A transparent record of how we design, build, test, break, and fix our software. From Anubis to a unified platform — no cherry-picking, the mistakes stay in.
What is Pantheon
One Install. All Deities.
Sirsi Pantheon is a unified DevOps intelligence platform. It brings together infrastructure hygiene, QA/QC governance, and persistent AI knowledge into a single, lightweight binary. Install once — get every deity.
𓂀 Anubis is the foundational module — the reason Pantheon exists. Weigh. Judge. Purify. It scans, judges, and purges waste across workstations with 58 rules, ghost app detection, file dedup that's 27× faster than naive hashing, and a policy engine for fleet enforcement.
𓆄 Ma'at governs quality — test coverage thresholds, canon document verification, and CI pipeline monitoring. 57 tests across 3 governance domains.
𓁟 Thoth gives AI assistants persistent memory — 98.7% less code to read per session, $4.08 saved per conversation.
Built in Go. Released under the MIT license. Runs entirely on your device — no data ever leaves your machine.
Our Philosophy
We looked inside the binary and found the truth: all 20 modules compile to 186 KB. That's 1.5% of the binary. The other 98.5% is Go's runtime, paid once. Adding a new deity costs 2–30 KB. Pantheon isn't a rename — it's a recognition that the architecture was already unified.
The Data Behind the Decision
When we analyzed exactly what's inside the binary, we discovered something remarkable. All 20 deity modules — 13,813 lines of application code — compile to just 186 KB. That's 1.5% of the 12 MB binary.
The insight: The Go runtime taxes every binary ~10 MB regardless of what's inside. Once you pay that cost, each additional module is nearly free. Doubling the module count would add ~100 KB. The binary already was a unified platform — calling it "Anubis" was limiting what it actually is.
Anubis isn't being demoted. It's the foundational module — the reason Pantheon exists. Every scan rule, every safety protection, every feature started with Anubis. Pantheon is what happens when one deity proves the architecture works and the others join in.
All sizes measured via go tool nm -size and ls -lh on darwin/arm64, March 2026.
Every sprint, every problem, every solution — documented as it happened.
The idea started with a 3-hour manual cleanup of a Mac that recovered 47 GB of wasted space. Virtual machines, old AI models, dead app leftovers — all hiding in plain sight. "Why am I doing this by hand?"
Built the scan engine with 58 rules across 7 categories — developer tools, AI models, virtual machines, IDEs, cloud services, storage, and general waste. Added ghost hunting: finding leftovers from apps you already uninstalled.
Added the smart features: a neural file classifier that understands what files are (not just their extension), an AI IDE integration so coding assistants can use Anubis, and a policy engine for teams to enforce hygiene standards.
Built the file deduplication engine — finds duplicate files across your Mac. Started with a simple approach that read every file completely. During testing, the folder picker didn't work at all.
Deep review of every component. The insight: "measure twice, cut once" — read just the beginning and end of each file instead of the whole thing. This made scanning 27 times faster.
Added trash-first deletion — files always go to your Mac's Trash (you can "Put Back"), with a complete audit log of every action.
Built Thoth — a system that gives AI coding assistants persistent memory. Instead of re-reading thousands of lines of code every conversation (expensive and slow), the AI reads a compact summary. 98% less wasted computation.
Named after the Egyptian god of knowledge. Free and open source — works with any project, any language.
Hit the "90% weighted average" wall. Realized we couldn't test signal handling or root kills without actually killing real processes. Introduced the Interface Injection Standard (Rule A16) and established the Antigravity IPC Bridge.
Now we can verify Safety Design 100% deterministically. If a root-kill fails, we know exactly how Pantheon reacts, without ever risking the host system.
Found along the way: an ARP parsing edge case where macOS "incomplete" entries could confuse the IP extractor, a comment-count mismatch in the rule registry, and confirmed that every default profile includes the "general" scan category (good product design).
Verified all platform builds with GoReleaser snapshot — 12 binaries across macOS, Linux, and Windows (both Intel and ARM). All within size budget. Updated launch copy, investor demo script, and all public-facing stats.
Released v0.4.0-alpha with Homebrew tap integration —
brew tap SirsiMaster/tools && brew install sirsi-pantheon.
GoReleaser auto-publishes the formula to homebrew-tools/Formula/ on every tagged release.
During release execution, discovered that Ma'at's pre-push hook took 55 seconds per push — running full test coverage across all 17 packages even when only 1-2 files changed. After 4 pushes (3 min 40 sec of waiting), implemented diff-based coverage.
