Wiki Log, April 2026

[2026-04-30] ingest | Awesome DeepSeek Agent (deepseek-ai curated list)

Ingested github.com/deepseek-ai/awesome-deepseek-agent — official deepseek-ai org's curated list of 16 AI agent / coding-assistant tools with per-tool guides for using DeepSeek-V4-Pro and DeepSeek-V4-Flash. The Old New Thing URL passed alongside this was already covered by the previous batch as itanium-too-few-parameters and was skipped to avoid duplicate work. The list splits along three lines: (a) mainstream coding agents that bolt onto any OpenAI-API-compatible endpoint (Claude Code, GitHub Copilot, Copilot CLI, Kilo Code, OpenCode, Crush, Langcli); (b) explicitly DeepSeek-aware terminal agents that tune around V4's behaviors — long <think> blocks, reasoning-effort knob, the trained-in style markers documented in deepseek-v4-roleplay-instruct (Pi, Oh My Pi, Reasonix, Deep Code); (c) China-facing chat-platform agents on Feishu/WeChat/Telegram (OpenClaw, AstrBot, WorkBuddy/CodeBuddy, nanobot). The China-facing tier is the most distinctive — it tells you DeepSeek treats Chinese work-messaging surfaces as a first-class deployment target in a way Western LLM ecosystem catalogs don't. Reasonix is the only entry that bills itself "DeepSeek-native" with a cache-first loop and MCP-native architecture. Worth flagging that the list contains a tool literally called OpenClaw — distinct from the existing wiki page build-your-own-openclaw, which is a tutorial on building a clone of an OpenClaw-style agent, not the actual project. Entries arrive via PR with a working guide following an existing structure, so being on the list is a soft signal of "this works against V4 today" rather than a quality benchmark. Cross-referenced into deepseek entity (added Agent ecosystem section, bumped updated). Pi family already tracked as pi-coding-agent. Pages created: awesome-deepseek-agent (summary). Pages updated: deepseek, index, log, ingested-urls.

[2026-04-30] ingest | Seven-source batch — DOS history, Itanium ABI, Zig anti-AI, SOB benchmark, PS5 Linux, GooseRelay, GTFOBins

Ingested seven sources in one batch on 2026-04-30. (1) opensource.microsoft.com — Microsoft's announcement of the DOS-History/Paterson-Listings MIT licensing: Yufeng Gao + Rich Cini's multi-year recovery effort scanning, OCR'ing, and transcribing Tim Paterson's printed source listings of 86-DOS 1.00, several PC-DOS 1.00 dev snapshots, CHKDSK, and the MASM assembler itself. Joshua Carsbrook developed printer-listing-specific OCR for the dense assembler printouts. Paterson preserved hand-written notes alongside code, framed as "a printed commit history of a Git repository." Donated to Interim Computer Museum. Earliest DOS source recovered to date — predates the 2018 1.25 release by about a year. Pages created: early-dos-paterson-listings (summary). Cross-referenced from sdl3-dos-port. (2) The Old New Thing — Raymond Chen's post on what happens when you call a function with too few parameters. Three failure modes catalogued: stack imbalance under callee-clean conventions, parameter-slot scratch reuse (compiler eliding a dead-local into the missing parameter's slot, then writing into the caller's frame), and uninitialized register reads on register-passing ABIs. Itanium gets the dedicated treatment: NaT-bit propagation through speculative loads can crash the function when an unused parameter's address is taken (NaT consumption exception on spill), and the register-stack engine architecturally enforces parameter count — reading registers outside the current frame is undefined, writing them required to raise Illegal Operation. The standard "calling with wrong arity is UB" claim made architectural rather than conventional. Pages created: itanium-too-few-parameters (summary), oldnewthing-blog (entity for Raymond Chen). (3) Simon Willison on Zig's anti-LLM policy and the Bun fork. The visible cost: Bun's 4× compile-speed improvement (parallel semantic analysis, multiple LLVM codegen units) explicitly will not be upstreamed because Zig has a strict ban on LLM-authored contributions. Loris Cro's "contributor poker" essay frames the rationale: review time is investment in growing trustworthy contributors, not just landing artifacts; an LLM-authored PR receives the investment without producing the long-run benefit. SW adds: if a maintainer can fire up their own LLM on the same problem, why review the PR? Cleanest articulation seen of an LLM-blanket-ban that survives even if LLMs become indistinguishable from senior engineers. Pages created: simonw-zig-anti-ai (summary), contributor-poker (concept), simon-willison-blog (blog entity). Pages updated: zig (added anti-LLM section + sources). (4) Interfaze's Structured Output Benchmark (SOB) — open benchmark for LLM structured output across text/image/audio with seven metrics, 21 models, schema-complexity-weighted leaderboard. Headline finding is the JSON-Pass-vs-Value-Accuracy gap: every frontier model clears 95%+ on JSON Pass but exact leaf-value accuracy sits 15-30 points lower (Schematron-8B has the worst gap at 25.6 pp, Qwen3.5-35B the tightest at 16.8 pp). Schema-only benchmarks have been hiding a third of the actual error rate. Audio is much harder than text/image (best Value Accuracy 23.7% vs 83.0%). Per-modality leaders differ — GLM-4.7 leads text, Gemma-4-31B images, Gemini-2.5-Flash audio. Interfaze-Beta sits at #6 on the unified leaderboard. The "structured hallucination" failure mode (type-correct, schema-valid, plausible, wrong) is the hardest to catch — guardrails don't help. Pages created: structured-output-benchmark (summary), json-pass-value-accuracy-gap (concept), interfaze (entity). (5) Tom's Hardware on TheFlow's public release of ps5-linux-loader — Andy Nguyen's hypervisor exploit + Ubuntu 24.04 build script + M.2 install tools + fan/boost utility, working on PS5 Phat firmware 3.xx-4.xx. GTA V Enhanced 60 FPS at 1440p with ray tracing was the March demo. Boot path is fake-DNS + manual-page-redirect + TCP payload + power-button on solid-orange LED → solid-white LED. Soft mod, must be re-run each boot. Internal SSD never modified. Lands in the same window as the sdl3-dos-port and early-dos-paterson-listings retro-meets-modern thread, plus hold-on-to-your-hardware (DRAM/GPU shortage) and Valve's Steam Machine delay — used PS5 Phat on old firmware becomes a budget Linux gaming PC. Pages created: ps5-linux-loadr (summary), ps5-linux-loader (toolbox). (6) GooseRelayVPN — third entry in the Apps-Script-as-circumvention-relay cluster after masterhttprelayvpn and masterhttprelayvpn-rust. Architecturally different from those: tunnels raw TCP via SOCKS5, end-to-end AES-256-GCM (Google never sees plaintext), but requires a small VPS exit so a real net.Dial can happen somewhere. Trades the "no infrastructure" appeal of MasterHttpRelay for being able to carry SSH/IMAP/anything. ~30-line Apps Script forwarder is genuinely dumb — only handles the encrypted bytes, never decrypts. Multi-deployment scaling with health-aware blacklist + same-poll failover; README warns against more than 3-4 deployment IDs. Wire format spelled out (frame layout, batch seal, base64 HTTP body). Single-author (kianmhz), Persian-language README, Iran-facing audience. Watchlist with re-check 2026-07-29 — want to see if independent contributors arrive, mobile clients ship, AES integration gets reviewed. Pages created: goose-relay-vpn (toolbox). Pages updated: index, watchlist. (7) GTFOBins — long-running curated list of Unix executables abusable for shell escape, privilege escalation, file transfer, and post-exploitation. Joint effort by Emilio Pinna and Andrea Cardaci with hundreds of community contributors via GitHub PRs. ~400+ binaries from awk/find/vim through docker/kubectl/terraform/aws to modern entries like opencode/codex. Bidirectional security primitive — attacker reference and sysadmin checklist for the same techniques. Has JSON API and MITRE ATT&CK Navigator integration. Explicit that it's not a list of exploits — the binaries do exactly what they were designed to do; this is configuration-level abuse. Created concept page for the broader "living off the land" pattern (legitimate pre-installed tools instead of dropped malware), since LotL keeps coming up adjacent to credential-management content (cf. long-lived-keys, ephemeral-credentials). Pages created: gtfobins (entity), living-off-the-land (concept). Pages updated: supply-chain-security (cross-references to LotL/GTFOBins as the runtime analogue, bumped updated to 2026-04-30). Index, log, and ingested-urls all updated.

[2026-04-30] ingest | A Text Editor as a User Interface (ratfactor)

Ingested ratfactor.com/cards/text-editor-as-ui — Dave Gauer card-post on the gap between CLI flags and a full TUI: hand the user a temp file and shell out to $EDITOR. Precedents are well-known but rarely articulated as a pattern: crontab -e, git commit, visudo, vipw. Two examples in the post — a three-line shell script that pipes an editor session through rev, and a Ruby yt-dlp wrapper that persists settings across runs (download.txt, not mktemp), uses a # subdirA subdirB ... comment as inline documentation for valid choices, and timestamp-suffixes the output filename to avoid sequence-number bookkeeping. Key design notes Gauer draws out: full editor expressiveness for one line of script, no documentation needed because the user already knows their editor, persistence-by-default beats mktemp for personal tools, switch to .ini/.toml once parameters exceed three or four lines. Pulled out two reusable concept ideas: editor-as-UI as a named pattern (with the validate-then-re-edit idiom from crontab -e/visudo, the comments-as-docs idiom, the persist-vs-mktemp decision, and a "when it fits" / "when it doesn't" rubric) and the digital-garden framing of his card-format site. Created an entity for ratfactor — Gauer's site is a working example of the digital-garden format and likely to come up again; footer carries an explicit anti-LLM-training declaration which is part of his stance. Cross-references thin in the existing vault — closest is clean-code-coding-agents (same minimum-context, lean-on-existing-affordances instinct from the LLM angle). Pages created: text-editor-as-ui (summary), editor-as-ui (concept), ratfactor (blog entity). Pages updated: index, log.

