v0.2.1 — pre-JOSS-submission hardening

Security defaults

• run_shell MCP exposure is now opt-in via FORTRANSPIRE_ENABLE_SHELL=1. Closes the unauthenticated shell-execution path through ask_agent.
• MCP file/directory arguments are jailed to the workspace root via a new _jail() helper. Auto-off in stdio mode (IDE trust boundary), enforced by default in HTTP/SSE.

CI

• New .github/workflows/tests.yml: pytest on every push and PR (Python 3.11/3.12 matrix), gfortran -fopenacc equivalence harness on push and manual dispatch.
• Tests badge added to README.

Repo cleanup

• Paper artefacts moved to paper/ (paper.md, paper.bib).
• Container artefacts moved to containers/ (Dockerfile, Dockerfile.hpc, apptainer.def, docker-compose.yml).
• Two stray root markdowns relocated under docs/concepts/ and docs/integrations/.
• Dead code removed: fortranspire/main.py, fortranspire/brain/, broken Dockerfile.ci and Apptainer.analyze.

Paper

• 2289 → 1103 body words (-52%). Eight-stage table + LLM-intervention rationale + agent-loop motivation moved to docs/concepts/architecture.md.
• Five orphan bib entries dropped; zero broken citations.
• "Sovereign" now backed in docs/concepts/llm-endpoints.md by GDPR (EU 2016/679), AI Act (EU 2024/1689), and NIS2 Directive (EU 2022/2555).

Install path fixes

• README and docs/getting-started/with-mistral-vibe.md: uv tool install fortranspire → uv tool install 'fortranspire[mcp]'. The [mcp] extra is required for the stdio handshake.
• Hardcoded /opt/homebrew removed (breaks Intel Macs).
• .env edit reminder after cp .env.example .env.

Install

```bash
uv tool install 'fortranspire[mcp]'
```

Full changelog: docs/changelog.md.

v0.2.0 — stdio MCP transport, mistral-vibe integration, JOSS-ready

Highlights

• stdio MCP transport for fortranspire mcp — IDE-spawned servers (mistral-vibe, Claude Code, Claude Desktop) over JSON-RPC framing. No port, no auth, local-by-construction. Pytest smoke test verifies the 9-tool surface.
• mistral-vibe integration officielle — turnkey 60-second quickstart in docs/getting-started/with-mistral-vibe.md, full walkthrough on PHYEX in docs/integrations/mistral-vibe.md.
• JOSS paper muscled with Real-world demonstration (PHYEX kernel — 1 routine, ~$0.04 codestral, 2 s estimation), Quality control (equivalence harness on real gfortran -fopenacc kernels), and Limitations / Roadmap sections.
• README cleanup — 1353 → 332 lines. Rich Fortran patterns content (11 patterns with before/after code) migrated to docs/concepts/fortran-patterns.md.
• Infra cleanup — removed obsolete Dockerfile.ci and Apptainer.analyze (broken paths). ROADMAP neutralised for JOSS-friendly framing.
• Affiliation standardised across paper.md / CITATION.cff / .zenodo.json / codemeta.json — Independent Researcher, Toulouse, France.
• CI — new .github/workflows/docs.yml builds Sphinx in nitpicky mode on every PR touching docs.

Tooling

• 9 MCP tools exposed: analyze_kernels, explain_port_cost, build_call_graph, generate_docs, translate_kernel_gpu, translate_kernel, profile_kernels, ask_agent, agent_status.
• 4 of the 9 are no-LLM (deterministic Loki / AST work).

Install

uv tool install fortranspire

Full changelog: see docs/changelog.md.

v0.1.3 — single-step install via loki-ifs

Restores pip install fortranspire as a one-shot install by depending on the new loki-ifs PyPI distribution (a redistribution of ECMWF Loki @ 0.3.7 under our control).

Highlights

• Single-step install — pip install fortranspire resolves Loki automatically via loki-ifs >= 0.3.7. No more manual pip install git+... step.
• Import unchanged — code continues to write from loki import Sourcefile, …. loki-ifs is purely a PyPI distribution name.
• Submodule for devs — vendor/loki-ifs is a git submodule pointing at https://github.com/maurinl26/loki-ifs, opt-in via git submodule update --init --recursive. Excluded from the sdist.

Install

pip install fortranspire==0.1.3

With extras:

pip install "fortranspire[gpu]==0.1.3"   # Phase 1
pip install "fortranspire[mcp]==0.1.3"   # MCP HTTP/SSE server
pip install "fortranspire[all]==0.1.3"   # everything (except docs/tests)

Related releases

• loki-ifs 0.3.7 — https://github.com/maurinl26/loki-ifs/releases/tag/loki-ifs-0.3.7 (PyPI: https://pypi.org/project/loki-ifs/0.3.7/)

See docs/changelog.md for the full notes.

v0.1.2 — PyPI-ready (Loki dep externalized)

Patch release that unblocks the first PyPI upload.

Why this release

PyPI's PEP 715 enforcement rejected the v0.1.0 and v0.1.1 wheels with 400 Can't have direct dependency: loki @ git+https://github.com/ecmwf-ifs/loki@0.3.7. ECMWF Loki is not published on PyPI under that name (the PyPI loki is an unrelated astronomy package), and PEP 715 forbids git+https://… URLs in Requires-Dist.

