It took a packet sniffer to find out what your coding agent actually sends.

The replay half of the watch this newsletter set on 07-08 ("the agent's trail is an asset — the verb nobody's shipped is replay") now exists as one Go binary. Mindwalk reads Claude Code and Codex session logs straight off your disk — no arguments needed, it scans ~/.claude/projects and ~/.codex/sessions — and plays each session back as light moving across a 3D night map of the repository: searched files glow moss green, read files moon white, edits warm amber, everything the agent never touched stays dark. A playback deck lets you scrub the run over a histogram where observation stays cool and mutation glows warm; timeline marks flag context compactions, subagent launches, and user turns as click-to-jump targets, and a HUD strip folds error rate, churned files, and edits-after-last-verify into a review signal. MIT, 398 stars, created four days ago by cosmtrek — the author of air, the Go live-reload tool half the Go world runs. Fully local; the install script verifies checksums, Windows archives on releases.

Reach for it when the agent says "done" and you want to know whether its footprint matched the scope you had in mind — did it read the config before editing the handler, or edit first and wander after. Delete the JSONL spelunking session where you grep the transcript to reconstruct what happened. Tradeoff: this is replay-as-viewing, not replay-as-execution — you can see the run, not fork it from minute twelve and send it down a different path, so the other half of the replay watch is still open. And it's four days old; the map is only as good as the session logs the harnesses happen to write.

The pitch opens on the waste: an agent that summarized a thread yesterday will re-derive how to summarize threads tomorrow, burning frontier inference on a procedure with a known shape — cost, latency, and drift, every session. Skillscript's answer is that agents have no substrate to write themselves down in, so it ships one: a small declarative language where a learned procedure crystallizes into a named, typed, composable skill — cheap to execute, cheap to inspect, cheap to improve — plus the runtime, compiler, CLI, dashboard, and an MCP surface so your existing agent can call what it has crystallized. Skills compose (skills calling skills), can wake agents on schedule, and can route sub-steps to local models so the frontier model stops doing the rote parts. MIT, npm install -g skillscript-runtime, pre-1.0 and stated as such; created in May, launched on Show HN today at 52 stars.

Reach for it when your CLAUDE.md has a paragraph of standing instructions the model re-interprets slightly differently every session — that paragraph is a program in soft form, and this is the hard form. The contrast with SkillOpt (our 07-08 pick) is the interesting part: SkillOpt optimizes a prose skill file against a validation set; Skillscript argues prose is the wrong substrate entirely and the skill should be code. Delete the re-derivation tax on every routine task. Tradeoff: a new language is a real adoption tax at 52 stars and pre-1.0 — and the design bets the model will faithfully call the crystallized skill instead of freelancing, which is a bet prose-instruction veterans will recognize.

Persistent memory for coding agents with the one property most memory layers fumble: a deletion policy. capn init installs a session-start hook for Claude Code or Codex; the agent asks capn before searching the codebase (capn ask "where are payment webhooks handled?"), and after a hard-won discovery it saves the question plus the exact files that answer it. Every backing file is sha256-fingerprinted at save time — the moment any of them changes or disappears, the entry deletes itself before it can ever answer again. Saved answers are never edited: either the files haven't moved and the chart is true, or it's worthless and gets thrown out. The author's own eval — 60 real developer questions across five production codebases (Dub, Polar, PostHog, Twenty, Documenso), methodology and raw evidence published in the repo — measures 77% fewer tokens on repeat questions, paying for a charting session in about 1.6 recalls. MIT, npm install -g capn-hook, created July 4, 22 stars.

Reach for it when your agent spends the first ten minutes of every session rediscovering where things live in a codebase it explored yesterday. Delete the daily re-exploration bill — the worst case is a stale entry gets pruned and the agent explores cold, which is what it would have done anyway. Tradeoff: the 77% is the author's number, not independently reproduced (though the eval is published and rerunnable), and the whole integration is one polite note the model reads at session start — no wrapper, no forced behavior. Honest design; weaker guarantee.

zot is a single-binary Go coding agent whose README calls itself "yet another coding agent harness, lightweight and in beta forever" — self-awareness this newsletter can respect. What it shipped today is the interesting layer: .zot files, which package an agent's persona and standing instructions, its skills, its requirements (minimum runtime, OS, required binaries, model capabilities), and its tool permissions into one portable artifact. The permissions are a ceiling the runtime enforces, not a suggestion the model reads — a code-reviewer agent packaged read-only cannot write, however persuasively it's prompted. zot inspect and zot verify let you audit an artifact before zot run ./agent.zot executes it, and agents run from a source directory, an archive, or a GitHub URL. MIT throughout; the harness is 292 stars, created in April, and runs locally on your own model credentials.

Reach for it when "set up the review agent" is currently a README paragraph teammates follow by hand and get slightly wrong. This is the packaged-expertise thread (06-11 security, 06-19 standardization, 07-01 supply) moved one level up — not a skill you load into your agent, but the agent itself as a distributable, auditable unit, with the enforcement-beats-instruction pattern (deptrust, MakerChecker, Kastra) built into the format. Delete the tribal setup knowledge and the hope that the prompt's "please don't write files" holds. Tradeoff: a .zot is portable between zot users, not between harnesses — the format only means something inside a 292-star runtime, and packaged-agent formats are only as valuable as the population that can run them.

Friday's edition was about software growing a machine-legible second interface; FableCut built a video editor around that idea from scratch. aftr does the harder thing: it retrofits the second interface onto the incumbent. A Node controller speaks JSON over a WebSocket to a CEP panel inside After Effects, which executes commands as ExtendScript and returns JSON — and the whole command set is exposed as MCP tools, so claude mcp add --transport http aftr http://127.0.0.1:8787/mcp puts Adobe's compositor under your agent's hands. On top sits a spec-driven pipeline that builds clips and self-corrects by rendering, reviewing the output, and revising. The engineering hygiene is unusual for a day-old repo: a headless simulator plus a test suite proves the stack healthy with no After Effects installed, and it ships via npm, pip, or Docker. MIT, created today, five stars.

Reach for it when the motion-graphics step of your pipeline is a human clicking through AE for the fortieth title card. Delete the screenshot-driven computer-use loop — patching a comp through a typed command channel is what models are good at, which is the same concession Flint made about charts on Friday. Tradeoff: the slate's freshest bet by hours, and it presupposes After Effects — paid, desktop, panel deployed from a clone — so this only helps people already inside Adobe's walls; the self-correcting render loop is the author's own demo, not a benchmark.

Off the thread but worth knowing: Ant (antjs.org, 320 points, the day's biggest Show HN) is a new JavaScript runtime with its own engine, package manager, and registry at ants.land — off-vertical for this newsletter, but a from-scratch JS engine shipping in 2026 is worth a look on its own merits. Osaurus (github.com/osaurus-ai/osaurus, MIT, ~7,000 stars, created August 2025) resurfaced via a Product Hunt launch today — a native macOS harness for fully local agents, and the most complete installable answer yet to the local-model wave this newsletter has tracked since 06-30; established, not new, disclosed as such. And claude-code-proxy (github.com/raine/claude-code-proxy, MIT, Rust, 275 stars, created April) is the wire teardowns' mirror image: once you can see the traffic between Claude Code and its model, you can also swap what's on the other end — it's a local Anthropic-compatible proxy that drives the Claude Code harness with your ChatGPT Plus, Kimi, Grok, or Cursor subscription. It's real, shipped, and brew-installable, and it almost certainly runs against those providers' terms of service — the author's stated motive is that Anthropic keeps tightening usage limits while others are more generous, which tells you the demand-side pressure is the same one both teardowns measured in tokens.