Time-Series Skill Router
Turn the prepared problem contract into a minimal execution plan. Recommend skills only; do not run modeling or detection work unless the user also asks for execution.
Routing Workflow
Validate the input
- Read
problem-summary.txt completely.
- Require
task_type, problem, decision_or_output, dataset, variable roles, data_structure, validation_plan, leakage_risks, constraints, readiness, and open_questions.
- If the file is absent, stale, contradictory, or blocked on a fact that changes routing, run or recommend
forecasting-data-prep for ordered time-series tasks or ts-classification-data-prep for classification, then resume routing with the refreshed file.
Load the available skills
- Read
skills.catalog.yaml; treat its IDs, stages, paths, dependencies, and preconditions as the source of truth.
- Filter by the documented task, data shape, required output, constraints, and readiness.
- Read each serious candidate's
SKILL.md and only the references needed to verify support. Never invent capabilities from a skill name.
Build the smallest valid chain
- Include the preparation skill that produced the summary as the first step, even when already completed.
- Select one primary skill per required stage. Do not chain equivalent libraries for completeness.
- Add feature aggregation only when the downstream method needs extracted tabular features.
- Expand
depends_on entries before their dependent skill and satisfy catalog preconditions.
- Order work as preparation, optional feature construction, primary modeling or detection, then task-specific validation. Keep leakage checks active at every stage.
Resolve ambiguity
- Prefer the candidate that directly supports the documented structure and output with the fewest adapters.
- Respect an explicit library constraint when the skill supports the task.
- When evidence cannot distinguish candidates, choose one primary recommendation and list at most two alternatives outside the execution chain with the fact that would change the choice.
Return the plan
- Use the output contract below.
- Keep reasons specific to
problem-summary.txt; avoid generic capability summaries.
- Do not proceed past a
readiness: blocked item unless the plan first resolves that blocker.
Output Contract
Problem: <one-sentence interpretation>
Readiness: <ready or blocking fact>
Recommended order:
1. <skill-id> - <why it is required and its expected handoff>
2. <skill-id> - <why it is the best fit and its expected handoff>
Validation: <split/backtest, metrics, and leakage checks that remain mandatory>
Alternatives: <none, or up to two skill IDs with a concrete switch condition>
Assumptions: <none, or unresolved non-blocking assumptions>
Reference
Read references/routing-rules.md to map task types to stages and apply the minimality rules consistently.
1---2name: time-series-skill-router3description: Read a validated problem-summary.txt and skills.catalog.yaml to recommend the smallest ordered sequence of repository skills for forecasting, classification, pattern discovery, feature aggregation, change point detection, or anomaly detection. Use this skill after a data-preparation skill has documented the problem, dataset, variables, validation plan, constraints, readiness, and leakage risks, or whenever an agent must choose and order skills from this repository without redundant library alternatives.4---56# Time-Series Skill Router78Turn the prepared problem contract into a minimal execution plan. Recommend skills only; do not run modeling or detection work unless the user also asks for execution.910## Routing Workflow11121. **Validate the input**13 - Read `problem-summary.txt` completely.14 - Require `task_type`, `problem`, `decision_or_output`, `dataset`, variable roles, `data_structure`, `validation_plan`, `leakage_risks`, `constraints`, `readiness`, and `open_questions`.15 - If the file is absent, stale, contradictory, or blocked on a fact that changes routing, run or recommend `forecasting-data-prep` for ordered time-series tasks or `ts-classification-data-prep` for classification, then resume routing with the refreshed file.16172. **Load the available skills**18 - Read `skills.catalog.yaml`; treat its IDs, stages, paths, dependencies, and preconditions as the source of truth.19 - Filter by the documented task, data shape, required output, constraints, and readiness.20 - Read each serious candidate's `SKILL.md` and only the references needed to verify support. Never invent capabilities from a skill name.21223. **Build the smallest valid chain**23 - Include the preparation skill that produced the summary as the first step, even when already completed.24 - Select one primary skill per required stage. Do not chain equivalent libraries for completeness.25 - Add feature aggregation only when the downstream method needs extracted tabular features.26 - Expand `depends_on` entries before their dependent skill and satisfy catalog preconditions.27 - Order work as preparation, optional feature construction, primary modeling or detection, then task-specific validation. Keep leakage checks active at every stage.28294. **Resolve ambiguity**30 - Prefer the candidate that directly supports the documented structure and output with the fewest adapters.31 - Respect an explicit library constraint when the skill supports the task.32 - When evidence cannot distinguish candidates, choose one primary recommendation and list at most two alternatives outside the execution chain with the fact that would change the choice.33345. **Return the plan**35 - Use the output contract below.36 - Keep reasons specific to `problem-summary.txt`; avoid generic capability summaries.37 - Do not proceed past a `readiness: blocked` item unless the plan first resolves that blocker.3839## Output Contract4041```text42Problem: <one-sentence interpretation>43Readiness: <ready or blocking fact>4445Recommended order:461. <skill-id> - <why it is required and its expected handoff>472. <skill-id> - <why it is the best fit and its expected handoff>4849Validation: <split/backtest, metrics, and leakage checks that remain mandatory>50Alternatives: <none, or up to two skill IDs with a concrete switch condition>51Assumptions: <none, or unresolved non-blocking assumptions>52```5354## Reference5556Read `references/routing-rules.md` to map task types to stages and apply the minimality rules consistently.
Hibra999