go-ontology report for 2026-05-11

[github-actions]

Report date: 2026-05-11
Repo: geneontology/go-ontology
Window: 2026-04-27 – 2026-05-11

dragon-ai-agent is active and being used in collaborative back-and-forth with curators. 4 open agent-touched issues, 25 agent mentions, 0 merged PRs, 0 agent-authored items closed in this window.

Current activity is all in-flight: issues are open and stalled or awaiting human action, not closed out. A recurring pattern of correctness friction (wrong term hierarchy, editing files that should be auto-generated) is visible across multiple threads.

[github-actions]

Qualitative Review — go-ontology, 2026-05-11

What's working

dragon-ai-agent is regularly summoned by curators and responds quickly and concretely. In issue #31877, the agent completed a multi-step obsoletion task (mark term obsolete, remove taxon restrictions) across two separate invocations within the same session, correctly isolating its edits to the canonical source file (never_in_taxon.tsv) after a curator correction. In issue #32029, the agent added taxon constraints for two new insect-specific biosynthetic terms and cited appropriate PMIDs within minutes of being summoned.

In issue #27220, the agent was asked for a deep-research-style synthesis of a two-year-old discussion. It produced a structured proposal without making premature edits — an appropriate level of restraint for a complex, contested topic.

Curators are clearly comfortable using the agent as a first-pass worker and as a consultation tool.

What's not working / friction points

Repeated correctness errors on term hierarchy and reaction direction.
In issue #27593, the agent created a new ferric iron reductase term (first as GO:7770057, then after conflict re-created as GO:7770068), but the reaction direction and parent assignment were questioned by two curators (pgaudet, ValWood). The issue was re-opened and the agent was re-summoned on 2026-05-08 to review its own work. As of this report the review comment has been posted but no corrective PR has been issued — the item is awaiting a follow-up action. This is the longest-running friction point in the window.

File scope errors on taxon constraints.
In both issue #31877 and issue #32029, the agent edited auto-generated files (never_in_taxon.ofn, go_taxon_constraints.owl, or the OFN files) instead of only the source TSV. In both cases a curator had to intervene to explain the correct scope. The agent handled the correction well in both cases, but the same mistake is appearing in two separate threads — suggesting the scope rule is not reliably in the agent's system prompt or knowledge.

PRs opened but not merged in window.
Zero PRs were merged in the 14-day scan window. Several are open or stalled (e.g., PR #31973, PR #31997, PR #32045). The dossier doesn't surface PR-level merge blockers, but the combination of open correctness questions and awaiting-human signals suggests curator review is the bottleneck.

Stale open issue flow.
3 of 4 agent-touched issues are flagged as stalled. The nitrogen cycle issue (issue #27220) dates from 2024 with renewed activity in May 2026; it's still open after the agent's research comment. The ferric reductase issue (issue #27593) dates from April 2024. These are complex enough that slow closure is expected, but they highlight that agent involvement doesn't yet reliably push items to closure.

Recurring failure mode

The clearest recurring failure mode is auto-generated file contamination: the agent edits derived artifacts alongside source files, requiring curator rollback. This has now happened in at least two separate issues. A simple system-level note — "only edit src/taxon_constraints/never_in_taxon.tsv; the OFN and OWL files are regenerated by CI" — would likely eliminate this class of error.

Potential missed opportunities

The dossier surfaces five open issues with no agent involvement. Three are plausible candidates with concrete, bounded asks:

1. Issue #32064 — Obsoletion request: specific pathways for positive regulation of fatty acid beta-oxidation. Filed 2026-05-11 by pgaudet. The body requests obsoletions of named terms, and vanaukenk confirmed annotation removal is in progress. Once annotations are cleared, this would be a direct obsoletion task similar to what the agent already handles in issue #31877 — a well-scoped match for the agent's demonstrated capabilities.

2. Issue #32062 — Obsolete most L- and D-amino acid metabolic process grouping terms. Filed 2026-05-11 by sjm41. The body provides a checklist of terms to obsolete and lists which have zero or low annotation counts. The scope is larger but well-structured; the agent could work through the checklist term-by-term if directed. Human consensus on scope appears to be forming.

3. Issue #31172 — Obsoletion request: metabolic process — unnecessary grouping terms. Filed 2025-12-05, recently active. Has a checklist, annotation review, and obsoletion notice in the linked issues. More complex (includes restructuring proposals) but the annotation review and notice are done — the mechanical obsoletion work could be agent-delegated for a subset of the checked-off items.

Issue #31854 (phosphatidylinositol parent fix) and issue #31932 (pathway variants review) are also surfaced but are currently blocked on external decisions (ChEBI/RHEA liaison, and a Google Sheets review respectively) — less suitable as agent starting points right now.