fix(hooks): detect subagents from hook stdin (agent_id / agent_type, SessionStart source=fork), drop the fork feature-flag test

[theNetworkChuck]

Summary

isSubagentContext() keys on environment variables. That cannot work for forked subagents, and the workaround it shipped with (CLAUDE_CODE_FORK_SUBAGENT === '1') turned every main session into a "subagent" and silently disabled nine hooks (#1864, #1911, PR #1942).

This PR removes the fork-flag row (same hunk as #1942) and replaces it with the signal that actually distinguishes a fork: the hook's stdin JSON. Consumers that already parse stdin pass it to the guard; the two SessionStart consumers that did not read stdin now do, best-effort.

Why an env var can never mark a fork

Measured on Claude Code 2.1.241 by spawning a fork and dumping its environment from inside: a fork's environment is byte-identical to its parent's — same CLAUDE_CODE_SESSION_ID, same CLAUDE_PID, no CLAUDE_AGENT_TYPE / CLAUDE_CODE_SUBAGENT_NAME / CLAUDE_CODE_SUBAGENT_TYPE, and CLAUDE_CODE_FORK_SUBAGENT=1 in both. Nothing in process.env separates them, so with the flag row removed (#1942) the original #1831 leak (forks re-injecting main-session context) is open again.

What does mark a fork

The hooks reference documents agent_id and agent_type as common input fields "for subagent hooks", and lists fork as a SessionStart matcher (i.e. source: "fork" on that event's input). Observed on the same 2.1.241 fork: its PostToolUse input carries agent_type: "fork" plus an agent_id; an ordinary delegate carries agent_type: "general-purpose"; main-session input carries neither.

Change

lib/subagent.ts

• isSubagentContext(input?: unknown) — optional parsed-stdin argument; result is the existing env union OR the stdin stamp. Every zero-arg call site keeps compiling and behaves as before, minus the false positive.
• New isSubagentHookInput(input) — true when agent_id or agent_type is a non-empty string, or hook_event_name === "SessionStart" with source === "fork". Pure, zero-dep, never throws.
• Fork-flag row removed (identical to fix(hooks): CLAUDE_CODE_FORK_SUBAGENT is a feature flag, not a subagent marker — disables 9 hooks on every install #1942).

Consumers now pass their parsed stdin: ISASync, SystemChangeSurface (HookInput gains optional agent_id/agent_type), ConfigEvalFire, MemoryReviewFire, MemoryTurnStart, KittyEnvPersist (guard moved after its existing stdin read; source comment now includes fork), LoadContext (adds a best-effort stdin read). LoadMemory and MemoryDeltaSurface are unchanged: their guarded paths are standalone probes, and their real caller MemoryTurnStart is now guarded on stdin.

lib/hook-io.ts — HookInput documents the optional agent_id, agent_type, source fields.

No new dependencies; hooks stay importable before bun install.

Verification

All runs with CLAUDE_CODE_FORK_SUBAGENT=1 in the environment (what settings.system.json and Claude Code ≥ 2.1.232 both set), against a throwaway HOME, using stdin shapes copied from real captured hook input:

• Guard unit cases (18): main-session input with the flag set → false for PostToolUse / Stop / SessionStart-startup; fork PostToolUse (agent_type: "fork") and SessionStart source: "fork" → true; ordinary delegate stamp → true; each surviving env marker alone → true; null / string / array / empty-string / non-string / source: "fork" on a non-SessionStart event → false, none throw.
• Hook-level: MemoryReviewFire writes its per-session state on a main Stop and writes nothing on a fork-stamped Stop; LoadContext runs on source: "startup" and prints its skip line on source: "fork"; KittyEnvPersist same split; SystemChangeSurface.run() returns its surface line for a main write and null for a fork-stamped one.
• Each edited hook transpiles (bun build --target=bun --no-bundle).

Relationship to open work: stacks on #1942 (the removal hunk is the same; merging either first leaves a trivial conflict on that one hunk). Closes the re-opened #1831 concern and the "what is the reliable fork marker" follow-up noted on #1864 / #1911.