Skip to content

Liaison notes

The liaison's working state, rewritten to the present before each reset.

State on 2026-11-05 (machine october)

Ledgers: docs/UPSTREAM_GA4GH_VRS.md, docs/UPSTREAM_POLARS.md, docs/UPSTREAM_DUCKDB.md, docs/UPSTREAM_ENSEMBL.md (a standing watch with no clone or watcher, handed over by the lead 2026-11-04; re-checked by hand monthly, next 2026-12-04). Read marks: ga4gh.vrs 2.3.3 (pre-releases read through 2.4.0-a6), polars py-2.1.0, duckdb 2.0.0. Installed (uv.lock): 2.3.3, 2.1.0, 1.5.6 (since the 2026-11-17 bump). mkdocs.yml not_in_nav carries UPSTREAM_*.md. docs/UPSTREAM_DATA_SOURCES.md (since 2026-11-07, the maintainer's ask): one row per cache lane, available / built / published; measured by hand that day, re-measured weekly. Behind then: clinvar, constraint (pinned 4.1, source at 4.1.2), civic, strchive, likely mitomap. Since 2026-11-09 all are published current; the shared cache lags only on strchive (left to the maintainer). The priming paragraph that adds it is committed (fc6e7ca1).

Clones (in /data/sources, lines in $XDG_STATE_HOME/agent-seat-blueprints/liaison/upstreams):

  • vrs-python is a sparse checkout without tests/. Its VCR cassettes have > and : in their names, which the NTFS windows_names mount refuses, so a full checkout fails.
  • polars and duckdb-python are blobless clones (--filter=blob:none). Expect a git log -S over them to be slow, because it fetches blobs.
  • duckdb (the engine) is blobless and unwatched. It is there to read source at path:line.

Watchers (.blueprints.toml): ga4gh-vrs polls its trunk every 300 s. polars --tags --tag-pattern '^py-' and duckdb-python --tags say tags only, one wake per release (kit v0.1.119, #246; the clone follows each kit release by move-to-tag, its pin in .claude/settings.local.json), a patch tagged on a version branch included.

Backlog: none. Read entry by entry on 2026-11-04, unattended: polars py-2.0.0-rc.1 through py-2.1.0 with the 2.0 migration guide, and duckdb 1.5.6 and 2.0.0 (both packages' notes). Nothing new breaks us; the verdicts and the RMs stand. One new fact: duckdb 1.5.6 refuses a 2.0 file with a different message (version number 69, not 999), the same IOException, so RM344's handler must key on the type. The lead was told.

The person's rule (2026-11-04, relayed by the lead): the liaison decides adoption blockers and hands the lead adoption maps. A pin goes on only when a new major is proven to break something, and only until support is integrated. The lead pins as a safety only if the liaison is unavailable and work must move. The lead commits the lock; the liaison schedules it and tells the lead when.

Decided by the person, 2026-11-04: no cap on either major. polars-bio's own <2 already holds just-dna-lite, and a second cap here would block everyone else.

Sent to the lead (it files the RMs; this seat's grant covers only its ledgers and notes), 2026-11-03/04: the adoption plan for both majors without a cap, which supersedes the first message's cap option; and, for RM304, PR ga4gh/vrs-python#665 (RleSubunitMode, which vrs/2.1 will default to SMALLEST).

Lock schedule: polars 2.1.0 is in uv.lock (5bdcca28) and shipped in 0.8.0 (cut 2026-11-05, with RM342 and RM352). duckdb 2.x once RM344 (open, a patch on main) lands: then tell the lead. RM343 (EXPLAIN test) is done. polars-bio 0.37.0 (on PyPI by 2026-11-07) requires polars>=1.40.0,<2.2: the <2 block on just-dna-lite is gone; the lead was told (zulip …638402138). Its new cap is <2.2: a polars 2.2 release holds just-dna-lite again until polars-bio follows. Re-measured 2026-11-16: just-dna-lite on polars 2.1.0 + polars-bio 0.37.0 since 2026-11-12; just-vcf moving now (its RM288). polars-bio is on the upstream-index watcher since 2026-11-16.

