Skip to main content

​
Provider Quirks

Each agent CLI has its own behavior, flags, and limitations. Gump’s adapters handle the differences, but understanding them helps when debugging or choosing agents.

​
Claude Code

The most full-featured adapter. Supports all session modes, reports cost natively, and has granular token breakdowns per message.
  • Permission mode: Gump uses --dangerously-skip-permissions for headless auto-approval. This is the only mode that works reliably in automated workflows.
  • Max turns: Supported via --max-turns.
  • Session resume: --resume <uuid> continues the conversation.
  • Working directory: --cwd <path> sets the worktree.

​
Codex

Runs in sandbox mode with network disabled by default.
  • Permission mode: Gump uses sandbox mode (--full-auto). Network is disabled inside the sandbox — agents can’t make HTTP calls.
  • Max turns: Not supported as a CLI flag. Gump relies on the max_turns guard instead.
  • Session resume: resume <thread-id> — the thread ID format is specific to Codex.
  • Token reporting: Per turn, not per message. Less granular than Claude.
  • Shell classification: Codex emits shell events differently — turn classification is reconstructed (lower confidence).

​
Gemini CLI

Uses auto-approve mode with implicit session resume.
  • Permission mode: --yolo auto-approves all actions. No intermediate mode available in headless.
  • Max turns: Not supported as a CLI flag. Guard-only.
  • Session resume: --resume without an argument. Gemini resumes the last session for the current working directory. Since each run has its own worktree, this naturally targets the right session.
  • Working directory: No --cwd flag. Controlled via exec.Cmd.Dir.
  • Token reporting: Aggregated in the final result event, not per message.

​
Qwen CLI

Similar to Claude in many ways, but with a few differences.
  • Permission mode: --yolo for auto-approve.
  • Max turns: --max-session-turns <N> (note the different flag name from Claude’s --max-turns).
  • Allowed tools: Passed as separate arguments, not a comma-separated list. Gump restricts to filesystem and shell tools only (no web search, no web fetch) for reproducibility.
  • Working directory: No --cwd flag. Controlled via exec.Cmd.Dir.
  • Session resume: --resume <uuid> — UUID format, same as Claude.

​
OpenCode

The most different adapter. Uses file-backed stdout instead of piped streams.
  • Stdout handling: OpenCode blocks on StdoutPipe(). Gump writes stdout to a temp file and reads it from there.
  • Working directory: Native --dir flag.
  • Session resume: --session <ses_id> — session IDs use a ses_ prefix format.
  • Token reporting: Aggregated per step_finish event, not per message.
  • Exit codes: OpenCode is the only provider where exit code 1 is meaningful (indicates a step failure, not a crash).

​
Cursor CLI

Shares the Claude Code adapter with minor additions.
  • Context files: Uses both CLAUDE.md and .cursorrules for context.
  • Behavior: Otherwise identical to Claude Code — same flags, same session format, same token reporting.

​
Cross-provider limitations

  • Session resume across agents: Blocked. Claude session IDs don’t work with Qwen, and vice versa. Gump validates format at the engine level.
  • Cost reporting: Only Claude and Cursor report cost natively. For other providers, Gump estimates cost from token counts and published pricing.
  • Tool restrictions: Each provider has different tool names. Gump maps them internally but can’t add tools the provider doesn’t support.

​
Provider volatility

Agent CLIs are moving fast. Flags change, output formats evolve, features appear and disappear. Gump’s minimal contract with providers is: filesystem access + Git. If a provider breaks its structured output, Gump falls back to compat mode (raw stdout + git diff + fresh sessions). The workflow still runs — metrics become less granular until the adapter is updated.