Stats
Actions
Tags
Help us improve
Share bugs, ideas, or general feedback.
From deepseek
Delegate tasks to DeepSeek V4 API for high-quality code analysis at lower cost than Sonnet. V4-Pro matches Opus 4.6 on SWE-bench (80.6%); V4-Flash handles mechanical checks at Haiku-tier cost. Use when offloading bulk diff review, code pattern analysis, or mechanical checks from Anthropic quota. Includes a generic agent runner that routes specific dm-review agents (pattern-recognition, code-simplicity, doc-sync, test-coverage) through DeepSeek when DEEPSEEK_API_KEY is set. Invoke with /deepseek for direct delegation.
npx claudepluginhub design-machines-studio/depot --plugin deepseekHow this skill is triggered — by the user, by Claude, or both
Slash command
/deepseek:deepseek-delegateThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
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.
Provides a checklist for code reviews covering functionality, security, performance, maintainability, tests, and quality. Use for pull requests, audits, team standards, and developer training.
Share bugs, ideas, or general feedback.
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.
| 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. |
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.
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
Load templates from ${CLAUDE_SKILL_DIR}/references/prompt-templates.md. Key principles:
-p flag for short prompts. For large inputs, pipe via stdin.| Agent | File | Purpose |
|---|---|---|
| deepseek-bulk-analyst | plugins/deepseek/agents/review/deepseek-bulk-analyst.md | Full-diff review using 1M context |
| deepseek-code-analyst | plugins/deepseek/agents/workflow/deepseek-code-analyst.md | Code pattern analysis and refactoring review |
DeepSeek API key must be set:
export DEEPSEEK_API_KEY="sk-..."
# Resolve the wrapper via the plugin cache (works from any CWD)
WRAPPER_PATH=""
for CACHE_ROOT in "$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.