knowledge 2026-07-27: CC v2.1.216–220 guardrails, Opus 5, Fable 5 resolved, 2 backlog papers read

[espi]

Routine self-improvements

• Applied (auto-gated): none this pass. No self-edit: commit; update-knowledge/SKILL.md was not modified.
• Suggested (needs your confirm): none. Research surfaced no stale internal cross-reference, broken relative path, or typo in the skill file. (I did verify the skill's one relative link — ../../../knowledge/archive/resolved-caveats.md — still resolves on disk.)

This PR touches only knowledge/. The routine never merges to main — a human reviews and merges.

---

Knowledge summary (window: Jul 20–27, 2026)

Seven-day pass, five parallel research agents (tooling / ecosystem / key voices / guardrails & cost / verification & skills). A modest but real window: the core is a Claude Code guardrail cluster that maps straight onto this repo's hard stops, plus model/billing resolutions and two long-standing backlog papers finally read. Several primaries returned 402/403 to automated fetch (x.com especially) — noted per-claim in sources.md.

Step 1 note: no update-knowledge PR was open. The only open PR is #12 (artifact-audit), which does not touch knowledge/, so no merge conflict is expected with this PR.

New / changed

• Claude Code v2.1.216–220 (Jul 20–25) — guardrail-tightening cluster (High, changelog read directly; whats-new/2026-w30 unpublished so the changelog was the sole primary):
  • --max-budget-usd now halts background subagents (v2.1.217) — closes a gap where the dollar ceiling didn't reach backgrounded fan-out. Directly reinforces §6 hard stop knowledge: 2026-06-22 update pass (7-day delta from Jun 15) #3.
  • Concurrent-subagent cap, default 20 (CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS, v2.1.217).
  • Subagent-nesting default flipped twice in one week (v2.1.217 off → v2.1.219 depth-3, CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH) — flagged as a re-verify item.
  • Dynamic Workflows default to a "medium" size guideline (~<15 agents, workflowSizeGuideline, v2.1.219); auto-mode checks moved to the classifier + /code-review no longer auto-launches + context: fork skills run in background (v2.1.218); sandbox.network.strictAllowlist, sandbox.filesystem.disabled, DirectoryAdded hook, symlink-write fix.
• Claude Opus 5 (claude-opus-5) — shipped Jul 24 (v2.1.219) as the default Opus model; 1M context, $5/$25 per MTok (unchanged from Opus 4.8), fast mode $10/$50, new xhigh tier; Opus 4.7 removed from fast mode. High.
• Fable 5 metered billing went live Jul 20 as planned (resolving the Jul 7→12→19 slips): Premium keeps it included to 50% of the weekly limit; Pro/Standard move to $10/$50 usage credits with a one-time $100 credit. High — resolves the primer's "live-moving deadline."
• Osmani "Software Factories, Light and Dark" (Jul 22, High, primary fetched): "harnessing loops at scale"; the back-pressure rule — "you can only hand a loop as much autonomy as you can cheaply and reliably verify, and not one inch more" (reuses Huntley's Jan-2026 term); "comprehension debt"; the "dark factory" image. Maps onto the repo's verification-gated-autonomy non-negotiables.
• Amp self-scheduling (Jul 21) — agents re-wake themselves with full context, no published re-wake cap; added to the peer-harness section as a live uncapped-loop example. High feature / Medium cap-absence.
• MCP 2026-07-28 release candidate — stateless core, first-class Tasks extension (bounded long-running async), MCP Apps, OAuth/OIDC, 12-month deprecation policy. High.
• Backlog papers read & promoted: "When Agents Do Not Stop" (arXiv:2607.01641) — IAL-Scan static analysis, 91.9% precision across 6,549 repos, now cited in §6 as empirical support for the max-iteration/stall hard stops; SkillCoach (2607.01874) — process rubrics complement a deterministic check. Plus new SkillCorpus (2607.15557, v4 Jul 23) and pre-window Recursive Self-Improvement (2607.07663) whose "verification hierarchy / self-confirming loops" finding is a citable primary for the "never declare done on self-assessment" rule. High (abstracts read directly).

Corrected / clarified

• SKILL.md governance softened (primer §4): two more independent agents confirm the LF's AAIF stewards MCP / AGENTS.md / goose, not SKILL.md (Anthropic-authored, community-maintained); cross-tool execution of SKILL.md has thin primary evidence — AGENTS.md is the real broad convention.
• Agent SDK billing split — the recurring "went live July 10" secondary claim surfaced again (one agent, Medium) and was again rejected: a second agent independently confirmed "still paused" (High), matching last pass. A textbook "don't trust search snippets" case. Status: paused, no revised plan (Medium).

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

• SkillCoach + "When Agents Do Not Stop" — both now read in full.
• Fable 5 metered-billing deadline — resolved (went live Jul 20).

Every currently-open re-verify item (per step 2 — carried forward each pass until resolved)

• SKILL.md cross-tool execution contested (AGENTS.md may be the real convention)
• Agent Skills governance — converging toward "not AAIF-governed," still open
• "costliest thing is managing the loop" is a paraphrase, not a sourced Cherny quote
• "5 tips for running agents autonomously" — real in substance, not a Cherny numbered list
• /goal "Codex invented it, Claude copied in 11 days" — single secondary
• $47K / 11-day loop and overnight-billing figures — self-reported anecdotes
• $500M in one month uncapped incident — unnamed, no primary
• June-15 billing split — still paused; re-check when a revised plan is announced
• Microsoft dropping Claude Code — multi-outlet but no primary read
• Huntley's Loom "orchestrator" framing — README more modest than the talk
• AI Engineer World's Fair 2026 sessions (Yegge / Osmani) — no verbatim recovered
• Cobus Greyling "HarnessX" essay — not directly fetched
• "Graph engineering" — still no primary definition; now well-characterized as a contested meme (Turing Post Jul 20 + Hamel Husain Jul 18)
• roborev.io/changelog 403s to automated fetch
• EvoAgentBench + SkillCheck — too new/thin to promote
• "Loop Engineering Is Dead" Medium piece — opinion, no primary data
• New this pass: roborev-dev vs kenn-io repo identity; GuardFall (single-aggregator, unverified); AgentGuard/LoopGain re-findability; Portkey (→Palo Alto) / Helicone (→Mintlify) ownership changes; subagent-nesting default flipped twice in one week

Not promoted

Linear "Loops" product (Jul 21, adjacent-but-distinct); Opus 4.7 fast-mode removal (captured with Opus 5); SkillCloak / Mitiga "Breaking Skills" / GuardFall supply-chain items (pre-window or single-aggregator — GuardFall parked in the re-verify backlog instead).

🤖 Generated with Claude Code

https://claude.ai/code/session_01YNFAE8jEJsNcMSDPukf6zH

---

Generated by Claude Code