Slow Triage
Accessible visual transcript / 60 seconds

Every screen, in text.

The silent video renders actual captured CLI output into readable scenes. Caption-only context and output excerpts are identified below. Reading holds are added; there is no simulated typing or GUI recording.

Fully Autonomous: generated with OpenAI-powered coding agents. All issue data and evaluation fixtures are synthetic. Only a local output-directory path was redacted in the shareable transcript copies; commands and numerical results are unchanged.

Back to video ↗Capture manifestRedaction record

Slow Triage

A real CLI run, rendered for reading · caption

Open embeddings + a real Mastra workflow
Synthetic issue data only
Captured command output; no fabricated typing
CLI demo, not a GUI screen recording
Reading holds added; this is not realtime playback
Fully Autonomous: generated with coding agents

Commands actually executed

Recorded subprocess argv; no simulated typing or shell prompt · caption

node scripts/triage.mjs data/sample.json
  ../entry_package/recording/triage-result.json

node scripts/demo-cli.mjs
  ../entry_package/recording/cli

node scripts/demo.mjs
All three recorded processes exited with code 0

Run the local issue queue

Captured stdout; ready = above threshold, still needs review · stdout

{
  "summary": {
    "total": 10,
    "ready": 6,
    "review": 4,
    "duplicateCandidates": 1
  },
  "timings": {
    "totalMs": 217
  }
}

Every Mastra stage executed

Captured CLI stdout excerpt, formatted JSON · stdout with caption

execution.status: success
validate-export: success
embed-locally: success
compare-and-abstain: success
build-review-queue: success

Verified 7 pinned local model artifacts.
Model artifacts already present; no paid inference API

Repeat the actual CLI

Captured repeatability report excerpts, formatted JSON · stdout

{
  "originalFileUnchanged": true,
  "decisionsAndScoresRepeatedExactly": true,
  "firstProcessTimings": {
    "totalMs": 208
  },
  "repeatProcessTimings": {
    "totalMs": 214
  }
}

Apply two suggestions and undo

Actual loopback API demonstration, launched from the CLI · stdout with caption

{
  "applied": 2,
  "restored": 2,
  "originalFileUnchanged": true
}

The browser interface was not exercised.

Keep the weaker results visible

Actual printed readout of preserved synthetic evaluations · stdout

PRESERVED SYNTHETIC EVALUATION - not rerun here
Developer classification: 9/20 (45.0%)
Developer keyword baseline: 90.0%
Independent classification: 15/40 (37.5%)
Independent keyword baseline: 24/40 (60.0%)
Answered 8; abstained 31; rejected 1
Duplicate pairs: TP 7; FP 1; missed 1
Duplicate precision 87.5%; recall 87.5%
Nearest positive query matches: 15/16
Broad classification was worse than keyword baselines.

What this demonstration establishes

Private prototype at recording time, October 19 2026 UTC · caption

Actual local inference and workflow execution
Review-first suggestions and reversible local output
Classification underperformed keyword baselines
No measured human time saving or real-world benchmark
Rendered-browser E2E remains unverified
At recording time: no public service or repository,
account or entry had been created.

Complete captured stdout

These are the byte-identical reviewed, shareable transcript copies used for the video. The measured-results command printed preserved synthetic evaluations; those evaluations were not rerun during this capture.