spider-strategy（蜘蛛策略）

🕷️ SPIDER v2.0 — Agentic Portfolio Operator

The first Senpi agent that operates above the position level.

Every other Predator decides "open or close this trade?" every few minutes. Spider builds a web — one anchor point, several radial threads — then waits. The web does the work. Vibrations come to it.

What changed in v2.0

v1.0 was an Elite-convergence single-position scanner. It contributed to the same fleet pathology that v2.0 is designed to correct: too many fills, weak signals, repeated DSL stops. v2.0 is a structural rewrite from a different thesis entirely.

The shape of the strategy

Spider runs a two-leg portfolio:

1. Anchor — a single high-conviction long (e.g. HYPE 3x), held minimum 7 days
2. Basket — 3–5 shorts on the highest-funding alt-coins, risk-parity sized to 40% of anchor notional

The basket isn't a directional hedge — it's a funding harvest. Shorts on coins paying the most positive funding to longs collect rate while providing partial downside cushion to the anchor.

Net portfolio: long-biased, partially hedged, funding-positive. Designed to throw off carry whether the anchor moves up modestly, sideways, or even slightly down.

Why this shape sidesteps three fleet diseases

What Spider captures that no other Predator does: funding harvest. Over 7-day holds, funding rates dwarf intraday price noise on most alts.

v2.0 fleet-learning constraints

The Apr 2026 fleet hypothesis was concrete:

"Bleeders are over-trading — relying on weak signals and hitting DSL repeatedly. The way to win big is high conviction + longer hold times. Most bleeders have positive gross edge eaten by fees."

Cheetah: $1,317 gross, $1,600 fees, $283 net loss. Roach: $1,720 gross, ~$2,000 fees, $280 net loss. Same pattern across 6+ agents. Spider v2.0 bakes the fix into the runtime:

1. Trade frequency ceiling

Max 12 fills per rolling 7d window. Soft warning at 8. Beyond 8 the agent must justify every additional fill in its rationale. Beyond 12 is technically not blocked but every fill demands written justification — behavioral constraint.

2. Basket member strictness gate

Basket members must individually score ≥ 6.5 to qualify. If fewer than 3 candidates clear the floor, Spider runs anchor-only mode — no basket that week. Better to hold a pure anchor than pad with weak-signal shorts.

3. Fee health monitoring

Every weekly ATTRIBUTION record carries gross_pnl, fees_paid, funding_harvested, and fee_pct_of_gross as the primary health metric. If rolling 4-week fee_pct_of_gross exceeds 25%, a FEE_CEILING_WARNING event fires and the agent must address it in next week's rationale.

4. Forced-exit tracking

Every leg close writes an EXIT event with a typed exit_reason:

If more than 1 safety_stop fires in any rolling 30d window, a SAFETY_STOP_WARNING event fires — the agent has lost control of position sizing or is ignoring score collapses until the hard stop bails it out (the bleeder signature).

The daily decision loop

Each cycle (13:00 UTC, skipping Thursday morning UTC because Arena resets) Spider:

1. Reads four scanner outputs:

- anchor_candidates — top 5 from top-15-volume universe, scored on Arena leader exposure (40%) + SM consensus delta (30%) + funding favorability (15%) + 30d relative strength (15%) - basket_candidates — top 10 from top-50-volume universe, scored on funding descending (40%) + SM rotation negative (25%) + relative weakness vs anchor (20%) + squeeze risk inverse (15%) - portfolio_state — current legs, days held, P&L, funding accrued, fees_paid_per_leg, fills_in_rolling_7d - fleet_overlap — what other Predators are positioned in (concentration awareness, never a hard veto)

1. Outputs one of five decisions with a reported fill_budget_remaining_7d:

1. Writes a DECISION record to state/spider-log.jsonl with full rationale + counterfactuals (alternatives considered + why rejected).

GO_TO_CASH is a first-class output. Spider is allowed — and expected — to refuse to trade when conviction is insufficient.

Cold start — the most important section

Spider does not trade for the first 7 days of its life.

Phase A — Warmup (days 1–7)

Spider boots in GO_TO_CASH. Runs the full decision loop daily. Produces zero trades. Every day it logs to spider-log.jsonl: *"I would have picked anchor=X, basket=[A,B,C,D], conviction=7.2, because..."*

The operator reviews these records before any capital moves. The agent earns trust through observable judgment, not through marketing.

