routing-tables

Routing Tables

This skill is the human-readable companion to the routing data in assets/routing-config.json and assets/gate-criteria.json. It centralizes current routing tables and enum values from Code-Conductor and Code-Critic so agents can load concise prompt context while future deterministic consumers read the same data from JSON.

When to Use

• When an agent needs the canonical current specialist-dispatch, review-mode, CE-surface, enum, or gate data
• When a deterministic consumer should read the shared routing assets instead of copying inline tables from an agent prompt
• When reviewing or updating routing metadata and you need one human-readable summary aligned with the JSON assets

Schema Overview

assets/routing-config.json contains six top-level sections:

• specialist_dispatch: file and task routing entries from Code-Conductor's Agent Selection table; deterministic FilePattern consumers should use the optional file_patterns arrays rather than parsing file_type_or_task
• review_mode_routing: selector-line marker to review-mode mapping from Code-Critic, including conflict resolution
• surface_identification: CE Gate surface to tool mapping from Code-Conductor; No customer surface carries the exact status marker template in status_result_template
• skill_mapping: delegation guidance from Code-Conductor's Skill Mapping table; this is a reference list, not a strict deterministic router
• nl_intent_routing: natural-language intent keys and patterns mapped to Claude and Copilot slash-command surfaces via claude_command and copilot_command; null means no platform equivalent. This supports Phase 1 directive prose and Phase 2 forward-compatible lookup, not runtime hook wiring
• enums: canonical current enum values used by review and routing outputs

For deterministic consumers, Invoke-RoutingLookup preserves the legacy first-match behavior. Use Invoke-RoutingLookupAll when a Pattern lookup needs to detect ambiguous natural-language matches and present every matched intent to the user.

assets/gate-criteria.json contains four top-level sections:

• scope_classification: abbreviated-vs-full pipeline gate criteria and tier table
• express_lane: low-risk fix gate criteria for routing findings directly to a specialist
• review_completion: pre-PR review-completion gate that requires prosecution, defense, and judgment completion to all be recorded as true for the current review cycle
• post_fix_trigger: conditions that require post-fix targeted prosecution after accepted review fixes

Specialist Dispatch

Specialist dispatch stays file- and task-oriented. Test files and fixtures route to Test-Writer. New behavior in src/**/*.ts and src/**/*.tsx routes to Code-Smith, while restructuring those same source files routes to Refactor-Specialist. PowerShell production scripts (*.ps1) also route to Code-Smith. Visual polish work on UI source files routes to UI-Iterator, and documentation files route to Doc-Keeper.

For deterministic lookups keyed by FilePattern, use the structured file_patterns arrays in the asset. The prose-bearing file_type_or_task field remains the human-readable summary and should not be parsed by consumers.

Planning artifacts and plan markers route to Issue-Planner. CE Gate evidence capture, customer framing, journeys, and scenarios route to Experience-Owner. Read-only quality review goes to Code-Critic, scored judgment goes to Code-Review-Response, and systemic gap analysis goes to Process-Review.

Review Mode Routing

Review mode routing is driven by explicit top-level selector lines of the form Review mode selector: "{marker}". No selector line means normal code prosecution with five parallel passes. Design review perspectives switch to design or plan prosecution with three parallel passes. Defense, CE prosecution, GitHub proxy prosecution, post-fix code prosecution, and lite code prosecution each run as single-pass modes.

When multiple selector lines appear, the conflict rule applies in strict priority order: defense, then CE, then proxy, then design, then post-fix code prosecution, then lite code prosecution, then standard code prosecution. One explicit override exists: Review mode selector: "Use code review perspectives" beats Review mode selector: "Use design review perspectives" and forces normal code prosecution.

Handshake disposition by command

Moved here from the root guidance file (CLAUDE.md) by issue #998, which needed the space for the standing statement of what a finished run is true of. The dispositions are unchanged by the move; the root file keeps the one-line summary and points here.

Command	Handshake
/orchestra:review	Required
/orchestra:review-lite	Required
/orchestra:review-prosecute	Required
/orchestra:review-defend	Required
/orchestra:review-judge	Optional

Surface Identification

CE Gate surface identification maps the customer surface to the expected validation tool. Web UI uses native browser tools first, with Playwright MCP as fallback. REST or GraphQL uses terminal HTTP calls. CLI and SDK surfaces use representative terminal invocations. Batch and pipeline surfaces use representative test data. No customer surface returns the documented status marker ⏭️ CE Gate not applicable — {reason} from structured data so consumers do not need a separate hardcoded case.

