Reconciliation Break Tracer
Trace ledger breaks to source entries and state the operational root cause.
The skill explicitly says it diagnoses rather than posts adjustments, and limits writing adjustments to a resolver, reducing direct external-effect risk. However, it reads sensitive GL and subledger data without documenting least privilege, user confirmation, data-flow disclosure, access control, retention, or rollback; the boundaries of actions such as suppress and adjust also depend on external controls, so points were deducted.
The trace steps, comparison fields, cause statements, and JSON schema are internally coherent. However, the skill depends on internal-gl and subledger MCPs and provides no availability checks, missing-field handling, abnormal-input behavior, timeout or ambiguous-match handling, or diagnostic failure messages. Under static calibration it cannot exceed 10, so it receives 8.
The audience, input of one already-classified break, and output fields are reasonably clear, and invocation is scoped to after gl-recon classification. It does not define false-trigger or non-fit cases, handling for non-unique matches, Chinese-language output, or mainland-China environment fit, so points were deducted.
The document has concise structure covering invocation context, procedure, cause format, examples, output schema, and a scope limitation. The repository supplies Apache-2.0 licensing, official provenance, and general CI validation, but the skill lacks its own versioning, changelog, maintenance owner, installation/dependency notes, and FAQ, so it is not full marks.
The skill directly addresses tracing a reconciliation break to source transactions and produces actionable owner, clear-date, and action fields. Nevertheless, there are no verified real outputs or key-path tests; expected clear dates and root-cause conclusions require review and depend on MCP matching and institution-specific rules. Static calibration caps this at 7, so it receives 6.
The SKILL.md provides auditable method and output requirements, while the repository includes general plugin validation, secret scanning, and read-only CI permissions. There is no skill-specific test suite, real execution evidence, MCP response sample, or cross-source corroboration, so the conclusion rests mainly on one document and receives 4.
- Confirm least-privilege access, retention, and audit controls for the internal-gl and subledger MCPs before use, because the workflow handles sensitive financial data.
- Define explicit stop and escalation behavior for missing fields, ambiguous matches, unavailable MCPs, and indeterminate clear dates.
- Treat suppress, adjust, and raise-ticket as workflow recommendations requiring independent authorization and review; the skill itself provides no rollback or approval control.
- Do not use the root-cause sentence or expected_clear_date for ledger adjustments or close decisions without human verification.
What it does & when to use it
This Agent Skill investigates the root cause of a general-ledger reconciliation break. It takes one break row classified by gl-recon, retrieves the related GL posting and subledger transaction, and compares their key attributes. It returns a root-cause statement, owner, expected clear date, and recommended action. It diagnoses breaks only and does not post adjustments.
Reads a single break row containing the key, GL values, subledger values, bucket, and likely cause; calls the internal-gl MCP to retrieve the journal entry or posting behind the GL line, including entry ID, posting date, source system, batch ID, and preparer; calls the subledger MCP to retrieve the matching transaction, including trade ID, trade and settlement dates, counterparty, source feed, and FX rate; compares posting date, FX rate and date, account mapping, quantity sign, and amount sign; produces one root-cause sentence and returns the owner, expected_clear_date, and action.
- A reconciliation analyst has a break classified by gl-recon and needs to trace both sides to source records.
- An accounting team needs to distinguish a timing, FX, mapping, or duplicate-posting issue.
- A reference-data team needs a documented diagnosis for a security-to-GL-account mapping mismatch.
- An operations team suspects a duplicated subledger trade and needs a suppression recommendation.
Pros & cons
- Provides a concrete two-sided trace from ledger lines to source records.
- Uses explicit field comparisons to isolate timing, FX, mapping, and duplicate-posting causes.
- Returns structured ownership, clearance-date, and action fields.
- Clearly separates diagnosis from adjustment posting.
- Requires access to both the internal-gl MCP and the subledger MCP; configuration and permissions are not documented.
- Depends on gl-recon classification and is not a replacement for that preceding workflow.
- Processes one break row at a time; batch capability is not described.
- The source provides no test suite, platform validation results, or retry behavior.
How to install
The Skill is located at plugins/agent-plugins/gl-reconciler/skills/break-trace/ in the repository. In Cowork, open Settings → Plugins → Add plugin, paste https://github.com/anthropics/financial-services, and select the relevant agent or vertical. In Claude Code, run claude plugin marketplace add anthropics/financial-services, then claude plugin install gl-reconciler@claude-for-financial-services. The README does not document standalone installation of this leaf Skill.
How to use
After gl-recon classifies a break, provide one break row with its key, GL values, subledger values, bucket, and likely cause, then request source tracing, attribute comparison, and the root-cause JSON output. The source does not document an exact invocation or trigger syntax.