From 12e548e638e95c8623aecaf6246ae8a6805558d4 Mon Sep 17 00:00:00 2001 From: Claude Date: Mon, 27 Jul 2026 05:18:24 +0000 Subject: [PATCH] =?UTF-8?q?knowledge:=20Jul=2020=E2=80=9327=20pass=20?= =?UTF-8?q?=E2=80=94=20CC=20v2.1.216=E2=80=93220=20guardrails,=20Opus=205,?= =?UTF-8?q?=20Fable=205=20resolved?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Seven-day update-knowledge pass (five research agents). Substantive but modest window; the core is a Claude Code guardrail cluster that maps onto this repo's hard stops, plus model/billing resolutions and two backlog papers finally read. - Primer §4/§6: CC v2.1.216–220 — --max-budget-usd halts background subagents (v2.1.217), concurrent-subagent cap default 20, subagent-nesting default flip (off→depth-3), Dynamic Workflows "medium" default, auto-mode classifier moves. - Primer §4: Claude Opus 5 (Jul 24, default Opus model); Fable 5 metered billing went live Jul 20 as planned (resolves the live-moving-deadline item). - Primer §3/sources: Osmani "Software Factories, Light and Dark" (Jul 22) — back-pressure rule, comprehension debt. - Primer §4: Amp self-scheduling (Jul 21) with no re-wake cap — uncapped-loop hazard added to peer-harness section. - Primer §6/sources: "When Agents Do Not Stop" (arXiv:2607.01641) as empirical support for the hard stops; SkillCoach + SkillCorpus + Recursive Self-Improvement papers added. - Softened SKILL.md governance claim (not AAIF-governed; AGENTS.md is the real cross-tool convention); rejected the recurring "Agent SDK split went live Jul 10" claim again. - Archived resolved caveats (two papers read, Fable 5 deadline); refreshed the standing re-verify backlog with new items. Co-Authored-By: Claude Opus 4.8 Claude-Session: https://claude.ai/code/session_01YNFAE8jEJsNcMSDPukf6zH --- knowledge/00-primer.md | 122 +++++++++++++--- knowledge/CHANGELOG.md | 103 ++++++++++++++ knowledge/archive/resolved-caveats.md | 19 +++ knowledge/sources.md | 196 +++++++++++++++++++++----- 4 files changed, 380 insertions(+), 60 deletions(-) diff --git a/knowledge/00-primer.md b/knowledge/00-primer.md index 7ac89b6..10637bf 100644 --- a/knowledge/00-primer.md +++ b/knowledge/00-primer.md @@ -1,6 +1,6 @@ # Loops: the primer -> The canonical briefing for this repo. Last substantive update: 2026-07-20. +> The canonical briefing for this repo. Last substantive update: 2026-07-27. > Companion: [`sources.md`](sources.md) (every claim's source + confidence), > [`CHANGELOG.md`](CHANGELOG.md) (dated updates). @@ -76,7 +76,18 @@ for ~3 months, with no single published cost figure. covered — plus Sonar's 2026 State of Code report (42% of committed code AI-generated/assisted) and GitLab's June 2026 AI-accountability research. Names three costs of over-delegation: **cognitive surrender**, **cognitive - debt**, and **"orchestration tax."** + debt**, and **"orchestration tax."** His follow-up, **"Software Factories, + Light and Dark"** (Substack, July 22, 2026), reframes a software factory as + *"harnessing loops at scale"* and makes the same verification-gated-autonomy + argument this repo enforces, in one rule: *"Back pressure is the rule that + you can only hand a loop as much autonomy as you can cheaply and reliably + verify, and not one inch more"* (borrowing Geoffrey Huntley's Jan-2026 "back + pressure" term). The "dark factory" (lights physically off, only machines on + the floor) is his image for full autonomy; he warns it's earned per-task by + cheap verification, not switched on wholesale, and names **"comprehension + debt"** — *"the widening gap between how much code exists and how much any + human still understands"* — as its cost. High-stakes paths (auth, billing) + keep human gates regardless of speed. - **Peter Steinberger (@steipete)** — the tweet that lit the fuse (~Jun 7 2026): *"you shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents."* Companion point: **wrap repeated or hard @@ -208,22 +219,27 @@ generalize, not as the only place they exist. MCP, which is how you find what a loop is actually costing. V2.1.174 added a per-skill/agent/plugin/MCP attribution breakdown (cache misses, long context, 24h/7d) to the VS Code Account dialog. -- **Model availability** — **Claude Sonnet 5** (`claude-sonnet-5`) became Claude - Code's **default model** in v2.1.197 (June 30, 2026) — native 1M-token - context, promotional pricing $2/$10 per MTok input/output through August 31, - 2026. **Claude Fable 5** (`claude-fable-5`; 1M context, 128k output, $10/$50 - per MTok I/O) launched June 9, 2026 (v2.1.170), was briefly suspended June - 12–13 following a US government export-control directive, and is **back on - the platform** as of June 22, 2026. Fable 5 was slated to move off - included-subscription access onto metered usage credits ($10/$50 per MTok - I/O) on July 7, 2026, but after user pushback Anthropic extended included - access twice within days — first to July 12, then again on July 13 to - **July 19, 2026** — so metered billing for Fable 5 now begins July 20 - rather