docs(readme): correct the subagent table — Claude is not disabled

Adversarial review of the section I added an hour ago found it wrong in the
generalisation and incomplete in the table.

It said connect "turns off each host's own subagent feature" and listed Gemini
and Codex. There are THREE agent-policy functions, and the third does something
categorically different: configureClaudeAgentPolicy (connect.ts:1026) writes
env.CLAUDE_CODE_SUBAGENT_MODEL = "sonnet" to ~/.claude/settings.json. Claude
Code keeps its subagents — they are pinned to an economy model, and fan-out is
discouraged by policy TEXT, not by config.

So a reader would have concluded connect disables Claude subagents when it does
not, which is exactly the misreading this section exists to prevent.

Also corrected the Codex path: $CODEX_HOME/config.toml, default ~/.codex, not
an unconditional ~/.codex/config.toml. CODEX_HOME is honoured when set.

Every path and behaviour in the table is now checked against source rather than
inferred from a constant name.

3743 passed / 0 failed (suite exit 0).

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: Synalux AI

README.md

@@ -58,14 +58,16 @@ features.
 
 ### What `prism connect` changes about host subagents
 
-`connect` turns off each host's own subagent feature, so bounded work routes to
-`prism_infer` on your machine instead of the host spawning agents of its own.
-Prism's local workers stay available over MCP either way.
+`connect` steers bounded work to `prism_infer` on your machine rather than to
+host-spawned agents. What it writes differs per host, and **it does not disable
+subagents everywhere** — Claude Code keeps them and is pointed at an economy
+model instead. Prism's local workers stay available over MCP in every case.
 
-| Host | Setting written | Why it looks different |
+| Host | Setting written | Effect |
 |---|---|---|
-| Gemini CLI | `experimental.enableAgents = false` in `~/.gemini/settings.json` | Gemini exposes one boolean, so that is all there is to set |
-| Codex | `features.multi_agent = false` in `~/.codex/config.toml`, plus a bounded fallback: 2 threads, depth 1, cheap subagent model, 900s cap | Codex exposes tuning, so re-enabling deliberately lands somewhere bounded rather than unlimited |
+| Claude Code | `env.CLAUDE_CODE_SUBAGENT_MODEL = "sonnet"` in `~/.claude/settings.json` | Subagents stay **enabled**, pinned to an economy model. Fan-out is discouraged by policy text, not by config |
+| Gemini CLI | `experimental.enableAgents = false` in `~/.gemini/settings.json` | Subagents **off**. Gemini exposes one boolean, so that is all there is to set |
+| Codex | `features.multi_agent = false` in `$CODEX_HOME/config.toml` (default `~/.codex`), plus a bounded fallback: 2 threads, depth 1, cheap subagent model, 900s cap | Subagents **off**, with a bounded profile underneath so a deliberate re-enable lands somewhere sane |
 
 Two things worth knowing: