Fix Dev Reviews

Fix Developer Reviews Skill (apply human review findings to an EARS file)

You are closing the gap between what a developer found by reviewing an EARS file against the codebase and the EARS specification itself. After /springboot-to-ears produced the spec, a developer read it next to the source and recorded findings the extraction missed — directly in the EARS file, inside two dedicated sections:

• ## Review by Developer (code) — backticked source-code pointers to things that exist in the code but are missing from the spec (file, file#method, or file:line-range), each with an optional note and an optional → Module: placement override. A code entry asserts either that a rule is missing (→ a new EARS statement) or that a whole entity / enum-state is missing (→ a new ### <Name> block in Domain Entities / Domain Concepts, plus any rule-bearing fields as statements).
• ## Review by Developer (business requirements) — business decisions/requirements the developer wants captured that need a human answer (no code expected); optional | context: pointer and optional → Module: override.

This skill reads those two sections, investigates each unchecked entry, proposes resolutions, and — only after the user approves — applies them and marks each entry done. The EARS statements stay code-free (springboot-to-ears Rule 2): no inline citations, no class/method names in statement prose. File paths do, however, legitimately appear in the body as the per-heading > **Source files:** blockquote annotation that every ## Module: and content ### Subsection carries (springboot-to-ears Rule 14) — this skill preserves those annotations, consults them as a supplemental locate hint, and keeps them accurate when it inserts new content. The developer's source pointers themselves live only in the review sections (developer input) and in this skill's report file.

This skill does two tasks in two gated phases:

1. Find — parse both review sections, investigate each unchecked entry against the code (driven by its pointer), and write a DEV-REVIEW REPORT proposing a resolution per entry. Then STOP and wait for the user.
2. Fix — only on the user's approval, apply the approved entries to the EARS file (confident rule findings as new code-free EARS statements; confident structural findings as a new ### <Name> block — full property/state table — in Domain Entities / Domain Concepts, plus any rule-bearing fields as statements; ambiguous findings and all business-requirement entries as [NEEDS REVIEW] Open Questions), check each applied entry off (- [ ] → - [x]) with a → resolved: annotation that names the landing location and the EARS line number, then refresh the Extraction Summary.

Scope is developer-seeded only. This skill acts on the entries the developer wrote in the two review sections — nothing else. It does not re-scan the whole controller to discover new gaps (that is ears-gap-fix), and it does not report orphan EARS statements or value mismatches.

Companion skills (all in this project's .claude/skills/):

• springboot-to-ears — produced the EARS file and emits the two empty review-section placeholders. This skill reuses its output rules by reference; read .claude/skills/springboot-to-ears/SKILL.md for the cited Rules/Steps/Procedures.
• ears-gap-fix — the agent-discovered gap closer. This skill reuses its Step 1 gates and its Step 2a large-file extraction protocol by reference; read .claude/skills/ears-gap-fix/SKILL.md.
• resolve-open-questions — the breadcrumb resolver. Run it after this skill's Phase 2 to chase any new [NEEDS REVIEW] Open Questions the business-requirement entries produced. This skill reuses its Extraction-Summary refresh mechanics (its Step 6) by reference.

Never read or rely on skills outside this repository.

Step 1 — Invocation Contract & Validation Gates

Invocation:

/fix-dev-reviews <path-to-EARS-file.md> [<Controller.java> ...]

The first (and only required) argument is the EARS file. Any trailing arguments are controller source files passed as an optional override for the controller set (used only to scope smart auto-locate, Step 3); when supplied, they replace the header-parsed set entirely.

Resolve the controller set exactly as ears-gap-fix Step 1 does: if controller args were passed, that list is the set; otherwise read the > **Source Controller(s):** line in the EARS header preamble and extract every backtick-wrapped project-relative path. Never scan the repo to guess a controller.

Validate before doing any work and exit early with the matching message:

When the gates pass, read the EARS file end-to-end once so you know which ## Module: / ### subsections exist (including the ## Domain Entities and Properties and ## Domain Concepts and States reference sections and the ### <Name> blocks already under them) and where the two review sections sit. Note that each ## Module: heading and every content ### Subsection carries a > **Source files:** blockquote line (springboot-to-ears Rule 14); record those lists — they are a supplemental locate hint (Step 3), a placement cross-check (Step 3), and the annotation-upkeep target (Step 5). Derive <Domain> from the EARS title (e.g. # Loan Proposal — OTC … (EARS) → Loan Proposal — OTC).

Step 2 — Phase 1: Parse Both Review Sections

Within each review section, every list item matching ^- \[[ x]\] is one entry. Skip every entry already marked - [x] (done on a prior run — do not re-investigate). For each unchecked (- [ ]) entry, build a record:

Multi-line entries are allowed (the developer may wrap a note); the entry runs until the next - [ ] / - [x] line or the next ## heading. Record entries in document order. Number them DR-001, DR-002, … for the report.

