version: "1.0.0" name: assessment-report-converter description: | Convert an arbitrary CSV report (e.g. a Black Duck export or a custom migration-issue inventory) into a schema-valid assessment report.json the modernization pipeline can consume — so it appears in the assessment UI and migration solutions resolve automatically. This skill is LLM-driven: you read and interpret the CSV yourself and author report.json by hand against the schema. A small helper script only does deterministic lookups (migration solutions, the ruleId for a solution) and validates the finished report. There is NO "convert everything" script and NO assumed column layout. Triggers: "convert csv to assessment report", "import csv report", "turn this spreadsheet into a report.json", "Black Duck csv to report", "build report.json from csv", "migrate a third-party assessment export". NOT for: AppCAT-style analysis from source (use assessment), generating a modernization plan (use create-modernization-plan), or editing a report.json the pipeline already produced.

Assessment Report Converter (CSV → report.json)

What this skill does

You take a row-oriented CSV of migration issues and produce a canonical `report.json` conforming to the assessment report schema. Once written to the versioned reports directory, the report is picked up by the assessment UI and by the solution-resolution logic — exactly like a native AppCAT report.

This is an LLM-driven conversion, not a fixed mapping script. CSV layouts vary between tools and change over time, so you read the file, decide what each column means, classify every row, and author report.json. You do not rely on hard-coded column names.

A small helper script — scripts/report_tools.sh (bash + jq) or its PowerShell twin scripts/report_tools.ps1 (PowerShell 7+) — provides only the deterministic pieces you should never guess:

• `list-solutions` — discover the available migration solutions.
• `rules-for-solution` — get the canonical ruleId(s) for a chosen solution.
• `upgrade-solutions` — the canonical JDK / Spring Boot / Spring Framework / Jakarta EE upgrade solutions + ruleIds.
• `validate` — structural + consistency validation of the finished report.json.

flowchart LR
    A[Read CSV] --> B[Understand columns] --> C[Find projects] --> D{Audit each row}
    D -->|CVE / CWE| E[Security finding]
    D -->|Needs upgrade| F[Upgrade incident]
    D -->|Other| G[Solution → ruleId → incident]
    E --> H[Assemble report]
    F --> H
    G --> H
    H --> I[Validate] --> J[Save report.json] --> K[Summarize]

Input parameters

• csv-path (mandatory): Path to the source CSV file.
• workspace-path (optional): Output root. Defaults to the current directory. The report is written to

{workspace-path}/.github/modernize/assessment/reports/report-{reportId}/report.json.

• producer (optional): A label identifying the source tool, stored in report.producer (e.g. "Black Duck").

Defaults to "CSV import".

When to use this skill

Use this when you have a CSV — not an AppCAT report.json — and you want it to behave like a real assessment report: a Black Duck / third-party export, or a hand-maintained spreadsheet of migration issues. Do not use it to run analysis from source code (that is the assessment skill).

The helper script

The helper ships as two interchangeable implementations of the same CLI — use whichever fits the machine: scripts/report_tools.sh (bash + jq) and scripts/report_tools.ps1 (PowerShell 7+, no extra dependencies). Both behave identically and never read the CSV. Run them from the scripts/ directory.

# bash + jq
./report_tools.sh list-solutions [--query KW] [--type Formula|Chat] [--ids-only]
./report_tools.sh rules-for-solution <solutionId>
./report_tools.sh upgrade-solutions
./report_tools.sh validate <path/to/report.json>

# PowerShell 7+
pwsh ./report_tools.ps1 list-solutions [--query KW] [--type Formula|Chat] [--ids-only]
pwsh ./report_tools.ps1 rules-for-solution <solutionId>
pwsh ./report_tools.ps1 upgrade-solutions
pwsh ./report_tools.ps1 validate <path/to/report.json>

• list-solutions prints the matching solutions from scripts/solution-mapping.json as

whole JSON objects (each always has solutionId, name, type, and tooltip; some also carry effort / prompt); --query filters by substring on id/name/tooltip.

• rules-for-solution prints { solutionId, ruleCount, rules: [{ruleId, sourceCategory}], preferredRuleId }. An empty

list means the solution has no rule (e.g. a security-only solution) — do not invent a ruleId.

• upgrade-solutions prints, per component (jdk, spring-boot, spring-framework, jakarta-ee), the solutionId

and its resolved preferredRuleId.

