Rendering Modes

PromptRunner uses AgentSessionManager's rendering pipeline to display LLM session output. Three modes are available:

Studio Mode (Recommended)

Studio mode produces clean, human-readable output matching the quality of the Claude Code and Codex CLIs.

log_mode: :studio

Tool Output Verbosity

Control how much tool output is shown:

tool_output: :summary   # One-line summaries (default)
tool_output: :preview   # Summary + last 3 lines
tool_output: :full      # Complete output

Example Output

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  Prompt 01: PubSub Integration
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

  ● gpt-5.3-codex session started

    I'll implement the PubSub integration. Let me read the
    existing sink modules first.

  ✓ Read lib/rendering/sink.ex (72 lines)
  ✓ Read lib/rendering/sinks/callback.ex (50 lines)

    Now I'll create the test file.

  ✓ Wrote test/pubsub_sink_test.exs (138 lines)
  ✓ Ran: mix test test/pubsub_sink_test.exs (exit 0, 3.2s)

  ● Session complete (end_turn) — 847/312 tokens, 6 tools

  ✓ Prompt 01 completed

Compact Mode

Dense single-line token format for log monitoring:

log_mode: :compact

r+ gpt-5.3-codex >> Implementing... t+Read t-Read tr:{...} r-:end

Verbose Mode

Line-by-line bracketed format for debugging:

log_mode: :verbose

[run_started] model=gpt-5.3-codex
[tool_call_started] name=Read id=tu_001
[tool_call_completed] name=Read output=...
[run_completed] stop_reason=end_turn tokens=847/312

CLI Overrides

elixir run_prompts.exs --log-mode studio --tool-output preview

Failure Detail Toggle

Use log_meta to control terminal error detail:

log_meta: :none   # summary only (default)
log_meta: :full   # include provider stderr details when available

log_meta: :full only affects failure rendering and does not change normal token/tool event rendering.