[2026-04-30] ingest | vera (programming language designed for LLMs)

Ingested github.com/aallan/vera — Alasdair Allan's programming language designed for LLMs to write, not for humans. Premise: LLM code-gen failures are mostly not syntactic but coherence-over-scale, and naming-related (misleading names, reused names, lost references — cites arXiv 2307.12488). Three load-bearing design choices: (1) No variable names. Bindings are referenced by typed De Bruijn index — @Int.0 is the most recent Int in scope, @Int.1 the one before, types disambiguate so adding a new Int doesn't shift what @String.0 means. (2) Mandatory contracts. Every function has requires, ensures, effects between signature and body. Z3 proves Tier-1 (decidable arithmetic, comparisons, booleans, ADTs, termination) statically; undecidable contracts become Tier-3 runtime checks. (3) Algebraic effects in signatures. effects(<Http, Inference>) on the function arrow, callers must permit the full effect row, pure is the default and enforced. Compiles to WebAssembly via wasmtime, with mandatory parity tests between CLI and browser bundles. Diagnostics are written for the model that wrote the code: every error has a stable code (E001–E702), an explanation, a concrete fixed example, and a spec reference, plus --json output for agent feedback loops. Repo ships SKILL.md, AGENTS.md, CLAUDE.md, DE_BRUIJN.md as separate agent-onboarding documents. VeraBench (companion repo, 50 problems × 5 tiers × 6 models): Kimi K2.5 reportedly hits 100% run_correct vs 86% Python / 91% TypeScript; flagship-tier 93% Vera vs 93% Python — single-run high-variance, but the early signal is favorable. Status: v0.0.127, 810+ commits, 127 releases, 261 stars, MIT, single author (Alasdair Allan, the IoT/maker writer). Repo ~10 weeks old, very active. Added to watchlist with re-check 2026-07-29 — single author, deliberately swimming against ecosystem gravity, want to see if anyone besides Allan writes nontrivial Vera, if VeraBench gets independent reproductions, and if the verified-MCP-server roadmap milestone lands. Created a concept page for typed De Bruijn indices since the idea is independent of Vera and worth referencing on its own (relates to vamp-ai-frontend, clippy-stricter-config, clean-code-coding-agents — three different axes of the same "languages designed against LLM failure modes" question). Pages created: vera (toolbox), typed-de-bruijn-for-llms (concept). Pages updated: index, watchlist, index, log.

[2026-04-30] ingest | Your CEO is suffering from AI psychosis (HandyAI)

Ingested handyai.substack.com/p/your-ceo-is-suffering-from-ai-psychosis — a power-user newsletter (Cursor/Claude Code daily) arguing that executive AI enthusiasm has decoupled from output. Two named cases anchor the post: (1) Garry Tan (YC CEO) open-sourced gstack two days before SXSW, claimed 37,000 LOC/day across five projects, hit 20k stars; developer Gregorein audited the actual output and found Tan's website was making 169 server requests vs HN's 7, shipping 28 test files to production, loading 78 unused JS controllers, serving uncompressed 2MB PNGs that should be 300KB, and including an empty 0-byte file with a rich-text editor on a read-only page. On the SXSW panel Tan said he'd been sleeping four hours/night and called it "cyber psychosis." (2) Karpathy on No Priors said he was "in a state of psychosis" and hadn't written a line of code since December. Both being held up as patterns to imitate. The platforms feeding this: Paperclip (30k stars, "operating system for AI organizations" with Board-of-Directors metaphor), Autoflowly (3-agent CTO/CMO/CFO startup-from-prompt), AgentShelf, Alacritous ($3000/mo), RuFlow (60+ pre-built agents) — all "project management theater performed by language models." Hard data: NBER study of 6,000 CEOs/CFOs shows ~90% of firms report zero AI productivity impact; average employee uses 1.5 hr/wk, average CEO under 1 hr/wk; 1-in-5 AI investments has ROI, 1-in-50 transformational, 95% of pilots fail. NYT coined "tokenmaxxing" — internal leaderboards for token consumption (OpenAI engineer at 210B tokens/week; Anthropic Claude Code user at $150k/mo; Shopify and Meta made AI usage a perf-review factor). Stanford-in-Science sycophancy study tested 11 models — affirm user actions 49% more than humans, including for deception/harm/illegal acts; 2,400-person follow-up showed sycophancy makes users more convinced, less skeptical, less empathetic, more dependent, while rating sycophantic responses as more trustworthy. Author's prescription is PM-flavored: spec before agent, measure output not activity, kill the token leaderboard, audit the agent fleet, stay skeptical of own enthusiasm. Connects strongly to existing ai-great-leap-forward (executive-suite version of the corporate-mandate failure mode), subprime-ai-crisis / ai-bubble-pale-horses / ai-subsidy-economics (macro-finance), no-silver-bullet-llms (DORA-style productivity decay numbers), cult-of-vibe-coding / peril-of-laziness-lost (engineering-discipline angle), average-is-all-you-need (optimistic counterpoint). Created garry-tan as an entity since the gstack episode is a recurring reference point. Pages created: ceo-ai-psychosis (summary), tokenmaxxing (concept), ai-sycophancy-loop (concept), garry-tan (entity), handyai-blog (blog entity). Pages updated: ai-great-leap-forward (cross-ref), index, log.

[2026-04-30] ingest | Your Clippy config should be stricter (emschwartz.me)

Ingested emschwartz.me/your-clippy-config-should-be-stricter/ — Evan Schwartz (runs Scour, the personalized-feed product) lost his Friday email-digest job to a Tokio panic: a UTF-8-oblivious &s[..200] truncation hit a non-character-boundary byte. The post argues for tightening Clippy beyond defaults to recover the "if it compiles, it works" property, especially relevant in the coding-agent era where seasoned engineers' reflexes don't apply. Key meta-points: (1) Clippy's category groupings don't map cleanly to "don't panic in production" — the restriction category is documented as "emphatically not to be enabled wholesale" because it contains contradictory lints; selective opt-in is the only sensible path. (2) Lints that don't fire today are still useful as tripwires for tomorrow. (3) Workspace inheritance is opt-in per crate (lints.workspace = true); use cargo-workspace-lints or a CI script. (4) Test relaxation lives in clippy.toml (allow-unwrap-in-tests = true, etc). (5) warn + cargo clippy -- -D warnings on CI beats deny for local-iteration ergonomics. Six categorized lint groups: don't-panic (string_slice, indexing_slicing, unwrap_used, panic, todo, unimplemented, unreachable, get_unwrap, unwrap_in_result, unchecked_time_subtraction, panic_in_result_fn); don't-fail-silently (let_underscore_future, let_underscore_must_use, unused_result_ok, map_err_ignore, assertions_on_result_states); async (await_holding_lock, await_holding_refcell_ref, if_let_mutex pre-2024, large_futures); unsafe-memory (mem_forget, undocumented_unsafe_blocks, multiple_unsafe_ops_per_block); numbers (float_cmp, lossy_float_literal, cast_sign_loss, invalid_upcast_comparisons); easy wins (rc_mutex, dbg_macro, infallible_try_from); and the agent-specific allow_attributes + allow_attributes_without_reason combo that forces explicit #[expect(..., reason = "…")]. Caveats: expect_used is borderline (the .expect message often documents the invariant); arithmetic_side_effects was 85% noise / 15% signal in his codebase. References the November 2025 Cloudflare unwrap outage as the high-profile reminder. Connects to existing clean-code-coding-agents from a different direction — narrowing bad-pattern templates is the same work as tight function signatures. Pages created: clippy-stricter-config (summary), clippy (entity), emschwartz-blog (blog entity). Pages updated: index, log.

[2026-04-30] ingest | A Haskeller's case for Zig (Pure Systems)

Ingested pure-systems.org/posts/2026-04-29-functional-programmers-need-to-take-a-look-at-zig.html — a 10+ year Haskeller's evaluation of Zig along three axes he uses for new languages: (1) domain expressiveness / amount of noise, (2) type-system programmability for correct-by-construction systems, (3) "mean-free path" — vacuum-physics borrowing for "lines of code before the program diverges from your mental model." Argues GC was a 1995-machine tradeoff (CPU and memory comparable then; CPU now ~10,000× faster while memory access lagged), so the forest-of-pointers pattern caps performance and breeds developers who never learned what programs do on the machine. References Stephen Diehl's "where will the next language come from" — academia/industry/hobbyists all lack the incentive — and credits Zig + Andrew Kelley with actually trying. The body walks through Haskell ↔ Zig translations: newtype = singleton struct; sum types = union(enum) (with an explicit Maybe(T) comptime function as illustration); typeclasses = comptime functions taking a type, asserting @hasDecl for required methods, returning a dictionary struct (clunky on purpose — explicit dispatch is "no spooky action at a distance"). The interesting hot take is that Zig 0.16's new IO interface (passing Io and allocator via init) is independent rediscovery of Haskell's Reader/IO monad shape — author frames monads as "algebraic description of imperative programming as a computational context," not arcane math. (Summary flags this framing as a stretch — passing an io parameter is dependency injection unless you also have do-notation and composition — but the underlying observation about explicit effect carriers holds.) New blog worth tracking: pure-systems.org has the FP-meets-systems angle. No prior Zig, comptime, Haskell, GC, or language-design pages — this lays foundation. Pages created: zig-functional-programmers (summary), mean-free-path-language (concept), comptime (concept), zig (entity), pure-systems-blog (blog entity). Pages updated: index, log.

