Ears Gap Fix

EARS Gap-Fix Skill (find + fix missing-from-EARS business rules)

You are closing the coverage gap between a Spring Boot controller's source code and an EARS specification file that was reverse-engineered from it by /springboot-to-ears. The EARS file is deliberately code-free in its body (no inline citations, no source-file index), but its header preamble records the controller(s) it was generated from in a > **Source Controller(s):** line. This skill reads that header to recover the controller path(s) and re-walks the code to find every business rule — and every data-model element — the EARS file failed to capture. (The user may still pass controller path(s) at invocation to override the header.)

The input EARS file follows the current /springboot-to-ears output format: every ## Module: heading and content ### Subsection carries a > **Source files:** blockquote annotation (Rule 14); the data model lives in dedicated ## Domain Entities and Properties and ## Domain Concepts and States reference sections (each a set of ### <Name> blocks); the file ends with a richer ## Extraction Summary, a ## Business Rules Summary, and the two ## Review by Developer sections that belong to fix-dev-reviews. The EARS statements stay code-free — file paths appear in the body only inside > **Source files:** annotations and Open-Question footers.

This skill does two tasks in two gated phases:

1. Find — scan the controller's full call chain, build an inventory of every business rule and every catalog element (entity property, enum state), diff it against the EARS prose and the Domain Entities/Concepts blocks, and write a GAP REPORT of items present in code but missing from the spec. Then STOP and wait for the user.
2. Fix — only on the user's approval, apply the approved gaps to the EARS file (confident rules as new code-free EARS statements; confident structural gaps as a new/extended ### <Name> block plus rule-bearing fields as statements; ambiguous ones as [NEEDS REVIEW] Open Questions), then refresh the Extraction Summary.

Scope is missing-only. This skill reports rules and data-model elements that are in the code but have no corresponding EARS statement or catalog row. It does not report "orphan" EARS statements (in spec, not in code) and does not report value "mismatches" (a statement that exists but cites a wrong constant). Those are explicitly out of scope.

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

• springboot-to-ears — produced the EARS file. This skill reuses its extraction rules by reference. Read .claude/skills/springboot-to-ears/SKILL.md for the cited Steps/Rules/Procedures. Note: that skill now auto-resolves its own Open Questions on the first pass (its Step 6.5), so a fresh EARS file usually arrives here with only genuinely-unresolvable [NEEDS REVIEW] residue.
• resolve-open-questions — the manual/later-pass resolver. It consumes the [NEEDS REVIEW] Open Questions this skill adds (and any human-added breadcrumbs); run it after this skill's Phase 2 to chase the new questions.
• fix-dev-reviews — the developer-seeded counterpart. Where this skill discovers missing rules and data-model elements by re-walking the controller, fix-dev-reviews applies findings a developer recorded by hand in the spec's two ## Review by Developer sections. It reuses this skill's Step 1 gates and Step 2a large-file protocol by reference; this skill reciprocally mirrors its structural ### <Name> block building (its Step 3) and its Rule 14 > **Source files:** annotation upkeep (its Step 5). This skill never reads, consumes, or modifies the two ## Review by Developer sections — those belong to fix-dev-reviews.

Never read or rely on skills outside this repository.

Step 1 — Invocation Contract & Validation Gates

Invocation:

/ears-gap-fix <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; when supplied, they replace the header-parsed set entirely. Validate before doing any work and exit early with the matching message. Do not auto-discover or guess a controller when none can be resolved.

Resolve the controller set (after the EARS file/.md gates pass, before scanning):

1. If the user passed one or more controller arguments, that arg list is the controller set — use it verbatim and ignore the header.
2. Otherwise, locate the > **Source Controller(s):** line in the EARS header preamble and extract every backtick-wrapped, comma-separated project-relative path. That list is the controller set.
3. Never scan the repo to guess a controller. If the header line is absent (and no args were given), or any resolved path does not exist on disk, exit with the matching message below.