• validate runs structural + cross-field consistency checks

(incident→rule references, required domain/category rule fields, enum values, security-finding shape and dedupe, and metadata.domains↔content consistency). Exit code is 0 valid, 1 invalid, 2 when the report can't be read/parsed. It accepts a --schema <path> flag for CLI compatibility but ignores it — the checks are self-contained.

Reference material

Detailed reference material is consolidated in the Reference section at the end of this file. The workflow below links to the relevant part at the step where you need it — you do not have to open any separate files:

• Report structure — the top-level report.json shape and the authoring conventions the schema can't express (the schema owns the field shape).
• Assessment domains — the metadata.domains values and when to include each.
• Security severity mapping — normalizing a source CVE/CWE severity to the security mandatory|potential|optional scale.
• Rule classification — the required domain / category rule fields that group and render a rule's incidents.

Workflow

Step 1 — Understand the report structure (do this first)

Read the authoritative schema scripts/assessment-report.schema.json — it is the source of truth for the field shape. The report is a single root object { version, producer, metadata, projects, rules, security? }, and additionalProperties is false almost everywhere (there is no summary object). The report structure reference adds the authoring conventions the schema can't express. You must know the required fields before authoring anything.

Step 2 — Read and understand the CSV

Read the CSV file directly (headers + a representative sample of rows). Then interpret the columns by meaning, since names and order vary by tool. Identify whichever of these the CSV actually carries (any may be absent):

• a project / module / service identifier,
• an application name,
• an issue title and description,
• a severity / criticality / priority,
• a category / domain / type,
• a CVE / CWE identifier,
• an affected component / package / library and its version,
• a file path and line,
• an effort / story-point estimate,
• a reference / URL.

When a concept is missing, leave the corresponding report field empty or at its default — never fabricate data.

Step 3 — Determine the projects

Work out how many projects the CSV describes and group rows accordingly:

• If a project/module/service column exists, group rows by it — one project per distinct value, using that value

as project.path.

• If the CSV has no project information, create a single project with path "." and minimal properties

(appName ""; leave optional jdkVersion / frameworks / languages / tools off or empty).

• Populate project.properties only from what the CSV actually provides; otherwise keep them empty. Only appName

is required.

Every incident you create later belongs to exactly one of these projects.

Step 4 — Audit each row, within its project

Classify every row into exactly one of three branches and attach the result to the row's project. Every rule you add to rules{} must set its domain and category fields (that is what groups and renders its incidents) — see rule classification. Incidents do not require any label.

1. CVE / CWE → security finding (directly).

When a row carries a CVE-… / CWE-… identifier (or unambiguously describes one), add a security finding to report.security[]. Capture as much of the column as possible: id (the CVE/CWE token), title, category, severity (the security scale mandatory | potential | optional — normalize the source severity per security issue severity mapping), description, evidence.files (affected paths), evidence.explanation, and optional storyPoint. Merge by id — one finding per CVE/CWE; accumulate evidence files and keep the strongest severity (mandatory > potential > optional). Security findings are not incidents; they live only in report.security[] (there is no summary to count them in).

1. Needs a major-component version upgrade → upgrade incident.

If the row implies upgrading a major component — JDK, Spring Boot, Spring Framework, or Java EE / Jakarta EE (e.g. a CVE against spring-boot, an out-of-support runtime, an explicit "upgrade JDK") — run the upgrade-solutions command, pick the component, and use its preferredRuleId. Add that rule to rules{} (severity: "mandatory", domain: "java-upgrade", a category, a reasonable effort) and add an incident to the project — one upgrade rule per component per project, one incident per triggering row. A CVE that implies an upgrade produces both a security finding (branch 1) and an upgrade incident — that is what makes the upgrade resolve as a migration solution.

1. Any other issue → solution → ruleId → incident.

Find the migration solution that fits the issue with the list-solutions --query <keyword> command, then get its canonical ruleId with rules-for-solution <solutionId> (use preferredRuleId). Add that rule to rules{} (with domain, category, severity, effort) and an incident to the project. If no solution fits, you may still record the issue with a clear synthetic ruleId (it just won't carry an automatic Formula solution) — or leave it for the "remaining" list in your summary. Either way, report it.

Process rows in parallel

Classifying a row is independent work, so for large CSVs do it concurrently rather than one row at a time:

• Batch the rows (e.g. 20–50 per batch, or one batch per project) and dispatch the batches in parallel — launch

several Explore/worker subagents at once, each auditing its batch into a partial result (security findings, upgrade hits, and ordinary incidents with their resolved ruleIds). Ask each worker to return structured JSON; it does not write files.