than the originally announced date. Track this as a live-moving - deadline, not settled. **Claude Mythos 5** (`claude-mythos-5`) - — limited availability via Project Glasswing since June 9; same pricing and - context. **Claude Opus 4.1 is deprecated** (retiring August 5, 2026). All - other models (Opus 4.8, Haiku) unaffected. +- **Model availability** — **Claude Opus 5** (`claude-opus-5`) shipped **July + 24, 2026 (v2.1.219)** as Claude Code's **default Opus model** — 1M-token + context, 128k max output, priced **$5/$25 per MTok I/O (unchanged from Opus + 4.8)**, fast mode $10/$50 (~2.5× faster), and a new `xhigh` reasoning tier. + `/fast` now covers Opus 5 and Opus 4.8; **Opus 4.7 was removed from fast + mode** (Jul 24). *(High — changelog read directly, multiple secondaries.)* + **Claude Sonnet 5** (`claude-sonnet-5`) remains the subscription **default + model** (since v2.1.197, June 30, 2026) — native 1M-token context, + promotional pricing $2/$10 per MTok I/O through August 31, 2026. **Claude + Fable 5** (`claude-fable-5`; 1M context, 128k output) launched June 9, 2026 + (v2.1.170), was briefly suspended June 12–13 (US export-control directive), + and returned June 22. After three deadline slips (Jul 7 → 12 → 19), **Fable 5 + metered billing went live July 20, 2026 as planned**: Max & Team Premium keep + Fable 5 included up to 50% of the weekly usage limit (stated permanent); + Pro & Team Standard move to usage credits at $10/$50 per MTok I/O (2× Opus + 4.8), softened by a one-time $100 credit claimable Jul 20–Aug 2. *(High — + changelog + corroborating secondaries; the earlier "live-moving deadline" + is now resolved.)* **Claude Mythos 5** (`claude-mythos-5`) — limited + availability via Project Glasswing since June 9; same context. **Claude Opus + 4.1 is deprecated** (retiring August 5, 2026). Other models (Opus 4.8, Haiku) + unaffected. - **Diagnostic flags** — `--safe-mode` / `CLAUDE_CODE_SAFE_MODE=1` (v2.1.169+) disables all customizations (skills, hooks, MCP, plugins, themes) for debugging without affecting auth. `fallbackModel` setting (v2.1.166+) chains up to three @@ -242,11 +258,18 @@ generalize, not as the only place they exist. version-controlled, testable, loaded on demand. `/loop` itself is one. As of v2.1.178, skills in **nested `.claude/skills/` directories** load automatically; on name clash, both appear as `:`. Since Dec 2025 - the **Agent Skills spec is an open standard** (Anthropic-authored, now under - the Agentic AI Foundation / Linux Foundation), adopted by Codex CLI, Copilot, - Cursor, VS Code, and ~40 additional products as of June 2026 — a skill - written here is portable across those tools; custom commands - (`.claude/commands/`) have been folded into skills. **Microsoft's Agent + the **Agent Skills / `SKILL.md` spec is an Anthropic-authored open format**, + community-maintained (agentskills.io) and adopted as a *format* by Codex CLI, + Copilot, Cursor, VS Code, and ~40 products — but two independent research + passes now agree it is **not** itself an AAIF/Linux-Foundation-governed + project (the LF's Agentic AI Foundation stewards **MCP, `AGENTS.md`, and + goose**, not `SKILL.md`), and there is **thin primary evidence that non-Claude + CLIs actually *execute* a `SKILL.md`** vs. merely accepting the format — + `AGENTS.md` is the genuinely broad cross-tool convention. So a skill written + here is portable *in principle* but assume per-tool testing, not drop-in, and + don't lean on its "governed open standard" status (both tracked as re-verify + items in `sources.md`). Custom commands (`.claude/commands/`) have been folded + into skills. **Microsoft's Agent Skills for .NET** (Microsoft Agent Framework) exited experimental preview to **stable/GA on July 7, 2026** — a first-party .NET implementation of the same SKILL.md-based open format, another concrete adoption data point beyond the @@ -321,6 +344,34 @@ generalize, not as the only place they exist. self-trigger after edits — a loop that relied on that implicit behavior must now call them explicitly as part of its verification step (see primer §5A). +- **v2.1.216–220 (Jul 20–25)** — a cluster of changes that tighten the + fan-out blast radius, several mapping straight onto this repo's §6 hard + stops. **`--max-budget-usd` now halts background subagents** (v2.1.217, Jul + 21): once the cap is hit, new subagent spawns are denied and running + background agents are stopped — previously the dollar ceiling didn't reach + backgrounded fan-out, so this closes a real gap in hard stop #3. Same release + added a **concurrent-subagent cap, default 20** (`CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS`), + "so one message can't fan out unbounded background agents," and made + subagents **not** spawn nested subagents by default. **v2.1.219 (Jul 24)** + then flipped that: subagent nesting defaults to **depth 3** (was 1), disabled + via `CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1` — the default flipped twice + inside one week, so re-verify the