Upgraded Python library versions to the latest compatible releases.

-WorkList use case: now uses ScheduleProcedure date, better test data

What's new

Multi-instance connection pool

Every execution tool now accepts an optional server parameter. Pass a server name and
the call routes to that instance without touching the default connection:

{ "tool": "iris_execute", "server": "prod", "code": "Write $ZV" }

iad maintains its own server registry at ~/.config/iris-agentic-dev/servers.json.
Passwords go to the OS keychain — never to the config file — using the same service name
as VS Code Server Manager, so both tools share credentials automatically.
Five new tools manage the pool: iris_servers, iris_add_server, iris_remove_server,
iris_test_server, and iris_import_servers.

Persistent WebSocket terminal sessions

Three new tools give Claude a persistent ObjectScript terminal over WebSocket:
iris_ws_open, iris_ws_exec, iris_ws_close. Variables and process state persist
between calls. Before this, every iris_execute call was a fresh context — a sequence
like "set X, do some work, read X back" had to be one call or use a global to carry
state. Now each step can be separate. Requires IRIS 2023.2+ (Atelier V7 API).

Administration and cross-instance comparison

Twenty-two new tools across global management, namespace/database admin, observability,
and cross-instance comparison. Most of this section is ported from Pierre Abdelsayed's
Server Manager MCP work — the global confirmation pattern, namespace/database admin,
observability tools, HL7 schema tools, Mermaid diagrams, and resolve_storage all
originate from his design. The data safety gates (PHI policy, system globals blocklist,
environment template) are also his, ported to Rust in an earlier release.

• Global management: global_preview + global_kill. Preview returns a confirmation
  
  
  token; kill requires it. The token is bound to the specific global and server, expires
  after 5 minutes.
• Namespace/database: iris_namespace_list, iris_namespace_create,
  
  
  iris_database_list, iris_database_stats.
• Observability: journal_search, query_audit_log, stream_inspect, my_access,
  
  
  capability_matrix.
• HL7 schema: hl7_schema_list, hl7_schema_inspect. Return HL7_NOT_AVAILABLE
  
  
  cleanly on IRIS builds without EnsLib.HL7.Schema (requires HealthShare or IRIS for Health).
• Visualization: mermaid_class, mermaid_production, resolve_storage.
• Cross-instance comparison: compare_document (unified diff of a single document
  
  
  across two servers), compare_namespace (full namespace diff).

MCP ToolAnnotations

All 64 tools now carry MCP ToolAnnotations. 57 are tagged read_only_hint = true
(queries, introspection, list tools). The 7 destructive tools — global_kill,
iris_admin, iris_credential_manage, iris_lookup_manage, iris_namespace_create,
iris_remove_server, skill_forget — are tagged destructive_hint = true. Claude Code
and other MCP clients that read these hints can show a confirmation step before
destructive calls and run read-only tools without prompting.

Notable fixes

Production management and skill tools broken on Atelier REST connections

iris_production, iris_production_item, and all skill tools returned errors or empty
results on connections without IRIS_CONTAINER set: VS Code extension, remote servers,
anything not running in a named local container. These tools have working Atelier REST
paths; the docker-exec fallback was routing around them. All affected tools now use
Atelier REST.

Breaking changes

None. All existing tool calls work identically. The server parameter defaults to
omitted, so hot-reload and active-connection behavior are unchanged.
iris_select_container still works.

Publish in Package Manager

Initial Release

Fixed the versioning issues.

• reorganize for Caché / Ensemble use only
• make it independent of other packages
• there's no ZPM on Caché >> all extras removed
• tie demo to Caché in Docker
• simplify container build
• include and run demo code in Docker build

Python Package Updated

Initial Release

Fixed timeout for the end-2-end run with model training. Ceiling corrected to 60s from 45s, updated to accommodate GaiaQualityScored's INSERT...SELECT with two PREDICT() calls over 74,998 rows was added. No code or model changes.

Today

Whoop and bundles support

Fix last week

make date filtering work with the FHIR server

date filters

submission

Adds a third custom IntegratedML IRISModel, this one a classifier, and the machinery to work out who is right when it contradicts the published DR3 label.

GaiaVariableType predicts variable-star type from 27 light-curve statistics computed off the per-epoch arrays: amplitude and scatter, skew and kurtosis, the Abbe parameter, a Stetson index, a Lomb-Scargle period and peak power, BP-RP colour, and ESA's own reject rates as features rather than a target. Labels come from gaiadr3.vari_classifier_result, so the ground truth is ESA's published classification. Held-out macro F1 0.938 over nine classes, 0.887 (YSO) to 0.991 (AGN). Registered through pathtoclassifiers, the mirror of the two NGBoost regressors already here, so the repo now exercises both AutoML pools. Run with do ^Classify.

Where the classifier and the catalogue disagree, do ^Adjudicate settles it in three stages of increasing cost, so nothing reaches a language model that arithmetic could have settled. Centroid distance decides 64.3% in the model's favour; k-means collapses about 50 label pairs into 6 recurring degeneracies. Only what survives goes to SIMBAD and an AI Hub agent, which weighs the published literature.

SIMBAD re-publishes Gaia's own variability catalogue, so every object-type query filters bibcode 2022yCat.1358 out. Treating those as independent evidence would mean checking ESA's labels with ESA's labels.

Well-studied stars route to a recursive pass over the bibliography (rlm-core) instead of truncating it to a keyword-ranked top 20. Also adds three test suites, including a cold-start regression that restarts the container before asserting PREDICT() returns rows.

The ObjectScript edition now processes the 20 independent Gaia DR3 input files in parallel using IRIS %SYSTEM.WorkMgr.

Each input file is processed as an independent work unit. The parent routine merges worker results in deterministic file and row order, preserving byte-identical output with the previously verified serial implementation.

The updated edition continues to produce 57,099 qualifying sources, passes all 16 validation cases, and matches the independent Python reference oracle.

In local testing, the median internal processing time improved from approximately 101 seconds for the serial implementation to approximately 33.6 seconds for the parallel implementation. Official moderator results may differ.

Similar to previous code but with unzip and warm start optimizations if allowed

added fhir server settings

udapte with WebApplication

add json support

add streams support for messages content

extra payload data

Added fix for handling possible stale .so file from previous builds

Initial Release

Initial Release

Initial Release

Separated gzip decompression from actual file processing, as realized the timed portion of RunScript for the benchmark, does need to include this step, implemented and added it as a pre-step

1. Improved performance -
   1. Using local container storage
   2. Using all cores
2. Shortened "short profile"
   1. Removed output file header line
   2. Changed some conditions