[FEATURE] Opt-in context inheritance for Agent-spawned subagents (inherit_context flag)

Preflight Checklist

• [x] I have searched existing issues (#6825, #4908, #50093, #40459, #41356) and this proposes a concrete solution to a problem previously closed without resolution
• [x] This is a single feature request
• [x] I am using the latest version of Claude Code

Problem Statement

Agent-spawned subagents silently discard all project context, causing permanent knowledge loss in multi-agent workflows.

When a parent session delegates work via the Agent tool, subagents start with zero project context:

• No ~/.claude/CLAUDE.md
• No .claude/rules/
• No project memory
• No hooks fire (SessionStart, PreToolUse, PostToolUse — confirmed by #69260)

This is by design for token efficiency. But there is no opt-in path for workloads where context inheritance is critical.

Real-World Impact (Forensic Data)

I operate a hooks-based memory system where the parent session is instructed to persist important findings to project memory. After a 9-day period of active development on a project:

• 296 subagent sessions were spawned via Agent tool (sdk-cli/queue-operation)
• 4 interactive sessions existed (cli/typed) — all from the final day
• Zero project memories were persisted across 9 days of work
• The parent sessions saved brief summaries to their own project directory (CWD-based), not the target project
• Subagents performed the substantive work (research, bug fixes, PR submissions) but couldn't persist anything

Result: When I opened the project directly the next day, Claude had zero recall of 9 days of work — including merged PRs, workflow patterns, and architectural decisions.

Observable Convention Violations from Bare Subagents

Because subagents never see project rules, every convention stored in CLAUDE.md or memory is systematically ignored:

• Subagents made commits with the wrong git identity (no rule telling them to check per-repo config)
• PR comments contained AI-generated text verbatim (project has an explicit "never paste AI text into OSS PR comments" rule — subagent never saw it)
• Memory files written in wrong format, later judged "invalid" and deleted by a subsequent session
• Retry loops against rate-limited services (a feedback memory says "never retry against server X" — subagent didn't have it, triggered a ban)

This is not occasional drift — it's systematic blindness across 70-95% of actual execution (the subagent portion).

The Token Efficiency Trade-off Is Currently All-or-Nothing

Prior issues (#6825) correctly identified that inheriting everything (60K+ tokens) for a trivial echo success subagent is wasteful. Anthropic's response was to strip ALL context from subagents.

But this created the inverse problem: subagents that DO need project rules get nothing. There's no middle ground.

The community has converged on brittle workarounds:

• Manually embedding rules in every Agent prompt (verbose, error-prone, duplicative)
• External hook-patching tools (j-p-c/alzheimer, rtk-ai/rtk)
• Binary patches to the omitClaudeMd flag (#40459)

Proposed Solution

Minimal Ask: context_files Parameter

The simplest implementation that solves the core problem — a file-injection primitive:

Agent(
    prompt="...",
    context_files: ["~/.claude/CLAUDE.md", ".claude/rules/safety.md"]
)

The caller explicitly lists which files get injected into the subagent's system prompt. This gives:

• Full caller control — no surprise token bloat
• Selective inclusion — safety rules without 60K of memory
• Zero infrastructure change beyond "read these files and prepend to system prompt"
• Backward-compatible — omitting the parameter preserves current behavior

Full Solution: inherit_context Flag with Presets

Building on context_files, add convenience presets:

Agent(
    prompt="...",
    inherit_context: "none" | "rules" | "full"
)

| Level | What's Inherited | Token Cost | Use Case |
|-------|-----------------|-----------|----------|
| "none" (default) | Nothing — current behavior | ~3K base | Quick lookups, grep, simple ops |
| "rules" | CLAUDE.md + .claude/rules/ + project memory index | ~10-30K typical | Agents that must follow project conventions, safety rules, memory instructions |
| "full" | Everything the parent has (CLAUDE.md, rules, memory, MCP tool descriptions) | ~60K+ | Long-running research agents, workflow agents that need full context |

Why a per-call flag, not a global setting:

• Different subagents in the SAME session need different levels. An Explore agent grepping for a symbol doesn't need 60K of context. A research agent building a PR needs all the project rules.
• The parent agent is best positioned to decide — it knows what work it's delegating.

Hook Inheritance (Independent, Complementary)

Separately from context, hooks are stripped from subagents (ref: #69260). A companion flag:

Agent(
    prompt="...",
    inherit_hooks: true | false
)

This would allow PreToolUse safety hooks to fire for subagent tool calls. The token cost is zero (hooks are code, not prompt tokens) — only execution time.

Hooks and context inheritance are independent but synergistic: if hooks fired for subagents, a SessionStart hook COULD inject context files — giving a userland escape hatch even without native context_files support.

Implementation Priority Ladder

For Anthropic, an incremental path rather than a big-bang:

1. Hooks fire for subagents (#69260) — zero token cost, high safety value, unblocks workarounds
2. context_files parameter on Agent tool — minimal implementation, caller-controlled, solves the core problem
3. Convenience presets ("rules" / "full") — sugar on top of (2)

Why This Matters More Than It Did in 2025

The Agent tool is used aggressively by newer models (Opus 4.x, Sonnet 4.x). In observed workflows:

• 70-95% of sessions in a project are subagents
• Subagents do the majority of substantive work
• The parent is increasingly an orchestrator, not an executor

Without context inheritance, the majority of actual work happens in a context-free void. This makes:

• Memory systems useless (can't save, can't recall)
• Safety hooks bypassable (don't fire in subagents)
• Project conventions invisible (not in prompt)
• Multi-session continuity impossible (knowledge dies with each subagent)

References

| Issue | Topic | Status |
|-------|-------|--------|
| #6825 | Configurable inheritance (proposed includes: syntax) | Closed (dup → #4908) |
| #4908 | Scoped context passing | Closed (NOT_PLANNED) |
| #50093 | Optional inherit CLAUDE.md + rules | Closed (NOT_PLANNED) |
| #40459 | omitClaudeMd:true regression analysis | Closed (NOT_PLANNED) |
| #41356 | Memory rules violated by subagents | Closed (NOT_PLANNED) |
| #69260 | PreToolUse hooks don't fire for subagents | OPEN |
| #55648 | Subagent skips task when writing memory | Closed |

Environment

• Claude Code version: 2.1.179+
• Platform: macOS (Darwin 24.6.0)
• Model: Opus 4.6 (via Bedrock)
• Workflow: ~300 subagent sessions across 9 days, 4 interactive sessions