Swing RSI Capitulation — Strategy Control

Strategy Header

identity · status · last backtest

Strategy

Swing RSI Capitulation

4-24h hold · 1.5-3% targets

Mode

LIVE · RUNNING

OKX EEA · signal loop alive

30d Tripwire

POSITIVE_EXTEND

all 3 KPIs clear · sample < 30

Last Backtest

2026-04-21

30d · 12 symbols

Expectancy

+18.35 bp

per probe ($0.13)

PnL (30d)

+$0.0048

20 trades

Strategy Spec

signal · exits · risk · execution

Admission Signal (NEW)

family

RSI mean reversion from capitulation

RSI lookback

60 bars (≈1h equiv)

oversold

RSI < 22 in last 30 bars

trigger

cross up through 30

drift band

−8% ≤ ret_4h ≤ −1%

vol gate

24h range ≥ 3%

cooldown

120 bars (2h) per symbol

Preserved 7-Sprint Stack

fractal regime

≥2/3 scales agree

liquidity inference

score < 0.80

exploration wallet

24h PnL > −$5

pre-probe guard

candle not collapsing

Execution

entry

post-only maker @ best bid · 30m timeout

probe

$0.13 (maker)

add1

$0.10 (maker, +0.5% continuation)

max assembled

$0.23

Exit Stack

TP1

+1.5% · take 75% · maker

BE

+0.02% after TP1 fires

TP2

+3.0% · runner · maker

trail arm

+2.0% · 1.0% giveback

SL

−1.2% · taker market

adverse cap

−2.0% · taker

struct break

close < 4h low − 0.5%

Hold States

mild timeout

12h scratch if flat

max hold

24h hard cap

Swing Tripwire

expectancy-primary · WR secondary

Rule	Threshold	Observed (30d)	Fires
expectancy per probe	> 0 bp	+18.35 bp	clear
TP1 capture rate	≥ 40%	45.00%	clear
win rate	≥ 40%	45.00%	clear
sample size	≥ 30 trades	20	need more

Verdict: POSITIVE_EXTEND_SAMPLE Extend to 90 days to cross sample threshold → EDGE_CONFIRMED on hold.

Backtest · 30d Canonical (v4)

RSI 22/30 · ret_4h ∈ [−8%, −1%]

Signals

26

Fills

20

77% fill rate

Add1 fired

11

55%

TP1

45%

TP2

15%

Timeouts

0

root cause defeated

Exit Reasons

Reason	n	%	Maker/Taker
TP2	3	15%	maker
TRAIL_STOP	1	5%	taker
BE_STOP	5	25%	taker
SL	5	25%	taker
KILL_STRUCT_BREAK_4H	5	25%	taker
SCRATCH_MILD	1	5%	taker
TIMEOUT_FLAT	0	0%	—

Winners hold ~430 min (7h), losers ~250 min (4h). Winners-longer-than-losers ✓ · hold-state logic functioning as designed.

Calibration Trajectory

each pass tightened toward positive

Pass	RSI (ovr/trig)	ret_4h filter	n	TP1	WR	exp (bp)	Verdict
v1	30 / 35	≥ −6%	53	17.0%	17.0%	−60.1	DISABLE
v2	30 / 35	≥ −6%	144	27.1%	31.9%	−43.7	DISABLE
v3	25 / 32	≥ −6%	81	29.6%	33.3%	−16.6	DISABLE
v4 ★	22 / 30	[−8%, −1%]	20	45.0%	45.0%	+18.4	POSITIVE
v5 sens	24 / 32	[−8%, −1%]	26	34.6%	34.6%	−11.9	DISABLE

Each tightening of the signal improved expectancy. v4 converged on positive via deep-oversold RSI + capitulation-band drift filter. v5 sensitivity (loosened RSI by 2 points) dropped expectancy 30 bp — result is parameter-sensitive.

90-Day Validation

in progress · results pending

Full 90d

PENDING

fetching data

Train (first 45d)

PENDING

—

Test (second 45d)

PENDING

—

Pass/fail criteria:

Full 90d: expectancy > 0 bp across ≥ 50 trades → EDGE_CONFIRMED
Train / Test split: both halves must be positive. Positive train + negative test = in-sample fitting, not real edge.
Any failure returns the arm to DISABLE and research continues.

Seven-Arm Research History

swing is the first positive

Arm	n	TP1	Exp (bp)	Verdict
CDV baseline (taker, 45-120m)	10,444	0.3-1.2%	~−19	DISABLE
CDV probe-press	251	40.6%	−13.8	DISABLE
CDV probe-harvest (taker)	1,358	37.2%	−34.4	DISABLE
CDV probe-harvest + maker	2,243	33.2%	−19.7	DISABLE
CDV confirmed-only (pre-entry gate)	103	31.1%	−21.2	DISABLE
OFI continuation (1m microstructure)	112	15.2%	−19.2	DISABLE
Swing RSI capitulation (4-24h)	20	45.0%	+18.4	POSITIVE

Every prior arm landed in the −14 to −34 bp band regardless of execution tweaks or signal direction. Widening the horizon from intraday to swing cleared the band for the first time — confirming the "horizon × cost mismatch" diagnosis.

