.title { text-wrap: balance } Spending for October, generated by piping hledger → R Over the past six months, I’ve tracked my money with hledger—a plain text double-entry accounting system written in Haskell. It’s been surprisingly painless. My previous attempts to pick up real accounting tools floundered. Hosted tools are privacy nightmares, and my stint with GnuCash didn’t last. But after stumbling on Dmitry Astapov’s “Full-fledged hledger” wiki1, it clicked—eventually consistent accounting. Instead of modeling your money all at once, take it one hacking session at a time. It should be easy to work towards eventual consistency. […] I should be able to [add financial records] bit by little bit, leaving things half-done, and picking them up later with little (mental) effort. – Dmitry Astapov, Full-Fledged Hledger Principles of my system I’ve cobbled together a system based on these principles: Avoid manual entry – Avoid typing in each transaction. Instead, rely on CSVs from the bank. CSVs as truth – CSVs are the only things that matter. Everything else can be blown away and rebuilt anytime. Embrace version control – Keep everything under version control in Git for easy comparison and safe experimentation. Learn hledger in five minutes hledger concepts are heady, but its use is simple. I divide the core concepts into two categories: Stuff hledger cares about: Transactions – how hledger moves money between accounts. Journal files – files full of transactions Stuff I care about: Rules files – how I set up accounts, import CSVs, and move money between accounts. Reports – help me see where my money is going and if I messed up my rules. Transactions move money between accounts: 2024-01-01 Payday income:work $-100.00 assets:checking $100.00 This transaction shows that on Jan 1, 2024, money moved from income:work into assets:checking—Payday. The sum of each transaction should be $0. Money comes from somewhere, and the same amount goes somewhere else—double-entry accounting. This is powerful technology—it makes mistakes impossible to ignore. Journal files are text files containing one or more transactions: 2024-01-01 Payday income:work $-100.00 assets:checking $100.00 2024-01-02 QUANSHENG UVK5 assets:checking $-29.34 expenses:fun:radio $29.34 Rules files transform CSVs into journal files via regex matching. Here’s a CSV from my bank: Transaction Date,Description,Category,Type,Amount,Memo 09/01/2024,DEPOSIT Paycheck,Payment,Payment,1000.00, 09/04/2024,PizzaPals Pizza,Food & Drink,Sale,-42.31, 09/03/2024,Amazon.com*XXXXXXXXY,Shopping,Sale,-35.56, 09/03/2024,OBSIDIAN.MD,Shopping,Sale,-10.00, 09/02/2024,Amazon web services,Personal,Sale,-17.89, And here’s a checking.rules to transform that CSV into a journal file so I can use it with hledger: # checking.rules # -------------- # Map CSV fields → hledger fields[0] fields date,description,category,type,amount,memo,_ # `account1`: the account for the whole CSV.[1] account1 assets:checking account2 expenses:unknown skip 1 date-format %m/%d/%Y currency $ if %type Payment account2 income:unknown if %category Food & Drink account2 expenses:food:dining # [0]: <https://hledger.org/hledger.html#field-names> # [1]: <https://hledger.org/hledger.html#account-field> With these two files (checking.rules and 2024-09_checking.csv), I can make the CSV into a journal: $ > 2024-09_checking.journal \ hledger print \ --rules-file checking.rules \ -f 2024-09_checking.csv $ head 2024-09_checking.journal 2024-09-01 DEPOSIT Paycheck assets:checking $1000.00 income:unknown $-1000.00 2024-09-02 Amazon web services assets:checking $-17.89 expenses:unknown $17.89 Reports are interesting ways to view transactions between accounts. There are registers, balance sheets, and income statements: $ hledger incomestatement \ --depth=2 \ --file=2024-09_bank.journal Revenues: $1000.00 income:unknown ----------------------- $1000.00 Expenses: $42.31 expenses:food $63.45 expenses:unknown ----------------------- $105.76 ----------------------- Net: $894.24 At the beginning of September, I spent $105.76 and made $1000, leaving me with $894.24. But a good chunk is going to the default expense account, expenses:unknown. I can use the hleger aregister to see what those transactions are: $ hledger areg expenses:unknown \ --file=2024-09_checking.journal \ -O csv | \ csvcut -c description,change | \ csvlook | description | change | | ------------------------ | ------ | | OBSIDIAN.MD | 10.00 | | Amazon web services | 17.89 | | Amazon.com*XXXXXXXXY | 35.56 | l Then, I can add some more rules