[2026-04-30] ingest | Linux 7.0 cuts PostgreSQL throughput in half (The Coder Cafe)

Ingested read.thecoder.cafe/p/linux-broke-postgresql — Salvatore Dipietro (AWS) posted to LKML on 2026-04-03 that the same pgbench simple-update workload on a 96-vCPU Graviton4 halved in throughput between Linux 6.x (98,565 TPS) and Linux 7.0 (50,751 TPS), with perf showing 56% of CPU spinning in s_lock from PostgreSQL's StrategyGetBuffer. Root cause: Linux 7.0 removed PREEMPT_NONE on modern arches, leaving PREEMPT_LAZY (introduced 6.12) as the de-facto server default. PostgreSQL's StrategyGetBuffer holds a single global spinlock while reading/writing the shared buffer pool — and at 120 GB shared_buffers × 4 KB pages, the pool has up to 31 million potential first-touch faults sprinkled throughout a run. Under PREEMPT_NONE the faulting holder kept running until release. Under PREEMPT_LAZY the scheduler can preempt it mid-fault, so the lock stays held for fault duration + scheduler delay, multiplied by every backend currently spinning (here, hundreds). The article's fix is huge pages: 2 MB cuts faults to ~61,440, 1 GB to ~120, and TLB pressure drops enough that StrategyGetBuffer runs at full speed. Tradeoffs: pre-allocation reservation + internal fragmentation. Peter Zijlstra (Intel, authored the preemption change) proposed PostgreSQL adopt Restartable Sequences (rseq); the community pushed back on principle ("don't break userspace") and on cost-shifting. No prior pages on Linux preemption, spinlocks, huge pages, or PostgreSQL. Cross-referenced from existing linux-kernel-pgit. Created blog entity for The Coder Cafe — it's a Substack with one-concept-per-post systems content (past posts: soft/hard lockup, instruction pipelining, SMT) and likely to come up again. Pages created: linux-7-postgres-regression (summary), linux-preemption-models (concept), spinlock (concept), huge-pages (concept), postgresql (entity), the-coder-cafe (blog entity). Pages updated: linux-kernel-pgit (cross-ref), index, log.

[2026-04-29] ingest | Your Terminal is Burning Battery Like It's Mining Bitcoin (frr.dev)

Ingested frr.dev/posts/terminal-gpu-battery-macbook-ghostty-iterm2/ — Fernando's blog post on a concrete macOS battery problem hit while running Claude Code. Activity Monitor's Energy tab showed Ghostty at 3,600 accumulated units over 12 hours vs Brave 125, Zoom 99, Claude desktop 46 — a terminal pulling 30× more than a browser. Two compounding causes: (1) every character that hits the screen runs through the GPU rendering pipeline, and Claude Code streams output continuously (spinners, log lines, tool results); (2) animated spinners defeat App Nap because the OS classifies the window as active. Three fixes ranked by effort: stock Terminal.app (CPU rendering, naps correctly, ugly but works); iTerm2 with a "Battery" profile that disables GPU rendering, optionally auto-switched on pmset -g batt via \033]50;SetProfile=Battery\a in .zshrc; Ghostty has no GPU toggle, only mitigations (cursor-style-blink = false, minimize windows). 1978 framing: a VT100+CRT pulled 30W; a 2026 MacBook Air exceeds that to display git status. Highly relevant to the user (uses Claude Code in terminal). No prior terminal pages in vault — created summary plus a concept page on GPU-accelerated terminals (Metal pipeline, App Nap interaction, idle redraw amplification, when GPU is genuinely worth it). Pages created: terminal-gpu-battery-cost (summary), gpu-accelerated-terminal (concept).

[2026-04-27] ingest | lfk (Lightning Fast Kubernetes navigator)

Ingested github.com/janosmiko/lfk — Bubbletea TUI for Kubernetes that replaces k9s' single-list view with a yazi-inspired three-column Miller layout (parent / current / preview) and an owner-based hierarchy: Cluster → Resource Type (grouped) → Resource → Owned Resources → Containers. Namespaces are not a navigation level — they're a top-right indicator with \ to switch and A to toggle all-namespaces. The flagship feature is owner walking: a Deployment shows its Pods through ReplicaSets (flattened), CronJobs show Jobs, Services show Pods matching the selector, ArgoCD Applications show managed resources from status or label discovery, Helm releases show resources via app.kubernetes.io/instance. Built-in integrations talk to the Kubernetes API directly — no per-tool CLI for KEDA pause/unpause, External Secrets force-refresh, Argo Workflows submit/suspend/stop/resubmit, ArgoCD sync/terminate/refresh, cert-manager, FluxCD, PVC resize. Only kubectl is required; helm and trivy are optional. Embedded PTY tabs stay running on tab switch (rare in TUI tools — most kill it). Inline log viewer streams across init-container transitions, staying attached as each finishes. The Secrets resource type is a real operational concern on big clusters because list calls return full bodies (tens of MB on heavy Helm clusters) — secret_lazy_loading: true switches to metadata-only listing with on-hover decode. 460+ ghostty themes, runtime picker (T), and OS dark/light auto-switching via CSI 996/2031 (Ghostty, kitty ≥0.27). Command bar (:) overloads four syntaxes: resource jumps, built-ins, :k/:kubectl passthrough with flag completion, :! shell. Six-week-old repo, 320 stars, 30 releases (v0.9.30 today), single author janosmiko, Apache-2.0 — added to watchlist with re-check on 2026-07-26 to see if v1.0 ships, release cadence holds, and a second contributor appears. No prior k9s page in vault; cross-referenced as the comparison anchor in prose. Pages created: lfk. Pages updated: index, watchlist.

[2026-04-25] ingest | You Don't Want Long-Lived Keys (Argemma Blog)

Ingested argemma.com/blog/long-lived-keys/ — short, opinionated post by Kelby Ludwig (Argemma security engineer). Three compounding risks of long-lived keys: people leaving with knowledge, brute-force probability over time, cryptographic wear-out (AES-GCM example: forgery attacks become realistic past 2^32 messages). Rotation is the standard mitigation but failure modes are familiar — emergency expedited rotations after leaks, ex-employee documentation rot, outage-during-rollout, no graceful degradation. Argues for replacing long-lived keys with ephemeral credentials (≤1 day) where rotation is structural rather than scheduled. Three concrete patterns: EC2 Instance Connect (replaces static SSH keys with auth-checked per-session creds), PyPI Trusted Publishers (OIDC from GitHub Actions instead of stored token), SSO (signed assertions instead of guessable per-app passwords). Acknowledges the SSO IdP signing key is itself long-lived — that's the second half of the argument: where long-lived keys are unavoidable, concentrate them. Fewer of them is easier to reason about than many distributed across teams; security rigor works better as someone's main job than as a tertiary product-team concern. For residual long-lived keys: scope tightly (e.g. data encryption key per shard), reason about lifetime mathematically (with the same confidence as "supercomputer billions of years"), rotate at least quarterly to keep the muscle exercised. Strong overlap with open-source-security-astral (which already documents Trusted Publishing in practice) and supply-chain-security (registry tokens as the canonical wrong-shaped credential) — added cross-references in both directions and bumped supply-chain-security's updated date. Created blog entity for argemma.com since the rest of the post list looks worth tracking (recent CVE notes, Go vulnerabilities, RAG threat models, timing attacks). Pages created: long-lived-keys (summary), ephemeral-credentials (concept), argemma-blog (entity). Pages updated: supply-chain-security (cross-references), index, log.

[2026-04-25] ingest | DeepSeek-V4 Roleplay Thinking-Mode Markers (victorchen96 GitHub)

Ingested github.com/victorchen96/deepseek_v4_rolepaly_instruct — Chinese-language README-only repo documenting two undocumented control markers that switch DeepSeek-V4 reasoning models' <think> style between roleplay-immersive first-person inner monologue and pure scene-and-strategy analysis. Markers are Chinese strings (【角色沉浸要求】... and 【思维模式要求】...) appended to the end of the first user message; subsequent turns inherit the style via conversation history. Author insists end-of-first-user-message is "the training-time injection position" — the FAQ is the most interesting line in the document because it implies the markers were trained, not engineered. Probabilistic, not enforced — supports re-rolling. Scope: DeepSeek's app/web Expert Mode and deepseek-v4-flash / deepseek-v4-pro API; web Fast Mode unsupported. No prior DeepSeek pages in vault — created an entity hub page since DeepSeek's V3/R1/V4 line is significant and likely to come up again. Created a concept page for the broader pattern of trained-in style toggles since this is genuinely distinct from prompt engineering (position-sensitive, wording-sensitive, probabilistic, persists in conversation history). Wiki pages written in English per global rule, with the Chinese marker strings preserved verbatim in code blocks since they're the literal text required to trigger the behavior. Pages created: deepseek-v4-roleplay-instruct (summary), controllable-thinking-style (concept), deepseek (entity).