• Share the deterministic lookups. Run the upgrade-solutions command and the list-solutions /

rules-for-solution queries once up front (results are stable) and pass them to the workers, so parallel batches don't repeat the same lookups or race on them.

• Keep workers side-effect free, then merge sequentially in one place so shared state stays correct:
• Security findings — merge by id (one finding per CVE/CWE; union evidence.files, keep the strongest severity).
• Upgrade rules — collapse to one rule per (project, component); keep every triggering incident.
• `incidentId`s — assign "<ruleId>/<n>" after the merge, never inside a worker (so ids are deterministic

regardless of batch order).

• If the CSV is small, just process the rows sequentially — the parallel split only pays off at scale.

Step 5 — Assemble the report and finalize metadata

There is no summary object to compute — assemble the top-level document and fill metadata:

• projects[] = your projects, each with properties (only appName required) and its incidents[].
• rules{} = every distinct rule you referenced, keyed by ruleId, each with id, title, severity, effort,

domain, and category.

• report.security[] = the deduped findings (omit or leave empty when there are none).
• metadata.domains = the assessment domains your report actually has content for, consistent with your rule.domain

values — see assessment domains.

• metadata.mode (optional) = "full" when any security finding exists, else "issue-only".
• metadata.status = "completed"; metadata.targetIds = the target ids in scope (may be empty).
• metadata.id (and the report-directory id) = analysisStartTime formatted yyyyMMddHHmmss (UTC); use the current UTC

time when the CSV has no timestamp.

Step 6 — Write and validate

Write report.json to the versioned location (Output location), then validate and fix until clean:

# bash + jq
./report_tools.sh validate "<workspace>/.github/modernize/assessment/reports/report-<id>/report.json"
# …or PowerShell 7+
pwsh ./report_tools.ps1 validate "<workspace>/.github/modernize/assessment/reports/report-<id>/report.json"

Resolve every reported consistency error before finishing.

Step 7 — Summarize for the user

Report a concise conversion summary:

• Converted rows — number of security findings (CVE/CWE), upgrade incidents (with components), and ordinary issue

incidents (with the solutions they mapped to); plus the project / rule / incident counts and the report path + id.

• Remaining rows — rows you could not confidently map, and why (no matching solution, ambiguous column, missing id).
• Suggestions — concrete next steps (e.g. pick a specific solution for a remaining row, add a project column to the

CSV, supply severities), so the user can close the gaps.

Output location

• {workspace-path}/.github/modernize/assessment/reports/report-{reportId}/report.json
• reportId = the report's metadata.id from Step 5 (analysisStartTime as yyyyMMddHHmmss, UTC).
• Consider copying the original CSV next to report.json as source.csv for provenance.

Success criteria

• ✅ report.json is written to the versioned reports directory and the validate command reports VALID.
• ✅ Projects reflect the CSV (one per module/service, or a single project with appName: "" when none is given).
• ✅ CVE/CWE rows are security[] findings (deduped by id) with a mandatory|potential|optional severity; mode is full.
• ✅ Rows implying a JDK / Spring Boot / Spring Framework / Jakarta EE upgrade add a mandatory upgrade incident whose

ruleId came from the upgrade-solutions command.

• ✅ Other issues map to a solution's canonical ruleId (via list-solutions + rules-for-solution) wherever one fits.
• ✅ Every rule carries domain and category fields; every incident's ruleId resolves to a rule in rules{}; CVE/CWE

findings stay in security[], not incidents[].

• ✅ metadata.domains matches the report's content and the emitted rule.domain values (enforced by validate).
• ✅ The user gets a summary of converted rows, remaining rows, and suggestions.

Troubleshooting

• `validate` reports a consistency error (e.g. an incident ruleId with no matching rule, or a

domains/security mismatch) — the message names the exact field; fix that field.

• Don't add fields the schema doesn't allow (e.g. a summary object, or issues/storyPoints on a project). The

helper's validate is lenient about extra keys, but the app's importer enforces additionalProperties: false and will reject the report — this schema has no summary. Keep to the documented shape.

• A row has a CVE/CWE *and* needs an upgrade — emit both: a security[] finding and an upgrade incident. They are

not duplicates; the finding documents the vulnerability, the incident drives the upgrade solution.

• No solution fits an issue — list-solutions --query returns nothing useful. Record the issue with a synthetic,

descriptive ruleId (no Formula will attach) or list it under "remaining" with a suggestion.