Skill Mapping

The skill mapping table is guidance for delegation prompts rather than a pure lookup table. It links each reusable skill to the kind of work where Code-Conductor should explicitly instruct a specialist to load it. The mappings cover review, customer experience, design exploration, documentation, frontend design, implementation discipline, parallel execution, planning, property-based testing, refactoring, research, review judgment, skill creation, software architecture, debugging, TDD, UI iteration, UI testing, and web app testing.

Consumers should treat this section as a reference list that explains when a skill is relevant, not as a hard decision engine that replaces agent judgment.

Enums

The enum section captures the canonical current values used by the review pipeline:

• severity: critical, high, medium, low
• points: critical and high map to 10; medium maps to 5; low maps to 1
• confidence: high, medium, low
• category: architecture, security, performance, pattern, implementation-clarity, script-automation, documentation-audit
• blast_radius: localized, module, cross-module, system-wide
• authority_needed: yes, no
• systemic_fix_type: instruction, skill, agent-prompt, plan-template, none

The category list is canonical-only. Legacy simplicity is intentionally excluded.

Gate Criteria

Scope classification is an AND gate: all five abbreviated-tier criteria must hold, otherwise the issue stays on the full pipeline. The stored tier table captures which phases run in full and abbreviated execution.

Express lane is also an AND gate. A finding must be low severity, strictly mechanical, non-logic-changing, test-neutral, backward-compatible with respect to stored IDs and schema, and limited to one file. It only applies to main review and post-fix review routing, and Tier 1 re-validation is still required after the fix lands.

Review completion is an AND gate. It fails closed unless the current review cycle records all three stage booleans as true: prosecution_complete, defense_complete, and judgment_complete. review_mode is tracked alongside the state as label metadata, but it is not part of the gate criteria.

Post-fix targeted prosecution uses OR semantics for its trigger conditions: one accepted critical or high finding is enough, and any accepted fix that modifies control flow is also enough. If no findings were accepted and applied, the post-fix review is skipped.

Interpretation Rules

Use exact current values from the JSON assets when a mode, enum, or routing value needs to be preserved across prompts or scripts. Match review markers literally. For specialist dispatch file routing, use explicit file_patterns keys instead of parsing prose. Treat skill_mapping as explanatory metadata, not a replacement for human judgment. Treat scope_classification, express_lane, and review_completion as all-criteria gates, while post_fix_trigger activates when any trigger condition is met.

When an entry includes prose qualifiers, preserve them rather than compressing them away. The goal of these files is stable extraction of current rules, not reinterpretation of the orchestration policy.

Fallback Guidance

Future deterministic consumers should prefer the JSON assets first. If a later script is unavailable, agents can still read the JSON directly. If a case is not covered by the structured data or a prompt needs human-readable context, fall back to this skill summary.

Routing authority remains with the calling agent. This skill provides shared data and summaries; it does not own orchestration decisions.

Intent Routing Mechanics

Plugin processes are the default chat experience. Natural-language requests matching the nl_intent_routing table route to the corresponding slash command with a visible confirmation; /raw opts out.

Activation order: (1) VS Code dropdown for VS Code users; (2) slash commands for both platforms; (3) natural-language with auto-routing confirmation; (4) @-mention is NOT recommended (unreliable in every plugin surface tested).

Slash commands diverge between Claude (commands/*.md) and Copilot (.github/prompts/*.prompt.md); the nl_intent_routing table carries both column names so the canonical command name is platform-portable.

First match per command-family per conversation asks explicitly, with options Run /X for this (Recommended), Continue as raw chat, and Don't ask again for this command-family this conversation; Claude confirmation phrasing should use Run /X?. Subsequent same-family matches use inline confirmation: Routing to /X — say /raw to opt out, otherwise proceed.

Tier hint: When a proposed command's model: frontmatter differs from the user-session model, append a one-line tier hint, e.g. Will run on sonnet + high per command frontmatter.

No-match and disambiguation: No-match answers normally; first no-match per conversation appends Tip: type /help for plugin slash commands, or /raw to suppress these hints. Ambiguous-match uses a text-only disambiguation prompt, e.g. Did you mean /orchestra:review (local code) or /review-github (GitHub PR)?

Gotchas

Trigger	Gotcha	Fix
Treating a routing table entry as binding policy	The skill centralizes canonical data, but the calling agent still owns the final routing decision and fallback behavior	Use the skill for shared data, then keep decision authority and override logic in the agent
Editing only the prose summary or only the JSON asset	Human-readable guidance and deterministic consumers drift apart	Update the JSON assets and the corresponding SKILL.md summary together in the same change