[2026-04-25] ingest | SDL3 DOS port (libsdl-org/SDL #15377)

Ingested PR #15377 — merged 2026-04-23, ~6,000 LOC across 58 files. SDL3 now has native DOS support via DJGPP. Combined work of @icculus, @madebr, @glebm, @jayschwa, @ccawley2011, finished by Anders Jenbo (AJenbo). Tested extensively against DevilutionX inside DOSBox, no real-hardware testing yet. What works: VGA + VESA 1.2+ video with hardware page-flip and vsync (and VBE state save/restore on exit), Sound Blaster 16/Pro/2.0/1.x audio via IRQ-driven DMA double-buffered auto-init, PS/2 keyboard with extended scancodes, INT 33h mouse, gameport joystick via INT 15h, PIT-based timer at ~1.19 MHz via DJGPP's uclock(), filesystem via DJGPP's searchpath() / POSIX fallback. The architecturally interesting piece is threading: full SDL mutex/sema/TLS/condvar primitives built on top of a cooperative scheduler using setjmp / longjmp with stack patching (conceptually borrowed from the SDL3 PS2 port), with explicit yield points in the event pump and delay functions to keep audio threads responsive in a no-preemption environment. Build is via CMake cross-compilation toolchain i586-pc-msdosdjgpp.cmake plus a dedicated DJGPP CI job. Missing: audio recording, native SDL_TIME (uses DJGPP gettimeofday instead), SDL_LoadObject (no DOS dlopen). Cross-references the dosbox-detection-callback-opcode post since the development was DOSBox-tested. Pages created: sdl3-dos-port (summary), sdl (entity).

[2026-04-25] ingest | Stop using JWTs (samsch gist)

Ingested gist.github.com/samsch/0d1f3d3b4745d778f78b230cf6061452 — short, widely-cited opinion gist arguing JWTs should not be used as session tokens. Four reasons: (1) JWT spec is designed for ~5-minute tokens, sessions need longer; (2) "stateless" auth is a lie — you need a store for revocation/rotation, so just use a session ID; (3) a JWT carrying just a session ID is bigger and slower than a session cookie; (4) the JWT/JOSE spec family has a poor security track record (alg: none, key-confusion attacks). Recommends regular cookie sessions for browser auth, and PASETO for the residual short-lived signed-token cases (SSO transports, password-reset links, email-verification, capability tokens). Pre-empts three rebuttals: Google uses JWTs only as SSO transports not for sessions, "stateless is better" links a sister "Stateless is a lie" gist, "I don't know how to set up sessions" points at framework middleware (express-session + connect-session-knex). Also flags localStorage/sessionStorage as the wrong place for any auth credential (XSS exposure). No prior auth pages in the vault — created three concept pages because the JWT-for-sessions debate is a recurring topic and these are reusable references. Pages created: samsch-jwt-auth-gist (summary), jwt-for-sessions (concept, anti-pattern), cookie-session-auth (concept, recommended pattern), paseto (concept, the short-lived-token alternative).

[2026-04-25] ingest | Detecting DOSBox via Callback Opcode (datagirl.xyz)

Ingested datagirl.xyz/posts/dos_inside_the_box.html — December 2025 post by Snow (new blog entity, datagirl.xyz, code on git.2ki.xyz) on reliably detecting DOSBox from inside a DOS program. Approach: DOSBox invented a custom x86 instruction FE /7 imm16 (the CallBack opcode used internally by virtual COM files like MOUNT.COM to call into the host) — on real x86 since the 80186, that opcode raises #UD (interrupt 06h). Install an INT 06h handler, fire db 0xfe, 0x38, 0x00, 0x00, check whether the handler ran. Brittle alternatives (BIOS string at FE00:0061, Z: drive serial, DOSBox-X's MOUNT/VER quirks) are all spoofable; the callback opcode is inherent to how DOSBox works because removing it would break the virtual command set. Bonus content: the author's first 86Box test gave a false positive, traced to a long-standing PCem-inherited bug where any FE with non-zero /n was decoded as DEC — fix already merged in 86Box PR #6561, with hardware cross-checking by Linear (linear.network). Closes by noting NTVDM / Win9x DOS Prompt detect via a single INT 2Fh, and DOSEMU has many host-callback COM files. Pages created: dosbox-detection-callback-opcode (summary), datagirl-xyz (blog entity).

[2026-04-25] ingest | T.Context in Go Testing (Jonathan Hall)

Ingested boldlygo.tech/archive/2025-04-09-t.context — short reader-question post by Jonathan Hall (new blog entity, Boldly Go newsletter) following up on his earlier context.TODO() piece. Bruno Schaatsbergen asked why Hall hadn't recommended Go 1.24's new T.Context() / B.Context() / F.Context() methods; Hall says he forgot, and the actual answer is: yes, use them by default — they replace the t.Cleanup(cancel) boilerplate by tying context cancellation to the lifetime of the specific test/benchmark/fuzz target. Exception: anything whose lifetime outlives a single test (a shared test server, a Testcontainers Docker container started for the whole test binary). For those, use context.Background() at the TestMain / package level. The general rule is to match the context's lifetime to the lifetime of the thing that owns it. Pages created: t-context-go-testing (summary), boldlygo-tech (blog entity).

[2026-04-25] ingest | Message Passing Is Shared Mutable State (causality.blog)

Ingested causality.blog/essays/message-passing-is-shared-mutable-state — anonymous essay using Tu et al's 2019 ASPLOS paper (171 real Go concurrency bugs across Docker/Kubernetes/etcd/gRPC/CockroachDB) to confirm Edward Lee's 2006 prediction that the shared-memory vs. message-passing debate is a false dichotomy. ~58% of blocking bugs in those flagship Go projects were caused by message passing, not shared memory; Go's deadlock detector caught only 2 of 21 blocking bugs. Central argument: Go channels aren't really channels — no distinct producer/consumer endpoints, just a concurrent queue any goroutine holding a reference can mutate, structurally identical to Java's BlockingQueue from java.util.concurrent. Same shared mutable structure, same bugs. Even Erlang's actor model needed ETS tables (shared-mutable escape hatch) for performance, reintroducing the bugs the model was supposed to prevent. Essay ends as setup for a follow-up on alternative foundations. New blog entity (no author byline visible). Created a separate concept page for the four-mode bug taxonomy (deadlock/leak/race/protocol violation) since it's a reusable lens for any channel-based language. Pages created: message-passing-shared-mutable-state (summary), message-passing-vs-shared-memory (concept), go-channel-bug-patterns (concept), causality-blog (blog entity).

[2026-04-25] ingest | AtomicApp

Ingested atomicapp.ai — Ken (kenforthewin)'s self-hosted personal knowledge base that ships the LLM-augmented features as the product, not as Claude Code skills bolted onto Obsidian. Five months old (repo created 2025-11-25), 1.3k stars, MIT, very active (pushed today, 100+ tagged releases). Stack: Rust + actix-web backend, SQLite + sqlite-vec datastore (both content and embeddings), Tauri v2 + React 18 desktop, SwiftUI iOS app, Docker/Fly.io headless server, browser extension web clipper. AI providers BYO via OpenAI-compatible endpoint (OpenRouter / Ollama / anything). Built-in MCP server (in-process bridge on desktop, HTTP /mcp with Bearer auth remote) so Claude / Cursor can search and write "atoms" directly. Atoms are markdown notes with hierarchical tags, source URLs, auto-chunking, and per-chunk embeddings — taxonomy is auto-extended by the LLM, no folders. Features: semantic search, LLM-synthesized wiki articles per tag with inline citations, agentic chat scoped to tag-or-library, force-directed spatial canvas, auto-tagging, web clipper, MCP. Closest in-vault comparison is the obsidian-ai-complete-guide / llm-wiki-pattern setup which builds the same outcome from Obsidian + Claude Code on a plain markdown vault — AtomicApp collapses it into one binary, more turnkey but trades the grep-able plain-files substrate for SQLite. Watchlist profile (single author, young) but counter-signals are strong. Pages created: atomicapp. Pages updated: index, watchlist.

[2026-04-25] ingest | Type Systems Vocabulary (Lewis Campbell)

Ingested lewiscampbell.tech/blog/250817.html — short pet-peeve post by Lewis Campbell (new blog entity) on type-system terminology. Two orthogonal axes: static/dynamic (when types are checked) and strong/weak (how much implicit coercion). C is static + weak, Python is dynamic + strong, neither pair travels together. Dynamic ≠ untyped (untyped means no checking at all, statically or dynamically). "Strict typing" and "loose typing" treated as amateur substitutions that conflate the two axes into a meaningless single slider. Created a separate concept page for the four-quadrant matrix since no existing wiki page covered type-system axes. Pages created: type-systems-vocabulary (summary), type-system-axes (concept), lewis-campbell (blog entity).

[2026-04-25] ingest | Stash

Ingested alash3al.github.io/stash — Mohammed Al Ashaal's brand-new (repo created 2026-04-24, ingested one day later) persistent memory layer for AI agents. Go single binary, Apache-2.0, Postgres + pgvector backend, MCP server with 28 tools, model-agnostic via OpenAI-compatible endpoints (OpenRouter/Ollama/vLLM/Groq). Three v0.x releases shipped in first 24h, 96 stars. Distinguishing piece is the nine-stage consolidation pipeline: episodes → facts → relationships → causal links → patterns → contradictions, plus three "NEW" stages for goals, failure patterns, and hypothesis scans. This goes substantially further than hippo-memory (stops at episodes-to-patterns) and mempalace (verbatim storage, no consolidation). Hierarchical namespaces (/users/, /projects/, /self/) with recursive reads — the /self namespace is unusual, the agent uses its own memory layer to maintain a self-model of capabilities, limits, and preferences (closer to autonomous-research-agent design). Ships an autonomous-research loop intended for cron, where the agent orients/researches/thinks/invents/consolidates/reflects every 5 minutes. Author is established (247 public repos, 557 followers in Hurghada, Egypt) but the project itself is one day old — clear watchlist profile, especially for stages 4–9 which are the kind of claims that need real-world reps to verify. Three-way comparison table now lives on mempalace (MemPalace vs Hippo vs Stash). Pages created: stash. Pages updated: agent-memory-decay (added Stash as longer-pipeline implementation), mempalace (replaced two-way table with three-way), hippo-memory (added related-tools section), index, watchlist.

