MIKE+ Audit

One JSON per case, written last. It is what makes "we ran X and got Y" checkable a month later, and it is the record a platform layer (SWMM / MIKE+ / LSTM side by side) can index.

Tool

• mike_manifest_write — {out_json, case?, model?, inputs?[], results?[], figures?[], run?, edits?[], tool_calls?[], notes?} → writes the manifest and returns {manifest, missing_files[], n_files, run_status}.
  • model: the .sqlite COPY that was run.
  • inputs: rain .dfs0 (incl. ones made by mike_rain_to_dfs0), CSVs, hotstart files.
  • results: the result_files returned by mike_run (or the two res1d of a compare).
  • figures: every PNG from mike_plot_*.
  • run: pass the whole mike_run return; the QA fields (status, completed, errors, warnings, issues, log_file, elapsed_s, active_simulation, result_files) are kept, the log tail is dropped.
  • edits: the mike_set_values returns (they carry table / muids / before / after).
  • tool_calls: ordered {tool, args, ok} records of what you called.
  • notes: the user's plain-language goal + anything not done and why.

Manifest shape (schema: agentic-mike/manifest/1)

created_utc, engine=mikeplus, case,
model{path, exists, bytes, sha256}, inputs[], results[], figures[]  (same shape),
run{status, errors, warnings, ...}, edits[], tool_calls[], notes,
environment{python, platform, mikeplus_mcp}

Conventions

• Write it even when the run failed: run.status, errors and notes are the evidence of why.
• missing_files must be [] before you say the case is complete; a listed path means you referenced something that is not on disk.
• Put it at runs/<case>/manifest.json; one case = one manifest (overwrite on re-run of the same case, or use a new case name).
• Never edit a manifest by hand; re-run the tool.

Orchestration

... mike_run -> mike_results_* -> mike_plot_* ...
mike_manifest_write {out_json: runs/<case>/manifest.json, model, inputs, results, figures, run, edits, notes}