to my checking.rules: if OBSIDIAN.MD account2 expenses:personal:subscriptions if Amazon web services account2 expenses:personal:web:hosting if Amazon.com account2 expenses:personal:shopping:amazon Now, I can reprocess my data to get a better picture of my spending: $ > 2024-09_bank.journal \ hledger print \ --rules-file bank.rules \ -f 2024-09_bank.csv $ hledger bal expenses \ --depth=3 \ --percent \ -f 2024-09_checking2.journal 30.0 % expenses:food:dining 33.6 % expenses:personal:shopping 9.5 % expenses:personal:subscriptions 16.9 % expenses:personal:web -------------------- 100.0 % For the Amazon.com purchase, I lumped it into the expenses:personal:shopping account. But I could dig deeper—download my order history from Amazon and categorize that spending. This is the power of working bit-by-bit—the data guides you to the next, deeper rabbit hole. Goals and non-goals Why am I doing this? For years, I maintained a monthly spreadsheet of account balances. I had a balance sheet. But I still had questions. Spending over six months, generated by piping hledger → gnuplot Before diving into accounting software, these were my goals: Granular understanding of my spending – The big one. This is where my monthly spreadsheet fell short. I knew I had money in the bank—I kept my monthly balance sheet. I budgeted up-front the % of my income I was saving. But I had no idea where my other money was going. Data privacy – I’m unwilling to hand the keys to my accounts to YNAB or Mint. Increased value over time – The more time I put in, the more value I want to get out—this is what you get from professional tools built for nerds. While I wished for low-effort setup, I wanted the tool to be able to grow to more uses over time. Non-goals—these are the parts I never cared about: Investment tracking – For now, I left this out of scope. Between monthly balances in my spreadsheet and online investing tools’ ability to drill down, I was fine.2 Taxes – Folks smarter than me help me understand my yearly taxes.3 Shared system – I may want to share reports from this system, but no one will have to work in it except me. Cash – Cash transactions are unimportant to me. I withdraw money from the ATM sometimes. It evaporates. hledger can track all these things. My setup is flexible enough to support them someday. But that’s unimportant to me right now. Monthly maintenance I spend about an hour a month checking in on my money Which frees me to spend time making fancy charts—an activity I perversely enjoy. Income vs. Expense, generated by piping hledger → gnuplot Here’s my setup: $ tree ~/Documents/ledger . ├── export │   ├── 2024-balance-sheet.txt │   └── 2024-income-statement.txt ├── import │   ├── in │   │   ├── amazon │   │   │   └── order-history.csv │   │   ├── credit │   │   │   ├── 2024-01-01_2024-02-01.csv │   │   │   ├── ... │   │   │   └── 2024-10-01_2024-11-01.csv │   │   └── debit │   │   ├── 2024-01-01_2024-02-01.csv │   │   ├── ... │   │   └── 2024-10-01_2024-11-01.csv │   └── journal │   ├── amazon │   │   └── order-history.journal │   ├── credit │   │   ├── 2024-01-01_2024-02-01.journal │   │   ├── ... │   │   └── 2024-10-01_2024-11-01.journal │   └── debit │   ├── 2024-01-01_2024-02-01.journal │   ├── ... │   └── 2024-10-01_2024-11-01.journal ├── rules │   ├── amazon │   │   └── journal.rules │   ├── credit │   │   └── journal.rules │   ├── debit │   │   └── journal.rules │   └── common.rules ├── 2024.journal ├── Makefile └── README Process: Import – download a CSV for the month from each account and plop it into import/in/<account>/<dates>.csv Make – run make Squint – Look at git diff; if it looks good, git add . && git commit -m "💸" otherwise review hledger areg to see details. The Makefile generates everything under import/journal: journal files from my CSVs using their corresponding rules. reports in the export folder I include all the journal files in the 2024.journal with the line: include ./import/journal/*/*.journal Here’s the Makefile: SHELL := /bin/bash RAW_CSV = $(wildcard import/in/**/*.csv) JOURNALS = $(foreach file,$(RAW_CSV),$(subst /in/,/journal/,$(patsubst %.csv,%.journal,$(file)))) .PHONY: all all: $(JOURNALS) hledger is -f 2024.journal > export/2024-income-statement.txt hledger bs -f 2024.journal > export/2024-balance-sheet.txt .PHONY clean clean: rm -rf import/journal/**/*.journal import/journal/%.journal: import/in/%.csv @echo "Processing csv $< to $@" @echo "---" @mkdir -p $(shell dirname $@) @hledger print --rules-file rules/$(shell basename $$(dirname $<))/journal.rules -f "$<" > "$@" If I find anything amiss (e.g., if my balances are different than what the bank tells me), I look at hleger areg. I may tweak my rules or my CSVs and then I run make clean && make and try again. Simple, plain text accounting made simple. And if I ever want to dig deeper, hledger’s docs have more to teach. But for now, the balance of effort vs. reward is perfect. while reading a blog post from Jonathan Dowland↩︎ Note, this is covered by full-fledged hledger – Investements↩︎ Also covered in full-fledged hledger – Tax returns↩︎