[2026-04-24] ingest | Gova

Ingested gova.dev — nv404's two-day-old (repo created 2026-04-22) declarative GUI framework for Go. Pitch: write desktop apps as plain Go, compile a single static binary for macOS/Windows/Linux, no Chromium, no JS runtime. Distinguishing feature is the explicit reactive Scope — state/signals/effects live on a value you can see and pass around, closer to SwiftUI's shape than React's hooks. Views are plain Go structs, props are fields, defaults are zero values, composition is function calls, compiler checks the UI. Portability caveat: "native dialogs" means NSAlert/NSOpenPanel/NSDockTile via cgo on macOS but Fyne fallback on Windows/Linux — the non-macOS story leans on Fyne. Requires Go 1.26+ and cgo on all platforms. Binary ~32 MB default, 23 MB stripped, ~80 MB idle RSS. CLI has gova dev (hot reload), gova build, gova run. 78 stars, MIT, no tagged releases, single author, pre-1.0 API. Classic watchlist profile — added to watchlist with 2026-07-24 re-check. Pages created: gova. Pages updated: index, watchlist.

[2026-04-24] ingest | MasterHttpRelayVPN-RUST

Ingested @therealaleph's Rust port of @masterking32's MasterHttpRelayVPN (upstream Python tool ingested three days earlier on 2026-04-21). Same architecture — MITM local HTTPS, reshape into JSON RPC, send through Google Apps Script relay fronted as www.google.com — but as a single static binary with egui desktop UI and a real Android build via VpnService + tun2proxy. Notable differences from the Python original: zero runtime deps (critical for users whose package mirrors are blocked), hybrid routing where Google-owned destinations skip the Apps Script detour and tunnel directly through Google edge (preserves the 20k-calls/day quota for non-Google traffic), portable-atomic for 32-bit MIPS/ARMv5 router support, dual glow/wgpu rendering backends for older Windows hardware. v1.2.11, 257 stars, MIT. Carved out a new concept page for the broader serverless-relay pattern (distinct from classic domain-fronting because the "front" and "back" are the same provider — Google→Google, Cloudflare→Cloudflare — so it survived the 2020 CDN fronting enforcement that killed most other tricks). Pages created: masterhttprelayvpn-rust (toolbox), serverless-relay-transport (concept). Pages updated: masterhttprelayvpn (cross-reference to Rust port + bumped updated), domain-fronting (added Rust port + concept page, added source).

[2026-04-24] ingest | SuperHQ + new watchlist convention

Ingested superhq-ai/superhq — three-week-old (created 2026-04-04) Rust desktop app for orchestrating sandboxed AI coding agents (Claude Code, Codex, Pi) on macOS Apple Silicon, AGPL-3.0, 194 stars, explicit "very early alpha." Built on GPUI (Zed's UI framework) and a separate sibling crate shuru-sdk (673 stars, Apache-2.0, macOS+Linux microVM sandbox). Notable design: an auth gateway — host-side reverse proxy that injects API credentials into outgoing requests so the agent inside the VM never sees real keys or OAuth tokens (handles ChatGPT Plus/Pro OAuth refresh too). User asked for a separate watchlist for projects too early/uncertain to commit to; created watchlist as the convention (frontmatter tag watchlist, table with what we're waiting on + next-check date + signals to evaluate). Updated sandboxing-ai-agents to add a microVM bullet under Layer 1 and a new "Layer 1b: credential isolation at the wire" subsection. Pages created: superhq (toolbox), watchlist (overview). Pages updated: sandboxing-ai-agents, index, index.

[2026-04-23] ingest | Forge (Andrew Nesbitt)

Ingested Nesbitt's March 2026 post on Forge — a Go CLI and library providing one unified interface over GitHub, GitLab, Gitea/Forgejo, and Bitbucket Cloud. Auto-detects the forge from the git remote, reuses existing tokens from gh/glab, covers repos/issues/PRs/releases/branches/CI/labels/milestones/deploy-keys/secrets/comments. Notable framing: AI coding agents today mostly hardcode GitHub's API, so a forge-abstraction library lowers lock-in. Third Nesbitt article after git-magic-files and features-to-steal-from-npmx; fits his ongoing "normalize across ecosystems" pattern (previously: Libraries.io/ecosyste.ms for registries, git-pkgs for lockfiles). Repo at github.com/git-pkgs/forge, Go/MIT, 67 stars. Pages created: forge (summary), forge (toolbox). Pages updated: nesbitt-io (added article to list).

[2026-04-22] ingest | Impeccable (Paul Bakaus)

Ingested impeccable.style — Paul Bakaus's skill bundle + CLI + Chrome extension that loads design vocabulary (typography, color, layout, motion, hierarchy, interaction, quality + Nielsen heuristics) into agent context via /impeccable teach, and lints output against 25 deterministic anti-pattern rules (purple gradients, nested cards, overused Inter, gradient text, low contrast). Installs across Cursor/Claude Code/Gemini/Codex/etc with one command. Notable design choice: detector is rules-based, not LLM-based, to sidestep the same training-data bias that causes the monoculture. 21.3k stars, Apache-2.0. Pages created: impeccable.

[2026-04-22] ingest | Garbage Collection Without Unsafe Code (Nick Fitzgerald)

Ingested Fitzgerald's 2024 write-up of safe-gc, a Rust garbage collection crate with zero unsafe code. Central design choice: Gc<T> is not deref-able — all access goes through heap[&gc_ref], which reduces GC access to ordinary Index/IndexMut and lets the borrow checker enforce lifetimes normally. Heap is HashMap<TypeId, Box<dyn ArenaObject>>, one arena per type, mark-and-sweep. Copying collector was tried first and abandoned due to borrowing conflicts with forwarding pointers. Falls-out-for-free safety: Drop can't access the heap (no finalization footguns), Trace doesn't need to be unsafe (wrong trace = logical bug, not UB). Pages created: safe-gc (summary), safe-gc (toolbox).

[2026-04-22] ingest | Seven-source batch — AI-agent infra and adjacent

Large batch covering open-source/LLM-collaboration and agent-infrastructure tooling. Two dpc.pw posts by Dawid Ciężarkiewicz (new blog entity): one arguing LLMs make external PRs net-negative for small maintainers, one walking through his Bubblewrap+Nix+tmux isolation setup. Five tools: cal.diy (Cal.com's explicit personal-use-only open-source spin-out), GoModel (Go LiteLLM alternative, 10+ providers), ctx (local conversation-bound context manager for Claude Code/Codex), Brex's CrabTrap (LLM-as-judge HTTP proxy for agent security), and Charlie Labs Daemons (Markdown-defined self-initiated background AI processes). Two new concept pages: sandboxing-ai-agents (four-layer taxonomy tying bubblewrap/hazmat/crabtrap/syscall-rewriting together) and agents-vs-daemons (Charlie's role-shaped vs task-shaped distinction). Pages created: i-dont-want-your-prs, bubblewrap-dev-env (summaries), dpc-pw (blog entity), cal-diy, gomodel, ctx, crabtrap, charlie-daemons (toolbox), sandboxing-ai-agents, agents-vs-daemons (concepts).

[2026-04-22] ingest | Features everyone should steal from npmx (Andrew Nesbitt)

Ingested Nesbitt's April 2026 post on npmx.dev, Daniel Roe's alternative frontend over the npm registry launched January 2026. Two theses: the launch demonstrated real competitive pressure (npmjs.com shipped long-requested dark mode within the month) and the feature catalog is reusable as a spec for any modern package-registry UI (transitive install size, install-script disclosure, multi-forge repo stats, ATProto-backed social, local-CLI admin proxy, etc.). Pages created: features-to-steal-from-npmx (summary), nesbitt-io (blog entity, 2nd post from him), npmx (entity), alternative-frontend-pattern (concept — Invidious/Nitter/npmx/nugx). Pages updated: git-magic-files (linked back to Nesbitt blog entity).

[2026-04-21] ingest | Russia VPN bypass state (April 2026)

Ingested a Claude-produced synthesis of ntc.party community discussions on current circumvention methods in RF. Folded in as analysis snapshot rather than a share-link summary since the content decays fast. Includes tier-ranked VLESS variants (Reality+XHTTP+steal-oneself on top), non-VLESS methods (AmneziaWG 2.0 leading, WireGuard/OpenVPN/old-AmneziaWG dead after Feb 13 2026), TSPU escalation timeline, and a hosting-selection section (PQ.Hosting/ishosting/u1host tier 1, Aeza/Russian hosters tier 2, Hetzner/OVH/DO flagged tier 3). Pages created: russia-vpn-bypass-state-2026-04 (analysis), tspu (entity). Pages updated: whitelist-internet-blocking, vpn-configs-for-russia (cross-references).