5-LLM Command Layer

research spec · not wired to live engine

WIRED LIVE · 5/5 LLMs RESPONDING. Full pre-entry cycle (Regime → Signal → Risk → Execution) + post-exit Review runs end-to-end. Latest stub cycle: 4 ideas → 3 risk decisions → 2 approved plans · total cost $0.0076 · latency 21s. Not yet attached to the live trading engine — runs standalone via python openclaw_v1/spot_aggro/llm_sequence/run_sequence.py {status|smoke|cycle|review}.

Operating principle: Use 5 LLMs as a command layer above the bot. One reads market regime, one finds setups, one controls risk, one builds execution plans, and one learns from finished trades — while the bot itself does the actual trading.

LLM 1 — Regime

every 4h

Job

Decide what market you are in: trend up / trend down / chop / high vol / low vol.

Output contract

            REGIME = TREND / CHOP / REVERSAL

            ALLOW_TRADING = YES / NO

            ALLOW_SYMBOLS = [top symbols only]

Why

Prevents the bot from trading the wrong strategy in the wrong market.

LLM 2 — Signal

on candidate

Job

Read the slower setup and build trade ideas: 4h trend, 1d structure, support/resistance, momentum, mean-reversion condition, funding / market context.

Output contract (per symbol)

            side = LONG / SHORT / SKIP

            entry_zone = [lo, hi]

            invalidation = px

            tp1, tp2 = px

            confidence = 0..1

Why

Idea generator. Only runs on symbols LLM 1 allows.

LLM 3 — Risk

on idea

Job

Reject bad ideas. Checks: RR too poor, invalidation too wide, symbol weakness, correlation with open trades, daily-loss cap nearness, same-side crowding.

Output contract

            verdict = APPROVE / REJECT / REDUCE_SIZE

            max_positions = N

            daily_risk_state = GREEN / AMBER / RED

Why

Keeps the bot from doing stupid size and duplicate exposure.

LLM 4 — Execution

on approval

Job

Turn approved idea into an exact execution plan: maker-only, entry ladder, timeouts, post-only flags, TP ladder, stop logic, cancel conditions.

Output contract (example)

            entry_1 = 30% · entry_2 = 30% · entry_3 = 40%

            post_only = YES · timeout = 30m

            cancel_if = price_runs_away

            tp1 = 50% · tp2 = 30% · runner = 20%

Why

Makes the idea executable by the bot.

LLM 5 — Review / Learning

after each exit + weekly

Job

After trade closes, review what happened. Did signal work? Did execution help or hurt? Did regime match? Was stop too tight? Was TP too far? Should the symbol be downgraded?

Output contract

            action = KEEP / DOWNGRADE_SYMBOL / DISABLE_SETUP / ADJUST_THRESHOLD

            weekly_report = markdown text rolling up 7 days

Why

This is how the system improves without guessing.

Full bot sequence (one decision cycle)

Step	Actor	Action	Cadence
1	Bot	Market scan — OHLCV, trend data, symbol stats, volatility, spread/liquidity	continuous
2	LLM 1	Regime verdict · allow_trading flag · allow_symbols list	every 4h
3	LLM 2	Build trade ideas only for allowed symbols	on candidate signal
4	LLM 3	Filter weak / duplicate ideas, set size	on each idea
5	LLM 4	Convert approved idea into exact order plan	on approval
6	Bot	Place order · manage fills · manage stops · manage TP · cancel if invalid	live, no LLM
7	LLM 5	Score the trade · update symbol / setup / regime quality	after exit + weekly roll-up

Deterministic bot owns

exchange connection
price ingestion
order placement
position tracking
stop loss
take profit
daily stop
blacklist
logging

5 LLMs own

regime understanding
signal ranking
risk filtering
execution shaping
post-trade learning

Practical rule. Do not let all 5 LLMs vote on every order — slow and messy. Use: LLMs 1-3 before entry, LLM 4 at order-plan time, LLM 5 after exit. That is the clean sequence.

LLM layer implementation status · WIRED 2026-04-21

LLM 1 · Regime

LIVE

anthropic · claude-haiku-4-5

LLM 2 · Signal

LIVE

gemini · gemini-2.5-flash

LLM 3 · Risk

LIVE

mistral · mistral-small-latest

LLM 4 · Execution

LIVE

openrouter · claude-haiku-4-5

LLM 5 · Review

LIVE

anthropic · claude-haiku-4-5

CLI: python openclaw_v1/spot_aggro/llm_sequence/run_sequence.py status · smoke · cycle · review
Artifacts: runtime/llm_sequence/runs/*.json
Note: OpenAI key in .env returns 401 (rotated/stale). Roles reassigned to 4 working providers; OpenAI slot vacant.

Exact Bot Rules

profit ladder · runner · loss rules · sizing · $10 trade

Principle: Sell most at TP1, protect the trade, then let a small remainder run with a trailing stop. Lock profit early · don't close the whole winner too soon · let best trades keep growing · cut weak trades fast.

TP1 · +1.50%

60% off

Sell 60% of position

Move stop on remaining 40% to:

entry price, OR
entry + 0.20%

TP2 · +2.00%

25% off

If price keeps going and hits +2.00%:

Sell 25%
Keep last 15% running

Runner · 15%

trail 0.50%

Trailing stop = 0.50% below highest:

If price keeps rising → keep holding
If price drops 0.50% from peak → exit

Strength filter — before letting the runner keep going

Only allow the bot to keep the runner if at least 3 of 4 are true:

#	Check	Pass when
1	Higher highs	price still making HH on recent bars
2	Higher lows	price still making HL on recent bars
3	Buy pressure	buyers still stronger than sellers
4	No rejection	no sharp rejection candle / no breakdown

If less than 3 of 4 are true: exit the rest early. Do not wait.

Fast Kill

2 of 4 fire

Exit immediately if any 2 of these happen:

Price breaks structure
Buyers disappear
Momentum dies
Price cannot hold above entry

Hard Stop

−0.80%

If price hits −0.80%:

Exit full trade
No exceptions
No waiting for recovery

Entry sizing — for each $10 trade

Probe

$5

initial entry

Add 1

$3

trade confirmed

Add 2

$2

still strong post-TP1 / breakout

Leg	Size	Condition
Probe	$5	entry fills (maker-only)
Add 1	$3	only when trade is confirmed
Add 2	$2	only when trade is still strong after TP1 zone or breakout continuation

Simple bot logic — full sequence

#	Action	Trigger
1	Enter with $5 (probe)	signal + maker fill
2	Add $3 if trade looks strong	confirmation
3	Add $2 if still strong	TP1 zone / breakout continuation
4	At +1.50%, sell 60%	TP1 hit
5	Move stop to breakeven / small profit	right after TP1
6	At +2.00%, sell 25%	TP2 hit
7	Let last 15% run with 0.50% trailing stop	post-TP2 if strength filter passes
8	If strength disappears, exit immediately	any time <3 of 4 strength checks

Implementation status. These exact rules are not yet applied to the simulator. Current sim uses: probe $0.13 / add1 $0.10 (no add2) · TP1 +1.50% take 75% · BE +0.02% · TP2 +3.0% · SL −1.2%. To apply these rules: scale to $10 trade ($5/$3/$2 split), add TP2 +2.00% (take 25%) with 15% runner, drop TP2 from +3.0% to +2.0%, tighten SL from −1.2% to −0.80%, add 3-of-4 strength filter on runner, add 2-of-4 fast-kill. Will regenerate the 30d + 90d backtest after the rule swap.

One-line version: Sell most at TP1, protect the trade, then let a small remainder run with a trailing stop.

Crypto.com Confirmation Layer

second set of eyes · not the main brain

Principle: Use Crypto.com API as a second set of eyes, not as the main brain. OKX stays the execution venue. Crypto.com answers one question only — "is this move real?"

OKX — Primary / Execution

WebSocket market data + order-book depth (recommended by OKX)
Signal detection
Maker-only entry
TP / SL / trailing management
Position tracking

Crypto.com — Confirmation Only

REST + WebSocket: order book, ticker, candles, mark, index, funding, trades
Pre-entry truth check
Divergence monitoring while in trade
Symbol quality scoring
Data redundancy (pause if feed glitches)

Entry gate — 3 checks before the bot trades

#	Check	Pass rule	If fails
1	Price direction agreement	Both exchanges moving same way	skip
2	Order book agreement	Both books support the same side	skip
3	Divergence check	OKX moving AND CDC moving same direction	skip (fake move)

Entry rule: Trade only when OKX setup is confirmed by Crypto.com. Fewer trades, higher quality.

A · Cross-exchange truth check

If OKX says "go long" but Crypto.com looks weak, the move may be fake. Block entry.

B · Symbol blacklist

If a symbol keeps showing bad agreement between OKX and Crypto.com, block it for the session.

C · Funding / regime filter

Crypto.com provides funding and mark/index streams. Use them to avoid crowded or distorted perp conditions.

D · Data redundancy

If one exchange feed glitches, the bot pauses instead of trading blind.

Do NOT use it for

another random signal
another chart
another API with no filter logic

Use it FOR

confirmation
divergence block
symbol quality scoring
market regime check

Full bot sequence with Crypto.com

Step	Actor	Action
1	OKX	scan for setup
2	Crypto.com	confirm (direction + book + divergence)
3	Gate	pass / fail
4	OKX	execute maker entry
5	Crypto.com	monitor divergence while in trade
6	Bot	exit / blacklist if mismatch grows

Most useful rule. If OKX setup appears, but Crypto.com does not confirm price direction or order-book support, do not trade. That is the cleanest upgrade.

Practical constraints (from Crypto.com docs)

Constraint	Value	Implication
Market-data WS channel	separate from user WS	Run on its own connection, not mixed with order-side traffic
Subscription limit	400 per connection	Stay under 400 symbols × channels; shard if needed
Heartbeat requirement	reply every 30 s	Dedicated heartbeat handler — do not share with execution loop
Role	dedicated confirmation service	Isolate from execution path; pause = skip, not block

Implementation status. Crypto.com candle puller exists in the backtest cache (runtime/backtest/cache/cdc_*_1m.json) and data_puller.fetch_cdc_candles is already implemented. Not yet wired into the swing-rsi live gate — the current backtest runs with pull_cdc=False. Next step: enable CDC fetch in the simulator, add the 3-check entry gate, and re-run 90d validation. WebSocket live-gate layer (divergence monitoring while in trade) is a separate phase.

Layer status

CDC REST candles

AVAILABLE

cache + puller live

CDC entry gate

SPEC

not wired

CDC divergence monitor

SPEC

not wired

CDC WebSocket

SPEC

separate phase

Simple conclusion: Crypto.com API improves your bot by helping it reject fake setups. It should be your confirmation layer, not your main trading layer.

Layer 0 Governance Framework

last-door governor · 3+3 layers · hash-chained

Current State

APPROVED_FOR_PAPER

demo cycle result

L0 Chain

1

verdicts recorded

Test Chain

1

reports recorded

Gov Chain

1

approvals recorded

Testing (L1/L2/L3) measures

Runs numeric checks against hard thresholds. PASS / FAIL / INCONCLUSIVE — deterministic, replayable. No judgement, no discretion.

Governance (G1/G2/G3) judges

Operator-driven APPROVED / REJECTED / HELD on top of test PASS. Cannot APPROVE if test ≠ PASS (hard-preflight). Each layer bound to one maturity step.

Layer 0 ratifies

Final progression authority. Reads evidence, checks artifact pack, emits ONE of 7 verdicts. Hash-chained. No trading, no signals, no orders.

State machine — legal transitions only

RESEARCH  →  APPROVED_FOR_PAPER  →  PAPER  →  APPROVED_FOR_MICRO_LIVE
→  MICRO_LIVE  →  APPROVED_FOR_SCALE  →  SCALED_LIVE
safety: any → FROZEN | DISABLED | ROLLED_BACK

Layer 0 verdicts — the ONLY 7 allowed

Verdict	Effect	Ack required
APPROVE_NEXT_LAYER	advance to next maturity state	YES
APPROVE_SCALE_UP	MICRO_LIVE → APPROVED_FOR_SCALE	YES
HOLD_CURRENT_LAYER	stay, no state change	no
REJECT_AND_REWORK	send back to research/test	no
FREEZE_STRATEGY	any → FROZEN (pause)	YES
DISABLE_LIVE	any → DISABLED (hard stop)	YES
ROLLBACK_TO_SAFE_MODE	any → ROLLED_BACK	YES

3 Testing × 3 Governance layer pairs

L1 Logic / Decision Truth

Proves logic/LLMs/filters/state machine/fee model/tripwires are correct before live routing.

13 hard checks

expectancy ≥ 0 · TP1 ≥ 40% · walkforward both halves +ve · determinism · LLM contracts · tripwire defined

G1 Research Approval

L2 Execution Truth

Paper + live-connected execution. Proves bot handles real order-lifecycle without fakes.

15 hard checks

maker fill ≥ 50% · 0 orphans · 0 dup orders · slippage ≤ 7bp · CDC feed ≥ 95% healthy · live-paper drift ≤ 8bp

G2 Deployment Approval

L3 Capital Truth

Real capital, micro size. Only layer that gates scale-up.

11 hard checks

realized exp ≥ 0 · DD ≤ $5 · streak ≤ 6 · drift ≤ 10bp · regime match ≥ 60% · 0 tripwire breaches

G3 Capital Approval

Artifact pack — required per progression step

Artifact	L1 → Paper	L2 → Micro-live	L3 → Scale	Safety
canonical_config	✓	✓	✓	—
run_manifest	✓	✓	✓	—
test_report	✓	✓	✓	—
governance_verdict	✓	✓	✓	—
tripwire_sheet	✓	—	✓	✓
incident_log	—	✓	✓	✓
audit_trail	—	✓	✓	—
exchange_divergence_report	—	✓	—	✓
symbol_risk_report	—	—	✓	—
scale_recommendation	—	—	✓	—
rollback_record	—	—	—	✓

Failure model — every failure maps to one L0 verdict

Failure	Proposed L0 verdict
test_l1_fail	REJECT_AND_REWORK
test_l2_fail	REJECT_AND_REWORK
test_l3_fail	ROLLBACK_TO_SAFE_MODE
exchange_feed_drift	FREEZE_STRATEGY
okx_cdc_disagreement	FREEZE_STRATEGY
orphan_position	DISABLE_LIVE
duplicate_orders	DISABLE_LIVE
tripwire_breach_hard	DISABLE_LIVE
daily_loss_cap_hit	HOLD_CURRENT_LAYER
live_paper_drift_high	ROLLBACK_TO_SAFE_MODE

Governance layer implementation status · WIRED 2026-04-21

L0

LIVE

last-door

L1 Test

LIVE

13 checks

G1 Gov

LIVE

research

L2 Test

LIVE

15 checks

G2 Gov

LIVE

deployment

L3 Test

LIVE

11 checks

G3 Gov

LIVE

capital

CLI:

python openclaw_v1/spot_aggro/governance_l0/run_governance.py --strategy swing_rsi {status|propose|demo|approve|hold|reject|freeze|disable|rollback|verify}

Artifacts: runtime/governance_l0/artifacts/ and runtime/governance_l0/verdicts/
Blueprint: governance_l0_blueprint_2026-04-21.md

Telegram Notifier

every buy/sell · PnL every 3h

WIRED LIVE. Bot token + chat_id loaded from .env. 4 live sends verified: ping · BUY (SOLUSDT $5 @ 160.2) · SELL (SOLUSDT TP1 +$0.0748) · SELL (DOGEUSDT SL −$0.0446) · 3h PnL rollup. 3-hour PnL loop running in background.