Luckily, I speak Leet. – Amita Ramanujan, Numb3rs, CBS’s IRC Drama There’s an episode of the CBS prime-time drama Numb3rs that plumbs the depths of Dr. Joel Fleischman’s1 knowledge of IRC. In one scene, Fleischman wonders, “What’s ‘leet’”? “Leet” is writing that replaces letters with numbers, e.g., “Numb3rs,” where 3 stands in for e. In short, leet is like the heavy-metal “S” you drew in middle school: Sweeeeet. / \ / | \ | | | \ \ | | | \ | / \ / ASCII art version of your misspent youth. Following years of keen observation, I’ve noticed Git commit hashes are also letters and numbers. Git commit hashes are, as Fleischman might say, prime targets for l33tification. What can I spell with a git commit? DenITDao via orlybooks) With hexidecimal we can spell any word containing the set of letters {A, B, C, D, E, F}—DEADBEEF (a classic) or ABBABABE (for Mama Mia aficionados). This is because hexidecimal is a base-16 numbering system—a single “digit” represents 16 numbers: Base-10: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 16 15 Base-16: 0 1 2 3 4 5 6 7 8 9 A B C D E F Leet expands our palette of words—using 0, 1, and 5 to represent O, I, and S, respectively. I created a script that scours a few word lists for valid words and phrases. With it, I found masterpieces like DADB0D (dad bod), BADA55 (bad ass), and 5ADBAB1E5 (sad babies). Manipulating commit hashes for fun and no profit Git commit hashes are no mystery. A commit hash is the SHA-1 of a commit object. And a commit object is the commit message with some metadata. $ mkdir /tmp/BADA55-git && cd /tmp/BAD55-git $ git init Initialized empty Git repository in /tmp/BADA55-git/.git/ $ echo '# BADA55 git repo' > README.md && git add README.md && git commit -m 'Initial commit' [main (root-commit) 68ec0dd] Initial commit 1 file changed, 1 insertion(+) create mode 100644 README.md $ git log --oneline 68ec0dd (HEAD -> main) Initial commit Let’s confirm we can recreate the commit hash: $ git cat-file -p 68ec0dd > commit-msg $ sha1sum <(cat \ <(printf "commit ") \ <(wc -c < commit-msg | tr -d '\n') \ <(printf '%b' '\0') commit-msg) 68ec0dd6dead532f18082b72beeb73bd828ee8fc /dev/fd/63 Our repo’s first commit has the hash 68ec0dd. My goal is: Make 68ec0dd be BADA55. Keep the commit message the same, visibly at least. But I’ll need to change the commit to change the hash. To keep those changes invisible in the output of git log, I’ll add a \t and see what happens to the hash. $ truncate -s -1 commit-msg # remove final newline $ printf '\t\n' >> commit-msg # Add a tab $ # Check the new SHA to see if it's BADA55 $ sha1sum <(cat \ <(printf "commit ") \ <(wc -c < commit-msg | tr -d '\n') \ <(printf '%b' '\0') commit-msg) 27b22ba5e1c837a34329891c15408208a944aa24 /dev/fd/63 Success! I changed the SHA-1. Now to do this until we get to BADA55. Fortunately, user not-an-aardvark created a tool for that—lucky-commit that manipulates a commit message, adding a combination of \t and [:space:] characters until you hit a desired SHA-1. Written in rust, lucky-commit computes all 256 unique 8-bit strings composed of only tabs and spaces. And then pads out commits up to 48-bits with those strings, using worker threads to quickly compute the SHA-12 of each commit. It’s pretty fast: $ time lucky_commit BADA555 real 0m0.091s user 0m0.653s sys 0m0.007s $ git log --oneline bada555 (HEAD -> main) Initial commit $ xxd -c1 <(git cat-file -p 68ec0dd) | grep -cPo ': (20|09)' 12 $ xxd -c1 <(git cat-file -p HEAD) | grep -cPo ': (20|09)' 111 Now we have an more than an initial commit. We have a BADA555 initial commit. All that’s left to do is to make ALL our commits BADA55 by abusing git hooks. $ cat > .git/hooks/post-commit && chmod +x .git/hooks/post-commit #!/usr/bin/env bash echo 'L337-ifying!' lucky_commit BADA55 $ echo 'A repo that is very l33t.' >> README.md && git commit -a -m 'l33t' L337-ifying! [main 0e00cb2] l33t 1 file changed, 1 insertion(+) $ git log --oneline bada552 (HEAD -> main) l33t bada555 Initial commit And now I have a git repo almost as cool as the sweet “S” I drew in middle school. This is a Northern Exposure spin off, right? I’ve only seen 1:48 of the show…↩︎ or SHA-256 for repos that have made the jump to a more secure hash function↩︎

