v0.6.12

This release updates the standard CLIO agents' shared presentation, task and
dataset guidance while retaining each specialist's declared capabilities.

Evidence and visual review

Agents can choose diagrams, HTML or SVG when they help explain the material.
Layout follows the content, and visual review captures the actual view, inspects
the image and repeats after repairs. APPL-CORE uses natural citations and adapts
its analysis to unfamiliar read-only exports.

Search and managed work

The CLIO Search blueprint provides bounded read-only discovery. Shared guidance
covers managed execution/document tools, accepted background tasks and honest
onboarding checks before reporting readiness.

The changed standard packs are Base Agent 0.2.8, APPL-CORE 0.2.5, Factorio Flat
0.4.3 and EarthScope 0.4.3. See CHANGELOG.md and integration PR 91 for validation
and the preserved feature history.

Marketplace 0.6.11

CLIO beta 3 adds attention-evidence support to SPOTTER and canonical workspace-state handling to APPL-CORE.

• SPOTTER reads configured Flowcept attention captures using CLIO's shared profile reducer, carries capture references and uncertainty into reviewed findings, and rejects partial, ambiguous, mismatched or out-of-root captures.
• APPL-CORE keeps generated cards, loaders, audits and views in its owning workspace's Agent state.
• Both packs require CLIO 0.9.5b3 or newer. Older runtimes report the version requirement instead of activating a pack without the needed handoff.
• SPOTTER's implementation version matches its 0.4.0 blueprint; its managed interpreter and Marketplace CI use Python 3.13.

Validation: Marketplace policy, real blueprint/catalog validation, APPL-CORE pack tests, and SPOTTER tests/lint/type checking passed in GitHub CI. Fresh Delta GPU inference and both OPAL demonstrations remain separate live acceptance work.

v0.6.10

Domain agents now build on the standard agent

APPL-CORE Analyst, Factorio Flat and EarthScope Skills now have everything the standard agent (Base Agent, CLIO's default) has, plus their own domain:

• The same tools: shell, file reading and edits, image and PDF viewing, web fetch, and asking you a question. EarthScope Skills gains all of these; Factorio Flat gains file editing and web fetch.
• PDFs everywhere: every agent can read and check an existing PDF, and writes a PDF report when you ask for one. Reports are Markdown otherwise.
• Planning mode and model refresh: CLIO's built-in planning and update-models skills are available in every agent, not just the default one.
• The same working habits:
  • stay grounded in what was actually read;
  • use the fewest steps that do the job;
  • report failures plainly;
  • verify edits.

New agent

• APPL-CORE Analyst (0.2.0) analyses APPL-CORE plant-phenotyping exports. It needs CLIO 0.9.4.23 or newer.

Also

• Factorio Flat can show Abaqus geometry and results interactively. That includes topology-optimization views with stress and design-history sliders.
• Charts across the packs use one chart component. EarthScope Skills and Factorio Flat need CLIO 0.9.4.23 or newer.

Full details: CHANGELOG.md.

v0.6.5

Every agent in the marketplace now declares which interactive view catalogs it can use. This release pairs with clio-agent 0.9.4.17.

Agents declare their interactive view catalogs

Each agent lists the catalogs it may use to build interactive views, in the order it prefers them. Every agent here lists the CLIO workspace catalog, so all of them keep interactive tables, charts and maps. EarthScope Skills lists the CLIO workspace catalog first for general views and its own station catalog second for station maps. (#76)

Changed

EarthScope Skills now needs clio-agent 0.9.4.17 or newer, the first release that reads the new catalog list. The other agents still work with older clio-agent releases.

See the CHANGELOG for details.

v0.6.4

This release adds PDF workflows to the research packs, lets agents read PDFs natively on models that support it, and gives EarthScope Skills its own interactive station picker.

Read PDFs natively

The work-with-pdfs skill now reads a PDF directly with the new view_pdf tool when the connected model can read PDFs, a page range at a time for long documents. Base Agent and Factorio Flat declare the tool. On models without PDF support, the skill falls back to converting the PDF and inspecting rendered pages. Requires clio-agent 0.9.4.15 or newer.