[2026-04-21] ingest | vpn-configs-for-russia

Ingested igareck's free-subscription aggregator: hourly auto-tested VPN configs from inside Russia, split by black vs white list regime. Prompted new concept page on whitelist internet blocking — default-deny national filtering currently rolled out on Russian mobile carriers. CIDR and SNI bypass strategies are the only viable responses. Pages created: vpn-configs-for-russia (toolbox), whitelist-internet-blocking (concept). Pages updated: russia-v2ray-rules-dat (cross-reference).

[2026-04-21] ingest | Karing

Ingested Karing, a Flutter GUI client built on a modified sing-box core. Cross-platform (iOS/Android/macOS/Windows/Linux/tvOS), multi-format subscription aggregator, iCloud/WebDAV/LAN sync. Sits in the circumvention-client cluster alongside mihomo, 3x-ui, xkeen, happ-proxy, masterhttprelayvpn. Pages created: karing (toolbox), sing-box (entity). Pages updated: mihomo, xray-tutorial (cross-references).

[2026-04-21] ingest | MasterHttpRelayVPN

Ingested a freshly-published (Apr 20 2026) censorship-circumvention tool that uses Google Apps Script as a free relay to front HTTPS traffic as google.com. MITM-terminates local HTTPS, reshapes the request into a JSON RPC to script.google.com, returns the response through the same path. Occupies the "zero-budget" niche next to the VLESS cluster. Pages created: masterhttprelayvpn (toolbox), domain-fronting (concept). Pages updated: xray-tutorial, happ-proxy (cross-references).

[2026-04-18] ingest | A Simplified Model of Fil-C (Peter Cawley)

Ingested Cawley's step-by-step walkthrough of Fil-C's memory safety model: AllocationRecord companions for every pointer, parallel invisible_bytes for heap pointer capabilities, and a concurrent GC (FUGC). Production complications covered: threads, function pointers with uniform ABI, memmove alignment quirks. Fil-C as a concrete example of pointer provenance. Pages created: simplified-model-of-fil-c, fil-c, pointer-provenance.

[2026-04-18] ingest | Average Is All You Need (rawquery)

Ingested essay-plus-launch-post from the rawquery team. Central argument: LLMs make "average" output cheap, which lifts the floor without raising the ceiling. Demonstrated via a Stripe+HubSpot email-campaign attribution scenario done entirely in English via a CLI-first data platform. Pages created: average-is-all-you-need, llm-as-average-democratizer, rawquery.

[2026-04-16] ingest | wacli

Added wacli to toolbox — WhatsApp CLI built on whatsmeow for local message sync, offline FTS5 search, sending, and group management. 1,554 stars, Go, by steipete.

[2026-04-16] ingest | Many-Step Sequences in Go (Chris Lesiw)

Ingested article on state machine pattern for multi-step Go tasks. Each step returns the next step; driver is a simple loop. Method values bind state, generic parameter namespaces sequences, runtime function comparison enables independent testing. Published as lesiw.io/step. Pages created: go-step-sequences.

[2026-04-16] ingest | The Paper Computer (James Somers)

Ingested essay on hybrid physical/digital workflow: do thinking on paper, let AI digitize invisibly. Argues for spatial cognition (rooms as memory scaffold), richer physical affordances vs software, and single-purpose OS modes to escape bundled distraction. Pages created: paper-computer.

[2026-04-16] ingest | Fast Levenshtein Distance Using a Trie (Steve Hanov)

Ingested classic 2009 algorithm post on efficient fuzzy dictionary search. Key insight: words sharing prefixes share Levenshtein table rows; traverse trie DFS to maximize row reuse. 300× faster than naive (0.014s vs 4.5s for 98k words). Early termination prunes branches where min(row) > maxCost. Pages created: levenshtein-trie.

[2026-04-15] ingest | Rewriting Every Syscall in a Linux Binary at Load Time

Ingested deep technical post on syscall interception via binary rewriting. Walks .text with an ILD, replaces 0F 05 with INT3 + NOP, runs process in KVM with a minimal shim. Self-healing LSTAR handler catches JIT'd syscalls. Under 1µs overhead. Part 1 of 7-part series on minimal VM runtime for AI agent execution. Pages created: syscall-binary-rewriting. Pages updated: hazmat.

[2026-04-15] ingest | Things You Didn't Know About Indexes (jon.chrt.dev)

Ingested practical guide to database indexes: the read/write trade-off, why composite index column order matters, why functions defeat indexes, and three lesser-known types (functional, partial, covering indexes). Pokémon table examples throughout. Pages created: database-index-gotchas.

[2026-04-15] ingest | Direct Win32 API and Weird-Shaped Windows (Ivan Rogoz)

Ingested nostalgic/technical article on Win32 non-rectangular windows: window regions (HRGN), bitmap-derived regions, and layered windows with per-pixel alpha. Covers the message-driven model and why custom windows disappeared (cultural shift from "crazy skins" to "work reliably"). Pages created: win32-weird-shaped-windows. Pages updated: win32-stable-abi.

[2026-04-15] ingest | Retrofitting JIT Compilers into C Interpreters (Laurence Tratt)

Ingested detailed technical post on yk, a meta-tracing system that automatically derives JIT compilers from C interpreters. Explains the technique (trace the interpreter, not the guest), key optimizations (promotion, idempotent functions, backward codegen), and the trade-off (less peak performance than LuaJIT, but trivial version tracking). Pages created: retrofitting-jit-c-interpreters, meta-tracing, yk.

[2026-04-15] ingest | My AI-Assisted Workflow (Matteo Barbero)

Ingested structured AI-assisted development workflow: seven steps from free-form planning through PRD, vertical slice issues, tasks, implementation, code review, and final audit. Key insight: AI stress-tests thinking rather than replacing it, with AFK/HITL classification for tasks. Pages created: ai-assisted-workflow. Pages updated: human-in-the-loop, context-engineering.

[2026-04-15] ingest | Go Transaction Linter (léon h)

Ingested blog post about building a custom Go linter after shipping a transaction leak bug. The linter uses the go/analysis framework to detect when database operations use the outer repository instead of the transaction parameter inside callback-based transactions. Pages created: go-transaction-linter, go-analysis-framework, transactioncheck.

[2026-04-13] ingest | boringBar

Added boringbar to toolbox — macOS Dock replacement that filters windows by Space/desktop, with thumbnail previews and desktop switching. Commercial app ($40 perpetual or $7.99/year), requires macOS 14+.

[2026-04-13] ingest | Obsidian + Claude Code Complete Guide (Huashu)

Ingested Huashu's 67-page PDF on building AI-powered knowledge bases with Obsidian and Claude Code. Covers three-layer architecture (raw → wiki → output), SCHEMA.md for wiki consistency, six design principles, seven practical workflows including agentic note-taking. Validates the LLM Wiki pattern with production implementation details. Pages created: obsidian-ai-complete-guide. Pages updated: llm-wiki-pattern.

[2026-04-13] ingest | The Peril of Laziness Lost (Cantrill)

Ingested Bryan Cantrill's blog post arguing LLMs lack the programmer virtue of laziness — the drive to simplify born from finite human time. Garry Tan's 37k lines/day as reductio ad absurdum, Oxide's LLM-as-tool position. Pages created: peril-of-laziness-lost, bryan-cantrill. Pages updated: no-silver-bullet-llms, cult-of-vibe-coding, building-syntaqlite-ai, clean-code-coding-agents.

[2026-04-11] ingest | No Silver Bullet (Brooks, 1986 — primary source)

Ingested the original Brooks essay from worrydream.com PDF. Substantially expanded no-silver-bullet with the four essential properties (complexity, conformity, changeability, invisibility), the full candidate bullets list with Brooks' reasoning, and the "promising attacks on the conceptual essence" section (buy vs build, rapid prototyping, incremental development, great designers). Pages updated: no-silver-bullet, fred-brooks.

[2026-04-11] ingest | With AI, You Barely Need a Frontend Framework

Ingested dlants's post arguing that framework implicitness (React hooks, VDOM, reconciliation) is the enemy of AI-generated code, proposing "vamp" — a minimal explicit View/State/Message pattern in vanilla TypeScript. Pages created: vamp-ai-frontend, vamp.

[2026-04-11] ingest | Ursa: Iceberg-First Storage Engine for Kafka

Ingested topicpartition.io post on StreamNative's Ursa — a diskless Kafka storage engine that writes directly to object storage as Iceberg/Delta tables. 10x cost reduction claim via eliminating cross-AZ replication. VLDB 2025 best industry paper, not yet open-sourced. Also covers Kafka protocol-vs-codebase dynamics and adaptable topics. Pages created: ursa-kafka-storage.

[2026-04-10] ingest | Clean Code in the Age of Coding Agents

Ingested yanist.com post arguing clean code matters more with agents because context windows are finite — messy code wastes context on orientation. Lightweight but connects the Martin/Fowler "structure vs value" tradition to LLM-agent development. Pages created: clean-code-coding-agents.

[2026-04-10] ingest | XKeen

Ingested Corvus-Malus/XKeen: transparent proxy for Keenetic routers using Xray or Mihomo cores with GeoIP/GeoSite filtering. Whole-network proxying without per-device config. ~1,000 stars, community fork, v1.1.3.9. Pages created: xkeen.

[2026-04-10] ingest | ETH Zurich Geometric Quantum Gates

Ingested ETH Zurich Nature paper (April 8, 2026) on geometric-phase swap gates for neutral-atom qubits: 99.91% fidelity across 17,000 pairs, sub-millisecond, noise-resilient. Crosses surface-code error correction threshold. Pages created: eth-geometric-quantum-gates. Pages updated: crqc.

