deepseek-delegate

DeepSeek Delegation

Invoke DeepSeek V4 API as a supplementary provider for tasks where cost-per-token or capacity matters. DeepSeek is always supplementary -- Opus remains the primary thinker.

When to Delegate

Advantage	Use Case	Why DeepSeek
Sonnet-tier quality at lower cost	Code review agents, pattern analysis, refactoring suggestions	V4-Pro scores 80.6% SWE-bench Verified, matches Opus 4.6. $1.74/MTok input vs Sonnet pricing.
Bulk diff analysis	Large diffs (>5K lines, >20 files)	1M token context. No truncation needed. Doesn't burn Anthropic Max quota.
Mechanical checks	Anti-pattern scans, link rot, doc-sync verification	V4-Flash at $0.14/MTok for tasks that don't need deep reasoning.
Capacity relief	Pipeline runs burning through Max quota	Every token routed to DeepSeek is a token NOT counted against your Anthropic weekly limit.

When NOT to Delegate

• Tasks requiring Claude's conversation context (DeepSeek calls are stateless)
• Tasks requiring MCP server access (DeepSeek can't reach Claude's MCP servers)
• Adversarial review or ambiguity detection (Opus's reasoning edge matters here)
• Security-critical analysis (keep on Anthropic for liability and quality assurance)
• Interactive multi-turn workflows (each DeepSeek call is a fresh session)

Invocation Protocol

Load the full protocol from ${CLAUDE_SKILL_DIR}/references/invocation-protocol.md. It covers curl syntax, JSON escaping via python3, the wrapper script (references/deepseek-wrapper.sh) with its v4-pro -> v4-flash fallback chain, HTTP status handling, and response parsing.

Key rules: always set --max-time timeout, always use the wrapper for automated flows, always escape prompts via python3 JSON encoder (never embed raw user input in curl -d). All failures are graceful skips.

Model Selection

Load the full decision table from ${CLAUDE_SKILL_DIR}/references/model-selection.md. It maps task types to models (v4-pro, v4-flash), timeouts, and expected latency.

Rate limit fallback chain: v4-pro -> v4-flash -> skip

Prompt Engineering

Load templates from ${CLAUDE_SKILL_DIR}/references/prompt-templates.md. Key principles:

1. Self-contained prompts. DeepSeek has no conversation context. Every prompt must include all necessary information.
2. Structured output requests. Always tell DeepSeek what format you need. For dm-review integration, request P1/P2/P3 findings.
3. Escape handling. Use the wrapper's -p flag for short prompts. For large inputs, pipe via stdin.

Available Agents

Agent	File	Purpose
deepseek-bulk-analyst	plugins/deepseek/agents/review/deepseek-bulk-analyst.md	Full-diff review using 1M context
deepseek-code-analyst	plugins/deepseek/agents/workflow/deepseek-code-analyst.md	Code pattern analysis and refactoring review

Prerequisites

DeepSeek API key must be set:

export DEEPSEEK_API_KEY="sk-..."

# Resolve the wrapper via the plugin cache (works from any CWD)
WRAPPER_PATH=""
for CACHE_ROOT in "$HOME/.claude/plugins/cache/depot" "$HOME/.codex/plugins/cache/depot"; do
  WRAPPER_PATH=$(ls -t "$CACHE_ROOT"/deepseek/*/skills/deepseek-delegate/references/deepseek-wrapper.sh 2>/dev/null | head -1)
  [ -n "$WRAPPER_PATH" ] && break
done
if [ -z "$WRAPPER_PATH" ] || [ ! -x "$WRAPPER_PATH" ]; then
  echo "deepseek wrapper not found in plugin cache" >&2
  exit 1
fi

# Verify authentication
bash "$WRAPPER_PATH" -m v4-flash -p "test"

The invoking Claude Code session must have Bash permissions for curl commands. If the first invocation is blocked by permissions, report to the user and skip gracefully.