Dashboard

The local web UI for browsing runs, inspecting iteration traces, reading prompt diffs, and branching new experiments from any iteration.

Start the dashboard from your project directory:

aevyra-reflex dashboard

This starts a local server (default port 8128) and opens the UI in your browser. The page reads directly from .reflex/runs/ — reload manually to pick up new runs.

---

Runs Dashboard

The landing screen lists every optimization run stored in .reflex/runs/, most recent first. Click any row to open its detail view.

Columns

Column	Description
ID	Numeric run identifier (e.g. 001). Branch runs show a └─ connector and an ⎇ parent/iter badge.
Created	Timestamp the run was started, derived from the directory name.
Status	completed, interrupted, or running.
Strategy	The optimization strategy used: auto, iterative, structural, fewshot, or pdo.
Model	Target model(s) being optimized, as passed with -m on the CLI.
Iters	Number of iterations completed.
Baseline	Mean score before any optimization.
Best	Highest mean score reached. Shown in green.
Target	If no verdict: the score_threshold from config. If verdict: target_model @ threshold [json|run].
Dataset	Dataset filename. Click to expand and reveal the full absolute path.
Reasoning	Reasoning model used for optimization.

Branch runs in the list

Branch runs appear immediately after their parent in the runs list — not at the bottom — so the full experiment tree stays together. A └─ connector and an ⎇ 003/5 badge show the parent run and the iteration it was branched from.

---

Run Detail

Click a row to open the detail view for that run. The URL updates to #/runs/<id> so you can bookmark or share it.

Stats row

Six summary tiles appear at the top of the page:

• Baseline — mean score before optimization
• Best Score — highest score achieved, with improvement shown below (e.g. +0.2859 / 49.2%)
• Iterations — total number of iterations run
• Strategy — strategy used for this run
• Eval tokens — total tokens consumed by evaluation calls across all iterations
• Reasoning tokens — total tokens consumed by the reasoning model across all iterations

A Verified tile also appears when the final verification eval produces a different score than the best in-run score — this can happen when a prompt generalizes differently on the held-out verification pass.

Flow graph

The left panel shows the full optimization trace as a top-to-bottom sequential chain. Iterations always run sequentially — the graph reflects this with a single vertical layout.

Node types:

Baseline

Teal pill at the top. Shows the baseline mean score and the target threshold.

Phase header

Coloured label separating strategy phases. Shows the phase name and iteration count (e.g. structural ×3).

Iteration card

Shows the iteration number, mean score, delta vs. previous iteration, a score bar with the target threshold marker, per-metric breakdown, eval and reasoning token counts, and timestamp. A ★ marks the best iteration.

Best

Green pill at the bottom. Shows the best score and total improvement over baseline.

Phase colours:

Phase	Colour
structural	Teal
iterative	Blue
fewshot	Purple
pdo	Orange

Click any iteration card to load its detail in the right panel. The most recent iteration is selected by default.

Iteration detail

The right panel updates when you select an iteration card. It shows three sections:

Prompt — Three tabs let you switch views:

• Current — the full system prompt used for this iteration
• Diff — a line-by-line diff against the previous iteration’s prompt (or the initial prompt for iteration #1). Additions are green, removals are red.
• Initial — the original prompt before any optimization

Reasoning — The reasoning model’s explanation of why it made the changes it did. The strategy phase prefix (e.g. [structural]) is stripped so you only see the human-readable explanation.

Scores — Per-metric score breakdown (e.g. rouge: 0.812 · bleu: 0.743). Only shown when the iteration includes metric-level data.

---

Token usage

Every iteration tracks two token counts:

• Eval tokens — tokens used by the target model and eval scoring calls for that iteration.
• Reasoning tokens — tokens used by the reasoning model (Claude, Qwen3, etc.) to analyze failures and write the revised prompt.

Token counts appear on each iteration card in the flow graph and are summed into the two stat tiles in the stats row. Values are formatted as 1.2K, 3.4M, etc. for readability. No dollar estimates are shown — token prices change frequently and vary by provider tier.

---

Branch runs

Branch a new experiment from any iteration without re-running the baseline. This is useful when a run plateaued on one strategy and you want to try a different approach starting from the best prompt found so far.

Starting a branch

Hover over any iteration card in the flow graph — a ⎇ button appears in the top-right corner of the card. Click it to open the branch modal:

1. Choose a strategy — iterative, fewshot, structural, pdo, or auto.
2. Set max iterations for the new run (default: 10).
3. Click ⎇ Start Branch.

The new run starts immediately. The browser navigates to the live job stream so you can watch it run. When it completes it appears in the runs list, indented directly below its parent.

What gets reused

	Branch run	Fresh run
Initial prompt	Iteration N’s prompt	Your original prompt
Baseline score	Copied from parent	Re-evaluated from scratch
Strategy	Your choice	Config / CLI flag
Dataset	Same as parent	Specified on CLI

Branching skips the baseline eval entirely — the parent’s baseline score is copied into the branch run so the improvement delta is computed against the same starting point as the parent.

Branch lineage

The branch run’s ID cell in the runs list shows ⎇ parent_id/iter (e.g. ⎇ 003/5) to indicate it was branched from iteration 5 of run 003. The run always appears immediately after its parent in the list regardless of when it was created.

---

States

State	What you see
No runs yet	A placeholder with the CLI command to start a first run.
Server not running	An error card explaining the API request failed.
Run in progress	A blue running badge. Data reflects iterations completed so far.
Interrupted run	A yellow interrupted badge. Data up to the last checkpoint is shown.
Verdict target	Target column shows model @ threshold [json|run] instead of a plain number.
Branch run	└─ connector and ⎇ parent/iter badge in the ID column.