Auto Tiers Meet Remote Fast Mode and Weekly Cap Reset

Tuesday is a routing-and-capacity morning, so the useful work is how these four moves fit together. GitHub finally lets you steer Auto’s cost and quality bias, Anthropic ships a dense Claude Code remote and sandbox release while permanent weekly caps replace the summer boost, and OpenAI’s Astra prompt hygiene note is the right companion for anyone still carrying GPT-5-era scaffolding.

Treat today as a model-routing and agent-context day. Pick an Auto tier in VS Code or Copilot CLI and confirm which model each reply used, pin Claude Code to v2.1.272 and smoke-test remote fast mode plus one `omitClaudeMd` subagent, run `/usage` so your team sees the new weekly ceiling, and ask Astra (or your coding agent) to audit one repo’s skills and AGENTS.md against the Sep 11 guidance.

Top Stories

Copilot Auto adds efficiency, balance, and intelligence tiers
Practical dev impact: You no longer have to guess whether Auto is optimizing for cheap or for deep, because GitHub’s September 14 changelog adds three tiers that bias how Auto weighs cost, quality, and latency while still evaluating each prompt against the same model pool. Efficiency keeps spend low for docstring-class work, Balance is the everyday default, and Intelligence leans quality for hard tasks, yet a simple prompt can still land on a small model under Intelligence. Tiers are rolling out in VS Code, Copilot CLI, and the GitHub Copilot app. You are billed for the model Auto actually picks, and paid plans keep the 10% Auto discount. Docs note that Auto routes at natural cache boundaries so it does not thrash models mid-session. Set Balance for day-to-day, use Efficiency on high-volume chore queues, and reserve Intelligence for multi-file or debugging sessions, then hover (or watch the CLI footer) to see which model answered.

Claude Code v2.1.271: remote fast mode, omitClaudeMd, per-command domains
Practical dev impact: Pin `@anthropic-ai/claude-code` to `2.1.272` (or at least `2.1.271`) so remote and self-hosted sessions inherit host `/fast` (or a typed `/fast`) where your org allows it, custom and plugin subagents can set `omitClaudeMd` to skip user, project, and local CLAUDE.md while still loading managed policy, and sandboxed Bash, PowerShell, or Monitor commands can declare per-command `allowed_domains` so only the hosts each command needs are opened for it. The same release adds mouse support in fullscreen `/config`, `–drain-marker-file` for self-hosted runner drains, `–accept-command ` for exact plugin install or update acceptance, and chargeback `multiplier` up to 10 on gateway pricing. September 15’s v2.1.272 is a narrow reliability follow-up, so prefer an explicit pin in images and Actions rather than floating latest overnight. Start with one remote `/fast` smoke test and one subagent that sets `omitClaudeMd: true` on a research-only task.

Claude Code weekly limits: permanent +25%, about -17% versus the summer boost
Practical dev impact: Recalibrate team capacity plans today, because Anthropic’s Help Center says the May 13 through September 13 50% weekly boost ended, and starting September 14 weekly Claude Code limits sit 25% above the pre-promotion baseline for Pro, Max, Team, and seat-based Enterprise (Free and consumption Enterprise seats were never in the promo). Relative to the boosted ceiling teams planned around all summer, that is roughly a 17% cut, which Anthropic also stated publicly when clarifying the math. Five-hour session limits were never part of the boost. Run `/usage` in the CLI, decide whether heavy workflows need API-rate extras or staggered agent fan-out, and update runbooks that assumed the temporary 150-style ceiling rather than the new permanent 125-style one.

OpenAI: shrink skills and prune AGENTS.md for GPT-6 Astra
Practical dev impact: Before you blame Astra for tentativeness or skill thrash, audit the scaffolding you accumulated under weaker models, because OpenAI’s September 11 developers post says long skill descriptions get truncated when too many skills compete, contradictory “use when” text loads the wrong skill, and AGENTS.md rules that force a full doc tour before every edit burn context on typo fixes. Prefer short skill descriptions that name when the skill applies, progressive disclosure via a thin root router, and AGENTS.md lines that point to docs by task instead of requiring architecture.md on every change. Astra already tends to run tests and check work, so leftover “always verify” stacks can waste turns, while overly hard stop boundaries can make it pause where you wanted persistence. Ask Astra to audit one repo’s skills and AGENTS.md, then keep only instructions that still earn their tokens.

