git:20260613.4b1ceb3 to git:20260614.5208b4c

50 added, 12 removed. Audit A to A.

---

name: audit-xls

+ multi_ticker_semantics: single_target

description: Audit spreadsheet, formula error detection, hardcoded cell finder, cross-sheet reference audit, workbook auditor, Excel model audit, financial model QA, spreadsheet review, cell dependency trace, formula integrity check

temporal_scope:

default_quarters: 1

max_quarters: 1

description: "Typical lookback: 1 quarters, max: 1"

allowed_tools:

- search_companies

- get_company_financials

- get_calculation_tree

- validate_calculation

- list_sources

retrieval_scope: simple_lookup

min_tool_diversity: 3

---

# audit-xls

- **Agent Call Tracing**: The first tool you call will return a `_run_id` in its result. On every subsequent tool call, include HTTP header `X-Agentii-Trace: agent={skill_name}; parent={caller_name}; instance={instance_label}`. The MCP server will inject run_id, depth, and user_id automatically. When spawning parallel sub-agents of the same type, assign each a unique instance label (e.g., equity-research-1, equity-research-2). See `contracts/x-agentii-trace-header.md` for the full contract.

## Triggers

- Audit spreadsheet

- formula error detection

- hardcoded cell finder

- cross-sheet reference audit

- workbook auditor

- Excel model audit

- financial model QA

- spreadsheet review

- cell dependency trace

- formula integrity check

## Defaults

| Parameter | Default Value | Rationale |

|-----------|---------------|-----------|

| ticker | (required) | Stock symbol to analyze |

| lookback_quarters | 1 | Standard lookback for this skill type |

## Methodology

- This skill follows the agentii retrieval protocol. Retrieval scope: **simple_lookup**. Minimum tool diversity: 3 distinct tools.

+ ### 1. Retrieval Scope

- **Calculation arc audit pattern **: When auditing a financial model workbook, this skill MUST cross-validate workbook computed totals against the XBRL calculation linkbase. The audit pattern:

+ This skill operates with `retrieval_scope: simple_lookup`. It uses only profile/entity metadata tools — no document or XBRL retrieval at scale.

- 1. **Retrieve calculation tree**: call `get_calculation_tree/{ticker}?statement_type=<income_statement|balance_sheet|cash_flow>` to get the expected parent-child formula relationships from `gold.xbrl_calculations` (756K rows). Each edge carries a `weight` (+1.0 for additive, -1.0 for subtractive).

- 2. **Reconstruct expected totals**: for each parent concept, compute `expected_value = Σ(child_value × weight)` per the XBRL calculation linkbase.

- 3. **Compare workbook vs expected**: for each parent-child relationship in the workbook, compare the workbook's computed total against the XBRL-expected total. Flag discrepancies ≥1% of the parent concept value as audit findings.

- 4. **Cross-validate with `validate_calculation/{ticker}`**: the endpoint performs automated calculation validation against `gold.xbrl_calculations` — use it as a second pass to catch discrepancies the workbook audit may have missed.

+ ### 2. Retrieval Strategy

- Discrepancies are categorized: (a) **material** (≥5% of parent value) — refuse delivery; (b) **minor** (1-5%) — flag in audit findings with `severity: warning`; (c) **negligible** (<1%) — note in audit trail but do not block delivery.

+ Follows the retrieval strategy decision tree in `retrieval.md`. Primary branch: **(d) Simple Lookup**. Resolve the canonical ticker first (exact → fuzzy alias → share-class) before any data call.

+ ### 3. Temporal Scope

+ Default lookback: 1 fiscal quarter(s); maximum: 1. The default balances recency against the trend window this analysis requires.

+ ### 4. Tool Allowlist

+ Per frontmatter `allowed_tools`:

+ - `search_companies` — ticker resolution + company context (entity-alias fuzzy match)

+ - `get_company_financials` — consolidated IS/BS/CF highlights

+ - `get_calculation_tree` — XBRL calculation linkbase (weights)

+ - `validate_calculation` — XBRL calc-consistency validation

+ - `list_sources` — used by this skill per the retrieval strategy

+ ### 5. Protocol

+ 1. **Pre-flight**: `get_company_fiscal_calendar/{ticker}` then `get_ticker_coverage/{ticker}`.

