[MarkusNeusinger]

Summary

Two commits, one per plan step; each reviews on its own (agentic/specs/261024-agent-all-libraries.md). Say so if you want them as two PRs.

• Step 1a (commit 7fc230e), the runtime seam, no behaviour change for matplotlib and seaborn: the renderer holds one fixed render command per language in agents/renderer/commands.py, keyed by the job's (language, library) and never taken from the request; the probe harness gains an exec mode (harness.py --exec -- <command>) that runs a non-Python command under the same limits, start and end lines, launcher retry and kill path; GET /status reports the toolchains the image carries and a job in a language without one answers 503 runtime_unavailable. On the agents side the registry agents/anyplot/render/runtimes/runtime_for hands the pipeline, the theme toggle, the local backend and the eval harness one RuntimeAdapter per library, which also carries the language's code conventions, error grammar, export and the render time a repair reserves. The exported code's artifact name follows the language (plot.py, plot.R, plot.jl, plot.js, plot.tsx) through the wire, the BFF, the feedback case and the result card.
• Step 1b (commit 63d98e0), plotnine and altair through the Python runtime: one save rule per library (regions.SAVE_RULES) replaces the savefig rule in the regions, the validator, the normaliser, the readiness scan and the edit applier (plotnine's save and ggsave, altair's save with its second plot-<theme>.html and the PIL padding block). The validator gains both import profiles, the import surveys' bans and the routes their adversarial pass found (altair's image mark and from_dict, the narwhals frame behind limit_rows, transformed_data, plotnine's .environment), parses plotnine's aesthetic and facet strings as code and takes them only as literals it can read, and refuses formula= and from_formula. The harness admits the HTML write, hands plotnine's Environment.capture the plot's frame, lets Environment.eval run the walked strings, and attaches an Agg canvas so the probe measures plotnine's figure. The altair command row runs without RLIMIT_AS (vl-convert's V8 aborts under 64 GiB of address space). The renderer image installs both libraries and links DejaVu Serif and Sans Mono for vl-convert; ci-image.yml renders one plot of each through the harness. The eval fixtures gain 120 cases and two smoke cases.
• Commit b182dd4, the two blockers CI raised on 63d98e0, neither in the steps' own code: CodeQL's py/polynomial-redos (high) on policy._FENCE_TAG, whose two blank runs around an optional slash backtracked quadratically on user text; the blanks after the slash now sit inside the slash's own group, the matches are the same, and the CodeQL CLI (2.27.3, the workflow's suite and paths-ignore) no longer reports it locally. And the memory-floor executor test, which killed the run 60 ms after the launch and then compared the deletes with a launch log the loaded runner's fake launcher had not written yet; the floor is now crossed once the launch is in the log.
• Commit f4ceff0, five in-process tests for the seam's lines the suite reached only through a sandbox child: the local factory's runtime resolver, opening.assess for a library without a runtime, frame_for_readers, _renderer, and harness.main --exec with set_limits patched out (the exit codes, the start and end lines, the usage error). Patch coverage 96.4 %; what stays uncovered is the nine Protocol stubs in contract.py and four harness lines only the sandbox child runs.
• The phase-2 plan records the spike measurements, the Fable second opinion, the 1b decisions that settle its text, the eligibility sweep and the owner questions with their defaults.

Nothing changes for users until AGENT_LIBRARIES names plotnine or altair (the default stays matplotlib,seaborn; the plan's 80 % matrix gate decides).

Eligibility (agents.evals.eligibility, 2026-10-25)

Library	Files	Eligible
matplotlib	325	285 (87.7 %), unchanged against main
seaborn	325	287 (88.3 %), unchanged against main
plotnine	321	264 (82.2 %)
altair	325	269 (82.8 %)

The plan lists every blocked file's cause. A harness sweep with the test venv rendered 258 of the eligible plotnine files and 258 of the eligible altair files.

Plan

agentic/specs/261024-agent-all-libraries.md

Test plan

• uv run ruff check agents api core tests, ruff format --check, uv run mypy api core agents, uv run python -m tools.changelog check --base origin/main pass
• uv run pytest tests/unit/agents tests/unit/api tests/unit/core: 11,798 passed (the tests/unit/workflows and agentic/regen failures of the cloud container are environmental, as noted for 1a)
• New tests: the command table and its rows, the exec mode, the registry, the artifact names and their five copies, the save rules (regions, normaliser, readiness, edits), the plotnine and altair validator profiles with the reviewed submodule pins, the harness under the guard with plotnine and altair (HTML allowlist, frame reader, probe), the fixture generator
• Local Docker: the renderer image builds with both extras (1.13 GB, +107 MB); the three CI render steps, the convert smoke and the multi-script render pass against it; measured in the image: plotnine 3200x1800 with a probe (max_rss 196 MB, cpu 1.9 s), altair 3200x1800 plus the HTML (max_rss 309 MB, cpu 2.3 s)
• cd app && yarn lint && yarn fm:check && yarn type-check pass (1a); 1b changes no frontend file
• The fence tests and the executor bound tests pass five times under -n 4 load after b182dd4
• CI green on f4ceff0, the renderer-image job, CodeQL and the backend patch coverage included; the six Copilot threads of the first round are resolved
• Owner: the matrix run for plotnine and altair against the deployed 1b renderer (AGENT_LIBRARIES=matplotlib,seaborn,plotnine,altair), then _LIBRARIES in agents/cloudbuild.yaml; the plan's open questions, or the defaults stand

[Copilot]

🟡 Changes recommended

The review identified unresolved inconsistencies in the plan’s scope, runtime contracts, validation commands, implementation architecture, and changelog summary.

6 open issues

• Add the required roadmap status line · New
• Resolve mismatch between five and six planned PRs · New
• Pass library identity for JavaScript rendering · New
• Clarify registry scope for implemented runtimes · New
• Remove rejected Kaleido architecture from the plan · New
• Make library selection valid for matrix cases · New

What changed in this PR

This documentation-only PR adds a Phase 2 roadmap for expanding “Use with my data” from matplotlib and seaborn to all 15 catalogue libraries.

Changes:

• Defines runtime registry, renderer, harness, toolchain, security, and evaluation plans.
• Records spike results and open implementation decisions across Python, JavaScript, R, and Julia.
• Adds a corresponding changelog fragment.

File	Description
changelog.d/​agent-all-libraries.md	Changelog entry describing the Phase 2 all-library expansion plan.
agentic/​specs/​261024-agent-all-libraries.md	Detailed architecture, implementation steps, measurements, validation commands, and open decisions.

_{Review effort: Lite}

[codecov]

Codecov Report

❌ Patch coverage is 96.35974% with 17 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
agents/anyplot/render/contract.py	60.86%	9 Missing ⚠️
agents/anyplot/render/harness.py	90.24%	4 Missing ⚠️
agents/anyplot/code/validate.py	98.23%	3 Missing ⚠️
agents/anyplot/code/normalise.py	96.77%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!