What changed

• Loki is no longer a declared dependency. Users install it manually in a second step:

  pip install fortranspire
  pip install "loki @ git+https://github.com/ecmwf-ifs/loki@0.3.7"

  The parser already falls back to a regex frontend when Loki is missing, so the package imports cleanly without it. analyze is just more accurate when Loki is available.

• Local development uses PEP 735 dependency groups. Run uv sync --group loki to pull Loki in for local work; this metadata is not part of the PyPI release.

• README + installation docs rewritten with the new two-step install and an explanatory callout on why Loki is separate.

Install

pip install fortranspire==0.1.2
pip install "loki @ git+https://github.com/ecmwf-ifs/loki@0.3.7"

Or with extras:

pip install "fortranspire[gpu]==0.1.2"   # Phase 1 transformations
pip install "fortranspire[mcp]==0.1.2"   # MCP HTTP/SSE server
pip install "fortranspire[all]==0.1.2"   # everything except docs/tests

See the changelog for full details.

fortranspire 0.1.1 — silent unified CLI + docstring fix

Patch release. Two cosmetic issues found by the post-tag smoke test of v0.1.0.

Fixed

• Unified fortranspire <verb> CLI no longer prints the legacy agent-* deprecation notice when dispatching to gpu, translate, or profile. Embarrassing UX for the canonical entry point — fixed by splitting each into an internal _*_main() (shared by unified + legacy) and a run_* wrapper (only the wrapper carries the deprecation).
• Package docstring (fortranspire/__init__.py) was a pre-rename relic. Replaced with the current project description; added __version__ = "0.1.1".

Upgrade

pip install --upgrade fortranspire    # once 0.1.1 lands on PyPI

No breaking changes. Every command from 0.1.0 keeps working identically. The deprecation warning on the legacy agent-* scripts is still in place (and will fire at the right time).

Test plan

234/234 pass (231 in 0.1.0 + 3 new regression guards that fail against the 0.1.0 wheel and pass here).

Full changelog

See docs/changelog.md.

fortranspire 0.1.0 — first PyPI release

First public release of fortranspire on PyPI. The project was renamed from coding-agent / local-code-agent; this is its first release under the new name and the first artifact published to PyPI + Zenodo.

Install

pip install fortranspire                      # core: analyze, explain, format, graph, diff, report, bench
pip install 'fortranspire[gpu]'               # + LLM pipeline (Phase 1, doc, port-batch)
pip install 'fortranspire[mcp]'               # + MCP server (FastMCP)
pip install 'fortranspire[jax]'               # + Phase 2 (Fortran → JAX, experimental)
pip install 'fortranspire[all]'               # everything

What's in this release

Unified fortranspire <verb> CLI with 13 subcommands:

no-LLM, CI-friendly	LLM-driven	server
analyze (Loki + SARIF)	doc (inline !> + Sphinx)	mcp (FastMCP)
explain (cost/risk estimate)	gpu (Phase 1 port, --gpu-pragma acc|omp)
format (fprettify wrapper)	port-batch (parallel)
graph (Mermaid call-graph)	translate (Phase 2 JAX)
diff (semantic before/after)	profile (bench)
report (HTML audit)
bench (regression detector)

Highlights

• Auto-generated CPU↔GPU equivalence test harness per ported kernel — every Phase-1 port comes with a numerical-equivalence pytest file (np.testing.assert_allclose per INTENT(OUT|INOUT))
• OpenMP target alternative emission — fortranspire gpu --gpu-pragma omp for multi-vendor compatibility (gfortran 13+, nvfortran, ifx, AMD AOMP, IBM XL)
• Parallel batch port with per-file ContextVar output isolation — 100 routines in ~45 min on 4 workers vs ~3h sequential
• SQLite-backed LRU cache for LLM responses — identical re-runs return instantly, zero token cost
• Per-tenant observability tracer with per-model price table (Mistral La Plateforme) and LangChain BaseCallbackHandler for token capture
• Hardened MCP server — JSON token registry, sliding-window rate limit, HMAC-signed audit log
• Native Mistral SDK auto-selected on api.mistral.ai, ChatOpenAI fallback for vLLM / TGI / Ollama / Scaleway / OVH
• French translations of every LLM prompt — initial audience (Météo-France, CEA, EDF R&D, ENM Toulouse) is francophone
• JOSS submission package — paper.md, paper.bib, CITATION.cff, codemeta.json, .zenodo.json, .readthedocs.yaml, and a GitHub Action that renders the JOSS draft PDF on every push

Sovereignty

No Azure / AWS / GCP dependency on the LLM path. Default endpoint = Mistral La Plateforme (EU-hosted). Works against any OpenAI-compatible endpoint (Scaleway Generative APIs, OVH AI Endpoints, self-hosted vLLM / TGI / Ollama).

Full changelog

See docs/changelog.md. 19 issues closed during the development sprint (#1 → #20).

Citation

A DOI for this release is being minted by Zenodo via the GitHub integration. Once it's live, the badge in the README will update; alternatively, every release tag mints its own DOI for fine-grained citation. Use CITATION.cff (rendered as the "Cite this repository" button on the repo home) for citation metadata.

🤖 Built with Claude Code