go-flashduty PR commit that adds Artifacts.Download (feat(artifacts): stream file downloads; no overall timeout on byte transfers go-flashduty#115). Bump it to the release tag before release.

Verification

• go test ./... passes, including TestEveryOperationHasPathCommand and TestGeneratorTargetsFullSpec.
• make check-cards passes.
• golangci-lint reports nothing in the changed files. The 3 govet reflect.Ptr notes are in untouched files.
• New tests:
  • a console link resolves art → pf → token, and a JSON file whose server name is ../../etc/data.json is saved verbatim as data.json;
  • -o with --json;
  • -o - writes to stdout;
  • an argument with no ID is rejected;
  • a stalled body is cancelled with errTransferStalled and leaves no partial file;
  • a 200ms trickle survives an 80ms idle timeout.
• Ran the built binary against a dev deployment:
  • art_ ID → 18996-byte HTML saved under its server name;
  • console link with -o --json;
  • pf_ ID to stdout;
  • a missing ID returns ResourceNotFound, exit 1.

Follow-up commits

• A failed download goes through a temp file that is renamed into place. The earlier file is kept and no partial file is left behind. Devices and FIFOs such as -o /dev/null are written directly. The temp file is created 0666, so the umask decides the saved file's mode.
• Without --output, a server-chosen file name that already exists is refused. An explicit --output path is overwritten.
• A response body's idle clock runs only while a read waits on the network. Piping into a slow consumer such as | less is therefore not reported as a stall. Uploads keep reset-on-read, because the SDK reads the source only as fast as the network drains it.
• An artifact that resolves to an empty file_id gets a clear error.

Release order

1. Merge and tag go-flashduty#115.
2. Repin go.mod to that tag. The current pin is a branch pseudo-version.
3. Release the CLI.
4. Bundle the new fduty into the runner and sandbox images.