v0.7.1

A single natural-language sentence referring only to a WGS84 bounding box now drives an end-to-end SWMM workflow: synthesise the drainage network from public OSM + DEM data, run SWMM, write a deterministic audit dossier, and render a spatial network map — all via the standard runs/<date>/<id>/ layout. Synthesis comes from SWMManywhere (Imperial College London, BSD-3-Clause); aiswmm is the agent-side adapter. SWMM execution is byte-identical to v0.7.0 — Tecnopolo model.out SHA256 unchanged.

Added — three new LLM-facing typed tools (v0.7.1)

• map_run — render the spatial network layout (subcatchments + conduits + outfalls) of a SWMM model as a PNG. Sibling of plot_run at the LLM surface (plot_run = hydrograph; map_run = network map). In-process wrapper around the aiswmm map CLI verb so the agent can request a network figure in one step. 14 unit tests + family-mapping drift entry.
• plot_run.link — new link parameter renders a conduit Flow_rate hydrograph when set. Mutually exclusive with node (which still renders the node-attribute hydrograph). Plumbed through three layers: ToolSpec schema, _plot_run_args mapper (forwards link, suppresses node, picks a sensible default out_png filename by link id), and the swmm-plot MCP server's zod Args + CallToolRequestSchema handler (emits --link or --node to the underlying script, never both). 9 unit tests.
• read_rpt_summary — parses SWMM .rpt summary sections (Link Flow Summary / Outfall Loading Summary / Node Inflow Summary) into structured JSON rows sorted by the per-section peak/max column. Replaces the read_file-with-4000-char-cap workaround on 300+ KB rpts. The ToolSpec description explicitly steers the LLM to call the tool once per section needed ("CALL THIS TOOL ONCE PER SECTION YOU NEED — the tool is stateless") and to use it instead of read_file or search_files for rpt data. Verified on the 2026-05-28 Tecnopolo run: the LLM called read_rpt_summary four times in a single run with different section values. 25 unit tests.

Fixed (v0.7.1)

• MCP transport spawn ENOEXEC. A zero-byte .venv/bin/python stub (left by a half-finished venv or a test-fixture leak) caused the Node MCP launcher to assign env.PYTHON to an unusable executable. The downstream spawn(PY, ...) call returned ENOEXEC because the kernel cannot recognise an empty file as an executable. Fixed in two layers: agentic_swmm/utils/subprocess_runner.py:runtime_env() now pins PYTHON=sys.executable so the launcher inherits the correct interpreter, and scripts/run_mcp_server.mjs adds isUsableInterpreter() that rejects empty / non-executable candidates before assignment. Regression test plants a zero-byte stub and asserts the launcher does not select it.
• final_report.md "What you got" listed SKILL.md paths instead of real artifacts. _what_you_got was reading only result["path"], and only _read_skill_tool put a path there. Production handlers nested artifact paths under result["results"] / result["excerpt"] (MCP JSON) / result["args"] / result["summary"]. New recursive _mine_paths() harvests artifact paths from anywhere in the result payload, with planner-internal-fragment filtering and an introspection-tool skip set (read_skill, read_file, list_*, select_skill, search_files, capabilities). Reports now list the produced synth.inp, model.rpt, network_map.png, audit JSONs in their correct sections. 13 unit tests.
• --max-steps default 16 → 40 in both aiswmm agent and aiswmm chat. The 16-step ceiling cut the planner off mid-workflow because gpt-5.5 typically spends ~15 steps on introspection (list_skills, read_skill ×N, list_mcp_tools, select_skill) before the first real op. The new default leaves ~25 steps of headroom. Pinned by 4 tests.

Changed (v0.7.1)

• skills/swmm-anywhere/SKILL.md install + attribution language humanised. Same substance (BSD-3-Clause, Imperial College London, GitHub URL, citation request) reframed as one researcher crediting another rather than compliance copy. Upstream-attribution paragraph merged into the install section so it cannot be missed.
• .gitignore now covers *.egg-info/, caches (.cache/, cache/, .pytest_cache/), coverage output (.coverage, htmlcov/), Docker-mounted run outputs (docker-runs/), memory runtime side-files (command_trace.json, project_overrides.yaml, .last_refresh_error.json), spike research artifacts (scripts/spike_swmmanywhere/), and local-only experimental data dirs (data/Todcreek/of1/, examples/hand1/). package-lock.json now tracked for reproducible MCP-server npm installs.

Evidence (v0.7.1)

• Byte-identical reproducibility re-verified. Tecnopolo model.out SHA256 = 85c5514a81ea745ebb0c1c3e2aebb0c2cc0d5a6aa3ef00a0fa6c8f7b760be38c on v0.7.1, identical to the 2026-05-15 canonical lock-in across macOS native vs Docker stacks. Downstream tests that pin this SHA can upgrade v0.7.0 → v0.7.1 without re-baselining. See docs/byte-identical-reproducibility.md.
• Minimum NL prompt length for the full Tecnopolo chain is 11 words. examples/tecnopolo/tecnopolo_r1_199401.inp。run it and audit it and plot the result reaches synthesise-free SWMM execution + audit + plot end-to-end on the standard prepared INP. See docs/byte-identical-reproducibility.md for the recipe.
• Cross-session memory layer autonomously activated on a real run. The LLM planner, with no memory-related keyword in the user prompt, issued recall_session_history(case_name="tecnopolo") and recovered two prior Tecnopolo sessions from 12 days earlier — the first user-observable activation of the memory layer on a real workflow. See docs/v0.7.1-cross-session-memory-evidence.md.
• Natural-language SWMManywhere chain end-to-end on two independent regions. Greenwich Peninsula (1×1 km) canonical case study figures + NYC Midtown (1×1 km) cross-geography verification. See docs/v0.7.1-swmmanywhere-nl-driven-evidence.md.

Out of scope — next milestone

v0.7.1 ships the agent-side plumbing for SWMManywhere and the cross-session memory layer. The modelling-science quality bars are explicitly out of scope and tracked as next-milestone work: calibration of the synthesised network against observed flows, systematic continuity-error characterisation across bbox sizes, a memory-aware calibration loop, negative-precedent handling in memory recall, and time-decay weighting for stale precedents.

Attribution

Network synthesis in v0.7.1 is the work of SWMManywhere by Imperial College London (BSD-3-Clause licensed). Please cite SWMManywhere in any publication that uses or extends the SWMManywhere workflows shown in this release.

---