[2026-04-10] ingest | Ark ECS

Ingested mlange-42/ark: archetype-based ECS for Go with type-safe generics, first-class entity relationships, batch ops, event system, and optional systems layer. 237 stars, MIT/Apache-2.0, v0.8.0. Pages created: ark-ecs, entity-component-system. Pages updated: go-bt, index.

[2026-04-10] ingest | I Still Prefer MCP Over Skills

Ingested David Mohl's post arguing skills are for knowledge (teaching LLMs existing tools) and MCP is for access (live service connections). Covers portability, auth, context bloat tradeoffs and the hybrid pattern of annotating MCPs with skills. Pages created: mcp-vs-skills.

[2026-04-09] ingest | botctl

Ingested botctl.dev (montanaflynn/botctl): Go process manager for long-running autonomous Claude agents with BOT.md declarative config, harness loop state machine, hot reload, operator message injection, and a three-level skills system with on-demand loading. MIT, 68 stars, v0.3.5. Pages created: botctl. Pages updated: build-your-own-openclaw, index.

[2026-04-09] ingest | Let's talk about LLMs (Bennett)

Ingested James Bennett's 6,000-word application of Fred Brooks' No Silver Bullet to LLM coding hype. Argues LLMs primarily attack accidental difficulty and are therefore capped below 10x by Brooks' math, with empirical backing from DORA 2025, CircleCI 2026, METR, and the Cloudflare Next.js rebuild case study. Pages created: no-silver-bullet-llms, no-silver-bullet, james-bennett, fred-brooks. Pages updated: cult-of-vibe-coding, building-syntaqlite-ai.

[2026-04-09] ingest | Build Your Own OpenClaw

Ingested czl9707's 18-step Python tutorial for building a Claude Code-style agent from scratch. 4 phases: single-agent capability, event architecture, autonomous/collaborative, production readiness. MIT, 917 stars, uses LiteLLM, ships with pickle-bot reference implementation. Pages created: build-your-own-openclaw.

[2026-04-09] ingest | Happ Subscription Converter

Ingested bubasik's gist documenting PHP scripts that emulate the Happ Android VPN client to extract VLESS server lists from a Happ subscription and bypass HWID-based device limits. Technique: use a logging script to capture the real app's headers, then replay them from a PHP proxy. Pages created: happ-subscription-converter, happ-proxy, vless-subscription-format. Pages updated: 3x-ui.

[2026-04-09] ingest | Your File System Is Already a Graph Database

Ingested Rumproarious's April 2026 post arguing that a markdown vault with wikilinks is already a graph database, and that its real value is as a "context engineering system" for seeding future LLM conversations. Author runs this at ~52k files with a PARA-extended layout (/people/{slack_handle}, /daily/, /meetings/). Pages created: filesystem-is-graph-database, files-as-graph-database, context-engineering. Pages updated: llm-wiki-pattern, para-method.

[2026-04-09] ingest | The AI Great Leap Forward

Ingested Han Lee's essay comparing corporate AI mandates to the Great Leap Forward: demoware, fabricated metrics, institutional knowledge destruction, anti-distillation. Pages created: ai-great-leap-forward. Pages updated: ai-bubble-pale-horses.

[2026-04-09] ingest | Chiasmus — Neurosymbolic MCP

Ingested yogthos's post on Chiasmus, an MCP server combining tree-sitter + Prolog for formal code analysis. Pages created: chiasmus, neurosymbolic-ai.

[2026-04-09] ingest | IngoDB

Ingested Henrik Ingo's announcement of IngoDB — an AI-native adaptive LSM-tree database in Rust with reactive indexing. Pages created: ingodb. Pages updated: lsm-tree.

[2026-04-09] ingest | Open Source Security at Astral

Ingested Astral's blog post on their supply chain security practices for Ruff/uv/ty. Pages created: open-source-security-astral, supply-chain-security.

[2026-04-09] ingest | Little Snitch for Linux

Toolbox page for Little Snitch's Linux port — eBPF-based per-app network monitor with web UI. Pages created: little-snitch-linux.

[2026-04-09] ingest | go-bt

Toolbox page for rvitorper/go-bt — minimalist behavior tree library in Go. Pages created: go-bt, behavior-tree.

[2026-04-09] ingest | LSM Trees and NoSQL Storage

Ingested Ramarathinam Iyer's Medium explainer on LSM trees. Pages created: lsm-trees-nosql, lsm-tree.

[2026-04-09] ingest | Linux Kernel in PostgreSQL via pgit

Ingested Oliver Seifert's article on importing the full Linux kernel git history into PostgreSQL using pgit. Pages created: linux-kernel-pgit, pgit.

[2026-04-08] ingest | Zsh om[1] Glob Qualifiers

Ingested Adam Johnson's post on using zsh glob qualifiers to select the newest file inline. Pages created: zsh-glob-qualifiers.

[2026-04-08] ingest | Full Text Search with IndexedDB

Ingested singpolyma's blog post on building full-text search for Borogove using IndexedDB's multiEntry indexes. Pages created: full-text-search-indexeddb, borogove, indexeddb, inverted-index. Pages updated: hybrid-search.

[2026-04-08] ingest | hazmat

Toolbox page created for dredozubov/hazmat — macOS containment system for AI coding agents with three-layer enforcement (user isolation, Seatbelt, pf firewall). Pages created: hazmat.

[2026-04-08] ingest | developer-icons

Toolbox page created for xandemon/developer-icons — optimized SVG tech logo components for React/TypeScript. Pages created: developer-icons.

[2026-04-06] update | Wiki initialized

Created directory structure, schema (CLAUDE.md), index, and log. Ready for first ingest.

[2026-04-06] ingest | How We Built a Virtual Filesystem for Our Assistant (Mintlify)

Ingested Mintlify engineering blog post about ChromaFs. Pages created: mintlify-chromafs (summary), mintlify (entity), virtual-filesystem-for-llm-agents (concept), rag-limitations (concept). Source saved as mintlify-chromafs-blog.

[2026-04-06] ingest | Working on Products People Hate (Sean Goedecke)

Ingested blog post about engineer agency and working on unpopular products. Pages created: working-on-products-people-hate (summary), individual-engineer-agency (concept). Source saved as working-on-products-people-hate.

[2026-04-06] ingest | antonz.org (blog)

Added antonz-org as a tracked blog entity. Covers Go, SQL, CLI tools, interactive guides. Author: Anton Zhiyanov. Updated ingest skill to support blog tracking and new-post checking.

[2026-04-06] update | Inbox and URL tracking

Added inbox workflow: inbox/ for Obsidian web clips, inbox-processed/ for backup after ingest, inbox-assets/ for clip attachments. Created sources/ingested-urls.md to track URLs fetched from terminal that need manual backup. Updated CLAUDE.md and ingest skill.

[2026-04-06] ingest | ADK Go 1.0 Arrives (Google Developers Blog)

Ingested Google's ADK Go 1.0 announcement. Pages created: adk-go (summary), human-in-the-loop (concept), agent2agent-protocol (concept). Source saved as adk-go-10-arrives.

[2026-04-06] ingest | Loading... [13 kB] (maurycyz)

Ingested post about TCP slow start explaining why downloads stall at 13 kB. Pages created: 13kb-tcp-slow-start (summary), tcp-congestion-control (concept). Source saved as 13kb-tcp-slow-start.

[2026-04-06] ingest | Building a Second Brain: Overview (Tiago Forte)

Ingested BASB introductory guide. Pages created: basb-overview (summary), progressive-summarization (concept), para-method (concept), llm-wiki-pattern (concept). Source saved as basb-overview.

[2026-04-06] update | Added Toolbox section

New page type toolbox for interesting projects to try later. Added wirez as first entry. Updated CLAUDE.md schema.

[2026-04-06] ingest | gallery-dl

Added gallery-dl to toolbox — CLI image gallery downloader supporting dozens of sites.

[2026-04-06] ingest | claude-defuddle

Added claude-defuddle to toolbox — Claude Code skill that extracts clean Obsidian notes from URLs, offloading content processing outside the context window.

[2026-04-06] ingest | The Subprime AI Crisis Is Here (Edward Zitron)

Ingested Zitron's March 2026 analysis of AI industry economics. Emphasis on the predictive framework — collapse sequence and warning signs. Pages created: subprime-ai-crisis (summary), ai-subsidy-economics (concept), ai-bubble-pale-horses (analysis), edward-zitron (entity).

[2026-04-06] ingest | slogbox

Added slogbox to toolbox — circular buffer slog.Handler for Go with debug endpoints and black box auto-flush.

[2026-04-06] ingest | Thinking Space

Added thinking-space to toolbox — local-first markdown knowledge workspace with AI agents and self-modifying rebuild pipeline.

[2026-04-06] update | Added parent field and updated date convention

Inspired by thinking-space's hierarchical metadata. Added optional parent: frontmatter field for nesting pages under broader topics. Added convention to bump updated: date when revising existing pages. Updated CLAUDE.md and ingest skill.

[2026-04-06] ingest | Scotty (Spatie)

Added scotty to toolbox — SSH task runner with plain bash syntax, real-time streaming, pause/resume during deploys.

[2026-04-06] ingest | Gemma Gem

Added gemma-gem to toolbox — Chrome extension running Gemma 4 entirely on-device via WebGPU with browser automation tools.

[2026-04-06] ingest | qmd

Added qmd to toolbox — local-first markdown search with BM25 + vector + LLM re-ranking. MCP server for Claude Code integration. Could index this wiki for semantic search.

