Add Hermes Agent and Codex CLI providers (multi-provider runtime support)

[josephndex]

Description

Adds two new HookProvider integrations — Codex CLI and Hermes Agent — on top of a small runtime patch that lets AgentRuntime host more than one provider at once (previously hardcoded to exactly one). All three providers now share one AgentStateStore/office; agents from any of them render side by side, independently tracked.

Both new integrations are built against each CLI's actual current capabilities, checked directly rather than assumed:

• Codex CLI (0.144.1) has no hooks/plugin API. It does write a structured, append-only JSONL rollout per session (~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl — session_meta, response_item.function_call/function_call_output, event_msg.task_complete). codexTailer.ts polls that directory (byte-offset tracked per file, partial-line-safe) and re-emits parsed events through the existing POST /api/hooks/:providerId ingress — reading Codex's own structured format, not scraping terminal output.
• Hermes Agent (v0.20.0) turns out to have three separate hook systems, and only one of them actually covers CLI sessions: gateway hooks fire only for messaging-platform sessions (Telegram/Discord/Slack/WhatsApp/Teams), shell hooks cover CLI but only a small fixed event set, and plugin hooks (ctx.register_hook() under ~/.hermes/plugins/) carry the full documented lifecycle catalog this needed (pre_tool_call/post_tool_call/on_session_start/on_session_end/on_session_finalize/subagent_start/subagent_stop). installHooks() drops a small stdlib-only Python plugin there and enables it via the documented hermes plugins enable CLI — no new dependency in Hermes's venv, no hand-edited config.yaml.

The runtime patch

AgentRuntime/HookEventHandler previously took one provider: HookProvider at construction. This PR changes the constructor to providers: HookProvider[] and builds one HookEventHandler per provider (each still scoped to exactly one provider internally — no shared mutable "current provider" field, so concurrent events from different providers can't race). All handlers share the runtime's store/timer Maps/SessionRouter seam. The module-level file-watcher/team singletons (setHookProvider, setFileWatcherHookProvider, setTeamProvider) stay bound to providers[0] (Claude) only, since neither new provider relies on the JSONL-scanning fallback path — this keeps Claude's existing behavior byte-for-byte identical (see Test plan).

Type of change

• New feature

Test plan

• npm run build && cd server && npx vitest run → 419 passed (377 pre-existing, unmodified, confirming Claude support has zero regressions + 42 new).
• New: server/__tests__/codex.test.ts, server/__tests__/hermes.test.ts — normalizeHookEvent per event, formatToolStatus, reading-vs-writing tool classification.
• New: server/__tests__/agentRuntime.multiProvider.test.ts — simultaneous sessions from different providers landing in one store tagged by providerId, duplicate-event handling, malformed/wrong-typed events never throwing, isolated tool state between concurrent agents on different providers, unknown-providerId events dropped safely, dispose() tearing down every provider handler.
• Manual: ran a real codex exec session against a temp dir and confirmed the full parse → normalize → route pipeline fired end-to-end over the actual JSONL rollout Codex wrote (confirmed via server logs); confirmed the untracked-project-dir privacy gate (the same one Claude uses) correctly declined to visually adopt it, proving the gate applies uniformly across providers.
• npm run lint, npm run check-types, npm run asyncapi:generate (no drift), npm run e2e:inventory (no drift) all pass clean.

E2E coverage

• If user-visible behavior changed, an e2e test was added or updated under pixel-agents/e2e/tests/.
• Explain below.

What e2e tests cover this change? No Playwright e2e was added. The existing suite's e2e model (e2e/README.md → "Mocking model & rules") is built around mocking a real claude binary via mock-claude/VS Code Electron fixtures — extending that pattern to two more real CLIs (codex, hermes) that don't have an equivalent lightweight mock harness felt like a separate, larger effort better scoped as its own follow-up rather than folded into this PR. The agentRuntime.multiProvider.test.ts Vitest suite exercises the same runtime code path (handleHookEvent → HookEventHandler → AgentStateStore) that a Playwright e2e test would ultimately assert against, just without the browser/webview layer — happy to add real e2e coverage as a follow-up if maintainers would rather have it in this PR before merge.

Known limitations (documented inline in the code)

• Codex's JSONL rollout format has no explicit "session ended" record, so a session is heuristically closed after an idle timeout if the process doesn't clearly exit first.
• Hermes only loads plugins at process start (HookRegistry.discover_and_load()), so a Hermes session already running before the plugin is installed cannot retroactively receive events — this is a Hermes limitation, not something this PR works around by restarting sessions.