Conversation
Wrap Genomic Intelligence's hosted DNA-sequence models (promoter, splice, enhancer, chromatin, expression, annotation, plus a find-genes-then-expression composite) over the REST /v1 API and the keyless hosted MCP server. SKILL.md + references only, no vendored client — mirroring the alphafold-database skill. Registered in openclaw.plugin.json and the README skills table.
fetch_ensembl_sequence takes gene (fetch_region handles coordinates), load_demo_sequence requires name, and find_genes / find_genes_and_predict_expression take a sequence handle rather than a region. The annotation task is find_genes; there is no predict_annotation and no load_local_fasta on the hosted server. Also drop dangling references/errors.md pointers, add the 422 validation_failed case, and state the 9,198 bp window consistently.
869 appears four times in README.md; the initial commit bumped only the Skills Overview total, leaving the badge, tagline and intro contradicting it. Bump all of them, and mirror everything in README_zh.md (counts plus the genomic-intelligence row), which the repo keeps in sync with README.md.
…8-500,000 bp) Genomic Intelligence shipped a new expression contract to production on 2026-08-18 (gpu_service 2026.08.18.1). Expression now has its own published OpenAPI operation (POST /v1/tasks/expression/predict, schema ExpressionPredictRequest) instead of the generic templated one, so the skill's description of it was wrong in four ways: - the sequence bound is 9,198-500,000 bp, not "exactly 9,198 bp"; - a new required field tss_index (0-based, into the whitespace-stripped sequence, 4599 <= tss_index <= len-4599) selects the window the server slices, and is mandatory unless the sequence is exactly 9,198 bp; - options is now a closed object whose only, required, key is description; - violations are 422 validation_failed, never a padded/truncated 200. Also documents the response's new tss_index / scored_window / submitted_sequence_length provenance fields, and the gotcha that an in-range but wrong tss_index returns a confident 200 for the wrong window. The five other tasks and the composite workflow are unchanged. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… options)
Correct the skill against the settled /v1 contract (DEV 2026.08.19.4):
- Replace "all six tasks share one shape / one PredictRequest". Six literal
predict operations are published, one per task, each with its own request
schema, minLength and closed options object. The URLs are unchanged.
- Replace the "1-500,000 bp" length bound with the real per-task floors:
promoter 300, splice 100, enhancer 50, chromatin 200, annotation 1,000,
expression 9,198, composite 1,000; 500,000 cap throughout.
- Add the floor-is-not-regime rule: a request above the floor but below the
model's bio_spec.context_window_bp is accepted and scored against a padded
window. Enhancer's 50 bp bound vs its 249 bp context window is the sharp case.
- Correct "413 sequence too long": over-length is 422 validation_failed. 413
means the 16 MiB raw-body cap or the composite's sync_too_large.
- Document typed, closed per-task options (an unknown key is a hard 422
extra_forbidden), the closed 21-value error code enum, the defensive-details
rule, the Prefer header on all six operations, and the published composite
REST workflow.
- Note that tss_index errors report at loc ["body"], never body.tss_index, so
clients must match on error.code.
- Document the new bio_spec fields (request_max_bp, context_window_bp,
trained_window_bp), that max_seq_length_bp is legacy/ambiguous, that there is
no strand_sensitive flag, and that GET /v1/tasks/{task}/models returns a flat
object rather than the {data, meta} envelope.
- Add an unknown task -> 404 not_found note and a schema-version caveat, since
production may still serve 2026.08.18.1 until this build is promoted.
Verified against the live DEV schema (info.version 2026.08.19.4 (a5b1f88),
11 operations, no PredictRequest component).
Expression reports `context_window_bp: null`; 9,198 bp is `trained_window_bp`. The SKILL.md and references/tasks.md tables listed 9,198 under a "Model context window" column, conflating the two.
…property Every predict operation declares the Prefer header and both 200 and 202. Verified live against 2026.08.19.4: promoter submitted with Prefer: respond-async returns 202 and polls to a normal payload; annotation with no Prefer header returns 200 synchronously. Relabels the task table's "Mode" column "Recommended mode" and states the rule under the table, so the latency guidance stays without asserting a constraint the API never had. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Alexander Boldakov <alexboldakov@gmail.com>
error.request_id is not always populated — 413 sync_too_large omits it from the body. The X-Request-Id header is set on every response, so read the header as a fallback. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Alexander Boldakov <alexboldakov@gmail.com>
…08.19.5
Today's GI ship landed after this text was written, so several claims went
stale within hours:
- 422 details is the declared {"errors": [...]} object, not a bare array.