+ 2. **Lookup**: `get_company_profile/{ticker}` / `get_entity_knowledge` for the requested metadata field(s).

+ 3. **Output**: write the deliverable per `## Output File`, then append to `agentii.md`.

+ ## Deliverable Chain

+ **Inputs** → **Build** → **Validate** → **Output** → **Next**

+ 1. **Inputs**: resolved ticker + structured facts (`search_xbrl_facts`, `get_company_financials`) and any filing pages from the three-layer protocol.

+ 2. **Build**: assemble the workbook/deck spec and call `xlsx.build` / `pptx.build` (office plane).

+ 3. **Validate**: run `xlsx.audit` (or recalc) and the `## Validation Gates` below.

+ 4. **Output**: write the artifact path per `## Output File`.

+ 5. **Next**: append to `agentii.md`; hand off to a downstream pitch/review skill if requested.

## Validation Gates

1. **calculation arc cross-validation **: workbook computed totals verified against `gold.xbrl_calculations` weights. Compare `get_calculation_tree/{ticker}` expected values against workbook formulas. Flag discrepancies ≥1% of parent concept value. *If failed*: If material discrepancy (≥5%): refuse delivery. If minor (1-5%): flag in audit findings with `severity: warning`.

2. **hardcoded cell detection**: zero hardcoded values in cells tagged as formulas. Use `xlsx_audit` hardcoded-count output. *If failed*: If hardcoded_count > 0: refuse delivery with audit report listing each hardcoded cell location.

3. **cross-sheet reference integrity**: all cross-sheet references resolve to valid cell ranges. *If failed*: If broken references found: refuse delivery with broken reference map.

4. **tool diversity**: distinct MCP tools used in this invocation >= `min_tool_diversity` (3). *If failed*: flag as depth-insufficient in Coverage Gaps.

## Output File

Write the final deliverable to `{ticker}/{{YYYY-MM-DD_HHMM}}_audit-xls_{{affix}}.md` .

## Output Structure

- 1. Executive Summary

- 2. Data Sources (with agentii://source/ citation watermarks)

- 3. Analysis Results

- 4. Coverage Gaps (if any)

+ Write to `{ticker}/{YYYY-MM-DD_HHMM}_audit-xls_{affix}.md` (see `## Output File`).

+ 1. **Executive Summary** (≤200 words) — headline conclusions for the analysis.

+ 2. **Data Sources** — filings + structured endpoints used, with `{ticker} {citation_id} page<N>` citations.

+ 3. **Analysis** — the core findings, tables, and commentary for this dimension.

+ 4. **Key Metrics** — the quantitative results with QoQ/YoY context where relevant.

+ 5. **Coverage Gaps & Citations** — data not retrievable + citation index.

+ **Citation density**: ≥1 citation per 200 words; bare `page_no` integers are forbidden — always cite as `{ticker} {citation_id} page<N>`. After writing, append a YAML block to `agentii.md` per `contracts/agentii-md-schema.md`.

+ ## Preflight

+ !curl -s -o /dev/null -w "%{http_code}" --max-time 2 https://mcp.agentii.ai/mcp/health 2>/dev/null || echo "UNREACHABLE"

+ **Agent Call Tracing**: The first tool you call will return a `_run_id` in its result. On every subsequent tool call, include HTTP header `X-Agentii-Trace: agent={skill_name}; parent={caller_name}; instance={instance_label}`. The MCP server will inject run_id, depth, and user_id automatically. When spawning parallel sub-agents of the same type, assign each a unique instance label (e.g., equity-research-1, equity-research-2). See `contracts/x-agentii-trace-header.md` for the full contract.

## Error Handling

| Error | Action |

|-------|--------|

| Ticker not found | Suggest checking spelling or trying list_coverage |

| No data available | Flag in Coverage Gaps, proceed with available data |

| API key invalid | Direct user to agentii.ai/api-keys |

| MCP server unreachable | Retry once; if persistent, halt with AGENTII_MCP_UNREACHABLE |