fix(native-chat): chats ignore saved Arguments; they are for terminal launches

[brennanb2025]

ELI5

Settings → Agents → Arguments holds the command-line flags your terminal tabs (and the default, terminal-backed chat) start an agent with. The updated (structured) native chat (Settings → Experimental → "Use updated structured native chat") also tried to reuse them: it guessed, flag by flag, which ones a Claude or Codex chat could use. That went wrong in confusing ways:

• some flags were silently dropped;
• some were placed where Codex ignores them;
• some valid flags stopped the chat from starting.

Now the updated (structured) native chat doesn't read Arguments at all. They apply to terminal tabs and terminal-backed chats, as before. The program path (Command) and the agent's environment settings still apply to chats.

What Changed

Before: the updated (structured) native chat translated your terminal Arguments per agent.

• Claude: flags went to the SDK one value each, and some were skipped silently.
• Codex: an allow/deny list went in front of app-server, where Codex ignores most options.

Some results:

• --oss silently used the cloud provider;
• -m did nothing;
• --disallowedTools was dropped;
• --mcp-config a.json b.json refused the chat with "give --mcp-config more than one value".

After:

• A structured Codex chat starts as plain codex app-server.
• A structured Claude chat gets no Arguments-derived flags or extra folders.
• Structured Grok, OpenCode and OMP chats already ignored Arguments.
• Terminal-backed chats are unchanged. That is the default chat, plus WSL/SSH chats and chats created before the toggle; they still launch the CLI with Arguments.
• Your agents' own config files still apply in chat: ~/.claude settings and .mcp.json, and ~/.codex/config.toml.
• The model and effort come from the chat's picker.
• Command (the program path, fix(native-chat): refuse a Command that can't run instead of silently running the stock CLI #25667) and the environment setting are unchanged.

Settings → Agents: each agent that has native chat now shows one line under its Arguments field, in all 6 languages:

• Claude: "Used when Claude CLI is launched. Native chat does not use these; put its settings in ~/.claude/settings.json."
• Codex, Grok and OpenCode: the same sentence, with ~/.codex/config.toml, ~/.grok/config.toml and ~/.config/opencode/opencode.json.
• OMP: "Used when OMP CLI is launched. Native chat does not use these." (no file named)
• Agents without native chat get no line, and the row's existing summary is unchanged.

Unchanged:

• Permissions. Agent Permissions still stores its full-access choice inside Arguments, and chat still reads that one setting until the chat permission setting (feat(native-chat): each Claude/Codex chat gets its own permission mode #26116) replaces it.
• Older hosts. A chat connected to an older host can still receive that host's "Saved Arguments …" refusals, so the wording for them stays.

Mechanism: this removes the per-agent translators (claude-structured-launch-args.ts, codex-structured-launch-args.ts), the shared reader that fed them, the Arguments-only error type and its emitters, and the runtime plumbing that carried Arguments into chat. That includes a launchArgs field on chat records, which current builds no longer write (builds before #20944 did). The Codex model-list probe already ignored Arguments; a test now pins that.

Why

The common pattern is a per-agent program-path setting for chat, with no reuse of a terminal arguments field. Agent CLIs keep their real configuration in their own config files, which chats already load. Translating terminal flags into a chat session needs per-CLI tables that drift with every CLI release, and it fails in ways users can't see. Ignoring them is the standard behavior and the simplest.

Differences from the common pattern:

• Orca has one Arguments field shared with terminal tabs, where the common pattern has none, or a separate chat-only field. Intended: terminal tabs keep using it.
• Agent Permissions is still read from Arguments for chat. Temporary, until feat(native-chat): each Claude/Codex chat gets its own permission mode #26116 moves chat to its own permission setting.

Linked Issue

Merge order: after #26116 (chat permission setting). Then the Settings line is exactly true, because the chat reads its own permission setting instead of the Agent Permissions flag kept in Arguments.

Supersedes #26059 (Codex), #26127 (Claude) and #26367 (Grok). It reverses the Arguments half of #25721; the Command half stays.

Visual Proof

QA ran on Windows (low spec), driven by a non-Claude agent over the DevTools protocol against a hidden window, with stand-in CLIs. The box's real agent configs were unchanged at every step.

Settings → Agents: the new line under Arguments (untouched profile). Agents with native chat get the line; Gemini, which has none, gets nothing:

The updated chat ignores saved Arguments.

• A Codex chat with saved -m qa-x -c foo.bar=1 --profile work --oss starts as exactly app-server and replies. Before, --profile and --oss were refused or silently changed what ran.
• A Claude chat with saved --mcp-config a.json --worktree feat --append-system-prompt "- x" starts and replies, and none of those reach the CLI.
  
  

Permissions still follow Agent Permissions. Untouched Arguments give full access; Manual gives on-request / workspace-write:

With the updated chat toggle off, the terminal-backed chat still launches with the saved Arguments (verified from the launch command line; the terminal tab itself wasn't captured).

Testing

• Targeted tests (explicit files, final head before the end-of-PR main merge: 6083afa). The full affected set, 491 files, passed at 53b2364; the locale and i18n tests (76 files) passed at a1f08ab; the Claude start and Fast-mode tests (38 files) passed at 6083afa. Among them:
  • Codex chat argv is exactly app-server with non-empty saved Arguments;
  • Claude extraArgs carry nothing from saved Arguments;
  • the Codex model probe ignores saved Arguments;
  • an older host's Arguments refusal still renders.
• orca-ci-checks: CI's lint, code-quality, localization and typecheck steps, 23/23 pass.
• Real Claude CLI suite: runs once on the final head, after the end-of-PR main merge.
• Electron QA on Windows (low spec), driven by a non-Claude agent with stand-in CLIs. With the updated chat on, a Codex chat started as exactly app-server despite saved -m qa-x -c foo.bar=1 --profile work --oss, and a Claude chat's launch carried none of --mcp-config a.json --worktree feat --append-system-prompt "- x". Both replied. Permissions still followed Agent Permissions. With the toggle off, the terminal-backed chat still launched with the saved Arguments. See Visual Proof.

Review

Pending.

Agent skill upstream boundary

• Not applicable, or this change follows docs/reference/agent-skill-sharing-upstream-boundary.md and copies or mechanically translates no upstream skill-installer source, tests, fixtures, registry entries, path tables, comments, or documentation.

Notes

• Mixed versions: no new wire content. An older host may still send an Arguments refusal, and newer clients still render it.
• SSH/WSL: unchanged. Chats run where they ran.
• Mobile: unchanged.
• Persisted data: the unused launchArgs record field is dropped. Old records that carry it are read without it; native chat has no real users yet.

Checklist

• This PR is small and focused
• I explained what changed and why (ELI5, the user-facing before/after, the mechanism, and why over the alternatives)
• Before/after screenshots or videos attached for UI changes, or N/A with reason
• Self-reviewed for correctness, security, and performance
• Cross-platform, SSH/remote, and path/shortcut impact considered (or N/A)
• pnpm lint, pnpm typecheck, pnpm test, and pnpm build pass (or CI will cover; local preferred)

[brennanb2025]

Merge-order note: #26116 (per-chat permission modes) is merged as 2233fd3. It was rebuilt before merging: the mode is now a chat option applied at launch, so live Claude switching and the old Codex per-turn overrides are gone, and permission flags typed into Arguments are still stripped. This PR was written against the earlier version of #26116, so check it against current main before taking it out of draft.