shipped default before baking it into a + template. v2.1.219 also made Dynamic Workflows default to a **"medium" size + guideline (~<15 agents)** via a new `workflowSizeGuideline` key (still a + guideline a larger-scale prompt overrides, not a hard cap), added + `sandbox.network.strictAllowlist` (deny non-allowlisted hosts without + prompting) and a `DirectoryAdded` hook. **v2.1.218 (Jul 22)** moved several + auto-mode checks (dangerous-`rm`, background-`&`, suspicious Windows paths, + un-provable read-only Bash in plan mode) from permission dialogs to the + auto-mode classifier — changing how a loop in auto mode gets gated — and made + skills with `context: fork` run in the background by default, moved + `/deep-research` and `/code-review` to background subagents, and confirmed + **`/code-review` no longer auto-launches** (consistent with v2.1.215). + v2.1.216 (Jul 20) added `sandbox.filesystem.disabled` and fixed workflow / + scheduled-task writes to stop following a symlink at `.claude` (which could + redirect writes outside the project); v2.1.220 (Jul 25) was reliability fixes + only. *(High — changelog read directly; `whats-new/2026-w30` had not + published yet, so the changelog was the sole primary.)* ### Beyond Claude Code — the same loop on other harnesses @@ -347,6 +398,16 @@ as a map, not gospel — re-verify before betting on a specific flag. | **Claude Agent SDK** | `Stop` hooks | **`max_turns` + `max_budget_usd` real enforcement** | you host | baseline for a loop-of-loops | | **LangGraph · Google ADK · CrewAI · AG2** | build-your-own | opt-in; mostly no $ default | needs Temporal/Diagrid | framework substrate | +**In-window movement (Jul 2026):** **Amp** (Sourcegraph) shipped +**self-scheduling** (Jul 21, 2026) — an agent sets its own schedule and, when +it fires, "wakes up with its saved prompt and continues right where it left +off, with all of its context and history." The published feature page documents +**no cap on re-wake frequency** — a clean example of a self-perpetuating loop +shipping *without* this repo's hard stops, not with them: if you run it, the +iteration/budget ceiling is yours to add. Amp stays in the "weak native +guardrails" column of the matrix. *(High for the feature; Medium that no +internal cap exists — absence in docs ≠ confirmed absent.)* + Three things worth carrying as durable facts: - **The validator-judge stop is now cross-tool.** Both Claude Code `/goal` and @@ -531,6 +592,21 @@ first-class params (`max_turns`, `max_budget_usd`). 2026) expose org/workspace limits and per-user estimated cost, so a gateway can read spend and cut the loop off; third-party gateways (e.g. Databricks Unity AI Gateway) now hard-stop requests at a budget rather than just alert. + Claude Code itself moved closer to a real in-harness ceiling in **v2.1.217 + (Jul 21, 2026)**: `--max-budget-usd` now **halts background subagents** + (denies new spawns, stops running ones) when the cap is hit — previously the + dollar ceiling didn't reach backgrounded fan-out — alongside a default-20 + concurrent-subagent cap. Still set the ceiling explicitly; the mechanism, not + a default limit, is what shipped. + +That unbounded feedback paths are a *widespread, statically detectable* defect +now has empirical backing: **"When Agents Do Not Stop: Uncovering Infinite +Agentic Loops in LLM Agents"** (arXiv:2607.01641) defines infinite agentic loops +as unbounded repetition of model/tool/handoff calls when the feedback path isn't +bounded, and introduces **IAL-Scan** — a static analyzer that builds an "Agentic +Loop Dependence Graph" and flags paths able to hit expensive ops without a bound, +at **91.9% precision across 6,549 repos**. Independent evidence that hard stops +#1/#2 guard against a real and common failure mode, not a hypothetical one. **Enforce these tool-agnostically, at the gateway.** The cleanest place to put the hard stops isn't inside any one agent — it's the **LLM gateway every agent diff --git a/knowledge/CHANGELOG.md b/knowledge/CHANGELOG.md index 32f53d9..5db8e23 100644 --- a/knowledge/CHANGELOG.md +++ b/knowledge/CHANGELOG.md @@ -3,6 +3,109 @@ Dated record of substantive changes to `knowledge/`. The `update-knowledge` skill appends a new entry here on each research pass. Newest first. +## 2026-07-27 — Seven-day follow-up pass (Jul 20–27) + +Five parallel research agents across tooling & versions, ecosystem & techniques, +key voices, guardrails & cost, and verification & skills. A modest but real +window: the substance is a Claude Code guardrail cluster, Opus 5, the Fable 5 +billing resolution, one new Osmani essay, and two backlog papers finally read. +No `update-knowledge` PR was open (only the artifact-audit PR #12, which +doesn't touch `knowledge/` — no conflict expected). Several primaries returned +402/403 to automated fetch (x.com especially); noted per-claim. + +### New facts added + +- **Claude Code v2.1.216–220 (Jul 20–25)** — a guardrail-tightening cluster, + several mapping onto this repo's §6 hard stops: **`--max-budget-usd` now halts + background subagents** (v2.1.217) — closes a gap where the dollar ceiling + didn't reach backgrounded fan-out; **concurrent-subagent cap default 20** + (`CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS`, v2.1.217); **subagent-nesting default + flipped twice** (v2.1.217 off → v2.1.219 depth-3, `CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH`); + Dynamic Workflows default to a "medium" (~<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, and a symlink-write fix at `.claude`. **High** + (changelog read directly; `whats-new/2026-w30` unpublished, changelog sole + primary). +- **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 + three-slip deadline): Premium tiers keep 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" item. +- **Osmani "Software Factories, Light and Dark"** (Substack, Jul 22): a software + factory is "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"* (reusing Huntley's Jan-2026 term) — plus "comprehension + debt" and the "dark factory" image. Maps directly onto the repo's + verification-gated-autonomy non-negotiables. **High** (primary fetched, + verbatim quotes). +- **Amp self-scheduling** (Jul 21): agents schedule and re-wake themselves with + full context, with **no published re-wake cap** — a self-perpetuating loop + shipping *without* the three hard stops; added to primer §4's peer-harness + section as a live uncapped-loop example. **High** feature / **Medium** + cap-absence. +- **MCP 2026-07-28 release candidate** — largest revision since launch: + stateless core, first-class **Tasks** extension for bounded long-running async + work, MCP Apps, OAuth/OIDC auth, 12-month deprecation policy. **High**. +- **Two backlog arXiv papers read and promoted**: **"When Agents Do Not Stop"** + (2607.01641) — IAL-Scan static analysis, 91.9% precision across 6,549 repos, + now cited in primer §6 as empirical support for the max-iteration/stall hard + stops; **SkillCoach** (2607.01874) — process rubrics *complement* a + deterministic check. Plus **SkillCorpus** (2607.15557, v4 Jul 23, in-window) + and the pre-window **Recursive Self-Improvement** paper (2607.07663, Jul 8) + 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 + +- **Agent Skills / `SKILL.md` governance** — primer §4 softened: two more + independent agents confirm the LF's AAIF stewards **MCP / AGENTS.md / goose**, + *not* `SKILL.md` (Anthropic-authored, community-maintained), and cross-tool + *execution* of `SKILL.md` (vs. format compatibility) has thin primary + evidence — `AGENTS.md` is the real broad convention. The two governance + caveats were strengthened toward "not AAIF-governed," still carried as open. +- **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" at High, matching the last pass's + finding. A textbook primer §5A "don't trust search snippets" case. Status: + paused, no revised plan, **Medium**. + +### Archived (resolved — see 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). + +### Currently-open re-verify backlog (carried forward — every open item, per step 2) + +`SKILL.md` cross-tool *execution* contested; Agent Skills governance +(converging toward not-AAIF); "costliest thing is managing the loop" paraphrase; +"5 tips" not a Cherny numbered list; `/goal` "Codex→Claude 11 days" single +secondary; `$47K`/`$500M`/overnight anecdotes unverified; June-15 billing split +(still paused, re-check on revised plan); Microsoft dropping Claude Code (no +primary read); Huntley's Loom "orchestrator" framing; AI Engineer World's Fair +2026 sessions; Cobus Greyling "HarnessX"; **"graph engineering"** (still no +primary definition — now well-characterized as a contested meme via Turing Post +Jul 20 + Hamel Husain); roborev.io/changelog 403s; EvoAgentBench + SkillCheck +too new. **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 (evidence insufficient or peripheral) + +- **Linear "Loops"** product (Jul 21) — a product *named* Loops (self-driving + project workflows), adjacent to but distinct from loop engineering; Medium, + secondary. Noted, not promoted. +- **Opus 4.7 fast-mode removal** (Jul 24) — captured with the Opus 5 entry; + minor and peripheral to loops on its own. +- **SkillCloak / Mitiga "Breaking Skills" / GuardFall** supply-chain items — + all pre-window or single-aggregator; GuardFall added to the re-verify backlog + rather than the KB. + ## 2026-07-20 — Beyond Claude Code: tool-agnostic loop landscape Broadened the repo from Claude-Code-specific toward **principles-first, Claude diff --git a/knowledge/archive/resolved-caveats.md b/knowledge/archive/resolved-caveats.md index a6a290f..90b0ef5 100644 --- a/knowledge/archive/resolved-caveats.md +++ b/knowledge/archive/resolved-caveats.md @@ -9,6 +9,25 @@ not a to-do list. Each entry keeps the resolution date and a one-line reason it was archived rather than carried forward. +## Archived 2026-07-27 + +- **SkillCoach (arXiv:2607.01874) and "When Agents Do Not Stop" + (arXiv:2607.01641) — both read in full.