.title {text-wrap:balance;} GIT - the stupid content tracker “git” can mean anything, depending on your mood. – Linus Torvalds, Initial revision of “git”, the information manager from hell Like most git features, gitcredentials(7) are obscure, byzantine, and incredibly useful. And, for me, they’re a nice, hacky solution to a simple problem. Problem: Home directories teeming with tokens. Too many programs store cleartext credentials in config files in my home directory, making exfiltration all too easy. Solution: For programs I write, I can use git credential fill – the password library I never knew I installed. #!/usr/bin/env bash input="\ protocol=https host=example.com user=thcipriani " eval "$(echo "$input" | git credential fill)" echo "The password is: $password" Which looks like this when you run it: $ ./prompt.sh Password for 'https://thcipriani@example.com': The password is: hunter2 What did git credentials fill do? Accepted a protocol, username, and host on standard input. Called out to my git credential helper My credential helper checked for credentials matching https://thcipriani@example.com and found nothing Since my credential helper came up empty, it prompted me for my password Finally, it echoed <key>=<value>\n pairs for the keys protocol, host, username, and password to standard output. If I want, I can tell my credential helper to store the information I entered: git credential approve <<EOF protocol=$protocol username=$username host=$host password=$password EOF If I do that, the next time I run the script, it finds the password without prompting: $ ./prompt.sh The password is: hunter2 What are git credentials? Surprisingly, the intended purpose of git credentials is NOT “a weird way to prompt for passwords.” The problem git credentials solve is this: With git over ssh, you use your keys. With git over https, you type a password. Over and over and over. Beleaguered git maintainers solved this dilemma with the credential storage system—git credentials. With the right configuration, git will stop asking for your password when you push to an https remote. Instead, git credentials retrieve and send auth info to remotes. On the labyrinthine options of git credentials My mind initially refused to learn git credentials due to its twisty maze of terms that all sound alike: git credential fill: how you invoke a user’s configured git credential helper git credential approve: how you save git credentials (if this is supported by the user’s git credential helper) git credential.helper: the git config that points to a script that poops out usernames and passwords. These helper scripts are often named git-credential-<something>. git-credential-cache: a specific, built-in git credential helper that caches credentials in memory for a while. git-credential-store: STOP. DON’T TOUCH. This is a specific, built-in git credential helper that stores credentials in cleartext in your home directory. Whomp whomp. git-credential-manager: a specific and confusingly named git credential helper from Microsoft®. If you’re on Linux or Mac, feel free to ignore it. But once I mapped the terms, I only needed to pick a git credential helper. Configuring good credential helpers The built-in git-credential-store is a bad credential helper—it saves your passwords in cleartext in ~/.git-credentials.1 If you’re on a Mac, you’re in luck2—one command points git credentials to your keychain: git config --global credential.helper osxkeychain Third-party developers have contributed helpers for popular password stores: 1Password pass: the standard Unix password manager OAuth Git’s documentation contains a list of credential-helpers, too Meanwhile, Linux and Windows have standard options. Git’s source repo includes helpers for these options in the contrib directory. On Linux, you can use libsecret. Here’s how I configured it on Debian: sudo apt install libsecret-1-0 libsecret-1-dev cd /usr/share/doc/git/contrib/credential/libsecret/ sudo make sudo mv git-credential-libsecret /usr/local/bin/ git config --global credential.helper libsecret On Windows, you can use the confusingly named git credential manager. I have no idea how to do this, and I refuse to learn. Now, if you clone a repo over https, you can push over https without pain3. Plus, you have a handy trick for shell scripts. git-credential-store is not a git credential helper of honor. No highly-esteemed passwords should be stored with it. This message is a warning about danger. The danger is still present, in your time, as it was in ours.↩︎ I think. I only have Linux computers to test this on, sorry ;_;↩︎ Or the config option pushInsteadOf, which is what I actually do.↩︎

