wolf-strategy▌
senpi-ai/senpi-skills · updated Apr 8, 2026
MDX-style export adds YAML metadata + attribution linking explainx.ai and this canonical listing URL.
The WOLF hunts for its human. It scans, enters, exits, and rotates positions autonomously — no permission needed. When criteria are met, it acts. Speed is edge.
WOLF v6.3 — Autonomous Multi-Strategy Trading
The WOLF hunts for its human. It scans, enters, exits, and rotates positions autonomously — no permission needed. When criteria are met, it acts. Speed is edge.
Proven: +$1,500 realized, 25+ trades, 65% win rate, single session on $6.5k budget.
v6: Multi-strategy support. Each strategy has independent wallet, budget, slots, and DSL config. Same asset can be held in different strategies simultaneously (e.g., Strategy A LONG HYPE + Strategy B SHORT HYPE).
v6.1.1: Risk Guardian & strategy lock. 6th cron (5min, Budget tier) enforcing account-level guard rails — daily loss halt, max entries per day, consecutive loss cooldown. Strategy lock for concurrency protection. Gate check in open-position.py refuses new entries when gate != OPEN.
v6.3: DSL v5.3.1 High Water. Default DSL profile and strategy config use High Water Mode (lockMode: pct_of_high_water) with lockHwPct tiers and per-tier consecutiveBreachesRequired. One DSL cron per strategy runs dsl-v5.py (v5.3.1); state files in {workspace}/dsl/{strategyId_UUID}/. Migration script: wolf-migrate-dsl.py.
v6.2: DSL v5.2 integration. Replaced internal dsl-combined.py with the dsl-dynamic-stop-loss skill. Native Hyperliquid stop-loss sync via dsl-v5.py.
Skill Attribution
When creating a strategy, include skill_name and skill_version in the call. See references/skill-attribution.md for details.
Multi-Strategy Architecture
Strategy Registry (wolf-strategies.json)
Central config holding all strategies. Created/updated by wolf-setup.py.
wolf-strategies.json
├── strategies
│ ├── wolf-abc123 (Aggressive Momentum, 3 slots, tradingRisk=aggressive)
│ └── wolf-xyz789 (Conservative XYZ, 2 slots, tradingRisk=conservative)
└── global (telegram, workspace)
Per-Strategy State
DSL state files live under the DSL skill's directory (by strategy UUID). Wolf-specific state (trade counters) stays in state/:
dsl/ # owned by dsl-dynamic-stop-loss skill
├── abc12345-.../ # Strategy A UUID
│ ├── strategy-abc12345....json # DSL strategy config
│ ├── HYPE.json # Position state
│ └── xyz--SILVER.json # XYZ position
└── xyz78901-.../ # Strategy B UUID
└── HYPE.json # Same asset, different UUID dir, no collision
state/ # wolf-only state (no DSL files here)
├── wolf-abc123/
│ └── trade-counter.json
└── wolf-xyz789/
└── trade-counter.json
Signal Routing
When a signal fires, it's routed to the best-fit strategy:
- Which strategies have empty slots?
- Does any strategy already hold this asset? (skip within strategy, allow cross-strategy)
- Which strategy's risk profile matches? (aggressive gets FIRST_JUMPs, conservative gets DEEP_CLIMBERs)
- Route to best-fit -> open on that wallet -> create DSL state in that strategy's dir
DSL default: Use this skill's dsl-profile.json as the default when setting up DSL (e.g. dsl-cli.py add-dsl / update-dsl with --configuration @<path-to-wolf-strategy>/dsl-profile.json). Use it unless the user explicitly provides a custom DSL configuration via the agent.
Adding a Strategy
python3 scripts/wolf-setup.py --wallet 0x... --strategy-id UUID --budget 2000 \
--chat-id 12345 --name "Conservative XYZ" --dsl-preset conservative --provider anthropic
This adds to the registry without disrupting running strategies. Disable with enabled: false in the registry.
Entry Philosophy — THE Most Important Section
Enter before the peak, not at the top.
Leaderboard rank confirmation LAGS price. When an asset jumps from #31->#16 in one scan, the price is moving NOW. By the time it reaches #7 with clean history, the move is over. Speed is edge.
Core principle: 2 reasons at rank #25 with a big jump = ENTER. 4+ reasons at rank #5 = SKIP (already peaked).