** Flagged across the Jul 13 and Jul + 20 passes as "not yet read." Both abstracts were read directly this pass and + promoted to sources.md's academic section with their actual findings + (SkillCoach: process rubrics complement a deterministic check; "When Agents + Do Not Stop": IAL-Scan static analysis, 91.9% precision, empirical support + for the max-iteration/stall hard stops — now cited in primer §6). No further + action. +- **Fable 5 metered-billing deadline — resolved.** Tracked in primer §4 as a + "live-moving deadline" after three slips (Jul 7 → 12 → 19). Metered billing + **went live July 20, 2026 as planned**: Max & Team Premium keep Fable 5 + included up to 50% of the weekly usage limit (stated permanent), Pro & Team + Standard move to usage credits at $10/$50 per MTok with a one-time $100 + credit (Jul 20–Aug 2). Primer prose updated from "not settled" to the settled + outcome; the moving-deadline history lives in the CHANGELOG. No further + action. + ## Archived 2026-07-20 - **Roborev creator confirmed**: Wes McKinney (@wesmckinn, author of Pandas). diff --git a/knowledge/sources.md b/knowledge/sources.md index 4d17dff..d685c2f 100644 --- a/knowledge/sources.md +++ b/knowledge/sources.md @@ -6,7 +6,7 @@ identical verbatim across multiple independent sources; **Medium** = consistent across several secondary sources but primary not directly confirmed; **Low** = single source / unverified provenance. -Verified as of 2026-07-20. Re-check before relying on version numbers or dates. +Verified as of 2026-07-27. Re-check before relying on version numbers or dates. ## Foundations & lineage @@ -104,8 +104,22 @@ Verified as of 2026-07-20. Re-check before relying on version numbers or dates. - Peter Steinberger tweet (Jul 18, 2026): "Are we still talking loops or did we shift to graphs yet?" — teaser, possibly signaling a "graph engineering" framing shift; not independently corroborated as a real trend yet (see - re-verify list). — **Medium** (search-snippet sourced; direct fetch 403'd). - https://x.com/steipete/status/2078277297791189132 + re-verify list). Followed ~Jul 25 by a one-line "am I a graph engineer now" + post — still no long-form essay. — **Medium** (search-snippet sourced; direct + fetch 402/403'd). https://x.com/steipete/status/2078277297791189132 +- **Addy Osmani "Software Factories, Light and Dark"** (Substack, July 22, + 2026) — follow-on to "Own the Outer Loop"; a software factory is "harnessing + loops at scale," oversight calibrated per task by verification cost and + consequence. Verbatim: *"Back pressure is the rule that you can only hand a + loop as much autonomy as you can cheaply and reliably verify, and not one inch + more"* (reuses Geoffrey Huntley's Jan-2026 "back pressure" term, + ghuntley.com/pressure/); *"A dark factory runs with the lights physically off, + because the only things on the floor are machines and machines don't need + light to see"*; names **"comprehension debt"** — *"the widening gap between how + much code exists and how much any human still understands."* Maps directly to + this repo's verification-gated-autonomy non-negotiables. — **High** (primary + Substack fetched directly, verbatim quotes, date confirmed). + https://addyo.substack.com/p/software-factories-light-and-dark - Boris Cherny at Meta @Scale (June 22, 2026): "Two years ago, we wrote source code by hand. We started to transition so agents write the code. And now we're transitioning to the point where agents are prompting agents that then @@ -305,6 +319,40 @@ Verified as of 2026-07-20. Re-check before relying on version numbers or dates. v0.62.1 (Jul 14, 2026): persistent CI panel metrics + new export command; stable JSON contract for version info; Codex agent hook can invoke `roborev-fix` skill; blocks incompatible model pairings. +- **Claude Code v2.1.216–220 (Jul 20–25, 2026)** — **High** (changelog read + directly; `whats-new/2026-w30` not yet published, changelog was sole primary). + v2.1.216 (Jul 20): `sandbox.filesystem.disabled`; workflow/scheduled-task + writes no longer follow a symlink at `.claude`. v2.1.217 (Jul 21): + `--max-budget-usd` now **halts background subagents** (denies new spawns, + stops running ones) at the cap; new **concurrent-subagent cap default 20** + (`CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS`); subagents no longer spawn nested + subagents by default. v2.1.218 (Jul 22): several auto-mode checks + (dangerous-`rm`, background-`&`, suspicious Windows paths, un-provable + read-only Bash in plan mode) moved from permission dialogs to the auto-mode + classifier; skills with `context: fork` run in background by default; + `/deep-research` + `/code-review` run as background subagents; `/code-review` + no longer auto-launches. v2.1.219 (Jul 24): **subagent nesting defaults to + depth 3** (was 1; `CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1` to disable — note + the default flipped twice inside the week); Dynamic Workflows default to a + "medium" size guideline (~<15 agents) via `workflowSizeGuideline`; + `sandbox.network.strictAllowlist`; `DirectoryAdded` hook; Opus 5 added (below). + v2.1.220 (Jul 25): reliability fixes only. + https://code.claude.com/docs/en/changelog +- **Claude Opus 5** (`claude-opus-5`) shipped v2.1.219 (Jul 24, 2026) as the + **default Opus model** — 1M context, 128k output, $5/$25 per MTok I/O + (unchanged from Opus 4.8), fast mode $10/$50 (~2.5× faster), new `xhigh` + reasoning tier; `/fast` covers Opus 5 + Opus 4.8, Opus 4.7 removed from fast + mode. — **High** (changelog + secondaries). + https://code.claude.com/docs/en/changelog · + https://www.marktechpost.com/2026/07/24/meet-the-new-claude-opus-5-frontier-class-agentic-coding-and-computer-use-at-unchanged-opus-pricing/ +- **Fable 5 metered billing went live July 20, 2026 as planned** (resolving the + Jul 7→12→19 slips): Max & Team Premium keep Fable 5 included up to 50% of the + weekly usage limit (stated permanent); Pro & Team Standard move to usage + credits at $10/$50 per MTok I/O (2× Opus 4.8), one-time $100 credit claimable + Jul 20–Aug 2. — **High** (v2.1.219 changelog fixed the Fable plan-labeling, + corroborating live rollout; multiple secondaries). + https://fable5.app/fable-5-usage-limits/ · + https://usagebox.com/articles/claude-fable-5-usage-credits-switch-july-2026 v0.63.0 (Jul 16, 2026): CI quiet-hours throttling (with bypass for certain workloads); machine-readable launch receipts on `roborev run` for automation; tightened skill triggers to prevent unintended activation. @@ -405,14 +453,18 @@ Verified as of 2026-07-20. Re-check before relying on version numbers or dates. Anthropic or Claude Code change. **High** (primary OpenAI page read). https://openai.com/index/managing-ai-investments-in-agentic-era/ - **Anthropic Agent SDK billing split — still paused, no revised plan found** - as of Jul 20, 2026. No primary Anthropic announcement located in the Jul - 6–20 window revising the pause from June 15. One AI-generated search summary - claimed the split "went live July 10, 2026" — this is **unverified and - likely erroneous**: it appears only in synthesized search output, contradicts - multiple independent June reports that it remains paused pending a revised - plan, and no dated primary or secondary article confirms it. **Do not treat - as fact.** Status: paused-with-no-revision-announced remains **Medium** - confidence (absence of evidence, not evidence of absence). + as of Jul 27, 2026. No primary Anthropic announcement located in the Jul + 6–27 window revising the pause from June 15. The recurring search-sourced + claim that the split "went live July 10, 2026" **surfaced again this pass** + (one research agent reported it at Medium via a single secondary, + thenewstack.io/anthropic-agent-sdk-credits) and was **again rejected**: a + second agent independently confirmed "still paused" at High + (thenewstack.io/anthropic-pauses-...), and the last pass had already flagged + the "went live" framing as unverified and likely erroneous. **Do not treat as + fact.** This is a textbook primer §5A "don't trust search snippets" case — + two passes, two independent contradictions. Status: + paused-with-no-revision-announced remains **Medium** (absence of evidence, + not evidence of absence). Re-check when Anthropic announces a revised plan. - **Anthropic Claude Enterprise spend controls** (Jul 2, 2026): model-level entitlements, spend-threshold alerts at 75%/90% of an org's limit, per-user/ per-group cost analytics dashboard, Admin API endpoints for scripting @@ -529,8 +581,15 @@ thin/403'd; re-verify a specific flag against live docs before treating as High. separate validator. More sophisticated than a single CC session. **Medium** (secondary). https://byteiota.com/cursor-automations-always-on-ai-coding-agents-end-prompt-loop/ - **Amp** (Sourcegraph) — autonomous multi-step agent; **no documented** - iteration cap / stall detector / budget ceiling / scheduler (guidance is "keep - threads short"). MCP. **Less** on guardrails. **Medium/Low** (secondary). + iteration cap / stall detector / budget ceiling (guidance is "keep threads + short"). MCP. **Less** on guardrails. **Medium/Low** (secondary). **New Jul + 21, 2026**: shipped **self-scheduling** — an agent sets its own schedule and + on firing "wakes up with its saved prompt and continues right where it left + off, with all of its context and history." Published page documents **no + re-wake-frequency cap** — a self-perpetuating loop shipping *without* the + three hard stops; the ceiling is the operator's to add. Also new: Puck + meta-agent, Slack "summon Amp." — **High** feature / **Medium** cap-absence + (absence in docs ≠ confirmed absent). https://ampcode.com/news/schedule - **Aider** — bounded ~3× self-correction retry + `--auto-test` (real test in loop); no goal primitive, scheduler, or budget ceiling. A retry helper, not an autonomous loop harness. **Medium**. @@ -569,10 +628,16 @@ thin/403'd; re-verify a specific flag against live docs before treating as High. ### Cross-tool