When the gates pass, read the EARS file end-to-end once so you know exactly which rules it already states and which ## Module: / ### subsections exist. While reading, also record:

• the ## Domain Entities and Properties and ## Domain Concepts and States reference sections and the ### <Name> blocks already under them (and which properties / states each block lists) — this is the baseline for structural gap detection (Step 3);
• every > **Source files:** blockquote annotation on each ## Module: heading, content ### Subsection, and ### <Name> reference block (springboot-to-ears Rule 14) — these are a supplemental locate hint (Step 2/3), a placement cross-check (Step 3), and the annotation-upkeep target (Step 5);
• the location of the two ## Review by Developer sections — this skill never reads their entries or modifies them (it re-walks the controller; it does not consume developer-seeded findings).

Derive <Domain> from the EARS title (e.g. # Loan Proposal — OTC Business Requirements Specification (EARS) → Loan Proposal — OTC).

Step 2 — Phase 1: Build the Expected-Requirement Inventory (the scan)

Trace each resolved controller's complete call chain and enumerate every business rule the system enforces. Reuse springboot-to-ears methodology — do not invent a parallel one. Apply, by reference, against the resolved controller set (Step 1) only — not their siblings:

• Step 0 orientation, for the parts that lie in this controller's call chain only: abstract base entities (Step 0b), @ControllerAdvice handlers (0c), aspects (0d), filters (0e), security config + access-control library (0f, Procedure D), i18n message files (0h), the full enum scan (0l), and ID/number generators (0m).
• Step 2a–2i to follow endpoints → services (read the *Impl, not the interface) → lib validators (Procedure A) → injected sub-validators (recursively, Procedure A-4) → JPA repos / JDBC DAOs (Procedure C / Step 2d) → row mappers → Feign/ERP/external calls → ApplicationEvents → DTO field-level constraints (Step 2h).
• graphify (a graphify-out/ graph exists): run graphify query "<concept>" / graphify path "<A>" "<B>" to confirm call edges cheaply before deep-reading a file, so you follow the real chain and don't miss a delegate.
• The per-heading > **Source files:** annotations the spec already carries (Rule 14) are a cheap cross-check: a module/subsection's listed files tell you which source feeds which section, which both confirms your chain-walk reached the right files and pre-resolves where a found gap belongs (Step 3 placement). Treat them as a supplemental hint, never as the authoritative chain.
• The catalog baseline: the spec's ## Domain Entities and Properties and ## Domain Concepts and States blocks are the data-model the extraction already captured. As you read each entity (Step 2c) and run the full enum scan (Step 0l) the call chain already requires, hold the catalog alongside so Step 3 can diff persisted fields and enum values/states against it.
• The "named-not-read" prohibition (Step 2i): if you name a method, helper, validator, sub-validator, enum, or generator, you must open and read its body before relying on it.

Inventory item shape. Each item is one atomic, rule-bearing construct — never collapse or summarize (springboot-to-ears Rule 3). For each, record:

• source — file:line (or file:start-end) where the rule lives;
• kind — one of: rejection/guard (if/else if/case that throws or returns an error), endpoint, field-constraint, state-transition, async-side-effect, integration-call, access-rule, generated-id-format, catalog-element (a persisted entity property, or an enum value/state — the unit of a structural gap, Step 3);
• condition — the exact code condition / trigger;
• message — the verbatim i18n message or string literal if the rule produces one (Rule 5);
• entity — the exact reference entity/field that supplies the comparison value (Rule 4, e.g. "loan product details" vs "loan product policy" — they are distinct records).

For a catalog-element item, also record the owning entity/enum class and the business-term name of the property or state (Rule 4: "Proposed loan amount", never proposedLoanAmount), its business type (Amount / Text / Number / Boolean / Date / Reference / Enum / List), and any rule-bearing trait (a default, a required/optional constraint, a computed/derived value, an enum-state default).

message and entity are your strongest matching anchors in Step 3, because the EARS file quotes i18n messages verbatim and names the exact reference entity.