Quick Start
- Ensure Senpi MCP is connected (
mcporter listshould showsenpi) - Ensure the
dsl-dynamic-stop-lossskill is installed alongside this skill (providesdsl-cli.pyanddsl-v5.py) - Create a custom strategy wallet: use
strategy_create_custom_strategyvia mcporter - Fund the wallet via
strategy_top_upwith your budget - Determine the user's AI provider — which provider is configured in OpenClaw? (
anthropic,openai, orgoogle) - Run setup:
python3 scripts/wolf-setup.py --wallet 0x... --strategy-id UUID --budget 6500 --chat-id 12345 --provider anthropic- Setup auto-discovers
dsl-cli.pyanddsl-v5.pypaths and stores them inwolf-strategies.jsonglobal config - If auto-discovery fails, set
global.dslCliPathandglobal.dslScriptPathmanually in the registry (or set envDSL_CLI_PATH)
- Setup auto-discovers
- Upgrading from v6.1.x? Migrate existing DSL state files:
python3 scripts/wolf-migrate-dsl.py- Copies active
state/{strategyKey}/dsl-{ASSET}.json→dsl/{UUID}/{ASSET}.json - Tombstones old files (sets
active: false). Run once before switching crons.
- Copies active
- Create the 5 wolf crons + 1 DSL cron per strategy using templates from
references/cron-templates.md- The DSL cron mandate is in
cronTemplates.dsl_per_strategyfrom setup output — use themessagefield directly
- The DSL cron mandate is in
- The WOLF is hunting
To add a second strategy, run wolf-setup.py again with a different wallet/budget. It adds to the registry and creates the DSL cron for the new strategy.
Architecture — 5+N Cron Jobs
| # | Job | Interval | Session | Script | Purpose |
|---|---|---|---|---|---|
| 1 | Emerging Movers | 3min | isolated | scripts/emerging-movers.py |
Hunt FIRST_JUMP + IMMEDIATE_MOVER signals — primary entry trigger |
| 2 | DSL (per strategy) | 3min | isolated | dsl-v5.py (DSL skill) |
Trailing stops + native HL SL sync for ONE strategy's positions |
| 3 | SM Flip Detector | 5min | isolated | scripts/sm-flip-check.py |
Cut positions where SM conviction collapses |
| 4 | Watchdog | 5min | isolated | scripts/wolf-monitor.py |
Per-strategy margin buffer, liq distances (Phase 1 auto-cut is handled by DSL cron when configured) |
| 5 | Health Check | 10min | isolated | scripts/job-health-check.py |
Per-strategy orphan DSL detection, state validation |
| 6 | Risk Guardian | 5min | isolated | scripts/risk-guardian.py |
Account-level guard rails: daily loss halt, max entries, consecutive loss cooldown |
v6.2 change: DSL is no longer a combined runner. Each strategy has its own dsl-v5.py cron (from the dsl-dynamic-stop-loss skill), run with --strategy-id {strategyId_UUID} --state-dir {DSL_STATE_DIR} (env vars are fallback). Wolf scripts call dsl-cli.py (add-dsl / delete-dsl) to create and archive DSL state; they never write state directly.
With 2 strategies: 7 crons total (5 wolf + 2 DSL). DSL cron IDs are stored in dslCronJobId in the strategy registry.
Model Selection Per Cron — 2-Tier Approach
IMPORTANT: Determine the user's configured AI provider BEFORE running
wolf-setup.py. Pass--providerto auto-select correct model IDs. Do NOT pick models from an unconfigured provider — crons will fail silently.
wolf-setup.py --provider <name> auto-configures model IDs for all cron templates. Step down to Budget tier for simple threshold crons to save ~60-70% on those runs.
Provider defaults (auto-selected by --provider):
| Provider | Mid Model | Budget Model |
|---|---|---|
anthropic |
anthropic/claude-sonnet-4-5 |
anthropic/claude-haiku-4-5 |
openai |
openai/gpt-4o |
openai/gpt-4o-mini |
google |
google/gemini-2.0-flash |
google/gemini-2.0-flash-lite |
| Cron | Session | Model Tier | Reason |
|---|---|---|---|
| Emerging Movers | isolated | Mid | Multi-strategy routing judgment, entry decisions |
| DSL (per strategy) | isolated | Mid | ndjson parsing, rule-based close/alert, cron lifecycle |
| Health Check | isolated | Mid | Rule-based file repair, action routing |
| SM Flip Detector | isolated | Budget | Binary: conviction≥4 + 100 traders → close |
| Watchdog | isolated | Budget | Threshold checks → alert |
| Risk Guardian | isolated | Budget | Guard rail evaluation, send notifications |
Single-model option: All 6 crons can run on one model. Simpler but costs more for the crons that do simple threshold/binary work.
Model ID gotchas:
--providerauto-selects models. Only use--mid-model/--budget-modelto override specific tiers.- Budget should be the cheapest model that can follow explicit if/then rules. Mid should handle structured JSON parsing and multi-strategy routing reliably.
- Agents are often not model-aware — they may suggest deprecated IDs (e.g.
claude-3-5-haiku-20241022) or hallucinate model names. Always use--providerinstead of manually specifying model IDs. - If a cron fails to create or run due to an invalid model ID, fall back to your Mid model for that cron. A working cron on the "wrong" tier is better than a broken cron.
