fix(plugin): suppress Codex auto-discovery of Claude hooks

[worktrunk-bot]

Problem

The Codex manifest (.codex-plugin/plugin.json) omitted a hooks key on the assumption — documented in plugins/worktrunk/CLAUDE.md — that "Codex ships no hooks." That assumption doesn't hold: when a plugin manifest has no hooks key, Codex auto-discovers hooks/hooks.json from the plugin root by convention (DEFAULT_HOOKS_CONFIG_FILE, the None branch of load_plugin_hooks). Since Claude and Codex share one payload dir, Codex picked up Worktrunk's Claude hooks file and surfaced the Claude events it recognizes (PreToolUse, PermissionRequest, UserPromptSubmit, Stop) in a Codex session — the symptom in #3362.

Per @ofek's follow-up, seeing anything about Claude while using Codex is poor UX regardless of whether the hooks function, so this takes the suppress direction rather than shipping Codex-native hooks.

Solution

Set hooks: {} — an empty inline hooks object — in the Codex manifest. An inline object takes Codex's Some(Inline) branch, which skips empty entries, and therefore never reaches the None auto-discovery branch, so no hooks load.

An empty hooks: [] array would not work: an empty path list falls back to auto-discovery (resolve_manifest_hooks). The inline empty object is what actually overrides discovery.

The change touches only the Codex manifest; the Claude manifest and its hooks/hooks.json are untouched. Also updates the now-stale reasoning in plugins/worktrunk/CLAUDE.md and src/commands/config/codex.rs (the old note said Codex has no Stop/turn-end event — current Codex does; the real reasons to suppress are the shared-file collision plus the missing SessionEnd-clear and the Claude-branding UX).

Testing

• Updated test_codex_plugin_metadata_is_valid_json to assert the manifest carries hooks: {} (encoding the correct belief — absence would re-open the bug) — passes locally.
• Verified against real Codex source (codex-core-plugins @ 98d28aa): added a test exercising load_plugin_manifest + load_plugin_hooks with hooks: {} and a present hooks/hooks.json — it yields zero hook sources, while Codex's own load_plugin_hooks_discovers_default_hooks_file confirms an absent key does discover the file. That upstream test can't run in this repo's CI, so the guarantee is pinned by source-level verification rather than a worktrunk-side runtime test.

Verification test (run against the cloned Codex workspace)

#[test]
fn empty_inline_manifest_hooks_suppress_default_discovery() {
    let (_tmp, plugin_root) = plugin_root();
    write_manifest(&plugin_root, r#"{ "name": "demo-plugin", "hooks": {} }"#);
    // The shared Claude hooks file sits at the default discovery path.
    write_hook_file(&plugin_root, "hooks/hooks.json", "PreToolUse", "echo claude");

    let (sources, warnings) = load_sources(&plugin_root);

    assert_eq!(warnings, Vec::<String>::new());
    assert_eq!(sources, Vec::<PluginHookSource>::new()); // discovery suppressed
}

test loader::tests::empty_inline_manifest_hooks_suppress_default_discovery ... ok

---

Closes #3362 — automated triage. A maintainer should confirm against a live Codex session before merge.

[worktrunk-bot]

The core fix is correct and I verified it against Codex source at the pinned 98d28aa: resolve_manifest_hooks deserializes {} as RawPluginManifestHooks::Inline → Some(Inline([empty])) (taking the Some branch, not None), and load_plugin_hooks's Some(Inline) arm hits if hooks_file.hooks.is_empty() { continue; }, so it emits zero sources and never reaches the None auto-discovery path. The hooks: [] caveat also checks out — an empty InlineList falls back through (!hooks.is_empty()).then_some(...) to None. Nice source-level pin.

One follow-up gap, though — the reasoning correction landed in plugins/worktrunk/CLAUDE.md and src/commands/config/codex.rs but two user-facing surfaces still carry the old "no turn-end event" claim this PR is retiring:

• src/cli/config.rs (after_long_help for the codex plugin subcommand — a primary doc source): "Codex exposes no turn-end hook event, so the Codex plugin omits them until it does."
• docs/content/claude-code.md ("Agent Integration"): "Codex omits activity tracking because its hooks have no turn-end event, so a 🤖 marker could never clear back to 💬."

I confirmed against codex-rs/hooks/src/schema.rs @ 98d28aa that HookEventNameWire now includes Stop (and SubagentStop), and the vocabulary has sessionStart but no session-end — exactly the "Stop exists, no SessionEnd to clear" story the updated CLAUDE.md tells. So these two strings are now not just stale but actively misleading: the second phrasing implies markers are omitted only until Codex ships a turn-end event, a condition that's now met while we deliberately keep omitting them.

The fix is the same rewrite already applied elsewhere (turn-end event exists; suppression stands for the shared-file + no-SessionEnd + Claude-branding reasons). Both are outside this diff, so no inline suggestion — editing after_long_help in src/cli/config.rs regenerates the help_config_plugins_codex snapshot and any command-page mirror, and docs/content/claude-code.md's edit auto-syncs to skills/worktrunk/reference/claude-code.md. I'm happy to push that as a follow-up commit if you'd like it folded into this PR rather than a separate one. (The CHANGELOG.md mention of "no turn-end hooks" is a historical release entry — left alone.)

Not approving since this is a self-authored PR; flagging for the maintainer who'll confirm against a live Codex session per the PR body.

[max-sixty]

@worktrunk-bot resolve conflicts

[worktrunk-bot]

I resolved the conflicts, but doing so surfaced that this PR is now fully superseded — the resolution collapses to zero net change against main.

While this PR was open, #3364 ("feat(plugin): ship Codex-native activity hooks", merged in 512a94de2 at 06:05Z) landed a superior fix for the same issue (#3362). Instead of suppressing hooks with hooks: {}, it ships Codex-native activity hooks inline in the manifest — which also takes Codex's Some(Inline) branch and blocks auto-discovery of the shared Claude hooks/hooks.json, so it closes #3362 the same way and adds the activity markers this PR deliberately omitted. It also already rewrote the user-facing "no turn-end event" strings I flagged in my earlier review (src/cli/config.rs, docs/content/claude-code.md).

All four conflicts were the two opposing approaches to the exact same lines (plugin.json hooks key, codex.rs install hint + comment, plugins/worktrunk/CLAUDE.md limitations section, config_show.rs assertions). The only coherent resolution is to take main's version everywhere — and once you do, the branch is byte-identical to main:

$ git diff origin/main --stat   # after resolving all 4 conflicts to main's side
$                               # (empty — nothing left)

So there's nothing to merge here: every change this PR made is already on main in a better form. I'm closing it as superseded by #3364 rather than pushing an empty merge commit. If you'd rather keep it open for any reason, just reopen and let me know.