/quote and discover.py priced a trade nobody would ever get (#183)

/doctor: stop asking a question /observe already answered (#179)

/ai: show what the bot has learned, not three example questions (#177)

Zero warnings on a correctly-configured box (#175)

/heal could not see the one thing blocking every trade (#171)

A false blocker is worse than a warning (#170)

Never fail silently, never name the wrong cause, and watch yourself (#169)

The 897ms was the sweep, plus 4 real bugs from review (#167)

Make the box ready for Yellowstone before the endpoint exists

Call him 35-bot, give him four routes and six requests a second

Two hops, and stop crying wolf about two venues that cannot be fixed

The bot was not losing auctions. It was outbidding itself out of them.

Your RPC key was in the logs, and a dead node held requests for 20s (#161)

The Jito tip was sized to fail the bot's own margin gate (#160)

/doctor: stop raising a problem from a number nothing applies (#159)

Stop the send path queueing behind the scanner for the same token bucket (#158)

Give the tip auction a memory, so /jito stops describing uptime (#155)

Retire the Orca and Meteora adapters on evidence, and probe what works (#152)

Make the warm cache stay warm, and stop the verifier inventing bad news (#150)

Stop reporting inferred gaps as money, and start recording history (#149)

Add /lighthouse: eleven screens, one verdict, one next action (#146)

Stop /doctor hiding what /why shows, and add the send-path benchmark (#145)

Break the feedback loop: freeze off the hot path, verified Apply, /decisions (#144)

Say what the percentage floor actually costs (#142)

Let the survey add what it finds, and stop repeating a read warning (#140)

Fix /tune, stop quoting a dead cause as live, price the flash fee into the pair survey (#137)

Make it trade: the blackout, the free flash loan, the noise, and the pipeline (#136)

Say which haircut is actually in force

Add /stats, make /start the buttons only, stop warning about correct behaviour

Learn the leg-2 haircut instead of guessing it

Measure decay, not the loan cap: the /decay numbers were a size artifact

/decay: report the recent slice, so a fix is visible before the p90 catches up

Telegram: Solana-first keyboard, compact /status, and delete a scary lie

/capture: of the signals we find, what share do we take, and where is the rest lost

Lock the bot to its operator, fix aws-secrets, and explain the no-trade

Failure-mode logging, drawdown alerts, route transparency, HTTP guard

Fix the 14-of-14 pruning outage; measure edge decay; public site for bot.elghaly.dev

/doctor: check the wallet that actually signs; rewrite /hunt in the operator's layout

Solana engine: route pruning, dynamic Jito tip, circuit breaker, key custody

Add /doctor Telegram command: plain-language "can it trade, and why not" diagnosis (#129)

Add plain-language trade journal report (/journal command + CLI)

AWS-standalone migration: drop Worker relay, add Solana RPC/WS failover (#124)

Retire the ARB (EVM) leg behind ARB_SCAN_ENABLED=false (#119)

Give Solana the same grounded, small-action AI copilot Arbitrum already has (#106)

Ground /ai's Solana claims: real leg numbers in context + anti-fabrication rules

Ground /ai_idea's proposed token address against the live chain (#70)

Ground /ai's flash-loan fee rate against the real configured value (#69)

Ground /ai's prompt against real bot state to reduce hallucinated advice (#66)

Add /testlog to test the payout/log Telegram channel on demand (#42)

Fix commands typed directly into a Telegram channel getting no response (#39)