Defensive handling of either shape is kept; only the prose was wrong.
- error.request_id is populated on every error path, including both 413
variants; success envelopes carry meta.request_id.
- bio_spec.max_seq_length_bp is retired and absent from the live response;
request_max_bp is the cap.
- Version notes hedging that PROD may still serve an older build are moot.
- Sync/async is a per-request delivery choice, never a task property; the
composite's 50,000 bp sync limit is the only forced case.
Also renames the client's synthesized non_json error code to http_error,
which is in the published 21-value enum; client-origin errors remain
distinguishable by carrying no request_id.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Signed-off-by: Alexander Boldakov <alexboldakov@gmail.com>
…omments Audited against AUTHORING.md. Removes every pointer at the private service codebase (retired `gpu_service/*` paths, internal constants and symbols) and the internal release stamps that went with them, restating each as what the Genomic Intelligence API serves, with the published OpenAPI document as the authority. Also normalizes the research-use wording, removes pinned model IDs and prediction values from prose, and makes comment voice consistent. No contract facts changed; every bound, field and status code re-verified against the live schema. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
splice is strand-specific and the wrong strand fails silently: a reverse-complemented sequence returns plausible sites at different positions, often at high confidence, so neither the score nor the count reveals the orientation. Without this stated, the natural inference from "strand-specific" is that a bad strand shows up as a weak or empty result, which is backwards. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…unction base The reference documented data.sites' start/end without saying what the pair means. It bounds one variable-width tokenizer token (4-10 bp observed), comes with a token_index, and the exon/intron junction sits inside it. Reading either coordinate as the boundary base is wrong by up to about 10 bp in either orientation, and the gff3 output carries the same span in columns 4-5, so a saved track intersected against reference annotation inherits the error. Verified against the live contract (info.version 2026.08.20.11) and the pinned HBB fixture, whose eight sites span 4, 5, 6, 7 and 8 bp. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Author
|
Could a maintainer approve the workflow runs here? As a fork PR they need a manual approval, and all 7 runs since the PR opened are at |
data.input.sequence_length was removed from the API at contract revision 13; the docs described it as the scored 9,198 bp. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LA7HntiXoc7ZpQz8boGk6b
Author
|
Small docs correction: the API dropped data.input.sequence_length (it was the 9,198 bp scored window and got mistaken for the submitted length). The submitted length is meta.sequence_length, and the two references now say so. No code change. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Add
genomic-intelligenceskillAdds a skill wrapping Genomic Intelligence (GI), a set of hosted transformer DNA
language models that predict regulatory features, gene structure, and expression
directly from DNA sequence. No local GPU or model weights are needed; the skill is
a thin client over a hosted, versioned inference API.
What it does
Give it a gene symbol, a genomic region, or a DNA/FASTA sequence and it returns
structured predictions across six tasks, plus a composite:
promoter: promoter regions (sliding window)splice: donor/acceptor splice sitesenhancer: developmental and housekeeping enhancer activitychromatin: chromatin state across hundreds of tracksexpression: sequence to expression, log(TPM+1)annotation: de-novo gene/transcript annotationShape
This mirrors the existing hosted-API skills (
alphafold-database,chembl-database,pubchem-database): aSKILL.mdwithnameanddescriptionfrontmatter plus a
references/directory, using inlinerequestswith novendored client.
There are two ways to call GI. The REST
/v1API needs aGI_API_KEYbearer. Thehosted MCP server at
https://mcp.genomicintelligence.ai/mcpworks keylessagainst a public demo quota, with the key optional for a higher quota.
references/documents the tasks, the REST contract and auth, the MCP tool flow,and Ensembl-based sequence acquisition, including the exact 9,198 bp TSS-centred
window the
expressionmodel requires.Registration
Appended
"./skills/genomic-intelligence"to theskillsarray inopenclaw.plugin.json, in alphabetical position. Added a row to the README skillstable under Scientific Databases (Genomics & Variants) and bumped the Skills
Overview counts from 43 to 44 for that category and from 869 to 870 for the total.
Validation
python scripts/validate_skill.py skills/genomic-intelligencereports "Skill isvalid!" and
python scripts/validate_skill.py openclaw.plugin.jsonreports"Manifest valid, found 870 skills". No secrets are committed; the key is read from
the
GI_API_KEYenvironment variable.The keyless MCP path was verified against the live endpoint: an anonymous
Streamable HTTP session with no
Authorizationheader completesinitializeandreturns a real
predict_promoterresult.Notes
For research and development use, not clinical or diagnostic decisions.
Docs: https://docs.genomicintelligence.ai