Practical Impact Analysis

Today’s through-line is control over spend and context, not another greenfield harness launch. Copilot Auto tiers finally expose the cost and quality dial that Auto’s black-box routing hid, so platform teams can standardize Efficiency for chore bots, Balance for default IDE chat, and Intelligence for incident or migration threads without abandoning Auto’s health-aware routing or the paid 10% discount.

Claude Code’s Tuesday release is the operational half of the same dial. Remote fast mode and `omitClaudeMd` cut latency and CLAUDE.md noise for subagents, while per-command `allowed_domains` tightens sandbox egress. Pairing that with the weekly-limit reset means you should not treat v2.1.271 as “turn everything on”: pin v2.1.272, enable fast mode where policy allows, and shrink concurrent workflow size on Pro (the release also lowers medium workflow guidance) so the new weekly ceiling lasts the week.

OpenAI’s Astra note closes the loop on prompt debt. If your Codex or Agents API sessions still carry GPT-5-era skill packs and always-read-the-monorepo AGENTS.md rules, you will pay for them twice: once in tokens, again when the model loads the wrong skill or stops early because a stale boundary says so. Clean the repo instructions in the same pass where you set Copilot Auto tiers and Claude Code pins.

If you only do three things this morning, set Auto to Balance (or Efficiency on a high-volume queue) and verify the model footer, pin Claude Code to v2.1.272 and run `/usage`, and delete or rewrite one bloated skill description plus one always-read AGENTS.md rule.

Tutorial

Pin Claude Code to v2.1.272 and smoke-test an `omitClaudeMd` research subagent. Export whatever auth your org already uses for Claude Code (Claude.ai login, API key, or gateway). Keep secrets in the environment, not in the agents JSON.

1. Pin `@anthropic-ai/claude-code@2.1.272` and confirm `claude –version`.
2. Optionally run `claude -p “/usage”` to confirm the post-promo weekly ceiling.
3. Launch a one-shot prompt with a custom research agent that sets `omitClaudeMd: true` so the subagent skips user, project, and local CLAUDE.md while managed policy still applies.
4. Confirm the research agent answers without dragging in project CLAUDE.md conventions you never intended for a one-shot summary. Next, try remote `/fast` where your org enables it, and add `allowed_domains` on any sandboxed network command that must reach a single host.

npm install -g "@anthropic-ai/claude-code@2.1.272"
claude --version

claude -p "/usage"

claude -p "What does package.json declare for test and lint scripts? Reply in three bullets." \
  --agents '{
    "research": {
      "description": "Read-only research. Use for summarizing repo facts without edits.",
      "prompt": "You only read files and answer. Do not edit, commit, or run destructive commands.",
      "omitClaudeMd": true
    }
  }'

npm install -g "@anthropic-ai/claude-code@2.1.272"
claude --version

claude -p "/usage"

claude -p "What does package.json declare for test and lint scripts? Reply in three bullets." \
  --agents '{
    "research": {
      "description": "Read-only research. Use for summarizing repo facts without edits.",
      "prompt": "You only read files and answer. Do not edit, commit, or run destructive commands.",
      "omitClaudeMd": true
    }
  }'

Recommended AI prompt

Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever you use.

You are my staff engineer for Copilot routing, Claude Code capacity, and coding-agent prompt hygiene on 2026-09-15. Copilot Auto now offers efficiency, balance, and intelligence tiers that bias cost, quality, and latency while still picking per prompt from the same model pool, with billing on the selected model and a 10% Auto discount on paid plans. Claude Code v2.1.271 adds remote or self-hosted fast mode, `omitClaudeMd` for custom and plugin subagents, and per-command `allowed_domains` in sandboxed auto mode. Pin through v2.1.272. Weekly Claude Code limits reset on Sep 14 to a permanent +25% over the pre-promo baseline (about -17% versus the summer boost). OpenAI urges short skill descriptions, progressive disclosure, and task-scoped AGENTS.md for GPT-6 Astra. Ask which Copilot surfaces, Claude Code pin, org fast-mode policy, and Codex or Agents repos we run. Then produce an Auto tier matrix for chore versus everyday versus complex work, a v2.1.272 pin plus omitClaudeMd and fast-mode smoke checklist, a weekly-limit capacity plan after the Sep 14 reset, and a one-repo skills and AGENTS.md prune plan for Astra. Keep it concrete and copy-paste ready.