knowledge: Aug 10–17 pass — CC v2.1.227–233, Practical Loop Engineering, Agent Plugins GA, roborev v0.65.0

[espi]

Seven-day update-knowledge follow-up pass covering Aug 10 – Aug 17, 2026 (last pass was Aug 10, merged via #15). Five parallel research agents; every load-bearing claim primary-verified. No prior update-knowledge PR was open — clean baseline (this branch sits at merged-main HEAD).

Routine self-improvements

Applied (auto-gated): none this pass. I reviewed update-knowledge/SKILL.md for stale cross-references, broken relative paths, and typos (the only may-auto-apply categories); step numbers (1–8) and the ../../../knowledge/archive/resolved-caveats.md path all check out on disk. Nothing qualified for the routine's own auto-apply lane.

Suggested → RESOLVED (human-authored, in this PR):

• The "Guardrails for this skill" reliance on CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION was stale — now fixed. Flagged for confirmation, then re-verified against the primary docs at the maintainer's request: the sub-agents page now states "There's no limit on the total number of subagents Claude can spawn over a session," and the env-vars reference no longer lists the variable at all. So it can no longer serve as the routine's fan-out backstop. With explicit maintainer go-ahead, the guardrail block (a protected region) was updated as a normal human-authored commit — not a self-edit: commit, so the self-edit-guard envelope is untouched (verified: guard passes, 0 self-edit commits). The backstop now points at the native caps that still exist — CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS (default 20, concurrency) and CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH (default 3, depth) — plus the existing ~5-agent limit. The corresponding re-verify backlog item is resolved and archived.
• Still needs you (settings-only, can't be done in a PR): drop CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION=12 from the Routine's env config — it no longer does anything; rely on the concurrency/depth defaults instead.

---

Knowledge changes

New facts added

• Claude Code v2.1.227–233 (Aug 10–14) — guardrail/subagent cluster; no new model, no new weekly digest (w33 still 404). Subagent forking on by default (v2.1.232); todo/task-tracking tools removed on Opus 4.8 / Sonnet 5 / Fable 5 / Mythos 5 and newer (restore with CLAUDE_CODE_ENABLE_TODO_TOOLS=1, v2.1.233); CLAUDE_CODE_TOOL_MEMORY_LIMIT (Bash memory cgroup cap, §6-relevant) + forward_user_identity (per-user spend attribution) (v2.1.233); /commit-push-pr no longer auto-approves dangerous git flags (v2.1.228); synced-skill prompt-injection hardening (v2.1.227, a §5A "Friendly Fire" mitigation). (High — changelog read directly.)
• Osmani "Practical Loop Engineering" (Aug 14) — states two of this repo's rules almost verbatim: "One sub-agent drafts the change. A separate one verifies it" and, for /goal users, "The evaluator sitting behind goal is not that checker… It doesn't look at the content." Added to primer §3. (High — read directly.)
• Agent Plugins 1.0 first cross-vendor adoption (Aug 12) — GitHub shipped it GA across VS Code, Copilot CLI, the Copilot SDK, and the Copilot app on all Copilot plans. Spec repo also read first-hand this pass (10 fields / 2 required, five founding maintainers + Vercel-lead + Google day-of, not an AAIF project) — resolves that backlog item. Primer §4. (High.)
• roborev v0.65.0 (Aug 17) — job-level CI cost exports, native browser app, configurable reasoning-effort tiers. Primer §5A + sources. (High.)
• Three in-window arXiv papers — spec-first convergence with a "two consecutive passes returning zero findings" stop rule (2608.12440); "Engineering Reliable Coding Agents" / verification-as-system-layer (2608.13867); "Why Does CLAUDE.md Keep Growing? Catastrophic Remembering" (2608.11095, on-thesis for this repo's own KB-discipline convention). (All High — abstracts + dates verified.)
• Ramp AI Index, August edition — whales-first spend (top 1% ~$7,400/employee/mo, >600:1 gap). Primer §5B. (High.)

Corrected

• Two changelog items were misattributed by one version by a research agent — synced-skill hardening is v2.1.227 and /commit-push-pr is v2.1.228 per the primary changelog (not v2.1.228/229). Fixed on verification.
• Anthropic Agent SDK billing split — still paused, no revised plan (confirmed again; the false "went live July 10" snippet did not resurface).

Archived (resolved → archive/resolved-caveats.md)

• Agent Plugins 1.0 spec not read directly → read first-hand this pass.
• "Loop Engineering Is Dead" backlash Medium piece → read directly, cites no data; SlopCodeBench (arXiv:2603.24755) recorded as the real empirical anchor for the code-erosion critique.
• CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION after the cap removal → re-verified against the primary docs (the follow-up above); env var gone, no per-session total cap.

Currently-open re-verify backlog (every open item, per step 2)

• SKILL.md cross-tool execution — unchanged (Claude Code + Gemini CLI execute; 16+ tools accept; AGENTS.md separate).
• NEW: Yegge "Model Welfare" (Part 2) exact date — byline month-only, ~Aug 10–11 by social signal; peripheral to loops regardless.
• Carried forward unchanged: "costliest thing is managing the loop" paraphrase; /goal "Codex→Claude 11 days" single secondary; $47K/$500M/overnight anecdotes (all recirculated, nothing new); June-15 billing split (still paused); Microsoft dropping Claude Code (no primary); Huntley's Loom framing; AI Engineer World's Fair 2026 sessions; roborev.io/changelog 403s; EvoAgentBench + SkillCheck.

Not promoted

Yegge "Model Welfare" essay (peripheral, unconfirmed date); Factory Droid v0.193–197 (minor stall-adjacent fix); DeepSeek V4-Pro tiered pricing (pricing, not a guardrail); Codex CLI v0.147.0 / MCP 2026-07-28 (no new in-window release).