Phase B — First strike

Strike condition (all three required):

1. Warmup complete (7 days observed)
2. Top anchor candidate scores ≥ 7.0 for 2 consecutive daily scans (persistence — never a single-day spike)
3. No catastrophic regime signal (BTC down >10% in 48h, funding regime broadly flipping, vol-regime expansion spike)

When strike fires, Spider enters in three staggered steps:

If at any pilot day the anchor drawdown exceeds 5% or the score collapses below 6.5 → abort: close everything, return to GO_TO_CASH, reset warmup to 3 days, retry.

The first trade Spider ever makes has the worst signal-to-noise ratio of any decision it will ever make. Pilot sizing is the agent encoding epistemic humility into the strategy.

Phase C — Steady state

Day 4 onward. Normal daily decision loop with full positions and real P&L feedback.

Signal source philosophy

Three populations, three uses:

The Predators read is a portfolio-risk discipline, not a quality judgment. Same check a real fund runs: *"before I add HYPE long, what's the firm's existing HYPE exposure across all books?"* Concentration only modifies sizing — it never blocks a trade.

The rationale log

Every scanner output, decision, pilot ramp, exit, abort, weekly attribution, and fleet-learning warning writes a typed JSONL record to state/spider-log.jsonl. The log:

• Survives session clears (lives on disk, not in LLM context)
• Is append-only (no record is ever mutated; attributions reference prior decisions by id)
• Captures counterfactuals on every decision (alternatives + score gap + why rejected)
• Calibrates the conviction threshold during warmup (top score percentiles)
• Surfaces fee, fill, and forced-exit pathologies via dedicated WARNING events
• Powers weekly retrospective scoring (ATTRIBUTION event with gross/fees/funding/net + benchmark comparison)

See references/rationale-log-schema.md for the full event taxonomy.

What's not yet built (and that's the point)

Spider v2.0 is the agent layer. Two runtime additions are needed for it to operate at full quality:

1. dsl_portfolio exit engine — portfolio-level exits with gross_exposure_cap, net_delta_target, weekly_drawdown_circuit. Until this ships, Spider enforces these in the LLM action layer (fragile but functional).
2. Data-layer primitives — funding regime context, vol regime classifier, fleet overlap aggregator, trade journal API. Spider is the first agent designed *for* these abstractions; without them it has to compute equivalents on the fly.

Spider will get strictly better as both ship. No code change required.

Install

mkdir -p /data/workspace/skills/spider-strategy/{config,scripts,state}

gh repo clone Senpi-ai/senpi-skills /tmp/senpi-skills
cp -r /tmp/senpi-skills/spider/* /data/workspace/skills/spider-strategy/

cp /data/workspace/skills/spider-strategy/config/spider-config.example.json \
   /data/workspace/skills/spider-strategy/config/spider-config.json

openclaw runtime start spider-tracker

Operator checklist for the first 7 days

• Day 1: confirm spider-log.jsonl is being written (one SCAN + one DECISION per cycle)
• Day 3: read the rationale text on each DECISION record. Does the agent's reasoning match what you'd want?
• Day 5: check warmup_status() — how many qualifying strike windows has the agent observed?
• Day 7: review WARMUP_END summary. Are the threshold-calibration percentiles in a sane range? If top score never exceeded 6.5, either the threshold is wrong or the regime is bad — both useful facts.
• After day 7: Spider can strike when the next qualifying signal appears. Could be day 8, could be day 30. Patience is the strategy.

Benchmark

Spider only justifies its complexity if it beats *"buy and hold the anchor it picks"* on a risk-adjusted basis. If it doesn't, the basket is noise and we kill it. Track this benchmark explicitly in weekly ATTRIBUTION records (benchmark_buy_and_hold_pnl field).

Secondary success criteria — Spider should also:

• Stay under 12 fills per rolling 7d window (no FILL_CAP_WARNING events)
• Keep fee_pct_of_gross under 25% on rolling 4w (no FEE_CEILING_WARNING events)
• Fire safety_stop no more than once per 30d (no SAFETY_STOP_WARNING events)

If all three guardrails hold, Spider is operating in the Vulture / Kodiak quadrant: low turnover, strict gates, asymmetric payoffs. If any breach, Spider is drifting toward the Cheetah / Roach quadrant: bleeding fees through positive gross edge.