Problem

Hermes' working directory (terminal.cwd → TERMINAL_CWD) is process-global. The gateway serves many concurrent sessions — the OpenAI-compatible /v1/chat/completions endpoint, messaging platforms — from a single process, all sharing that one cwd. An individual API-driven session has no way to operate in its own working directory: the file, terminal, and code_execution tools all resolve against the single global value.

This makes per-project or per-conversation workflows impossible for any client of the OpenAI-compatible API — every concurrent session is pinned to the same directory.

Prior art that doesn't close this gap

• terminal.cwd / TERMINAL_CWD — global, fixed at gateway startup.
• /set-workspace (feat(cli): add /set-workspace command to change session cwd #23535) — registered cli_only=True; runs in the interactive CLI REPL and mutates the global env var, so it neither reaches API sessions nor isolates concurrent ones.
• ACP sessions — these do get a per-session cwd, via register_task_env_overrides + per-task_id path resolution in file_tools (_resolve_path_for_task). That path is exclusive to the ACP adapter.
• Per-session cwd storage (feat: store working directory (cwd) per session #19292, Restore session cwd on resume #21427) — records/restores a session's cwd for identification and --resume, not for concurrent isolation.
• Gateway cwd fixes (fix(gateway): honor configured cwd in gateway sessions #29365, fix(gateway): re-read terminal.cwd on session reset; use TERMINAL_CWD in environment hints #27488) — make gateway sessions honor the configured global cwd; still a single value.

Observation

The per-session-cwd machinery already exists — for ACP. register_task_env_overrides plus per-task_id resolution gives an ACP session its own working directory while its tools run. It is simply not exposed to OpenAI-compatible API sessions.

Proposal

Allow an OpenAI-compatible API session to specify a working directory, routed through the same per-task override mechanism the ACP adapter already uses, so that:

• the file, terminal, and code_execution tools resolve against that session's directory;
• ideally, project context-file discovery (AGENTS.md / .hermes.md / .cursorrules) follows the same directory (cf. [Bug]: ACP session cwd is used for tool execution but not for project context file discovery #11515, the analogous ACP gap);
• concurrent sessions with different working directories do not interfere.

Open questions for maintainers

• Transport — a field on the chat-completions request body? a session-creation parameter? a gateway-routable /set-workspace?
• Should context-file discovery follow the per-session cwd, or stay on the global value?
• How should this compose with the session-cwd storage in feat: store working directory (cwd) per session #19292 / Restore session cwd on resume #21427?

Happy to implement this once there's a steer on the preferred transport.