SDx Evaluation Agent

An MVP evaluation agent for the SDx development flow. It reads engineer–agent traces from LangFuse, extracts evidence of agent quality across three signal classes, and writes a markdown learnings report. Reports are designed to be ingested by the SDx knowledge layer alongside requirements, specs, and FDA/ASPICE documents.

Demo target: F2F update on 2026-05-21.

---

What this repo contains

eval-agent/
├── .opencode/
│   ├── agent/evaluator.md                            # primary opencode agent
│   └── skills/langfuse-trace-analysis/SKILL.md       # on-demand skill
├── scripts/
│   ├── load_traces.py                                # LangFuse CSV → packet
│   └── generate_synthetic_traces.py                  # deterministic toy data
├── data/
│   └── synthetic-traces.csv                          # 1 session, 8 turns
├── eval-reports/
│   └── ses_synthetic_camera_obj_det_001_…_learnings.md   # worked example
├── docs/
│   ├── SYSTEM-EXPLAINED.md                           # plain-language docs
│   ├── slides.md                                     # Marp source
│   ├── slides.html                                   # rendered deck
│   └── slides.pdf                                    # exported deck
├── opencode.json                                     # agent registration
└── ACTION_LOG.md                                     # decisions + progress

---

How it works (one diagram)

                      I · Capture
                      ─────────────────
                      Engineer → SDX → LangFuse → trace export (.csv)

                      II · Evaluate
                      ─────────────────
                      load_traces.py → trace packet (.json)
                                       │
                                       ▼
                          evaluator agent ← trace-analysis skill
                                       │
                                       ▼
                                       learnings report (.md)

                      III · Improve
                      ─────────────────
                      team · prompts · agents · knowledge layer

The agent extracts three orthogonal signal classes from each session:

Class	Method	What it captures
Structural	count over turns	tool-call mix, finish reasons, tokens, system-prompt churn
Qualitative	model judgement, cited per turn	did the agent answer the request; did the plan match the ask
Implicit	pattern-match the six-item catalogue	user correction, retry, abandoned trajectory, follow-up edit, context bloat, tool failure echo

The implicit class follows the approach presented at the CMU FLAME event by the OpenHands team.

---

Quick start

1 · Run the agent in opencode

git clone https://github.com/ryanlingo/eval-agent.git
cd eval-agent
opencode
# press Tab, pick `evaluator`, then ask:
# "Evaluate session ses_synthetic_camera_obj_det_001 in data/synthetic-traces.csv"

The agent loads the langfuse-trace-analysis skill, reads the trace packet, applies the three extractors, and writes a report to eval-reports/<session_id>_<UTC_TIMESTAMP>_learnings.md.

2 · Run the loader directly (no LLM)

# List sessions in a CSV
python scripts/load_traces.py --csv data/synthetic-traces.csv --list

# Build a packet for one session
python scripts/load_traces.py --csv data/synthetic-traces.csv \
  --session ses_synthetic_camera_obj_det_001 > /tmp/packet.json

3 · Regenerate the toy dataset

python scripts/generate_synthetic_traces.py --out data/synthetic-traces.csv

The synthetic session is deliberately designed to surface three signals: a user correction (tr_0004), a tool failure echo (tr_0006), and a retry (tr_0005 → tr_0006).

4 · Render or update the slide deck

npx --yes @marp-team/marp-cli docs/slides.md -o docs/slides.html
npx --yes @marp-team/marp-cli docs/slides.md --pdf --allow-local-files \
  -o docs/slides.pdf

---

Report file naming

eval-reports/<session_id>_<YYYYMMDDTHHMMSSZ>_learnings.md

Examples for the same session over time sort together in a directory listing. UTC timestamps are second-precision so re-evaluations of the same session never overwrite older reports.

---

Documentation

• docs/SYSTEM-EXPLAINED.md — plain-language description of every component, the data flow, file naming conventions, and a glossary. Sentences are deliberately short; written with non-native English readers in mind.
• docs/slides.pdf — nine-slide deck for the F2F. Title · Problem · Method · Architecture · Components · Demonstration · Sample report · Integration · Roadmap.
• ACTION_LOG.md — append-only log of decisions, scope changes, and progress. Read this first if you're picking up the project mid-stream.

---

Scope and limitations

This is an MVP. It is not:

• a benchmark — no numeric scores;
• a real-time monitor — runs on past traces only;
• a cross-session aggregator — one report per session;
• a replacement for FDA or ASPICE reports — it is a new artifact that lives next to them in the knowledge layer.

The worked example uses a synthetic toy session generated from scripts/generate_synthetic_traces.py. Validation against real sessions is part of the roadmap on the final slide.

---

Roadmap

Before the demo (this iteration):

• Replace the CSV-export step with the LangFuse scoring/cost API.
• Run the agent against three real sessions to estimate signal recall.
• Cite reports from the eval-reports corpus inside one downstream agent run.

Beyond MVP:

• Cross-session rollups to detect agent-level regressions.
• Implement implicit-signal detectors as deterministic Python; reserve the LLM for qualitative judgement.
• Convert recommendations into prompt/skill diffs the team can review.
• Trial Laminar against LangFuse on the same sessions.