Step 2a — Large-File Extraction Strategy (MANDATORY; files up to ~10,000 LOC)

This is the heart of the skill. The whole reason it exists is that large validators (e.g. BufferLoanProposalValidatorImpl, StandardBufferLoanProposalValidatorImpl) get partially read and rules get silently dropped. A large file must be read in full with zero skipped rules. Use this structural-index-first, ledger-verified protocol (it extends springboot-to-ears Step A-2.1, which is the floor):

i. Measure first — always, before the first Read. Run wc -l <file> and record total T. State it in your notes. Any file with T > 500 enters this protocol; T > ~1500 (the single-Read truncation point) makes it non-negotiable.

ii. Build a structural index with one cheap grep pass (no reading yet). This makes coverage deliberate instead of luck-based:

grep -nE 'public |private |protected | void | boolean |^\s*(if|else if|case|default)\b|throwError|throwErrorByMessage|throw new |return (false|null)|@(Get|Post|Put|Delete|Request|Patch)Mapping|@Scheduled|@KafkaListener' <file>

From the output build two maps, each with line numbers:

• method map — the start line of every method (the modifier/signature lines);
• rule-site map — every if / else if / case / throw / throwError line. These two maps are the file's completeness baseline.

iii. Read by method range, not by arbitrary windows. Walk the method map in order; for each method, Read from its start line to the next method's start line, plus a ~30-line overlap so a body that closes just past the boundary is still fully seen. Only when method boundaries are genuinely unclear, fall back to fixed windows — and then overlap them by ~50 lines and extend the schedule until a chunk includes line T.

iv. Recurse with no depth limit. When a method calls a private helper, read that helper fully (Procedure A-3). When it delegates to an injected sub-validator/sub-service, open that concrete *Impl and run this same protocol on it (Procedure A-4). A helper or validate() defined in a superclass in another file is a separate-file lookup — open that file and run the protocol there. "Lives beyond the read window" is never a valid reason while unread lines remain (see vi).

v. Keep a completeness ledger per file. Maintain a checklist:

• every method-map entry marked READ;
• every rule-site-map line marked either EXTRACTED → inventory item #N or EXTRACTED → not a rule (e.g. a guard that branches but never rejects/throws). The file's scan is not done until every rule-site line is accounted for and your last read reaches the class's closing brace at/after line T. Carry the ledger totals into the GAP REPORT header, e.g. BufferLoanProposalValidatorImpl: 412 rule-sites · 412 accounted · 0 unread.

vi. Hard prohibition. It is forbidden to raise a [NEEDS REVIEW] (or skip a rule) with a reason like "method body beyond the read window", "file too large to fully extract", or "could not be read in one pass" while the file still has unread lines. The truncation reminder is a paging instruction ("Call Read with offset=N…"), not a stopping point. Read more chunks.

vii. Context-safety option for the largest files. If several ~10K-line files would otherwise swamp your context, you MAY delegate one large file per subagent (Agent tool) whose sole job is to return that file's structured inventory (file:line · kind · condition · verbatim message · reference entity) plus its completeness ledger. Give the subagent this exact protocol (ii–vi) and require the ledger back so you can trust coverage without re-reading. Keep the diff-against-EARS step (Step 3) in the main agent. This is an efficiency lever only — it never licenses skipping lines, and a subagent that returns unread > 0 must be sent back to finish.

Step 3 — Phase 1: Diff the Inventory Against the EARS File (missing-only)

For each inventory item, decide whether the EARS file already expresses it. The EARS file is code-free prose, so matching is semantic, anchored on the signals springboot-to-ears guarantees, in priority order:

1. Verbatim i18n message — if the item has a message, search the EARS text for that exact quoted string. A match is near-conclusive that the rule is captured.
2. Exact reference entity + condition — match the entity (Rule 4 precision: "loan product details" ≠ "loan product policy") together with the condition direction (exceeds / below / missing / duplicate) and the response verb (reject / set status / require role).
3. Behavioural paraphrase — for rules without a message, look for a statement whose trigger and response mean the same thing, even if worded differently.

A rule inventory item with no expressing statement anywhere in the file is a MISSING gap. (If a statement exists but looks value-wrong, that is a mismatch — out of scope; do not report it. If an EARS statement has no code behind it, that is an orphan — out of scope.)

