feat: add Hermes Agent provider

[AICodingJedi]

Adds a CodeBurn provider for the Hermes AI agent (https://github.com/AICodingJedi/hermes-agent).

Hermes stores session data in a SQLite database (~/.hermes/state.db) with cumulative token counts at the session level. This provider distributes tokens proportionally across assistant turns:

• Input tokens: equal split (shared conversation context)
• Output tokens: proportional by assistant content length
• Cache/reasoning tokens: proportional split

Features:

• Lazy-loaded via loadHermes() (same pattern as Cursor/OpenCode)
• Maps 25+ Hermes tools to CodeBurn display names
• Maps 11 model IDs to readable display names
• Deduplication keys prevent duplicate entries on re-scan
• Graceful fallback when better-sqlite3 is unavailable
• Supports HERMES_HOME env var for custom data directories

[AgentSeal]

Hey @AICodingJedi, thanks for adding Hermes support. The token distribution approach for cumulative session-level totals is a smart solve. A few things need fixing before we can merge:

Must fix:

1. Use src/sqlite.ts instead of rolling your own loader. We have a shared SQLite adapter that all providers use (Cursor, OpenCode). Your PR duplicates ~60 lines of loading logic. Import isSqliteAvailable, getSqliteLoadError, openDatabase from ../sqlite.js and delete your custom loadSqlite, isSqliteAvailable, getSqliteLoadError, openDatabase functions plus the SqliteDatabase type. See src/providers/opencode.ts for the pattern.

2. Add bashCommands: string[] to your yield. We added this field to ParsedProviderCall in v0.5.1. Your PR will fail TypeScript without it. If Hermes has bash/terminal tool calls you can extract commands from, wire it up via extractBashCommands from ../bash-utils.js. Otherwise just yield bashCommands: [].

3. Move calculateCost to a top-level import. You're doing await import('../models.js') inside the generator loop. Just add import { calculateCost } from '../models.js' at the top. No need for dynamic import here.

4. Remove unused imports. readdir, readFile, basename are imported but never used.

5. Add tests. We need at least: session discovery, archived/empty session handling, token distribution math, dedup, cost fallback, corrupt data. See tests/providers/opencode.test.ts for the pattern (creates real SQLite databases, 31 tests).

6. Fix dedup key. Your key includes derived token counts (${turnInputTokens}:${turnOutputTokens}). If the proportional split heuristic changes, the same messages get double-counted. Use hermes:${session.id}:${turn.assistantMsg.id} instead.

Style (per our CLAUDE.md):

7. Remove the section separator banners (// ----) and the file-header doc comment. We keep comments minimal -- only when the WHY is non-obvious.

8. Remove the as Function cast in openDatabase (goes away when you switch to sqlite.ts).

9. Add a newline at end of file (both hermes.ts and index.ts).

10. Don't move the opencodeProvider/opencodeLoadAttempted declarations in index.ts. Keep your changes to adding Hermes only -- no cosmetic reordering of existing code.

The core logic (session parsing, tool mapping, model names, token distribution) is solid. Once the above is addressed it should merge cleanly. Happy to help if you have questions on any of these.