PDF workflows in Factorio Flat

Factorio Flat can prepare existing PDFs as searchable text, structured JSON, and rendered page images, including a visual-only path for drawings, scans, and equations. A separate skill produces a PDF report only when a scientist asks for one.

Station picker for EarthScope Skills

EarthScope Skills ships its own interactive catalog: a station map and a multi-select station picker that hand the agent a structured selection instead of free text.

Fixed

• Base Agent can now inspect rendered pages of drawings and scans instead of reporting that it cannot see them.
• SPOTTER AI syncs its environment when its server starts.
• PDF guidance uses real workspace paths for both referenced and uploaded files, and no longer guesses units or dimensions from extracted labels.
• EarthScope Skills requires clio-agent 0.9.4.15 or newer instead of an unreleased version.

See the CHANGELOG.

v0.6.3

Marketplace 0.6.3 aligns the blueprint catalog with CLIO 0.9.2.

Included

• Factorio Flat includes Daisy Quach's six materials science specialists for materials, manufacturing, characterization, mechanical testing, fatigue and failure, and data analysis.
• EarthScope station selection preserves the exact Browser-selected station as structured context before staging and analysis.
• Base Agent and qualification blueprints cover current planning, workflow, resource, memory, provider, and delegated-agent contracts.

The release was live-qualified with the Marketplace commit fac72f0670607a08871a369040acd867e642ebec.

v0.6.2 — HDF5 expert, Factorio, Deep Researcher packs

Three new/updated Agent Blueprints land on main.

New packs:

• factorio — scientific-simulation research pack (Abaqus engineering workflow): a react main routing to research/simulation methodologists, an Abaqus engineer, evidence researcher/critic/leaf experts, and a virtual-lab expert, with a 4-skill library (frame_research_problem, formulate_abaqus_package, audit_scientific_package, maintain_scientific_dossier).
• deep-researcher — general-purpose research pack: main/researcher/critic expert hierarchy for open-ended research questions.

Updated pack:

• data-semantics — adds a dedicated hdf5 expert (tier-2 ReAct, child of main) with a curated 16-tool surface over clio-kit mcp-server hdf5 and a 21-skill HDF5/NetCDF library (datatypes, chunking, filters, file space, dimension scales, region references, VDS, SWMR, parallel I/O, map objects, CF compliance, cloud-optimized/HSDS/ROS3, the VFD family, VOL usage, and the benchmark-backed hdf5_clio_core_ingest ingest-decision guide). Requires bounded reads, exact tool evidence, guaranteed file closure, and descriptive failure/recovery guidance. Routes HDF5/NetCDF4/clio-core/CTE/CAE/chunking/filter/VDS/SWMR/VOL/parallel-I/O questions to the new child. Ships deterministic contract tests (data-semantics/tests/test_hdf5_pack.py) covering registration, routing, tool names, skill completeness, and cross-skill references.
  • This reconstructs and supersedes an earlier stale port of the same expert (#37, credited); the two competing PRs (#37, #50) are resolved in favor of #50, which was rebased against current main, adds the test suite, and expands the curated tool surface from 8 to 16 tools.

Housekeeping:

• feat/spotter-packs deleted — fully absorbed into main (zero unique commits/diff) since the spotter-ai implementation landed via subtree import in a prior merge.

Full history since v0.6.1: HDF5 expert pack, Factorio pack, Deep Researcher pack.

v0.5.2 — qwopus 7/7 EarthScope winning blueprints

EarthScope GNSS region blueprints synced to the exact experts that passed the full grind matrix 7/7 on qwopus (validated live; gemma & gpt-oss also 7/7).

Highlights:

• synthesis.md leaned 564→190 lines — prose-only contract, derive-don't-author, station-id = catalog code, copy-paths-whole, grounded-numbers-only, honest data-blocked / no-coverage negative. The key fix for small-context (9B) models.
• main.md completion grounding (staged station → route to analysis), negative path preserved.
• ndp_resource_resolver.md execute-don't-defer (stage the ranked candidate).
• earthscope_station_catalog.md tightened no-coverage rule (geo_filter column error ≠ no-coverage; retry with explicit lon/lat columns).

Use with clio-agent v0.5.2.