CCS Delegation

Delegate deterministic tasks to cost-optimized models via CCS CLI.

Core Concept

Execute tasks via alternative models using: - Initial delegation: ccs {profile} -p "task" - Session continuation: ccs {profile}:continue -p "follow-up"

Profile Selection: - Auto-select from ~/.ccs/config.json via task analysis - Profiles: glm (cost-optimized), kimi (long-context/reasoning), custom profiles - Override: --{profile} flag forces specific profile

User Invocation Patterns

Users trigger delegation naturally: - "use ccs [task]" - Auto-select best profile - "use ccs --glm [task]" - Force GLM profile - "use ccs --kimi [task]" - Force Kimi profile - "use ccs:continue [task]" - Continue last session

Examples: - "use ccs to fix typos in README.md" - "use ccs to analyze the entire architecture" - "use ccs --glm to add unit tests" - "use ccs:continue to commit the changes"

Agent Response Protocol

For /ccs [task]:

1. Parse override flag - Scan task for pattern: --(\w+) - If match: profile = match[1], remove flag from task, skip to step 5 - If no match: continue to step 2

2. Discover profiles - Read ~/.ccs/config.json using Read tool - Extract Object.keys(config.profiles) → availableProfiles[] - If file missing → Error: "CCS not configured. Run: ccs doctor" - If empty → Error: "No profiles in config.json"

3. Analyze task requirements - Scan task for keywords:

   • /(think|analyze|reason|debug|investigate|evaluate)/i → needsReasoning = true
   • /(architecture|entire|all files|codebase|analyze all)/i → needsLongContext = true
   • /(typo|test|refactor|update|fix)/i → preferCostOptimized = true

4. Select profile - For each profile in availableProfiles: classify by name pattern (see Profile Characteristic Inference table) - If needsReasoning: filter profiles where reasoning=true → prefer kimi - Else if needsLongContext: filter profiles where context=long → prefer kimi - Else: filter profiles where cost=low → prefer glm - selectedProfile = filteredProfiles[0] - If filteredProfiles.length === 0: fallback to glm if exists, else first available - If no profiles: Error

5. Enhance prompt - If task mentions files: gather context using Read tool - Add: file paths, current implementation, expected behavior, success criteria - Preserve slash commands at task start (e.g., /cook, /commit)

6. Execute delegation - Run: ccs {selectedProfile} -p "$enhancedPrompt" via Bash tool

7. Report results - Log: "Selected {profile} (reason: {reasoning/long-context/cost-optimized})" - Report: Cost (USD), Duration (sec), Session ID, Exit code

For /ccs:continue [follow-up]:

1. Detect profile - Read ~/.ccs/delegation-sessions.json using Read tool - Find most recent session (latest timestamp) - Extract profile name from session data - If no sessions → Error: "No previous delegation. Use /ccs first"

2. Parse override flag - Scan follow-up for pattern: --(\w+) - If match: profile = match[1], remove flag from follow-up, log profile switch - If no match: use detected profile from step 1

3. Enhance prompt - Review previous work (check what was accomplished) - Add: previous context, incomplete tasks, validation criteria - Preserve slash commands at start

4. Execute continuation - Run: ccs {profile}:continue -p "$enhancedPrompt" via Bash tool

5. Report results - Report: Profile, Session #, Incremental cost, Total cost, Duration, Exit code

Decision Framework

Delegate when: - Simple refactoring, tests, typos, documentation - Deterministic, well-defined scope - No discussion/decisions needed

Keep in main when: - Architecture/design decisions - Security-critical code - Complex debugging requiring investigation - Performance optimization - Breaking changes/migrations

Profile Selection Logic

Task Analysis Keywords (scan task string with regex):

Profile Characteristic Inference (classify by name pattern):

Selection Algorithm (apply filters sequentially):

profiles = Object.keys(config.profiles)
classified = profiles.map(p => ({name: p, ...inferCharacteristics(p)}))

if (needsReasoning):
  filtered = classified.filter(p => p.reasoning === true).sort(['kimi'])
else if (needsLongContext):
  filtered = classified.filter(p => p.context === 'long').sort(['kimi'])
else:
  filtered = classified.filter(p => p.cost === 'low').sort(['glm', ...])

selected = filtered[0] || profiles.find(p => p === 'glm') || profiles[0]
if (!selected): throw Error("No profiles configured")

log("Selected {selected} (reason: {reasoning|long-context|cost-optimized})")

Override Logic: - Parse task for /--(\w+)/. If match: profile = match[1], remove from task, skip selection

Example Delegation Tasks

Good candidates: - "/ccs add unit tests for UserService using Jest" → Auto-selects: glm (simple task) - "/ccs analyze entire architecture in src/" → Auto-selects: kimi (long-context) - "/ccs think about the best database schema design" → Auto-selects: kimi (reasoning) - "/ccs --glm refactor parseConfig to use destructuring" → Forces: glm (override)

Bad candidates (keep in main): - "implement OAuth" (too complex, needs design) - "improve performance" (requires profiling) - "fix the bug" (needs investigation)

Execution

Commands: - /ccs "task" - Intelligent delegation (auto-select profile) - /ccs --{profile} "task" - Force specific profile - /ccs:continue "follow-up" - Continue last session (auto-detect profile) - /ccs:continue --{profile} "follow-up" - Continue with profile switch

Agent via Bash: - Auto: ccs {auto-selected} -p "task" - Continue: ccs {detected}:continue -p "follow-up"

References

Template: CLAUDE.md.template - Copy to user's CLAUDE.md for auto-delegation config Troubleshooting: references/troubleshooting.md