ccpe-system/requirements/skills-vault/2026-06-19-routing-behavior...

127 lines
3.4 KiB
Markdown

# Skills-Vault Request: Routing Behavior Diff Audit
Date: 2026-06-19
Requested supplier repository:
```text
C:\Users\wangq\Documents\Codex\skills-vault
```
Requesting CCPE artifacts:
```text
agents/agent-specs/ccra-local-reviewer.agent.md
runtimes/hybrid/ccra-local-review.runtime.md
```
## 1. Capability Name
```text
routing-behavior-diff-audit
```
## 2. Product / Runtime Context
Local CCRA review must detect non-target behavior changes after selector, routing, QPI, or model-governance rule changes.
The first consumer is The Mindscape of Bro Tsong selector and regression workflow, but the operation should remain reusable for any rules-based routing system.
## 3. Repeated Action Or Friction Point
After a routing patch, humans often inspect only targeted cases. Local CCRA needs a full before/after behavior diff so it can identify collateral changes, unexpected no-call changes, category shifts, and cases that require calibration discussion.
## 4. Inputs
Expected input:
```text
before_results_path: file or directory containing prior routing outputs
after_results_path: file or directory containing post-change routing outputs
case_id_field: configurable field name
route_field: configurable field name
optional_expected_route_field: configurable field name
```
The Skill should support JSON and Markdown summary inputs when feasible, but JSON should be the first stable target.
## 5. Expected Outputs
Write or return:
```text
total_cases_compared
unchanged_cases
changed_cases
targeted_changes
non_target_changes
new_failures
resolved_failures
missing_before_cases
missing_after_cases
diff_table
machine_readable_summary
```
Preferred output files when called by a runtime:
```text
routing-behavior-diff.md
routing-behavior-diff.json
```
## 6. Allowed Actions
- Read specified before/after result files.
- Compare case IDs and routing classifications.
- Write diff reports to the caller-specified review run directory.
## 7. Forbidden Actions
- Do not edit selector rules.
- Do not edit regression cases.
- Do not update expected labels.
- Do not decide whether a patch should close.
- Do not infer product meaning beyond the deterministic diff.
## 8. Validation Expectation
The Skill should include tests for:
```text
unchanged full set
targeted route changes
non-target collateral changes
case missing in before
case missing in after
changed no-call status
UTF-8 paths and Chinese filenames
```
## 9. Reusability Rationale
This is a deterministic comparison operation. CCRA judgment about whether a change is acceptable belongs in the Agent / Runtime layer; the diff engine belongs in `skills-vault`.
Example non-CCRA consumers:
```text
knowledge-vault:
Compare topic routing or source-block classification before and after rule changes.
writing-workbench:
Compare reviewer assignment, source routing, or section classification after workflow rule changes.
video-workbench:
Compare scene, asset, or script-beat routing after production taxonomy changes.
ccpe-system:
Compare Agent, Skill, Runtime, or model-index classification results after taxonomy or rubric updates.
development repositories:
Compare issue triage, intent classification, feature-flag routing, or rules-based dispatcher behavior before and after patches.
```
## 10. Current Blocked Task
Local CCRA can manually read reports, but full collateral behavior detection should become a reusable automation Skill before formal high-volume use.