standards & portability - **MCP** — de-facto cross-tool standard for tool/context access (OpenAI, - Google, Microsoft, Anthropic; an AAIF project; 5,800+ servers; 2026 spec adds - stateless HTTP, MCP Apps, a **Tasks** extension for long-running work). - Standardizes *tool access, not the loop harness*. **High** adoption / **Medium** - figures. https://blog.modelcontextprotocol.io/posts/2026-mcp-roadmap/ + Google, Microsoft, Anthropic; an AAIF project; 5,800+ servers). The + **2026-07-28 release candidate** (largest revision since launch) finalized at + this window's edge: stateless core, a first-class **Tasks** extension for + long-running async work (directly relevant to *bounded* long-running loop + work), **MCP Apps** (server-rendered UI over the same JSON-RPC consent path), + OAuth/OIDC-aligned auth, and a formal 12-month deprecation policy; beta SDKs + for the RC shipped in-window. Standardizes *tool access, not the loop + harness*. **High** (primary release-candidate post + secondaries). + https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/ · + https://blog.modelcontextprotocol.io/posts/2026-mcp-roadmap/ - **Codex `/goal` = cross-tool validator-judge pattern** — same architecture as Claude Code `/goal` (distinct judge, evidence-based stop). Confirms the validator-model stop is an industry pattern, not a Claude feature. **High**. @@ -637,6 +702,37 @@ thin/403'd; re-verify a specific flag against live docs before treating as High. — arXiv:2606.08671, submitted June 7, 2026. Addresses loss of decision history across skill revisions. — **High** (arXiv HTML page read). https://arxiv.org/html/2606.08671 +- **"When Agents Do Not Stop: Uncovering Infinite Agentic Loops in LLM Agents"** + — arXiv:2607.01641 (~Jul 1–2, 2026; **read this pass**, resolving the + standing backlog item). Defines infinite agentic loops as unbounded + repetition of model/tool/handoff calls when the feedback path isn't bounded; + introduces **IAL-Scan**, a static analyzer building an "Agentic Loop + Dependence Graph" to detect paths that repeatedly hit expensive ops without a + bound — **91.9% precision across 6,549 repos**. Empirical support for this + repo's max-iteration / stall-detection hard stops (primer §6). — **High** + (abstract read directly). https://arxiv.org/abs/2607.01641 +- **"SkillCoach: Self-Evolving Rubrics for Evaluating and Enhancing Agentic + Skill-Use"** — arXiv:2607.01874 (submitted Jul 2, 2026; **read this pass**, + resolving the standing backlog item). Derives skill-grounded *process* rubrics + from real rollouts to evaluate skill selection/following/composition/ + reflection, catching failures outcome-only checks miss; keeps an external + verifier as a separate signal — i.e. process rubrics *complement*, don't + replace, a deterministic success check (consistent with primer §5A). — **High** + (abstract read directly). https://arxiv.org/abs/2607.01874 +- **"SkillCorpus: Consolidating and Evaluating the Open Skill Ecosystem for + Real-World LLM Agents"** — arXiv:2607.15557 (v1 Jul 17, **v4 Jul 23, 2026, + in-window**). Consolidates ~821K→96,401 curated skills across 16 categories + + a retrieval stack; reports up to +7.5pp task gains. Relevant to the + `SKILL.md` ecosystem/portability question. — **High** (abstract read + directly). https://arxiv.org/abs/2607.15557 +- "Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous + Research Loops" — arXiv:2607.07663 (submitted Jul 8, 2026). Posits a + **verification hierarchy** — formal verifiers (strongest) → … → self-assessment + (weakest) — and shows demonstrated self-improvement strength *tracks* this + hierarchy, with **self-confirming loops** as the failure mode when the ordering + is violated. A citable primary source for this repo's "an agent never declares + done on its own self-assessment" rule. — **High** (abstract read directly). + https://arxiv.org/abs/2607.07663 ## Known caveats / things to re-verify @@ -654,12 +750,16 @@ instead of deleting it or leaving it here indefinitely. per-tool execution behavior (discovery, honored frontmatter, permissions) differs. Treat a skill authored here as portable-with-testing, not drop-in. Re-verify which tools truly run `SKILL.md` vs. only `AGENTS.md`. -- **Agent Skills governance status unclear.** Earlier KB passes recorded Agent - Skills as an AAIF / Linux Foundation-governed open standard; the 2026-07-20 - survey found the LF only names **MCP / AGENTS.md / goose** as AAIF projects — - `SKILL.md` may still be Anthropic-spec + community, *not* neutrally governed. - If so, `SKILL.md` portability rests on vendor goodwill, not a neutral - standard. Re-verify against the LF/AAIF project list. +- **Agent Skills governance status — converging toward "not AAIF-governed."