Structural diff (catalog-element items). In parallel, diff the data model the code actually defines against the spec's reference sections:

• for each entity reached, compare its persisted fields against the matching ### <Name> block in ## Domain Entities and Properties — excluding the shared audit, soft-delete, and optimistic-locking fields, which are described once under Cross-Cutting → Audit and Record Lifecycle and are deliberately not repeated per entity;
• for each enum reached, compare its values/states against the matching ### <Name> block in ## Domain Concepts and States.

A persisted field or an enum value/state with no corresponding row is a structural gap:

• if a ### <Name> block for that entity/enum already exists, the gap is a row addition to that block — never a duplicate block (mirror fix-dev-reviews Step 2);
• if no block exists for it, the gap is a new ### <Name> block — sub-type entity (→ Domain Entities) or concept (enum/state → Domain Concepts). If the target section is genuinely absent from the file, downgrade the gap to ambiguous (Open Question); never invent a top-level section.

Classify each gap, mirroring fix-dev-reviews Step 3:

• confident rule — 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 via 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 — the entity/enum is located and its shape is unambiguous. You can build a finished ### <Name> block (full property/state table) or row addition now. Any rule-bearing field (a default, a required/optional constraint, a computed/derived value, an enum-state default) also yields a derived code-free statement; pure data-shape fields contribute only their table row.
• ambiguous — the rule exists but its value/intent cannot be pinned from code alone (config- or profile-driven, seeded in Flyway/DB, depends on an external constant you could not locate after applying the Rule 9 gates), or a structural gap whose entity/enum class cannot be located or whose meaning cannot be pinned, or a structural gap whose target reference section is absent. Propose a [NEEDS REVIEW] Open Question with both Rule 8 footers.

For every gap also resolve its target location:

• a confident rule lands in the existing ## Module: and ### subsection where the statement belongs (honor springboot-to-ears Rule 7 placement — Validation Rules module for validator branches, Sub-Validators module for sub-validator rules, Async module for side effects, etc.); a cross-cutting concern lands in the matching ## Cross-Cutting Requirements subsection (note these subsections carry no > **Source files:** line — a Rule 14 exclusion). Use a subsection's > **Source files:** list as the placement confirmation signal: the subsection whose listed files include the gap's source file is its natural home;
• a confident structural gap lands in ## Domain Entities and Properties (entity) or ## Domain Concepts and States (concept); its derived statements are placed by Rule 7 like any rule gap.

Never invent a new top-level section. If no subsection fits, prefer the closest existing one and note it; do not create new module headings. ## Business Rules Summary and the two ## Review by Developer sections are not edit targets.

Step 4 — Phase 1: Write the GAP REPORT, then STOP at the approval gate

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

# <Domain> — EARS Gap Report

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

## Coverage
- Rules inventoried: <N>
- Already covered in EARS: <C>
- Missing (gaps below): <G>   (confident rule: <Gc> · confident structural: <Gs> · ambiguous: <Ga>)

## Large-file ledgers
- <File>.java (<T> LOC): <sites> rule-sites · <accounted> accounted · <unread> unread
- … one line per scanned file > 1500 LOC …

---

## GAP-001  [MISSING · rule · confident]
Target: Module: <…>  ›  ### <…>
Source: src/main/java/.../<File>.java:1820-1834
Code condition: if (dto.getX() > policyDetails.getMaxX()) throwError("<verbatim message>")
Proposed EARS statement:
> If the requested ... exceeds the configured maximum on the loan product details, the
> <Domain> system shall reject the request with the message "<verbatim message>".

## GAP-002  [MISSING · structural · confident]
Target: ## Domain Entities and Properties  ›  ### <Name>   (new block | row add)
Source: lib/shared-dto/src/main/java/.../<Name>Dto.java
Proposed entity/concept block:
> ### <Name>
> > **Source files:** `lib/shared-dto/src/main/java/.../<Name>Dto.java`
> <one-line description>
> | Property | Type | Meaning | Constraints |
> |----------|------|---------|-------------|
> | <… every property, audit/soft-delete/version fields excluded …> | | | |
Proposed derived statement(s) (rule-bearing fields only):
> When a new loan proposal is submitted, the <Domain> system shall … → Module: <…> › ### <…>

## GAP-003  [MISSING · ambiguous]
Target: Open Questions
Source: src/main/java/.../<File>.java:245
Why ambiguous: <one line — what value/intent could not be determined and which gates were tried>
Proposed Open Question:
> N. **[NEEDS REVIEW]** <concrete question a domain expert can answer in 1–2 sentences>
>    *Where agent looked:* `<file>:245` — <what was read and why insufficient>.
>    *Hint for reviewer:* Likely answer in `<probable/file>` (<reasoning>). Try `<grep/find/graphify command>`.