A code entry asserts one of two things; classify each by note wording with a class-shape fallback:

• rule (default) — "a real rule lives here and is missing from the spec." Produces a new EARS statement.
• structural — "a whole entity or enum/state is missing from the spec." Treat a code entry as structural when either the note signals a missing entity / model / DTO or enum / state / concept (words like "entity", "model", "DTO", "enum", "status", "states", "add … with its fields/properties"), or the pointer resolves to a class whose shape is structural — a JPA @Entity, a *Dto/record/POJO that is just fields + accessors, or an enum / state type — and no ### <Name> block for it already exists under ## Domain Entities and Properties / ## Domain Concepts and States. Sub-type it entity (data class / @Entity / *Dto) or concept (enum / state type). If a matching ### <Name> block already exists, it is not structural — handle it as a rule entry (a missing field/rule on an existing block), never a duplicate block.

A business entry asserts "this decision/requirement must be captured", is never structural (structure is code-discoverable), and is always destined for a [NEEDS REVIEW] Open Question.

Step 3 — Phase 1: Investigate Each Entry

Drive the investigation off the entry's pointer. Reuse ears-gap-fix Step 2a (the large-file structural-index + rule-site-ledger protocol) and the springboot-to-ears "named-not-read" prohibition (Step 2i) — do not invent a parallel method.

For each entry capture, where applicable: the exact file:line(-range), the precise code condition / trigger, the verbatim i18n message or string literal it produces (Rule 5), and the exact reference entity/field that supplies the comparison value (Rule 4 — "loan product details" ≠ "loan product policy"). These are the anchors for the proposed statement and the report's proof.

For a structural entry, investigate the whole class instead of a single rule-site. Resolve and read the pointed class in full (reuse ears-gap-fix Step 2a large-file protocol if it is large; obey the springboot-to-ears "named-not-read" prohibition — never describe a field you did not read). Capture every persisted/serialised field (entity) or every value/state and transition (enum/state). Map each to business terms per springboot-to-ears Rule 4 (code-free naming): property names like "Proposed loan amount" — never proposedLoanAmount; Types in business terms (Amount / Text / Number / Boolean / Date / Reference / Enum / List), never Java types. Exclude the shared audit, soft-delete, and optimistic-locking fields — they are described once under Cross-Cutting → Audit and Record Lifecycle and are deliberately not repeated per entity (the Domain Entities section preamble states this). While reading, also note any rule-bearing field — a default, a required/optional constraint, a computed/derived value, an enum-state default — for the derived-statements step.

Classify each entry:

• confident (rule code entries) — the rule and its values/messages are unambiguous from the source. You can write a finished, code-free EARS statement now. Choose the EARS pattern by springboot-to-ears Step 4 priority (Event-Driven > State-Driven > Unwanted Behaviour > Optional

  Ubiquitous > Complex); most validator rejections are Unwanted Behaviour (If …, the <Domain> system shall reject …). Quote any i18n message verbatim. No IDs, no bullets, no code references (Rule 2).

• confident structural (structural code entries) — the class is located and its shape is unambiguous. You can build a finished ### <Name> block now (full property/state table) plus any rule-bearing fields as code-free statements. See the structural build/place rules below.
• ambiguous — a rule code entry whose value/intent cannot be pinned from code alone (config/profile-driven, seeded in Flyway/DB, depends on an external constant not locatable after applying the Rule 9 gates), a structural entry whose class cannot be located or whose meaning cannot be pinned, or any business entry. These become a [NEEDS REVIEW] Open Question with both Rule 8 footers. For a business entry the *Where agent looked:* footer records whatever context the | context: pointer yielded; if there was no pointer, it records that the entry is a developer-supplied business decision with no code source.

Build the structural block (confident structural entries). Follow the reference-file shapes exactly; each block is a content ### subsection, so it must carry its own > **Source files:** annotation (Rule 14):

• entity → target ## Domain Entities and Properties: ### <Entity> → > **Source files:** <backticked path(s)> → a one-line description → | Property | Type | Meaning | Constraints | table covering all properties (audit/soft-delete/ version fields excluded).
• concept → target ## Domain Concepts and States: ### <Concept> → > **Source files:** <backticked path(s)> → an optional one-line description → **<Label>:** + | <Name> | Business Meaning | table, plus a | From State | To State | Triggered by | transitions table where the type encodes transitions.

Derive rule-bearing fields into statements. For every rule-bearing field noted during investigation, additionally write a code-free EARS statement (Rule 2 prose, valid pattern, i18n quoted verbatim) and place it in the correct module by Rule 7. Pure data-shape fields contribute only their table row — no statement.