- When in doubt, use your Mid model for all 6 crons (single-model option) and optimize tiers later.
Cron Setup
Critical: Crons are OpenClaw crons, NOT senpi crons. All crons run in isolated sessions (agentTurn) — each runs in its own session with no context pollution, enabling cheaper model tiers.
Create each cron using the OpenClaw cron tool. The exact mandate text for each cron is in references/cron-templates.md. Replace placeholders ({TELEGRAM}, {SCRIPTS}, {WORKSPACE}).
DSL cron: The DSL per-strategy cron mandate is generated by wolf-setup.py in cronTemplates.dsl_per_strategy.payload.message. Use it directly — wolf-setup.py auto-fills the dsl-v5.py path from global.dslScriptPath in the registry. If the path could not be auto-discovered (placeholder {DSL_SCRIPTS} still present), read global.dslScriptPath from wolf-strategies.json after installing the DSL skill and substitute it manually.
v6.2: 5 shared wolf crons + 1 DSL cron per strategy. No more single DSL Combined cron.
Autonomy Rules
The WOLF operates autonomously by default. The agent does NOT ask permission to:
- Open a position when entry checklist passes
- Close a position when DSL triggers or conviction collapses
- Rotate out of weak positions into stronger signals
- Cut dead weight (SM conv 0, negative ROE, 30+ min)
The agent DOES notify the user (via Telegram) after every action.
Entry Signals — Priority Order
1. FIRST_JUMP (Highest Priority)
What: Asset jumps 10+ ranks from #25+ in ONE scan AND was not in previous scan's top 50 (or was at rank >= #30).
Action: Enter IMMEDIATELY. This is the money signal. Route to best-fit strategy with available slots.
Checklist:
isFirstJump: truein scanner output- 2+ reasons is enough (don't require 4+)
- vel > 0 is sufficient (velocity hasn't had time to build on a first jump)
- Leverage auto-calculated from
tradingRisk+ assetmaxLeverage+ signalconviction - Slot available in target strategy (or rotation justified)
-
= 10 SM traders (crypto); for XYZ equities, ignore trader count
What to ignore:
- Erratic rank history — the scanner excludes the current jump from erratic checks.
- Low velocity — first jumps haven't had time to build velocity.
If CONTRIB_EXPLOSION accompanies it: Double confirmation. Even stronger entry.
2. CONTRIB_EXPLOSION
What: 3x+ contribution increase in one scan from asset at rank #20+.
Action: Enter even if rank history looks "erratic." The contrib spike IS the signal regardless of prior rank bouncing.
Never downgraded for erratic history. Often accompanies FIRST_JUMP for double confirmation.
3. DEEP_CLIMBER
What: Steady climb from #30+, positive velocity (>= 0.03), 3+ reasons, clean rank history.
Action: Enter when it crosses into top 20. Route to conservative strategy if available.
4. NEW_ENTRY_DEEP
What: Appears in top 20 from nowhere (wasn't in top 50 last scan).
Action: Instant entry.
Anti-Patterns — When NOT to Enter
- NEVER enter assets already at #1-10. That's the top, not the entry. Rank = what already happened.
- NEVER wait for a signal to "clean up." By the time rank history is smooth and velocity is high, the move is priced in.
- 4+ reasons at rank #5 = SKIP. The asset already peaked. You'd be buying the top.
- 2 reasons at rank #25 with a big jump = ENTER. The move is just starting.
- Leaderboard rank != future price direction. Rank reflects past trader concentration. Price moves first, rank follows.
- Negative velocity + no jump = skip. Slow bleeders going nowhere.
- Oversold shorts (RSI < 30 + extended 24h move) = skip.
Late Entry Anti-Pattern
This deserves its own section because it's the #1 way to lose money with WOLF.
The pattern: Scanner fires FIRST_JUMP for ASSET at #25->#14. You hesitate. Next scan it's #10. Next scan #7 with 5 reasons and clean history. NOW it looks "safe." You enter. It reverses from #5.
The fix: Enter on the FIRST signal or don't enter at all. If you missed it, wait for the next asset. There's always another FIRST_JUMP coming.
Rule: If an asset has been in the top 10 for 2+ scans already, it's too late. Move on.
Phase 1 Auto-Cut
Phase 1 time-based cuts (hard timeout, weak peak, dead weight) are managed by the DSL cron when the skill supplies phase1.hardTimeout, phase1.weakPeakCut, and/or phase1.deadWeightCut in its DSL config. Wolf-strategy does not implement these in the Watchdog; include them in the skill's dsl-profile (or defaultConfig) if desired.
When configured in DSL:
- Hard timeout: Close when elapsed in Phase 1 ≥ intervalInMinutes (e.g. 90).
