Skip to main content

First Turn

This page walks through a first interactive session so you can see the harness in action.

Start Interactive Mode​

Start a local HTTP model service such as Ollama, then select it with the ollama:<model> provider form:

paddles --model ollama:qwen3

The TUI opens with a prompt composer at the bottom and a transcript area above.

Ask A Question​

Type a question about the workspace and press Enter:

How does the turn loop gather evidence?

Read The Trace​

The transcript shows each step as it happens:

• Classified · 0ms
  └ planned

• Checked action-selection capability · 120ms (+120ms)
  └ ollama:qwen3: configured

• Planning...

• Action step 1: search · 3.2s (+3.1s)
  └ Rationale: Looking for turn-loop implementation

• Action step 2: read · 5.8s (+2.6s)
  └ Rationale: Reading the recursive execution loop

• Action step 3: answer · 6.1s (+300ms)
  └ Rationale: Evidence is sufficient

• Final-rendered grounded answer · 8.4s (+2.3s)
  └ Sources: src/application/mod.rs, src/domain/ports/action_selection.rs

Each line is a real event from the harness — not decorative output. The action-selection client searched the workspace, read relevant files, and decided it had enough evidence before final rendering.

The muted "Planning..." row appeared because the action-selection step took more than two seconds. These in-flight indicators show what the harness is doing during long operations.

Tune Verbosity​

The default interactive stream runs at verbosity level 0. That view keeps the operator-facing steps visible while suppressing low-value bookkeeping for casual direct responses.

For example, a greeting at the default level stays short:

User
  └ Howdy

• Action step 1: answer directly · 3.2s (+3.1s)
  └ Rationale: The user offered a casual greeting with no repository task or request for evidence, so a brief direct reply is appropriate.

Paddles · 3.2s total (+6ms)
  └ Howdy

Raise verbosity when you want more internals:

  • -v adds info-tier final-rendering and action-selection summary rows.
  • -vv adds debug-tier routing and capability detail.
  • -vvv adds trace diagnostics such as boot and prompt-payload logging.

You can also set verbose = 0, 1, 2, or 3 in paddles.toml. CLI flags override config, and the same resolved level drives both the TUI transcript and the web UI event stream.

Use A Different Action-Selection Client​

You can route action selection and final rendering to different HTTP model clients:

paddles --model openai:gpt-5.4 --action-selection-provider ollama --action-selection-model qwen3

The final renderer uses the remote provider while action selection uses the local HTTP service for investigation.

One-Shot Mode​

For scripts, use --prompt to skip the TUI:

paddles --prompt "Summarize the current turn runtime"