Resolve each entry's target location.

• rule / confident → if the entry has an override (→ Module:), use it — unless that module/subsection does not exist in the file, in which case fall back to the closest existing subsection and note the substitution in the report. Otherwise pick the target by springboot-to-ears Rule 7 placement (Validation Rules module for validator branches, Sub-Validators for sub-validator rules, Async for side effects, etc.); use a subsection's > **Source files:** annotation as a confirmation signal — the subsection whose listed source files include the entry's pointer file is its natural home.
• structural → the block targets ## Domain Entities and Properties (entity) or ## Domain Concepts and States (concept); its derived statements are placed by Rule 7 like any rule entry. Both sections already exist in springboot-to-ears output, so adding a ### <Name> subsection is not a new top-level section. If the relevant section is absent, downgrade the entry to ambiguous (Open Question) rather than invent it.

Never invent a new top-level section. Ambiguous/business entries target ## Open Questions.

Step 4 — Phase 1: Write the DEV-REVIEW REPORT, then STOP at the approval gate

Write the report to docs/ears/<Domain>-FIX-DEV-REVIEWS-REPORT.md (overwrite if present). Use this format:

# <Domain> — EARS Developer-Review Report

Source EARS: <ears-path>
Controllers in scope: <controller path(s), comma-separated>
Generated: <Day, DD Month YYYY>

## Summary
- Entries found (unchecked): <N>   (code-rule: <Nr> · code-structural: <Ns> · business: <Nb>)
- Confident rule (→ EARS statement): <Gc>
- Confident structural (→ entity/concept block): <Gs>
- Ambiguous / business (→ Open Question): <Ga>

---

## DR-001  [code · rule · confident]
Entry: `BufferLoanProposalValidatorImpl.java#checkTenure` — tenure exceeds product details max not captured
Target: Module: <…>  ›  ### <…>     (developer override / agent-chosen)
Source: src/main/java/.../BufferLoanProposalValidatorImpl.java:1820-1834
Code condition: if (dto.getTenure() > details.getMaxTenure()) throwError("<verbatim message>")
Proposed EARS statement:
> If the requested loan tenure exceeds the maximum tenure configured on the loan product details,
> the <Domain> system shall reject the request with the message "<verbatim message>".

## DR-002  [code · entity · confident]
Entry: `GuarantorDto.java` — Guarantor entity missing from the spec
Target: ## Domain Entities and Properties  ›  ### Guarantor   (new block)
Source: lib/shared-dto/src/main/java/smartmf/shared/dto/GuarantorDto.java
Proposed entity block:
> ### Guarantor
> > **Source files:** `lib/shared-dto/src/main/java/smartmf/shared/dto/GuarantorDto.java`
> A guarantor attached to the loan proposal from the member's record.
> | Property | Type | Meaning | Constraints |
> |----------|------|---------|-------------|
> | <… every property, audit/version fields excluded …> | | | |
Proposed derived statement(s) (rule-bearing fields only):
> When a new loan proposal is submitted, the <Domain> system shall … → Module: <…> › ### <…>

## DR-003  [business · open question]
Entry: Should auto-disbursement be blocked for migration loans during month-end freeze?
Target: Open Questions
Why open: developer-supplied business decision; no single code source determines the intended policy.
Proposed Open Question:
> N. **[NEEDS REVIEW]** <concrete question a domain expert can answer in 1–2 sentences>
>    *Where agent looked:* `<context-file>:<line>` — <what was read> (or: developer-supplied business decision, no code source).
>    *Hint for reviewer:* Likely answer in `<probable/file>` (<reasoning>). Try `<grep/find/graphify command>`.

…

Then print a stdout summary (≤ 12 lines): the entry counts (code-rule vs code-structural vs business), and the confident-rule / confident-structural / ambiguous split. Do not edit the EARS file in this phase. End the turn by asking the user to review docs/ears/<Domain>-FIX-DEV-REVIEWS-REPORT.md and reply with exactly one of:

• apply all — apply every entry in the report;
• apply DR-003 DR-007 … — apply only the listed entries;
• abort — make no changes.

Phase 2 runs only after the user gives one of these in the conversation.

