Tool Shed stock-tool TRS version discovery

Investigated 2026-09-16 in response to Foundry PR #558’s review thread. Foundry issue #552 records the original failure; PR #558 permits explicit-version probes as a workaround.

Confirmed gap

Read-only requests to the public Tool Shed reproduced these results:

Current Galaxy dev (aa111536957) has the same defect. Both TRS tool metadata and version listing call get_tool() in lib/tool_shed/managers/trs.py, which unconditionally resolves repository metadata. A bare stock ID decodes to a GUID with no repository owner/name; guid_to_repository() then raises ValueError while unpacking that GUID. New unit tests reproduced this failure before the implementation changed.

Explicit-version fetches already branch on whether the ID contains ~ and use a lazily initialized stock-tool registry. This is an endpoint inconsistency, rather than missing stock wrappers. The TypeScript client’s getLatestTRSToolVersion() uses the last entry in the TRS list, so stock versions must be returned oldest first.

Implemented fix

Draft PR: Galaxy #23565, now titled [26.1] Expose stock tool versions through the Tool Shed TRS API and targeting release_26.1 following Marius’s approved review. Branch: toolshed_stock_trs, rebased tip f211f388f0, based on release tip 1f6bbd3c42c. Only the three feature commits were replayed; the diff contains the same five feature files, with release-branch type annotations retained when resolving the registry conflict.

• Move the existing stock source registry into cached galaxy.tools.stock.stock_tool_sources_by_id() and share it between explicit-version fetches and TRS metadata.
• Resolve bare IDs from that registry in TRS, publishing wrapper-declared versions, tool name/description, GALAXY descriptor type, and URLs preserving the configured server scheme and escaping IDs/versions.
• Sort stock versions with Galaxy’s version parser, so 1.10.0 follows 1.2.0.
• Return 404 for unknown stock IDs.
• Preserve the existing workflow-safe version fallback for explicit-version stock fetches.

Representative available versions in this checkout:

Once deployed, the existing client’s unversioned stock-tool lookup can discover a version via TRS and fetch its parsed model without probing candidate pins. The public Tool Shed still requires its current explicit-version workaround until deployment.

Validation and readiness

• Tool Shed unit suite plus stock-tool tests: 82 passed, 12 skipped.
• Tool Shed tools API suite against a temporary local server: 9 passed, including both TRS endpoints, unknown-ID 404s, URLs for IDs containing spaces, and successful exact-version fetches for every advertised version of the six representative tools.
• Mypy: no issues in the three changed modules checked.
• Repository-pinned Ruff, isort, Black, and git diff --check: passed.
• Follow-up: replaced the monkeypatched ordering test with the existing multiple_versions_sorted XML fixtures (1.9, 1.10), retaining the numeric-versus-lexical ordering check and verifying exact-version fetches through the real registry; 11 focused tests passed. The TRS unit and API tests use no monkeypatching.
• PR lint follow-up: both Python jobs failed only on mypy’s optional descriptor_type iteration in the unit test; formatting passed. Added an explicit non-None assertion, preserving the existing descriptor-value assertion. Reproduced that error with mypy from lib/, then confirmed the targeted check passes after the fix; all 11 focused tests and formatting/lint checks pass. Full local make mypy additionally reports 13 unused-ignore errors in unchanged code that were absent from the PR CI logs.

Lint fix pushed as 3d2873a811; replacement PR checks are pending, and fork Python lint is queued.

Release rebase validation at f211f388f0: 75 unit tests passed, 12 skipped; all eight tools API tests passed. Targeted mypy on the changed modules, TRS schema, and unit test passed, along with the release-pinned Ruff (0.15.13), isort (8.0.1), Black (26.3.1), and git diff --check. The branch was pushed with an exact force-with-lease check, and the PR target, title, and validation results were updated. Replacement CI remains pending; the PR stays in draft.

The initial push triggered 29 workflows: 27 queued and two skipped at the first check. After the fixture-based follow-up, current fork Tool Shed tests, unit tests, and Python lint remain queued. PR CI is also queued, with CircleCI in progress. The draft is recorded in vault/agents/gx_branches/MY_BRANCHES.md under waiting for CI. No review-thread reply has been posted.

Separate existing TRS limitations

Code inspection also found that the global TRS GET /tools endpoint returns an empty list, and no TRS individual-version or descriptor routes are declared in lib/tool_shed/webapp/api2/tools.py. Those broader TRS features are separate from the stock-tool version-discovery failure addressed here; this branch does not implement full TRS conformance.