🟢 Every BUY

LIVE

Fires on every entry fill (probe / add1 / add2). Shows symbol, leg, price, size, order_id. Bot calls events.trade_opened(...).

❌ Every SELL

LIVE

Fires on every exit fill (TP1 / TP2 / SL / BE_STOP / TRAIL_STOP / KILL_*). Shows entry→exit, PnL $ + %, hold time, reason. Bot calls events.trade_closed(...).

📈 PnL every 3h

LIVE

Background loop sends rollup every 10800s. Shows last 3h · today (UTC) · since start · WR · best/worst symbol · exit distribution.

Message formats

🟢 BUY · SOLUSDT
swing_rsi · leg=probe · LONG
price: 160.2
size:  $5.00
order: okx-demo-001
17:01:10 UTC

✅ SELL · SOLUSDT
swing_rsi · LONG · TP1
entry → exit: 160.2 → 162.6
pnl: +$0.0748 (+1.50%)
size: $5.00
hold: 7h 15m
17:01:18 UTC

❌ SELL · DOGEUSDT
swing_rsi · LONG · SL
entry → exit: 0.148 → 0.1458
pnl: −$0.0446 (−1.49%)
size: $3.00
hold: 3h
17:01:21 UTC

📈 PnL Report · 3h
swing_rsi  2026-04-20 17:04 UTC

Last 3 hours
  realized: +$0.0302
  trades: 2 close · 1 open
  WR: 50.0% (1W / 1L)
  exits: TP1:1 · SL:1
  best: SOLUSDT (+$0.0748)
  worst: DOGEUSDT (−$0.0446)

Today (UTC)
  realized: +$0.0302

Since start
  realized: +$0.0302
  trades: 2 · WR 50.0%

Background loop status

Channel

LIVE

@claw8403bot

Interval

3h

10800 s

Last send

just now

on-demand + loop

Log

JSONL

runtime/notifier/

        CLI:

        python openclaw_v1/spot_aggro/notifier/run_notifier.py test-send

        python ... send-buy  --symbol BTCUSDT --price 75421 --size 5.0

        python ... send-sell --symbol BTCUSDT --entry 75421 --exit 76550 --size 5.0 --pnl 0.75 --reason TP1 --hold-min 435

        python ... pnl-now

        python ... pnl-loop --interval-s 10800

        Artifacts:

        runtime/notifier/events.jsonl · trades.jsonl · pnl_state.json

Bot integration hook. The bot's deterministic execution layer should import openclaw_v1.spot_aggro.notifier.events and call trade_opened() on every fill and trade_closed() on every exit. That's the only change needed — the notifier owns Telegram + event log + PnL loop. All calls are best-effort (never raise) and fail-open (Telegram failure doesn't block the bot).

Per-Trade Notional

user-editable · live re-read on every entry

Default is $10/trade split across three legs: probe $5, add1 $3, add2 $2. Edit below, press Save, then run the shown command on your host — the executor re-reads the config file on every entry attempt, so the next entry uses the new sizing.

Probe ($)

initial entry — $0.10 to $100

Add 1 ($)

on confirmation — 0 disables

Add 2 ($)

post-TP1 strong — 0 disables

Total per trade

$10.00

hard cap from safety.py still applies: max gross $1000 across all positions

To apply to a running executor (one of):

# option 1 — CLI (easiest):
python openclaw_v1/spot_aggro/live_executor/run_live.py config --set-probe 5 --set-add1 3 --set-add2 2

# option 2 — Download the JSON above and save to:
runtime/live_executor/user_config.json

Note: this panel is static HTML on Cloudflare Pages — it cannot write to your laptop's filesystem directly. Your browser saves values to localStorage for convenience. To apply them to the executor, run the CLI command or download the JSON and drop it in runtime/live_executor/.

Laptop Health + One-Click Flipper

sleep/wake/network alerts · one-click live swap

Events that fire a Telegram message

Event	Trigger	Message
😴 Sleep	Kernel-Power event 42	"Laptop entering sleep" + network/power/executor state
⏰ Wake	Kernel-Power event 107	"Laptop woke from sleep" + current executor state
⚠️ Crash	Kernel-Power event 41	"Unexpected shutdown detected" (ran on next boot)
📡 Network	NIC state change (heartbeat detects)	"Network went DOWN / came back up"
🌐 Internet	ping 1.1.1.1 fails / recovers	"Internet unreachable / reachable again"
🔋 DC	Unplugged from AC	"Switched to battery" — sleep may override powercfg on low battery
🔌 AC	Reconnected to AC	"Back on AC power"
💓 Heartbeat	Every ~3h	"Host heartbeat — all checks ok"
🤖 Executor state	Scheduled Task state change	"Executor task state changed Running → Ready" (etc)

Desktop shortcuts (one-click actions)

🚨 FLIP TO LIVE

Swaps executor from DryRun to Live mode. Requires typing FLIP to confirm. Auto-reads AUTHORIZATION_PHRASE from run_live.py — you never paste it.

FLIP TO LIVE — Swing RSI.lnk

↩️ FLIP TO DRY-RUN

Rolls back to DryRun. No confirmation needed (safer state). Open live positions get flattened on exit.