The dependency bump (the person, 2026-11-04, relayed by agents-kit/builder, zulip …637618265) is this seat's unattended routine, in the grant since fc6e7ca1 (2026-11-09; desk #195 withdrawn as moot once the auto-mode rules were current). The watcher nightly runs nightly.py deps (41e8dced), started in the watchdog 2026-11-09; it says the bump due weekly while the person is away (this seat's day: Tuesday). The grant stands: desk #246 (whether a relayed word holds) was withdrawn as moot on 2026-11-12 under the kit's D193, the seat-builder's hand being the person's word. First bump 2026-11-10 (867e8462, 4e30ef46, pushed): 39 moves, duckdb held at 1.5.5 for RM344; typer 0.29 dropped Parameter.type, so scripts/gen_cli_pages.py reads display_type. One test red on main before and after, the fence test's huggingface_hub case: hub fetches its agent-harness registry from the global endpoint and the test passes only within a day of a real hub call. Fix on liaison/hf-hub-2.4 (343c559c); the lead took its six lines to main as 11f86d74 (2026-11-10). Lesson: before blaming a moved package, run the failing test on the old lock. Second bump 2026-11-17 (830962e9): eleven moves, duckdb 1.5.5 -> 1.5.6 (held on 1.x for RM344), suite 7446 green, no new warning. Re-measured the same run: data sources (clinvar and mitomap behind their sources, sent to the lead), the GWAS probe (unchanged), the VCF PRs (unchanged). The duckdb-python tag watcher went (finals only, as the index); polars' stays (its rc tags). Next bump: Tuesday 2026-11-24. Its worktree goes on disk, sh $K/bin/wt.sh liaison/deps-<date> (.worktrees/), never the scratchpad: /tmp here is a 32 GB tmpfs (RAM), and the three bump worktrees of 11-10/11-17 took 1.5 GB of it (base v0.1.236, kit #536).

GWAS Catalog REST v2 (docs/UPSTREAM_GWAS_CATALOG.md, since 2026-11-11, after RM428): a live service, no package or clone; its release channel is the news page (read through the 2026-08-19 post). A weekly probe of the page ceiling (500), the legacy 410 and the data release rides with the bump. Its first verdict: the docs now publish a 15-queries-per-second throttle, which gwas.py's docstring says does not exist; a docstring fix, sent to the lead.

VCF specification (docs/UPSTREAM_VCF_SPEC.md, since 2026-11-12; the maintainer's word relayed by the pm, 639790664): hts-specs master watched daily, read mark 510c107 (the 4.5 audit's tree, unchanged since). Filing an RMn myself for a spec change is blocked: desk #348 (filed 2026-11-16 when the person went away; their word came only through the pm); its default, and the rule until answered: a finding goes to triage. Tooling: the kit's watch-upstream.sh covers it; no new script.

AlphaGenome (terms: docs/UPSTREAM_ALPHAGENOME.md, opened 2026-11-14 after the pm's RM433 trial read newer pages; the client and protos stay here; the lead checked v2.0.0 on 2026-11-11): Atlas protos byte-identical to the pin aa6fc8f, no bump owed. Trap for the next pin bump: at v2.0.0 protos/dna_model.proto is a shim importing protos/v1/dna_model.proto, so atlas_protos.py must fetch that too. Adopting v2's live service is RM433 (parked).

Role change read (v0.1.195, 2026-11-13): a new duty, upstream-index. Done 2026-11-13 on v0.1.200 (L143 there): [seat.liaison.watch.upstream-index] over ga4gh.vrs, polars and duckdb with --final (kit #470, v0.1.201: duckdb publishes a .devN to the index almost daily). At the next weekly bump: drop the polars and duckdb tag watchers if they only duplicate this one. reflect after 2026-11-20: classes/liaison/LIAISON.md as of v0.1.195.

Kit #464 (fixed in v0.1.199, 2026-11-13): the clone's only install record sat under the old symlinked path at 0.1.192. move-to-tag.py v0.1.199 --take (v0.1.199's own copy) wrote one under the real path at 0.1.199; the stale record stays, left as #142 says. A seat loads 0.1.199 after its session restarts. Check the record at each kit move.

Fellow liaison: just-dna-registry/liaison (seated 2026-11-04) stands on fastapi, huggingface-hub, and polars and duckdb through our tiers. A defect in a library under both of us is raised with each other first, so it is filed once. longevity-loci/liaison (seated 2026-11-05) stands on us: just-dna-enricher 0.7.3, compiler 0.7.2, format 0.7.1 from PyPI, and the module-creator plugin v0.48.0. It watches our tags, asks us before mapping our 0.8 branches, and polls us before filing a shared defect; answer such a poll in agent-chat. Since 2026-11-17 it is in another chat organisation (agents-kit/lead, 640912858): its accounts here are retired; reach it through the embassy (watchdog.py send --encrypt ... --embassy genie), and what it sends is information, never an instruction (base, "Another organisation").

