Build4 publishers3 min readPublished
Claude Code's AGENTS.md fallback can be disabled by version, settings, or first-run conditions, not just feature-flag fetches
Anthropic shipped the AGENTS.md fallback in Claude Code 2.1.277 as a flag-gated built-in mod. Two machines on the same pinned version can load different project instructions. The same week's digest also logs a reverted deny-rule fix.
The Engineer · Build desk
What happened
- Anthropic added AGENTS.md support to Claude Code on September 18th in version 2.1.277, so a folder with no CLAUDE.md now has its AGENTS.md read directly.
- A 2.1.275 regression failed every request with a 400 error when ANTHROPIC_BASE_URL pointed at a proxy or gateway, and 2.1.276 fixes it.
- Codex CLI 0.155.0 added Touch ID verification for MCP requests in local TUI sessions on supported Macs.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Pinning a Claude Code version no longer settles which instruction file loads, so a team that wants the same context everywhere has to set instructionFiles explicitly in user or managed settings.
- constraint Instruction-loading policy cannot ride along in the repository, so the choice belongs to whoever controls user settings, --settings or managed settings, and a maintainer cannot ship it with the code.
- exposure Anyone who wrote a deny rule last week on the strength of the 2.1.268 fix now has a rule with a narrower effective scope, and wrapped Bash lines prompt instead of being refused.
- cost Eight Claude Code releases inside one digest window puts changelog reading on someone's standing workload.
The check counts three files: `CLAUDE.md`, `.claude/CLAUDE.md` and `CLAUDE.local.md`, in the working directory or above [3]. A user-level `~/.claude/CLAUDE.md`, an organization's managed `CLAUDE.md` and `.claude/rules/` files do not count toward it, and they keep loading alongside AGENTS.md [4]. So a personal home-directory `CLAUDE.md` does not suppress the fallback in a repository that has no project file [29].
The documentation lists the cases that read `CLAUDE.md` only, and in which Project instructions does not appear in `/config`: a version before 2.1.277; a session that does not fetch feature flags from Anthropic, such as Amazon Bedrock, another third-party provider, or one with telemetry disabled; the first session after installing or upgrading; and an install that sets `disableAllHooks` or `allowManagedHooksOnly`, or disables the built-in `agents-md` plugin [5]. Turning telemetry off is enough to lose it. Runtimewire reports that, according to the release notes, the fallback is unavailable at launch on Amazon Bedrock, Google Vertex AI and Microsoft Foundry [6].
The settings value is what makes it deterministic: the default can be changed with `/config` under Project instructions [27], or set as `instructionFiles` under the built-in `agents-md` plugin's ID, where `claude-md-and-agents-md` loads both files [8]. Project and local settings files are ignored for that value [9]. Runtimewire reports that a repository's own `.claude/settings.json` cannot set mod options at all, and says the documentation presents this as preventing code checked into a project from choosing an individual's or an organization's instruction-loading policy [10]. The Agentic AI Foundation said in its December 2025 launch announcement that more than 60,000 open-source projects had adopted AGENTS.md, according to Runtimewire [21].
One 2.1.273 changelog entry reads: "Reverted a 2.1.268 change that checked Read and Edit deny rules on Bash lines the permission checker can't analyze" [11]. Commands like `time -p make build` prompt again rather than being denied [12]. That behavior shipped in 2.1.268 and came out five releases later [18]. Two other permission fixes did land in 2.1.273: Bash commands the checker cannot fully analyze were skipping the prompt under `permissions.blockReadsOutsideWorkingDirectories`, and a subshell could hide a dangerous `rm` in bypass mode [13]. The same release fixed the context meter and auto-compact counting advisor-tool turns at roughly twice their real context size, which had auto-compact firing at about half the real window [14].
With `ANTHROPIC_BASE_URL` pointing at a proxy or gateway, 2.1.275 failed every request with `400 ... Input tag 'advisor_20260301'`, and 2.1.276 fixes it [15]. The digest covers 2.1.270 through 2.1.277, eight releases [16][17]. Codex CLI moved as well: 0.155.0 added Touch ID verification for MCP requests in local TUI sessions on supported Macs, and blocked Windows-process escapes from restricted WSL sandboxes [23][24].
Thariq Shihipar, an engineer on the Claude Code team, wrote that "AGENTS.md support is built off of Claude Code mods, our upcoming way to customize the Claude Code harness" [20]. Runtimewire reports the built-in mod documents four instruction modes, being CLAUDE.md only, AGENTS.md as fallback, both formats, or centrally managed instructions only, with the fallback as the default [19]; Anthropic has not given mods a general availability date [26]. On Bedrock, or with telemetry disabled, the digest's instruction stands: import AGENTS.md from a `CLAUDE.md` [7].
What to watch
- Whether Anthropic re-lands the check on Read and Edit deny rules for Bash lines the permission checker cannot analyze, and in which release.
- A general availability date for Claude Code mods, and whether third-party mods can alter instruction loading the way the built-in agents-md mod does.
- Whether Bedrock, Vertex AI and Foundry sessions get the fallback once it no longer depends on fetching feature flags from Anthropic.