FLIP TO DRY-RUN — Swing RSI.lnk

🛑 EMERGENCY STOP

Immediately stops executor. Flatten-on-exit cancels open orders + market-sells every position. Fires Telegram before exit.

EMERGENCY STOP — Swing RSI.lnk

⏸️ PAUSE

Block NEW entries. Existing positions keep running — TP/SL/trail/fast-kill all stay active. Creates a PAUSED flag file the executor reads before every probe. Fires Telegram.

PAUSE — Swing RSI.lnk

▶️ RESUME

Delete the PAUSED flag. New entries allowed again within 60s (next signal poll). Fires Telegram.

RESUME — Swing RSI.lnk

🕒 FIX CLOCK (auto-elevates)

Starts Windows Time service + forces NTP sync. Required once if OKX returns error 50102 Timestamp request expired — your clock drifted past OKX's 30s window. Triggers a UAC prompt (needs admin), runs, verifies offset, then exits.

FIX CLOCK — Swing RSI.lnk · script: Fix-SystemClock.ps1

🛡️ Governor · auto-heal + dead-process alerts

LIVE · 60s cadence

Separate 60-second Scheduled Task that cross-checks the executor every minute. Catches the case where Task Scheduler reports State: Running but the Python child has actually died (e.g. Ctrl-C / 0xC000013A). Heartbeat only caught it every 5 min by TaskState; Governor checks the actual Python PID from the wrapper log.

6 liveness checks per tick (short-circuit on first real fail)

#	Check	Fail behavior
1	Task exists	Diagnose `TASK_MISSING` + Telegram
2	PAUSED flag absent	Pause = intentional; governor does nothing
3	Task State = Running	Diagnose `TASK_NOT_RUNNING`
4	Python PID from wrapper log alive	Diagnose `PYTHON_DEAD`
5	Wrapper log age < 6 min	Diagnose `WRAPPER_STALE`
6	Internet reachable (ping 1.1.1.1)	Skip heal; wait for network

Auto-heal ladder (per consecutive-failure streak)

Streak	Action
1	`Start-ScheduledTask`
2	`Stop → Start`
3+	`Stop → purge orphan python procs → Start`

Rate-limit guardrail

Max 4 heal attempts per rolling 60-minute window. When exceeded, governor fires ⚫ Governor: giving up for 1 hour and holds off further heals — prevents restart storms when a deeper problem is loose.

Telegram severity ladder

Trigger	Message
healed (streak 1-2)	🟡 Governor: auto-healed executor
healed (streak 3+)	🔴 Governor: repeated heals — 3×
heal raised	❌ Governor: heal FAILED
healthy after outage	🟢 Governor: executor healthy again
budget exhausted	⚫ Governor: giving up for 1 hour
no internet	📡 Governor: internet unreachable

Live-test result

Killed Python PID 38076 manually at 23:20 UTC → governor detected within 60s → Start-ScheduledTask fired → task back to Running → Telegram 🟡 auto-healed sent. End-to-end ✓.

Script: Governor-Executor.ps1 · Task: SwingRsiGovernor · Audit: runtime/live_executor/logs/governor_YYYY-MM-DD.log · State: runtime/live_executor/governor_state.json

Pause vs Stop — which do I use?
PAUSE = soft. Block new entries, keep managing open trades. Click RESUME to unpause. Good for when you're away for a short stretch or news event.
EMERGENCY STOP = hard. Kill the executor entirely, flatten-on-exit cancels orders + market-sells positions. Use when something looks wrong.

Scheduled Tasks on this laptop

Task	Purpose	Trigger
`SwingRsiExecutor`	Main executor (DryRun default)	at logon + on demand
`SwingRsiExecutorWatchdog`	Restarts main if stale > 10 min	every 5 min
`SwingRsiHealth_Sleep`	Telegram on sleep	Kernel-Power event 42
`SwingRsiHealth_Wake`	Telegram on wake	Kernel-Power event 107
`SwingRsiHealth_Crash`	Telegram on crash-boot	Kernel-Power event 41
`SwingRsiHealth_Heartbeat`	Network/power/executor diff + periodic alive ping	every 5 min

Logs: runtime/live_executor/logs/health_YYYY-MM-DD.log · heartbeat_YYYY-MM-DD.log · wrapper_YYYY-MM-DD.log
State: runtime/live_executor/health_state.json (previous values for change detection)
Scripts: openclaw_v1/spot_aggro/live_executor/deploy/windows/*.ps1

v5 Ruleset · 20 symbols · 1H + 4H · 3-7 trades/day target

pullback · breakout-retest · maker-first

LIVE with v5 predicates. Signal loop now fetches 1H bars (300 × 1H = 12.5 days), applies 1H-native direction filter, then tries Setup A (Pullback) then Setup B (Breakout-Retest). Full 4H-aggregated version available in simulator_swing_v5.py for backtests.

Universe — 20 liquid spot pairs

        BTC · ETH · SOL · XRP · ADA · DOGE · LINK · AVAX · SUI · DOT

        LTC · BCH · TRX · UNI · APT · NEAR · ATOM · TON · PEPE · WIF

Setup A — Pullback Swing Long

Use when trend is already up.