** + Earlier KB passes recorded Agent Skills as an AAIF / Linux Foundation-governed + open standard; the 2026-07-20 survey and now **two independent 2026-07-27 + agents** all found the LF's Agentic AI Foundation names only **MCP / + AGENTS.md / goose** as its projects — `SKILL.md` is Anthropic-authored, + community-maintained (agentskills.io), *not* confirmed AAIF-governed. Primer + §4 softened accordingly this pass. Still an open re-verify (no primary AAIF + project list stating the negative was read), but the evidence is now + one-directional. If it holds, `SKILL.md` portability rests on vendor goodwill, + not a neutral standard. - The slogan "the costliest thing in AI coding is managing the agent loop" is a community paraphrase, not a sourced Cherny quote. - The "5 tips for running agents autonomously" is real in substance but not @@ -688,13 +788,20 @@ instead of deleting it or leaving it here indefinitely. that learn from execution traces — **Low-Medium**, existence corroborated by search but article not directly fetched, exact date unconfirmed. Not promoted; re-verify before citing. -- **"Graph engineering" as a successor term to "loop engineering"** — surfaced - from a single thin secondary (datasciencedojo.com, ~Jul 18) plus a - Steinberger tweet ("did we shift to graphs yet?") gesturing the same - direction. Two independent hints, neither confirmed as an established term. - **Human-reviewed 2026-07-20**: warm signals, still unconfirmed — no primary - essay found. **Low.** Carry forward to next pass; promote to the lineage - ladder (primer §2) only if a primary essay actually defines the term. +- **"Graph engineering" as a successor term to "loop engineering"** — the meme + is heating in secondaries but **still has no primary long-form definition** as + of 2026-07-27. This pass found: Steinberger has published only tweets (Jul 18 + "did we shift to graphs yet?"; ~Jul 25 "am I a graph engineer now"), no essay; + Hamel Husain's X Article "Loop Engineering Is Dead. Enter Graph Engineering" + (Jul 18, pre-window) is an opinion piece, not a definition; **Turing Post + FOD#159 "Is Graph Engineering Real?" (Jul 20, read directly, High)** argues + the term spread with *no consensus definition* ("graph" used for control + graphs, knowledge graphs, execution traces, and improvement loops + interchangeably); Louis-François Bouchard (Towards AI, in-window) and others + push back on the "loop engineering is dead" framing. Net: a **contested, + undefined meme, not a defined successor discipline**. **Low.** Carry forward; + promote to the lineage ladder (primer §2) only if a primary essay actually + defines the term. https://www.turingpost.com/p/is-graph-engineering-real-why-everyone-is-talking-about-it - **roborev.io/changelog** consistently 403's to automated fetch even as the GitHub releases page is readable — cross-check the two if a claim ever depends on changelog prose rather than release notes. @@ -708,10 +815,25 @@ instead of deleting it or leaving it here indefinitely. cited per summary, exact publish date unconfirmed) — not promoted to primer; useful as an early signal of backlash if corroborated later. https://medium.com/ai-engineering-simplified/loop-engineering-is-dead-heres-the-data-behind-the-ai-backlash-6d1b204e4b9a -- **SkillCoach: Self-Evolving Rubrics for Evaluating and Enhancing Agentic - Skill-Use** (arXiv:2607.01874, submitted Jul 2, 2026) — still not read in - full as of the Jul 20 pass; noted for next pass. -- **"When Agents Do Not Stop: Uncovering Infinite Agentic Loops in LLM - Agents"** (arXiv:2607.01641, ~Jul 1–2, 2026) — directly on-topic (AutoGPT- - style non-termination) but still not read in full as of the Jul 20 pass; - flagged to read fully next pass. +- **roborev repo identity** — search surfaced two repos described as + "continuous background code review": `roborev-dev/roborev` (the one the KB + tracks and reads directly) and `kenn-io/roborev`. Confirm which is canonical + (possible fork/rename) before citing the other. **Low.** Added 2026-07-27. +- **GuardFall** — a claimed "universal shell-injection design flaw affecting + >500k open-source deployments," attributed to July 2026, surfaced from a + single aggregator (adversa.ai roundup) with no primary source or precise + date. Verify against a primary before treating as real. **Low.** Added + 2026-07-27. +- **AgentGuard / LoopGain re-findability** — two research agents this pass + could not re-locate AgentGuard (KB cites github.com/bmdhodl/agent47, High + from README) or LoopGain (github.com/loopgain-ai/loopgain, Medium) via + search. A search miss is not disproof, but confirm both repos still exist at + the cited URLs next pass before relying on their flag names. Added 2026-07-27. +- **Gateway ownership changes** — Portkey was acquired by Palo Alto Networks + (closed ~May 29, folding into Prisma AIRS) and Helicone by Mintlify (~March + 2026); both are cited in primer §6. Pre-window context, but re-verify the + budget-enforcement flags still exist under new ownership before citing. + **Medium.** Added 2026-07-27. +- **Subagent-nesting default flipped twice in one week** (v2.1.217 off → + v2.1.219 depth-3, Jul 21–24) — re-verify the currently-shipped default before + baking it into a template. Added 2026-07-27.