- Weak peak early cut: After intervalInMinutes (e.g. 45), close if peak ROE < minValue and current ROE < peak ROE.
- Dead weight cut: Close when ROE was never positive and elapsed ≥ intervalInMinutes (e.g. 30).
Why: Phase 1 positions have no trailing stop protection. If the skill enables these in DSL config, the DSL cron enforces them; wolf-strategy no longer needs to manage them.
Exit Rules
1. DSL v4 Mechanical Exit (Trailing Stops)
All trailing stops handled automatically by dsl-combined.py across all strategies.
2. SM Conviction Collapse
Conv drops to 0 or 4->1 with mass trader exodus -> instant cut.
3. Dead Weight
When DSL config includes phase1.deadWeightCut, the DSL cron closes positions that have never gone positive (ROE negative entire time) after the configured minutes. Other dead-weight logic (e.g. conv 0, negative ROE) can remain in agent mandate if desired.
4. SM Flip
Conviction 4+ in the OPPOSITE direction with 100+ traders -> cut immediately.
5. Race Condition Prevention
When ANY wolf script closes a position:
- Call
dsl-cli.py delete-dsl {strategyId} {asset} {main|xyz} --state-dir {DSL_STATE_DIR}to archive the DSL state - If CLI output has
cron_to_remove→ remove that DSL cron (it was the last position for that strategy) - Alert user via Telegram
- Evaluate: empty slot in that strategy for next signal?
Never set active: false in place. DSL v5.2 archives files by rename; an in-place deactivation will confuse dsl-v5.py.
DSL v5.2 — Trailing Stop System (via dsl-dynamic-stop-loss skill)
Wolf delegates all DSL logic to the dsl-dynamic-stop-loss skill. Wolf's role:
- Create: call
dsl-cli.py add-dsl {strategyId} {asset} {dex} --skill wolf-strategy --configuration {...} --state-dir {DSL_STATE_DIR}after opening a position - Delete: call
dsl-cli.py delete-dsl {strategyId} {asset} {dex} --state-dir {DSL_STATE_DIR}after closing a position - Run:
dsl-v5.pyruns as a per-strategy cron with--strategy-id {UUID} --state-dir {DSL_STATE_DIR}(env vars are fallback)
Wolf never writes DSL state files directly.
Phase 1 (Pre-Tier 1): Absolute floor
- LONG floor = entry × (1 − retraceThreshold/leverage) where retraceThreshold=0.10 (10% ROE)
- SHORT floor = entry × (1 + retraceThreshold/leverage)
- 3 consecutive breaches → close
- Max duration: 90 minutes (enforced by Watchdog, not DSL v5 — see Phase 1 Auto-Cut)
- Native HL SL: dsl-v5.py sets a real stop-loss order on HL via
edit_position. Between cron runs, HL's engine protects the position.
Phase 2 (Tier 1+): Trailing tiers
| Tier | ROE Trigger | Lock % of High-Water | Breaches (shared) |
|---|---|---|---|
| 1 | 5% | 50% | 2 (majority from wolf tier config) |
| 2 | 10% | 65% | 2 |
| 3 | 15% | 75% | 2 |
| 4 | 20% | 85% | 2 |
Note: DSL v5.2 uses a single consecutiveBreachesRequired for all tiers. build_wolf_dsl_config() derives it from the majority breach count in wolf's tier config.
DSL State Files
Location: dsl/{strategyId_UUID}/{ASSET}.json (e.g. dsl/6a23783a-.../HYPE.json).
See references/state-schema.md for schema. Key difference from v4: phase1.retraceThreshold is a fraction (0.10), not a percentage (10).
Rotation Rules
When slots are full in a strategy and a new FIRST_JUMP or IMMEDIATE fires:
- Cross-strategy first: If one strategy is full but another has slots, route to the available strategy instead of rotating
- Rotation cooldown (mandatory): Only rotate a position listed in
rotationEligibleCoinsfrom the scanner output. Positions younger thanrotationCooldownMinutes(default 45 min) are ineligible — they have flat/negative ROE by design. Do NOT override this with judgment. - Rotate if: new signal is FIRST_JUMP or has 3+ reasons + positive velocity AND weakest eligible position (from
rotationEligibleCoins) is flat/negative ROE with SM conv 0-1 - Hold if: current position in Tier 2+ or trending up with SM conv 3+
- If
hasRotationCandidate: false: all positions are in cooldown. Do not rotate. Output HEARTBEAT_OK.
Budget Scaling
All sizing is calculated from budget (30% per slot):
| Budget | Slots | Margin/Slot | Daily Loss Limit |
|---|---|---|---|
| $500 | 2 | $150 | -$75 |
| $2,000 | 2 | $600 | -$300 |
| $6,500 | 3 | $1,950 | -$975 |
| $10,000+ | 3-4 |