[2026-04-06] ingest | OpenScreen

Added openscreen to toolbox — free open-source Screen Studio alternative for product demos.

[2026-04-06] ingest | mihomo (MetaCubeX)

Added mihomo to toolbox — rule-based proxy/tunnel tool (Clash fork) supporting many protocols, TUN mode, and DNS. Cross-referenced with wirez.

[2026-04-06] ingest | sheets

Added sheets to toolbox — terminal spreadsheet with vim keybindings by Maas Lalani.

[2026-04-06] ingest | Pi Coding Agent

Added pi-coding-agent to toolbox — minimal extensible terminal coding agent by Mario Zechner with tree-structured sessions and 15+ providers.

[2026-04-06] ingest | Signals: The Push-Pull Based Algorithm (Willy Brauner)

Ingested article on how signals work internally. Pages created: signal-push-pull-algorithm (summary), reactive-signals (concept), tc39-proposal-signals (concept). Source saved as signal-push-pull-algorithm.

[2026-04-06] ingest | Perfmon

Added perfmon to toolbox — tabbed TUI performance monitor with sparklines and configurable commands.

[2026-04-06] ingest | delta

Added delta to toolbox — syntax-highlighting git pager with side-by-side view, word-level diffs, and blame support.

[2026-04-06] ingest | Sam Altman May Control Our Future — Can He Be Trusted? (New Yorker)

Ingested Farrow & Marantz investigative profile. Pages created: altman-new-yorker-profile (summary), sam-altman (entity), openai (entity). Cross-referenced with ai-subsidy-economics, ai-bubble-pale-horses, subprime-ai-crisis.

[2026-04-06] ingest | proxmox-manager

Added proxmox-manager to toolbox — single Bash script TUI for managing Proxmox VMs and containers.

[2026-04-06] ingest | russia-v2ray-rules-dat

Added russia-v2ray-rules-dat to toolbox — auto-generated geoip/geosite routing rules for Russian-blocked domains and IPs. Pages updated: mihomo (cross-reference).

[2026-04-06] ingest | 3x-ui

Added 3x-ui to toolbox — Xray-core web management panel with multi-user support, traffic quotas, and subscription links. Pages updated: mihomo, russia-v2ray-rules-dat (cross-references).

[2026-04-06] ingest | Xray-core VLESS proxy tutorial

Ingested step-by-step tutorial on setting up Xray-core with VLESS protocol. Pages created: xray-tutorial (summary). Pages updated: 3x-ui (cross-reference). Full source saved to sources/xray-tutorial.md.

[2026-04-06] ingest | AmneziaWG Architect

Added amneziawg-architect to toolbox — browser-based obfuscation config generator for AmneziaWG (WireGuard fork with DPI evasion). Cross-references to 3x-ui, xray-tutorial, russia-v2ray-rules-dat, mihomo.

[2026-04-06] ingest | Preferring Local OSS LLMs (Ted Neward)

Ingested blog post arguing for local open-source LLMs over cloud AI. Pages created: titit-local-ai (summary). Pages updated: subprime-ai-crisis, ai-subsidy-economics (cross-references).

[2026-04-07] ingest | Git's Magic Files (Andrew Nesbitt)

Ingested catalog of committed files that control git behavior across a team. Pages created: git-magic-files (summary). Source saved as git-magic-files.

[2026-04-07] ingest | ultimate-gojust

Added ultimate-gojust to toolbox — comprehensive Go Justfile with version injection, cross-compilation, and dev tooling. Noted reusable patterns for Go build automation (ldflags, config cascade, cross-compile loop).

[2026-04-07] ingest | Veil

Added veil to toolbox — client-side dark-mode PDF reader that uses PDF.js operator lists to invert text while preserving images. Includes browser OCR and dark PDF export.

[2026-04-07] ingest | Multi-agentic Software Development is a Distributed Systems Problem (Kiran)

Ingested post applying FLP impossibility and Byzantine fault theory to multi-agent LLM coding. Pages created: log-distributed-llms (summary), distributed-consensus (concept), flp-impossibility (concept), byzantine-fault (concept). Pages updated: human-in-the-loop (added HITL-as-consensus-oracle section).

[2026-04-07] ingest | Agent Reading Test

Added agent-reading-test to toolbox — benchmark for AI agent web documentation reading with 10 tasks and self-scoring. Created llms-txt concept page covering the /llms.txt convention for agent-navigable websites.

[2026-04-07] ingest | Hippo Memory

Added hippo-memory to toolbox — biologically-inspired memory system for AI agents with three-layer architecture, decay mechanics, hybrid search, and cross-framework portability. Pages created: agent-memory-decay (concept), hybrid-search (concept), memory-conflict-detection (concept). Cross-referenced with qmd, human-in-the-loop, byzantine-fault.

[2026-04-07] ingest | macOS TCP Time Bomb (Photon)

Ingested deep-dive on XNU kernel's 32-bit tcp_now overflow: freezes TCP clock after 49.7 days, breaks TIME_WAIT GC, exhausts ephemeral ports. Pages created: macos-tcp-time-bomb (summary). Pages updated: tcp-congestion-control (added TCP timer internals section). Source from inbox.

[2026-04-07] ingest | Ghost Pepper

Added ghost-pepper to toolbox — local hold-to-talk STT for macOS using WhisperKit + Qwen cleanup, fully on-device. Cross-referenced with titit-local-ai.

[2026-04-07] ingest | The Cult of Vibe Coding Is Insane (Bram Cohen)

Ingested Cohen's critique of pure vibe coding and his audit-discuss-execute workflow for AI-assisted development. Pages created: cult-of-vibe-coding (summary). Cross-referenced with human-in-the-loop, byzantine-fault, individual-engineer-agency.

[2026-04-07] ingest | Hold on to Your Hardware (Marius)

Ingested analysis of AI data center demand cannibalizing consumer hardware supply: DRAM, NAND, HDD shortages, price spikes, rental dystopia scenario. Pages created: hold-on-to-your-hardware (summary). Pages updated: ai-subsidy-economics (added hardware market distortion). Cross-referenced with titit-local-ai, subprime-ai-crisis, ai-bubble-pale-horses.

[2026-04-07] ingest | CRQC Timeline (Filippo Valsorda)

Ingested Valsorda's reassessment of quantum computing threats after two April 2026 papers collapsed timelines. Google sets 2029 migration deadline. Pages created: crqc-timeline (summary), post-quantum-cryptography (concept), crqc (concept), ml-kem (concept), ml-dsa (concept), store-now-decrypt-later (concept).

[2026-04-07] ingest | The Pinnacle of Enshittification (Michał Górny)

Ingested Gentoo developer's comprehensive anti-LLM argument: infrastructure DDoS from scrapers, FLOSS community fracturing, copywashing, curl bug bounty killed by AI spam. Pages created: llm-enshittification (summary). Cross-referenced with ai-subsidy-economics, subprime-ai-crisis, hold-on-to-your-hardware, titit-local-ai, cult-of-vibe-coding.

[2026-04-07] ingest | Pion Handoff

Added pion-handoff to toolbox — WebRTC session handoff from browser to Go backend for recording, media injection, and protocol analysis. By Sean DuBois (Pion creator).

[2026-04-08] ingest | Eight Years of Wanting, Three Months of Building with AI (Lalit Maganti)

Ingested detailed account of building syntaqlite with AI: vibe coding failure, disciplined rewrite, addiction loop, expertise hierarchy, design vs implementation. Pages created: building-syntaqlite-ai (summary). Cross-referenced with cult-of-vibe-coding, human-in-the-loop, byzantine-fault.

[2026-04-08] ingest | MemPalace

Added mempalace to toolbox — AI memory system using palace metaphor with verbatim ChromaDB storage, temporal knowledge graph, and MCP integration. Includes comparison table with hippo-memory. Cross-referenced with agent-memory-decay, memory-conflict-detection, hybrid-search.

[2026-04-08] ingest | Emotion Concepts in Claude (Anthropic)

Ingested Transformer Circuits paper on emotion vectors in Claude Sonnet 4.5: 171 emotion representations, circumplex structure, invisible misalignment, emotion deflection, sycophancy link. Pages created: claude-emotion-concepts (summary). Cross-referenced with human-in-the-loop, byzantine-fault.

[2026-04-08] ingest | Bitcoin and Quantum Computing (Neha Narula)

Ingested risk analysis of Bitcoin's PQ migration challenge: probabilistic framework, Taproot's 4-year precedent, UTXO migration dilemma, absence of concrete plans. Pages created: bitcoin-quantum-computing (summary). Cross-referenced with crqc-timeline, post-quantum-cryptography, ml-dsa, store-now-decrypt-later.

[2026-04-08] ingest | Win32 Is The Only Stable ABI on Linux (ivyl)

Ingested 2022 post on glibc 2.36 breaking Linux games by dropping DT_HASH ELF sections. Pages created: win32-stable-abi (summary), abi-stability (concept).

[2026-04-22] ingest | Quartz (jackyzha0)

Added quartz to toolbox — Obsidian-compatible static-site generator for publishing Markdown vaults as digital gardens (TypeScript, MIT, 11.9k stars). Created digital-garden concept page covering the Garden vs Stream framing and how the format relates to Zettelkasten, PARA, and the LLM Wiki pattern. Cross-referenced with llm-wiki-pattern, files-as-graph-database, filesystem-is-graph-database, obsidian-ai-complete-guide, karing, mihomo, para-method.