.gitignore pattern pantheon was silently ignoring cmd/pantheon/seba.go
(matching any path containing "pantheon"). CI builds were failing since the Seba rebrand.
Fix: anchor with /pantheon to match root-level binary only.
IDE click latency hit 400ms+. Diagnosed to LSP processes (gopls, TypeScript) consuming P-core CPU time.
Built Sekhmet renice — deprioritizes LSPs to Background QoS via pantheon guard --renice lsp.
Protected language_server_macos_arm from accidental slay. Implemented scoped Horus indexing (856K→50K files).
Built the Pantheon VS Code Extension — 5 TypeScript modules replacing the JS scaffold.
Always-on Guardian (auto-renice every 60s), Ankh status bar (live RAM/CPU from ps every 5s),
7 Command Palette entries, and Thoth context compression from .thoth/memory.yaml.
Deployed the Deity Registry to Firebase Hosting at sirsi-pantheon.web.app. Wired custom domain pantheon.sirsi.ai via Firebase Hosting API + GoDaddy CNAME. Rebuilt the registry index with click-to-flip 3D cards showing developer info on the back.
Real numbers. No inflation.
Measured on a real folder with 709 files. Not a synthetic test.
How it works: instead of reading entire files (which can be gigabytes), we read just the first and last 4 KB. If those 8 KB match, then — and only then — do we read the full file. Most non-duplicates are eliminated instantly.
Case Study · We Eat Our Own Dog Food
We built Thoth because our own AI sessions kept failing. Then we measured the difference.
The problem: Without Thoth, our 23,177-line codebase consumed 139% of the AI's context window just to read at startup. The source code literally didn't fit. The AI had to read selectively, creating blind spots that caused hallucinated function names, duplicated code, and missed safety rules.
The result: Thoth compresses project knowledge into 300 lines of curated facts. The AI starts productive in 10 seconds instead of 3–5 minutes, with 137% more context available for real work. In our best session, we wrote 150 tests and made 11 commits — impossible without context preservation.
All metrics from real development sessions, verified via wc -l and git log.
Read the full case study →
Dogfooding · Process Improvement Before Release
We found a product-killing performance bug by using our own tool. Fixed it in the same session.
The discovery: During v0.4.0-alpha release execution, every
git push triggered Ma'at, which ran go test -cover ./... across
all 17+ packages. After 4 pushes (3 min 40 sec of waiting), the insight landed:
a governance tool that destroys productivity is worse than no governance at all.
The fix: Diff-based coverage.
Ma'at now uses git diff to detect which packages have changed .go files
and only tests those. Unchanged packages use cached coverage from
~/.config/pantheon/maat/coverage-cache.json.
Same verdicts, same accuracy — 1,500× faster.
The lesson: Most teams discover performance issues after users complain. Pantheon found and fixed this during internal development — a direct benefit of building in public and eating our own dog food.
Measured on Apple M1 Max, macOS Tahoe. All numbers independently verifiable. Read the full case study →
Architecture · Every Deity Wins If One Deity Wins
Three deities were independently walking the same directories. Now they share one index.
The recursive win: Horus indexes the filesystem once with parallel goroutines → Jackal queries the index in 5ms instead of walking for 15s → Ka queries in 1s instead of walking for 8.5s → faster pre-push gate → more pushes per session → more dogfooding → more performance discoveries → better product. The improvements compound. Every deity wins if one deity wins.
Phase 2 coming: macOS FSEvents, Linux fanotify, and Windows USN Journal will make the index incremental — tracking only what changed since the last scan. Target: 200–500ms for any deity.
Dogfooding · Product Thesis Validated
While analyzing Weigh output, we discovered that 97.6% of all detected waste was one unused application.
The recursive insight: Ma'at was slow → fixed it → enabled more pushes → discovered Weigh was slow → fixed Weigh → analyzed output → found 64 GB Docker ghost → removed it. The performance work enabled the disk discovery that validated the entire product.
This is exactly what Pantheon is built to do: surface hidden waste in 833ms, give actionable context (rule name, path, size), and let the user decide. The founder didn't know Docker was consuming 64 GB until Pantheon told him.
We'd rather tell you what isn't ready than let you discover it.
Part of the Sirsi Ecosystem
Pantheon is the unified DevOps intelligence platform from Sirsi Technologies. Every deity maintains its identity while sharing a single runtime and a single install.
Anubis (infrastructure hygiene), Ma'at (QA/QC governance), Thoth (persistent knowledge) — and future deities in any language. The architecture proved itself: adding a module costs 2–30 KB. The platform is ready.