2.1.220 silently added a system-prompt directive suppressing Agent-tool dispatch (absent in 2.1.218)

Summary

Starting in 2.1.220, the system prompt contains a directive instructing the model not to call the Agent tool unless the user requested it. This was not present in 2.1.218.

The effect is that Claude declines to dispatch subagents for workflows whose entire design is parallel dispatch — custom skills for code review, auditing, and plan review — even though invoking the skill is itself the request. The workflows do not error; they silently run degraded, or the model reports that it needs permission to fan out.

This arrived in a patch release with no signal, and it is not present in any configuration file on disk, which made it extremely difficult to diagnose.

Evidence it arrived in an upgrade

Measured across every version retained on this machine:

$ grep -ac "unless the user requested it" ~/.local/share/claude/versions/2.1.210   # Jul 14
0
$ grep -ac "unless the user requested it" ~/.local/share/claude/versions/2.1.218   # Jul 23
0
$ grep -ac "unless the user requested it" ~/.local/share/claude/versions/2.1.220   # Jul 24
2
$ grep -ac "unless the user requested it" ~/.local/share/claude/versions/2.1.221
2
$ grep -ac "unless the user requested it" ~/.local/share/claude/versions/2.1.222
2

Absent in 2.1.218, present from 2.1.220 onward.

The string appears in no config file on disk — not in ~/.claude/settings.json, not in project settings, not in any memory or rules file. We attempted several configuration-side fixes before locating it inside the binary, and initially mis-diagnosed it as a long-standing compiled constant because we compared only 2.1.220, 2.1.221 and 2.1.222 to each other and never checked 2.1.218.

Environment

| | |
|---|---|
| Claude Code | 2.1.220 and later (2.1.218 unaffected) |
| Platform | macOS 26.5.2 (25F84), arm64 |
| Node | v22.23.2 |
| Shell | zsh |
| Install method | native, autoUpdates: false |

Impact

Our review workflow depends on fanning out parallel reviewers. With the directive in place it silently stopped doing so. As a direct consequence our code-review gate was skipped on four commits that reached production; running the review retroactively surfaced three critical defects, including one in a drift-detection gate that caused it to falsely pass the exact condition it had been built to catch.

The failure mode is the dangerous kind: nothing errors, and the absence of review looks identical to a clean review.

What I'd like to see

1. Release-note system-prompt changes that alter agent behavior. A directive that suppresses tool dispatch is a behavioral breaking change, not an internal detail.
2. Make such directives inspectable — some supported way to see the effective system prompt, so this is diagnosable without byte-level comparison of binaries.
3. Provide a supported override for projects whose workflows legitimately depend on programmatic subagent dispatch.
4. Reconsider the directive's phrasing. Invoking a skill whose documented purpose is multi-agent fan-out should count as the user requesting it.

Related: I have filed a separate issue about the 2.1.222 worktree-isolation guard. Both are behavioral regressions that arrived silently via patch releases on a setup that had been stable for months.

tengu_heron_brook:    ABSENT from cachedGrowthBookFeatures
tengu_fennel_godwit:  false