4H trend	bullish
1H RSI zone	38-45 (aggro 35-38)
1H close	strong (upper third)
invalidation	no breakdown under prev low
reject	if RSI already rebounded >52 in last 3 bars

Setup B — Breakout-Retest

Use when market already broke upward.

4H structure	bullish
breakout	24h high × 1.025 in last 12h
retest zone	prior 24h high ±0.5%
1H RSI	45-55, turning up
retest holds	low ≥ prior_24h_high × 0.995

Direction filter (4H bias, 1H-native proxy)

price vs 80-bar SMA	must be above
ret_48h (1H)	≥ −2%
last 1H upper wick	< 2% of close
1H RSI (proxy for 4H)	≥ 40

Sizing + exits (operator spec)

Probe	$0.13
Add 1	$0.08 (on confirmation)
Add 2	$0.05 (on continuation)
Max assembled	$0.26
TP1	+1.5% · sell 60%
TP2	+2.0% · sell 25%
Runner	15% · trail 0.5% below high
Hard SL	−0.8%
Hold	1H: 6-24h · 4H: 1-3 days

Frequency target: 3-7 trades/day across the full 20-sym portfolio. Do not force 5 from one chart — come from the whole panel. Some days less, active days more. If you don't reach 5: add more liquid symbols first, then add shorts — don't weaken the setup.

Implementation status

Layer	v5 status	Note
Universe (20 syms)	LIVE	`run_live.py UNIVERSE`
Setup A predicate	LIVE	1H-native in `signal_loop.evaluate_signal`
Setup B predicate	LIVE	1H-native breakout+retest
Direction filter	LIVE	1H-native proxy (80-SMA + ret_48h + wick + RSI)
OKX candle fetch	LIVE	switched from 1m to 1H bars · 300 × 1H = 12.5d
Sizing $0.13/$0.08/$0.05	PARTIAL	Probe from `user_config.json` ($5/$3/$2 default); adjust via Per-Trade Notional card
TP1 +1.5% take 60% · BE entry+0.20%	LIVE	`order_manager.py` · v5 constants
TP2 +2.0% take 25% (of total)	LIVE	`_take_tp2_partial()` added; leaves 15% runner
Runner 15% · trail 0.5% below high	LIVE	Trail arms on TP2 hit; `TRAIL_GIVEBACK_PCT=0.005`
Hard SL −0.8%	LIVE	`SL_PCT=-0.008`
Fast-kill 2-of-4	LIVE	`_evaluate_fast_kill()` · 4 triggers × 6-min window · pre-TP1 only
Full 4H aggregation (simulator)	LIVE	`simulator_swing_v5.py` predicates from 1m cache
Full 4H backtest runner	LIVE	`run_swing_v5.py` · results in v5 Backtest card below
Symbol blacklist	PENDING	v5 strike rules per spec — next pass

Live status: full v5 ruleset (A+B predicates + direction filter + 20-sym universe + TP1 60% / TP2 25% / runner 15% trail 0.5% / SL −0.8% / fast-kill 2-of-4) is LOADED in the executor. Python PID 4940 restarted 23:34 UTC with v5 order_manager. First v5 trade will fire Telegram + place a real OKX order with the v5 exit stack. Backtest card below reports cache-based walk-forward.

v5 → v6 → v7 Backtest · walk-forward · 9/20 symbols cached · 90d

v7 single-variable change from v6: FLAT_TIME_EXIT_MIN=60 (was 18) + 3 guards (price<entry AND no-reclaim AND weak-momentum). All other v6 settings held constant (TP1 +0.8% / TP2 +1.4% / SL −1.0% / fast-kill 8min / Setup A 40-48 / Setup B 1.5% / 1.0%).

Metric	v5	v6	v7	v6→v7
n trades	61	103	101	flat
Trades/day (9-sym)	0.76	1.29	1.26	flat ✓
TP1 capture	21.3%	15.5%	40.6%	+25.1 pp ✓
TP2 capture	9.8%	6.8%	17.8%	+11.0 pp ✓
Win rate	21.3%	15.5%	40.6%	+25.1 pp ✓
Setup A fires	61	103	101
Setup B fires	0	0	0	freeze — no breakouts in cache
Expectancy (bp/probe)	−41.0	−18.3	−18.8	flat ✗
Total PnL	−$0.0325	−$0.0245	−$0.0247	flat

Exit distribution — v5 vs v6 vs v7

Reason	v5 %	v6 %	v7 %	v6→v7
FLAT_TIME_EXIT	—	55%	13%	−42 pp ✓
SL	57%	8%	25%	+17 pp (trades held longer hit SL)
BE_STOP	11%	9%	23%	+14 pp (more reaching BE)
TRAIL_STOP	10%	7%	18%	+11 pp (more runners)
FAST_KILL	21%	21%	21%	unchanged ✓

v7 thesis: 3 of 4 outcomes delivered.
✓ FLAT_TIME_EXIT rate fell below 55% → dropped to 13% (−42 pp)
✓ TP1 rate rose above 15.5% → jumped to 40.6% (+25 pp)
✓ Trades/day stayed similar → 1.29 → 1.26
✗ Expectancy did NOT improve — stayed at −18.8 bp (flat vs v6 −18.3)

Why: extending the time window let trades breathe — TP1 captures more than doubled, runners took +11 pp more trail wins, BE saves climbed +14 pp. But the trades that would have been flat-exited at 18m now split between TP1 (good) and SL (bad) — the two cancel at the PnL level.