Step 5 — Phase 2: Apply the Approved Entries

Run only when the user has approved (all or a subset). Apply only approved entries. Mirror the surgical editing discipline of resolve-open-questions (Step 5–6):

• confident rule entry → insert the proposed statement as a standalone prose paragraph (blank-line separated) into its target ### subsection, in springboot-to-ears Rule 2 style: complete sentence, subject the <Domain> system, no ID, no bullet, no code reference, i18n quoted verbatim. Place it logically (e.g. after related rejection paragraphs in the same subsection). Then, annotation upkeep (Rule 14): if the finding's source file is not already listed in that subsection's > **Source files:** blockquote, append it (comma-separated, backticked, project-relative, preserving existing order). Touch only the target subsection's annotation; never reorder or rewrite existing entries; never add a > **Source files:** line to a section that lacks one.
• confident structural entry → insert the proposed ### <Name> block into its target section (## Domain Entities and Properties for entity, ## Domain Concepts and States for concept), placed logically (after related blocks) or appended at the section end. The block carries its own > **Source files:** annotation as built in Step 3. Then insert each derived statement into its module exactly like a confident rule entry above, applying the same per-subsection annotation upkeep for each.
• ambiguous / business entry → append a new numbered item to ## Open Questions, continuing the existing numbering, with **[NEEDS REVIEW]** and both mandatory footers (*Where agent looked:* and *Hint for reviewer:*) exactly as springboot-to-ears Rule 8 requires, so /resolve-open-questions can act on it next. (Open-Question entries touch no > **Source files:** annotation.)
• Check off and annotate the entry (the audit feature). After the statement/block/Open Question is in place, edit the originating review-section line: flip - [ ] → - [x] and append, on a continuation line indented under it, the resolution annotation:
  • confident rule → → resolved: Module: <Module> › <Subsection> (EARS line <N>)
  • confident structural → → resolved: <Domain Entities|Domain Concepts> › <Name> (EARS line <N>) and, on further continuation lines, one → also: Module: <Module> › <Subsection> (EARS line <N>) per derived statement
  • ambiguous/business → → resolved: Open Question <N> (EARS line <M>) Compute the EARS line number after the edit lands (so it points at the inserted paragraph, block, or new Open Question). When applying several entries, apply edits one at a time and recompute line numbers so earlier insertions don't make later annotations stale. The (EARS line N) is a convenience snapshot — it may drift if the file is later hand-edited.
• Refresh ## Extraction Summary (identical mechanics to resolve-open-questions Step 6): recount EARS statements by leading word — If →Unwanted, When /After →Event-Driven, While →State-Driven, Where →Optional, The <Domain> system shall →Ubiquitous, While … when →Complex — and recount the marker rows ([NEEDS REVIEW], [DISABLED], [UNRESOLVED]). Overwrite only the right-hand count cells; the Total EARS statements row must equal the sum of the pattern rows. Update the **Extraction completed:** line to today's date in <Day, DD Month YYYY> format. If the file has no ## Extraction Summary, skip this. A new entity/concept block adds no EARS statement, so it does not move the pattern counts; its derived statements do, and are counted by their leading word like any other statement.

Print a stdout summary (≤ 12 lines): statements added (per target module), entity/concept blocks added, Open Questions added, entries checked off, entries skipped because the user did not approve them.

Step 6 — Idempotency & Edge Cases

Step 7 — Self-Review Checklist (before ending each phase)

Phase 1 (report):

• Every unchecked entry in both sections has a DR item; every - [x] entry was skipped.
• Every confident item has a file:line source proof and a resolved target location; placements were cross-checked against target subsections' > **Source files:** lists where available.
• Every confident rule item's proposed statement is code-free, uses a valid EARS pattern, and quotes any i18n message verbatim (Rules 2, 4, 5).
• Every confident structural item's proposed block covers all properties/states (audit/ soft-delete/version fields excluded), carries a > **Source files:** annotation, names a valid target section, and lists any rule-bearing fields as derived statements.
• Every ambiguous/business item's proposed Open Question carries both Rule 8 footers.
• No new top-level section was proposed; the EARS file was not modified; the report was written; the approval prompt was shown.

Phase 2 (fix):

• Only user-approved entries were applied.
• New statements are plain paragraphs in existing subsections (no IDs/bullets/code refs); new entity/concept blocks match the reference shape and carry their > **Source files:** line; new Open Questions are well-formed with both footers.
• Every applied statement's source file is present in its target subsection's > **Source files:** line (appended if it was missing); no other > **Source files:** annotation was modified, removed, or reordered.
• Each structural block covers all properties/states with audit/soft-delete/version fields excluded; its derived statements are code-free, rule-bearing only, and correctly placed.
• Every applied entry was flipped to - [x] and annotated with → resolved: … and an accurate (EARS line N) computed after the edit (structural entries also name each derived statement's location).
• No new top-level sections were invented; the two review sections themselves were not restructured (only entries checked off + annotated).
• ## Extraction Summary counts sum correctly and **Extraction completed:** is today's date.
• The diff is additive and surgical — only new statements/blocks/Open Questions, the source-file annotation append, the entry check-off + annotation, and the Summary refresh changed.