Humans do no operate on hexadecimal symbols effectively […] there are exceptions. – Dan Kaminsky When SSH added ASCII art fingerprints (AKA, randomart), the author credited a talk by Dan Kaminsky. As a refresher, randomart looks like this: $ ssh-keygen -lv -f ~/.ssh/id_ed25519.pub 256 SHA256:XrvNnhQuG1ObprgdtPiqIGXUAsHT71SKh9/WAcAKoS0 thcipriani@foo.bar (ED25519) +--[ED25519 256]--+ | .++ ... | | o+.... o | |E .oo=.o . | | . .+.= . | | o= .S.o.o | | o o.o+.= + | | . . .o B * | | . . + & . | | ..+o*.= | +----[SHA256]-----+ Ben Cox describes the algorithm for generating random art on his blog. Here’s a slo-mo version of the algorithm in action: ASCII art ssh fingerprints slo-mo algorithm But in Dan’s talk, he never mentions anything about ASCII art. Instead, his talk was about exploiting our brain’s hardware acceleration to make it easier for us to recognize SSH fingerprints. The talk is worth watching, but I’ll attempt a summary. What’s the problem? We’ll never memorize SHA256:XrvNnhQuG1ObprgdtPiqIGXUAsHT71SKh9/WAcAKoS0—hexadecimal and base64 were built to encode large amounts of information rather than be easy to remember. But that’s ok for SSH keys because there are different kinds of memory: Rejection: I’ve never seen that before! Recognition: I know it’s that one—not the other one. Recollection: rote recall, like a phone number or address. For SSH you’ll use recognition—do you recognize this key? Of course, SSH keys are still a problem because our working memory is too small to recognize such long strings of letters and numbers. Hacks abound to shore up our paltry working memory—what Dan called “brain hardware acceleration.” Randomart attempts to tap into our hardware acceleration for pattern recognition—the visiuo-spacial sketchpad, where we store pictures. Dan’s idea tapped into a different aspect of hardware acceleration, one often cited by memory competition champions: chunking. Memory chunking and sha256 The web service what3words maps every three cubic meters (3m²) on Earth to three words. The White House’s Oval Office is ///curve.empty.buzz. Three words encode the same information as latitude and longitude—38.89, -77.03—chunking the information to be small enough to fit in our working memory. The mapping of locations to words uses a list of 40 thousand common English words, so each word encodes 15.29 bits of information—45.9 bits of information, identifying 64 trillion unique places. Meanwhile sha256 is 256 bits of information: ~116 quindecillion unique combinations. 64000000000000 # 64 trillion (what3words) 115792089237316195423570985008687907853269984665640564039457584007913129639936 # 116 (ish) quindecillion (sha256) For SHA256, we need more than three words or a dictionary larger than 40,000 words. Dan’s insight was we can identify SSH fingerprints using pairs of human names—couples. The math works like this1: 131,072 first names: 17 bits per name (×2) 524,288 last names: 19 bits per name 2,048 cities: 11 bits per city 17+17+19+11 = 64 bits With 64 bits per couple, you could uniquely identify 116 quindecillion items with four couples. Turning this: $ ssh foo.bar The authenticity of host 'foo.bar' can't be established. ED25519 key fingerprint is SHA256:XrvNnhQuG1ObprgdtPiqIGXUAsHT71SKh9/WAcAKoS0. Are you sure you want to continue connecting (yes/no/[fingerprint])? Into this2: $ ssh foo.bar The authenticity of host 'foo.bar' can't be established. SHA256:XrvNnhQuG1ObprgdtPiqIGXUAsHT71SKh9/WAcAKoS0 Key Data: Svasse and Tainen Jesudasson from Fort Wayne, Indiana, United States Illma and Sibeth Primack from Itārsi, Madhya Pradesh, India Maarja and Nisim Balyeat from Mukilteo, Washington, United States Hsu-Heng and Rasim Haozi from Manali, Tamil Nadu, India Are you sure you want to continue connecting (yes/no/[fingerprint])? With enough exposure, building recognition for these names and places should be possible—at least more possible than memorizing host keys. I’ve modified this from the original talk, in 2006 we were using md5 fingerprints of 160-bits. Now we’re using 256-bit fingerprints, so we needed to encode even more information, but the idea still works.↩︎ A (very) rough code implementation is on my github.↩︎