knowledge: Aug 17–24 pass — /goal check-in backoff, Skills API GA, MCP roadmap, Osmani "Human judgment relocates", roborev v0.66.0, VB Pulse enforcement-gap survey

[espi]

Routine self-improvements

• Applied (auto-gated): none this pass. No self-edit: commit was made; update-knowledge/SKILL.md was not touched.
• Suggested (needs your confirm): none. I read SKILL.md in full for stale cross-references, broken relative paths, and typos (the only may-auto-apply categories). Step cross-references (2→8, 7→8, guardrails→7), the ../../../knowledge/archive/resolved-caveats.md path, and the CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION-removal note in "Guardrails for this skill" are all current and correct — nothing surfaced. Default-deny held.

---

Knowledge summary (Aug 17 – Aug 24, 2026)

Seven-day pass, five parallel research agents. A quieter-than-usual window; the real substance is one genuinely loop-relevant /goal change, the Claude API Skills-API GA, a new MCP roadmap, a new Osmani essay, roborev v0.66.0, a sharp enforcement-gap survey, and an in-window arXiv batch. Clean baseline — no update-knowledge PR was open (PR #16 merged; branch sits at that HEAD). The headline claim (the /goal backoff) was spot-verified against the primary changelog, catching a research agent's one-version-off attribution.

Branch note: this session was assigned the branch claude/focused-clarke-sbur5p rather than the skill's default claude/knowledge-update-YYYY-MM-DD. It still carries the claude/ prefix, so the Routine's "never touches main" push restriction holds; the name just differs from the convention.

What's new

• Claude Code v2.1.234–241 (Aug 17–23) — most loop-relevant, all in v2.1.239 (Aug 21): /goal repeat check-ins on long-running background work now back off (30 min → 1 h → every 2 h; opt out CLAUDE_CODE_GOAL_CHECKIN_MINUTES=0), and /goal restores its active goal on claude --resume. §6: --max-budget-usd//cost/status-line estimates now include the 1.1× US-only-inference premium for data-residency workspaces. Plus a "Concise" output style, ANTHROPIC_DEFAULT_MODEL, /permissions+/add-dir openable while working, a /design preview. No new model. Weeks 33 and 34 digests now published (w33 had 404'd). Primer §4. High.
• Claude API: Agent Skills / Skills API GA (~Aug 19–20) — /v1/skills GA, skills-2025-10-02 beta header retired; computer use GA, browser use tool, Files API GA, Python SDK v1.0. "Skills as a durable asset" now rests on a GA primitive. Primer §4. Medium-High.
• MCP "The New MCP Roadmap" (Aug 22) — loop-relevant priorities: maturing agentic messaging / hardening the Tasks extension, and progressive/lazy tool discovery. Primer §4. High.
• Osmani "Human judgment doesn't leave the software factory. It relocates." (Aug 21) — introduces a "verification budget" (cheap checks early, expensive late) and "mental model debt"; extends his arc. Primer §3. High.
• roborev v0.66.0 (Aug 22) — review daemon defers its own self-updates while reviews are in flight, global autofix guidelines, better security-review precision, zero-output-review CI-gate fix. Primer §5A. High.
• VentureBeat "VB Pulse" survey (Aug 20) — 107 enterprises; 21% can't stop a runaway agent's spend in real time. The clearest current field measurement of the alert-vs-enforcement gap §6 exists to close. Primer §6. High.
• Gartner "agentic AI costs to rise >fivefold by end of 2028" (via The Register, Aug 17). Primer §5B. Medium-High.
• In-window arXiv batch (dates verified on abs pages): promoted to §5A — AI-to-AI Code Reviews of GitHub PRs (2608.21311) and Artic / "NL Workflows Are Not Software Yet" (2608.21341, per-step local verification obligations); recorded in sources.md — Jagged Frontier (2608.18389, harness robustness), Harness Paradigm (2608.20622), When Agents Coordinate (2608.16801), TRUSS (2608.17588), Agent Lightning v1.0 (2608.17528), Terminal Agents survey (2608.20485), Agent-Friendly Docs (2608.20195), plus borderline Adversarial Review (2608.18167, v1 Aug 16, one day pre-window).
• Peer CLIs (sources.md, no new hard-stop primitive): Codex v0.148–0.149.1 (codex agents dashboard, codex queue), Goose v1.47.0 ("unrolled agent-loop state machine"), Gemini CLI → Antigravity CLI rename, LiteLLM v1.98.0 + OpenRouter Analytics (tracking, not enforcement).

Corrected / clarified

• /goal backoff version corrected to v2.1.239 (agent said v2.1.238/239) — the recurring one-version-off trap, caught by spot-verifying the primary.
• Codex v0.149.0 "2021" date rejected — a WebFetch hallucination; the releases index confirms Aug 2026.
• Agent SDK billing split — still paused, no revised plan; the false "went live July 10" snippet resurfaced and was again rejected.

Deprecated / archived (step 6)

• Yegge "Shape of Things to Come, Part 2: Model Welfare" exact date — resolved and archived: two agents place it ~Aug 3 (out of window, predates last pass), and it's a model-welfare framing, not a loop pattern. Not worth tracking further.
• Pruned the now-stale "w33 still 404 / changelog sole primary" parenthetical from the primer's v2.1.227–233 entry.

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

1. SKILL.md cross-tool execution (Claude Code + Gemini CLI execute; 30+ accept; AGENTS.md separate — no in-window change; Agent Skills is now a GA API).
2. "costliest thing is managing the loop" paraphrase (unsourced).
3. /goal "Codex→Claude 11 days" — single secondary.
4. $47K / overnight billing anecdotes — self-reported, unverified.
5. $500M in one month — unnamed company, no primary.
6. June-15 2026 billing split — still paused, no revised plan ("July 10 live" rejected again).
7. Microsoft dropping Claude Code — no primary read.
8. Huntley's Loom "orchestrator" framing — no datable in-window post.
9. AI Engineer World's Fair 2026 sessions — no transcript.
10. roborev.io/changelog 403s to automated fetch.
11. EvoAgentBench + SkillCheck too new (SkillCheck now v3.29/v3.30 but month-dated only — can't confirm in-window).
12. NEW: LoopsBench (arXiv:2608.00267, Jul 31, out of window but named after this repo's subject) — read next pass before promoting.

The human reviews and merges. Never merged to main automatically.

🤖 Generated with Claude Code

https://claude.ai/code/session_011yYNbfH28ecF6dyXKhJNQJ

---

Generated by Claude Code