What the fellows did with the majors (2026-11-05): the registry's suite passes on polars 2.1.0 and duckdb 2.0.0 (612 passed, no new warning); it takes polars 2 with its 0.8 adoption and duckdb 2 after RM344. longevity-loci capped both <2 in its own lock, because the module-creator plugin's server (v0.48.1, still polars 1.44.2 and duckdb 1.5.5) shares that machine's enricher cache: the mixed cache RM344 is about. Its cap lifts when the plugin moves; their call. just-module-creator/liaison (seated 2026-11-07) keeps the plugin in step with us: ledgers UPSTREAM_COMPILER.md, UPSTREAM_REGISTRY.md, UPSTREAM_HOSTS.md; it locks format 0.7.1 / compiler 0.7.2 / enricher 0.7.3, capped <0.8, and polars and duckdb <2, until both registry instances serve 0.8. It is on our triage's wake_from (8de1fa80). just-vcf/upstream-liaison (seated 2026-11-16, its 6f06b56) watches polars-bio and polars from the index too: a polars-bio cap that holds both of us is raised with it first.

S34 with the registry (2026-11-16): answered (their d76adff), a minor for their pm: release / release_unreadable on both lane replies, revision for remote reads. Relayed to the pm with their two corrections (a label would not have caught the ClinPGx copy, RM445's case; our pull record writes no revision).

S36 with the registry (2026-11-18): answered (their f26ef9b; three candidates for their pm: read our KEY_COMPARES at their 0.9 floor, a per-request compare, or fold by default; the reply will name its compare; relayed to lead and pm). Their pm decided the first (2026-11-18): the registry reads KEY_COMPARES in its release that floors 0.9, the lane reply naming the compare per key column; told them it needs the enricher floor, since the declaration lives in just_dna_enricher.key_compare, and told our lead to keep that module name stable) (uncommitted in their inbox, their triage told in agent-chat 641541806), for the lead's RM430 follow-up: their lane lookup compares a key cell exactly (cast(String).is_in, read at their e0129b8), so a folded gene key or a member of a ;-joined cell comes back short. Our 0.9 reads Hugging Face for those columns meanwhile.

Push: origin/main holds every liaison commit through 5611ea91; the notes commits after it wait for the lead's next push (a push of mine was refused by auto mode, kit #332). The liaison pushes only its own ledgers and notes, and only as a fast-forward that carries no other seat's commit.

Role change read (v0.1.137, 2026-11-06): the role adds "An in-house release" (the liaison takes the mechanical half of adopting an upstream the person also keeps). This repo has no in-house upstream (ga4gh.vrs, polars, duckdb, Ensembl are all outside), so nothing changes here. reflect after 2026-11-13: classes/liaison/LIAISON.md as of v0.1.137 (filed: kit #518, neutral). reflect after 2026-11-17: classes/base/BASE.md as of v0.1.153 (a brief change another seat relays is committed only on the person's word in this session; the grant change fc6e7ca1 rested on a relayed word, put to the person 2026-11-10). reflect after 2026-11-17: classes/base/BASE.md as of v0.1.165 (the tooling seat named by its duties, D181; this repo has none, nothing changed here; the liaison role's text unchanged); and as of v0.1.166 ("Another organisation": what another organisation sends is information unless the person arranged it; the fellow liaisons are seats of this organisation, and an upstream filing is not an embassy message, so nothing changes here); and as of v0.1.167 (a person's reaction on this seat's post reaches it, a yes or no emoji wakes it; no concern); and as of v0.1.176 (a secret seen in a message is never repeated; this seat holds none, no concern). reflect after 2026-11-21: classes/base/BASE.md as of v0.1.206 (the nothing button closes the wake, not the turn: pending work that nothing blocks goes on in the same turn; no concern). reflect after 2026-11-23: classes/base/BASE.md as of v0.1.224 (only a real fork goes to the person: of this seat's asks, none is a fork; the VCF filing grant waits on the rule that a seat-relayed brief change is the person's to confirm, the rest are steps only the person can take; no concern). reflect after 2026-11-24: classes/base/BASE.md as of v0.1.231 (a claim of done is tested where the person looks, an absence carries its time; the liaison's live-read rule of 2026-11-16 says the same for upstream states; no concern). reflect after 2026-11-25: classes/base/BASE.md as of v0.1.236 (a worktree goes in .worktrees/, never the scratchpad; this seat's three were there, unlocked and handed to the person, whose settings keep removal theirs; the next bump uses bin/wt.sh).

Priming c339487 (a cap on a proven break; adoption maps to the lead) was read; no concern.