Laminar Trace 39f38cb4

Single-run dashboard for agent behavior: harness activity, prompt carry-forward, and direct search, edit, and test burden.

Run
39f38cb4-a374-3834-7153-8d1443d04822
Success
yes
Steps
843
Wall-clock duration
12m 45.6s
Summed event duration
55m 37.5s
Input tokens
8,517,025
Output tokens
23,694
Carry-forward ratio
98.8%

Behavior and burden summary

Compact signals for long-run dynamics, with direct lower-is-better search and verification counts.

Dominant action
overhead
Dominant lane by time
workspace
LLM steps
103
Verification gap episodes
0
Max verification gap
0
Test run events
4
Controller events
13
Files before first edit
16
Edits before first pass
30
Tests before first pass
6
Tests after first pass
852
Max retry burst
1
Duration note
Wall-clock uses first start to last observed end; summed event duration adds step durations and can exceed wall-clock when spans overlap.
Top control events
finish (10), plan_update (4), wait (1)
Top tools
local_file_store (424), file_editor (102), terminal (98), task_tracker (4), finish (2)
Top files
/workspace/jinja (71), /workspace/jinja/src/jinja2/utils.py (52), /tmp/jinja2_download (26), /workspace/jinja/src/jinja2/nodes.py (12), nodes.py (10)

Search / edit / verify burden

Direct counts for files before the first edit, edits before the first pass, tests before the first pass, tests after the first pass, and retry bursts. Lower is usually better.

Harness lane occupancy

Wall-clock bursts for controller, model, workspace, and runtime so you can see which layer actually held the run over time.

Feedback → next meaningful move heatmap

Raw counts of what came after each feedback category once routine execute spans were collapsed into their observed outcomes. This keeps the heatmap aligned with the more legible categories like edit, file content, test failure, test pass, retry, wait, and finalize.

All anchors overview

One dense Sankey combining inspect, edit, test failure, and test pass anchors in a single view. It is less readable than the selector, but useful for spotting the overall balance of branches across all anchor types at once.

Selected anchor: next two meaningful moves

Choose a high-value anchor like inspect, edit, test failure, or test pass. The chart then shows the next two meaningful nodes after that anchor while skipping plan and overhead and collapsing routine execute steps into their observed outcomes. finalize is allowed as an endpoint, but not as a starting anchor.

Lane summary and handoffs

Observed active time, burst lengths, and layer-to-layer handoffs. Active-time shares are relative to full run time and can overlap when layers were active concurrently.

How lane timing works

Controller
Harness orchestration around the main solve loop. Examples: retry, wait, finish, finalize, task updates.
Model
LLM call spans: prompt-in, model-out, and prompt carry-forward effects. Example: one model response that sets up the next tool action.
Workspace
Task-facing work such as reading files, editing code, and ordinary terminal/tool actions. Examples: open file, patch code, run tests.
Runtime
Internal overhead and persistence activity rather than task-solving work. Examples: LocalFileStore and harness plumbing.
Active time
Union of observed wall-clock intervals for a lane, so overlapping events within the same lane are merged instead of double-counted.
Burst
A consecutive same-lane stretch in observed event order, shown as one occupancy bar segment.
Handoff
One lane burst followed by the next lane burst in trace order. The Sankey width shows how often that transition happened.

Run timeline

Each bar shows one event positioned by time, colored by dominant action type.

Prompt makeup per LLM call

Each bar is normalized to 100% so you can see whether a prompt was mostly fresh task/context, replayed prior context, cached prefix reuse, or tool output. Hover for raw token counts.

File transition graph (top files)

Observed movement between the most-touched files in the run, filtered down so the main repo navigation path stays readable.

Harness steering events

Only the harness-side steering actions: plan updates, retries, waits, interruptions, and finish. This is not ordinary file reading, editing, or testing work.

Step anatomy

Click a chart row or choose a step to inspect the exact command, feedback, prompt, and control metadata.

Cumulative progress

Tokens and progress measures over the trajectory, useful for spotting churn versus convergence.