…

Then print a stdout summary (≤ 12 lines): the coverage numbers, the ledger totals, and the count of confident-rule vs. confident-structural vs. ambiguous gaps. Do not edit the EARS file in this phase. End the turn by asking the user to review docs/ears/<Domain>-GAP-REPORT.md and reply with exactly one of:

• apply all — apply every gap in the report;
• apply GAP-003 GAP-007 … — apply only the listed gaps;
• 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 Fixes

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

• confident rule gap → 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 gap'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 (the Cross-Cutting subsections and summary sections deliberately have none).
• confident structural gap → if the gap is a new block, insert the proposed ### <Name> block (carrying its own > **Source files:** line) 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. If the gap is a row addition to an existing block, insert just the new | Property | … | / state row into that block's table (and append the gap's source file to the block's > **Source files:** line if missing). Then insert each derived statement into its module exactly like a confident rule gap above, applying the same per-subsection annotation upkeep for each.
• ambiguous gap → 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 gaps touch no > **Source files:** annotation.)
• Refresh ## Extraction Summary (the current table format). Overwrite only the right-hand count cells / ledger lines; never restructure the table:
  • Statement pattern rows — recount EARS statements by leading word against the current row labels: When /After → Event-Driven (When+After), While → State-Driven (While), If → Unwanted Behaviour (If), Where → Optional (Where), The <Domain> system shall → Ubiquitous (The system shall), While … when → Complex (While … when) — and update Total EARS statements accordingly.
  • Marker rows — recount [NEEDS REVIEW] items, [DISABLED] items, [UNRESOLVED] items.
  • Rule-site coverage rows — refresh Rule-sites inventoried, Rule-sites covered by a statement, and Rule-sites uncovered from this run's scan inventory and the gaps just applied (a confident gap applied moves a rule-site from uncovered to covered).
  • Per-file coverage ledger — update the **Per-file coverage ledger** bullet list from this run's large-file ledgers (one line per source file > ~500 LOC scanned), so its accounted / uncovered figures reflect the post-apply state.
  • A new entity/concept block (or row addition) adds no EARS statement, so it does not move the pattern counts; its derived statements do, counted by their leading word like any other.
  • Update the **Extraction completed:** line to today's date in <Day, DD Month YYYY> format.
  • If the file has no ## Extraction Summary, skip this (do not invent one); refresh only the rows and ledger lines that actually exist.

Print a stdout summary (≤ 12 lines): statements added (per target module), entity/concept blocks (or rows) added, > **Source files:** annotations updated, Open Questions added, and gaps 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 scanned file > 1500 LOC has a ledger line with accounted == sites and unread == 0.
• Every gap has a file:line source proof and a target location.
• Every confident rule gap's proposed statement is code-free, uses a valid EARS pattern, and quotes any i18n message verbatim.
• Every confident structural gap names a valid target section, proposes either a full ### <Name> block (all properties/states, audit/soft-delete/version fields excluded, with its own > **Source files:** line) or a single row addition, and lists rule-bearing fields as derived statements.
• Every ambiguous gap's proposed Open Question carries both Rule 8 footers.
• No orphan-EARS or value-mismatch items were reported (out of scope).
• The EARS file was not modified; the report was written; the approval prompt was shown.

Phase 2 (fix):

• Only user-approved gaps 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.
• No new top-level sections were invented; the two ## Review by Developer sections were not modified.
• ## Extraction Summary reflects this run: statement pattern + marker counts, the rule-site coverage rows, and the per-file ledger were all refreshed, and **Extraction completed:** is today's date.
• The diff is additive and surgical — only new statements/blocks/rows, the source-file annotation appends, and the Summary refresh changed.