Per operator's pre-committed decision rule: "If v7 still fails, then even with stop fixed and time fixed, Setup A still does not generate enough move." → Setup A signal quality is the final blocker. Setup B confirmed 0/0/0 across v5/v6/v7 — freeze as planned. Next decision: disable Setup B fully, freeze this swing panel, move to a different signal family or wider horizon.

Raw: runtime/backtest/runs/1776730155920_v5_swing.json

Live Executor

staged · two-key launch · not running

LIVE RUNNING · 2026-04-21 02:30 UTC. Scheduled Task SwingRsiExecutor in -Mode Live, Python PID alive inside the signal loop. Pre-flight passed: OKX creds ✓ · Telegram creds ✓ · OKX EEA public ping 91ms ✓ · authenticated balance call ✓. Two bugs fixed this session to reach this state: (1) clock drift +207s → Fix-SystemClock.ps1 synced NTP, offset now +4ms; (2) global www.okx.com returned 50119 API key doesn't exist because the account is on OKX EEA shard — client now reads OKX_HOSTNAME=eea.okx.com from .env.

What's built

okx_client.py — REST + HMAC auth, dry_run default, audit log
signal_loop.py — 1m candle poll, swing_rsi evaluator (reuses sim predicate)
order_manager.py — probe/add1/add2, TP1 maker (60%), TP2 / trail / SL taker
safety.py — $1000 gross cap, −$3 daily cap, −$15 session hard-disable, 5 concurrent max, 120s per-symbol cooldown
run_live.py — CLI: status · dry-run · live --authorize "…"
Notifier wired: every BUY, SELL, ERROR, gated-entry fires Telegram
HOST_247.md — VPS setup + systemd unit + watchdog + kill-switch

What's NOT built (v1 limits)

WebSocket fill confirmation (uses REST polling)
CDC cross-exchange gate (spec only — sec 7.7)
5-LLM gate on live entries (spec only — sec 7.5)
Partial-fill reconciliation vs exchange truth
24/7 host (your laptop sleeps — see HOST_247.md)

Hard safety caps (in code, not config)

Cap	Value	Behavior
max gross exposure	$1000	new entries refused if total open notional would exceed
daily loss cap	−$3	all new entries blocked until next UTC day (open positions managed)
session hard-disable	−$15	full shutdown + flatten-on-exit
max concurrent positions	5	6th entry refused
per-symbol cooldown	120s	no re-entry same symbol within 2 min
flatten on exit	ON	SIGINT/SIGTERM → cancel all + market-sell positions

Two-key launch

run_live.py will NOT place real orders unless you pass --live --authorize "…" with the exact phrase below. Default is dry-run.

          # 1. Check credentials + reachability (read-only)

          python3 openclaw_v1/spot_aggro/live_executor/run_live.py status

          # 2. Dry-run 24-72h (zero capital risk, real candles, real Telegram)

          python3 openclaw_v1/spot_aggro/live_executor/run_live.py dry-run --poll-s 60

          # 3. ONLY after (2) passes, promote to live:

          python3 openclaw_v1/spot_aggro/live_executor/run_live.py live \

              --authorize "<LITERAL PHRASE — see run_live.py constant AUTHORIZATION_PHRASE on the host>" \

              --poll-s 60

Op-sec note: the actual literal authorization string is defined in openclaw_v1/spot_aggro/live_executor/run_live.py as the constant AUTHORIZATION_PHRASE. It is intentionally not published in this public panel. Edit that constant to rotate.

24/7 hosting (required)

Your laptop will sleep. A live executor needs an always-on host. Three docs cover the end-to-end launch:

deploy/bootstrap.sh — one-shot Ubuntu VPS bootstrap
deploy/swing-rsi-executor.service — systemd unit (DRY-RUN default)
deploy/watch-executor.sh — 5-min watchdog + Telegram alerts
deploy/env.example — all required env vars
RUNBOOK.md — exact 10-step launch sequence
GAPS.md — v1 limitations, honestly
HOST_247.md — manual setup notes

Component status

OKX adapter

BUILT

REST + auth

Signal loop

BUILT

1m poll

Order mgr

BUILT

TP/SL/trail

Safety rails

BUILT

6 caps

Smoke tested

PASSED

auth ✓ · EEA shard · ping 91ms

Running

LIVE

signal loop · 60s poll

Artifacts: runtime/live_executor/logs/dry_run_YYYY-MM-DD.jsonl, live_YYYY-MM-DD.jsonl, safety_state.json
Governance: This arm has NOT passed L1/L2/L3 testing or G1/G2/G3 governance. Deploy under operator override, explicit rule skip recorded in session ledger.

Actions

research mode only

Decision tree

If 90d full shows expectancy > 0 bp across ≥ 50 trades AND walk-forward test half stays positive → enable paper trade at $0.13 probe for 14 days.
If paper confirms expectancy > 0 across ≥ 20 trades with live fills matching maker assumption → graduate to $1 probe notional.
If 90d fails or walk-forward test half is negative → result was noise; return to signal research.
Never deploy without both 90d extension AND walk-forward validation.