• `rules-for-solution` returns an empty list — that solution has no rule (often security-only). Don't fabricate a

ruleId; handle the issue via the security-finding path or pick a different solution.

• Wrong severity enum — rule/incident severity is the 4-value mandatory|potential|optional|information enum;

security-finding severity uses the same scale minus `information` (mandatory|potential|optional, since a finding is always at least optional). Don't use the old critical|high|medium|low|info values.

Reference

Consolidated reference material. The workflow above links here at the step where each part is needed.

Report structure

You author a single report.json conforming to the authoritative schema scripts/assessment-report.schema.json. The schema is one root object (draft-07, its sub-types live under definitions) with the top-level shape { version, producer, metadata, projects, rules, security? }. Read the schema for the exact required fields, types, and enums per object — it is the source of truth and the validate command checks it, so this section does not restate the field list. additionalProperties is false almost everywhere, so do not invent fields (there is no summary object, and projects have no issues / storyPoints).

Authoring rules of thumb — the conventions and cross-field rules the schema can't fully express:

• version is the string "1.0.0" (the schema accepts any string; this is the value to use).
• Rule/incident severity is the enum mandatory | potential | optional | information. Map source severities by meaning

(e.g. critical/blocker → mandatory, major/medium → potential, minor/low → optional, info → information).

• Classification is done with rule fields, not labels. Each rules{} entry sets domain

(cloud-readiness | java-upgrade | security) and category (a free string heading) directly — see Rule classification. rules{}.labels and incidents[].labels are optional free-form arrays; you normally leave them out.

• Security-finding severity uses the report criticality scale mandatory | potential | optional (the rule Severity

values minus information — a security finding is always at least optional) — normalize per Security severity mapping.

• metadata.domains records which assessment domains the report has content for — see

Assessment domains. It must be consistent with the rule.domain values you emit.

• incidentId convention: "<ruleId>/<n>" (n is a per-rule counter).
• locationKind = "source-file" when a file path is present, else "unknown".
• status is normally "completed".
• metadata.mode is optional; use "full" when the report has security findings, else "issue-only".

Assessment domains

metadata.domains is a string[] recording which assessment domains produced the report. Allowed values are cloud-readiness, java-upgrade, and security.

• The native Java default is ["cloud-readiness", "java-upgrade"].
• Set metadata.domains to exactly the domains your report has content for — don't list security with no

findings, or java-upgrade with no upgrade incidents. validate flags a security/report.security mismatch in either direction.

• metadata.domains must be consistent with the rule.domain field on your rules — every rule.domain value you

emit should appear in metadata.domains.

Security severity mapping

A security finding's severity uses the report criticality scale — mandatory | potential | optional. This is the rule Severity enum minus `information` (a security finding is always at least optional). The extension renders these values directly, so there is no separate security severity scale and no conversion step.

Map the source CVE/CWE severity by meaning to the nearest value (case-insensitive):

• When merging duplicate findings by id, keep the strongest severity (mandatory > potential > optional).
• Security findings live in report.security[] only; there is no summary object to key by severity. validate checks

each finding's severity is one of the three values and that findings are unique by id.

• Include the security domain in metadata.domains whenever report.security[] is non-empty (and only then).

Rule classification

In this schema a rule is classified with fields on the rule object, not with labels. Every rules{} entry is required to carry a domain and a category (the validate command enforces both, and the schema rejects a rule that is missing them).

Choosing `category` (and the source label). Each mapped rule also has a sourceCategory in the solution mapping, which you get from rules-for-solution <solutionId> (rules[].sourceCategory). The UI builds the group heading as category when it equals sourceCategory (or sourceCategory is empty / null), otherwise as category (sourceCategory). So:

• When a rule's sourceCategory is non-null, set its category to exactly that value — e.g.

mi-postgresql → rule azure-database-postgresql-02000, sourceCategory: "postgresql" → category: "postgresql", which renders as one clean Postgresql heading (mismatching it, e.g. category: "database", would render the doubled Database (Postgresql)).

• When sourceCategory is `null` (e.g. the JDK-upgrade rules), choose a sensible category yourself

(e.g. upgrade) — it renders as-is.

Labels are optional. rules{}.labels and incidents[].labels are optional free-form string[]s in the schema. Classification no longer depends on them, so you normally leave them out. The engine may still emit context labels (target=, os=, capability=) on native reports, but when hand-authoring a CSV conversion you do not need any label to make content render — the required domain/category fields do that.