trace-eval 0.2.0

Deterministic-first CLI for evaluating AI agent traces

trace-eval

Tell me why this agent run went wrong and what to change next.

A deterministic-first CLI for evaluating AI agent traces. No dashboards, no LLM-as-judge, no cloud dependency. Built for solo builders and small AI-native teams using coding/CLI agents.

What It Does

Run trace-eval on an agent trace file and get:

• A scorecard — 0-100 across 5 dimensions
• Root causes — critical and high-severity issues surfaced first
• Actionable suggestions — what to fix, not just that something broke
• Before/after comparison — see if your changes actually improved things

See It in Action

# 1. Install (uv or pip)
uv sync --all-extras

# 2. Validate a trace file
trace-eval validate trace.jsonl
# Schema validation PASSED — 8 events, field coverage bars printed

# 3. Run a scorecard
trace-eval run trace.jsonl
# ============================================================
#   TRACE-EVAL SCORECARD  Total: 32.4/100
# ============================================================
#
# LIKELY ROOT CAUSES:
#   - Use canonical retrieval entrypoint
#   - Stop accessing deprecated files
#   - Context pressure exceeded 90% — reduce prompt size
#
# DIMENSION SCORES:
#   reliability             5.0  (high)
#   efficiency             77.4  (medium)
#   retrieval               0.0  (high)
#   tool_discipline        80.0  (high)
#   context                32.0  (high)

# 4. Compare before vs after a fix
trace-eval compare before.jsonl after.jsonl
# Total score: 67.5 -> 99.3  Change: +31.9 (improved)
#
# FLAG CHANGES:
#   [RESOLVED] reliability_errors
#   [RESOLVED] retrieval_no_entrypoint
#   [RESOLVED] tool_retries

# 5. CI gate — fails the build below a threshold
trace-eval ci trace.jsonl --min-score 80
# PASS (exit 0) or FAIL (exit 1)

Good Run

trace-eval run examples/hermes_good.jsonl
============================================================
  TRACE-EVAL SCORECARD  Total: 98.9/100
============================================================

DIMENSION SCORES:
  reliability           100.0  (high)
  efficiency             94.5  (medium)
  retrieval             100.0  (high)
  tool_discipline       100.0  (high)
  context               100.0  (high)

Bad Run

trace-eval run examples/hermes_bad.jsonl
============================================================
  TRACE-EVAL SCORECARD  Total: 32.4/100
============================================================

LIKELY ROOT CAUSES:
  - Use canonical retrieval entrypoint
  - Stop accessing deprecated files
  - Context pressure exceeded 90% — reduce prompt size

DIMENSION SCORES:
  reliability             5.0  (high)
  efficiency             77.4  (medium)
  retrieval               0.0  (high)
  tool_discipline        80.0  (high)
  context                32.0  (high)

FRICTION FLAGS (sorted by severity):
  [CRITICAL] retrieval_no_entrypoint
    -> Use canonical retrieval entrypoint
  [CRITICAL] retrieval_deprecated_file @event 9
    -> Stop accessing deprecated files
  [CRITICAL] context_pressure_critical
    -> Context pressure exceeded 90% — reduce prompt size
  [HIGH] retrieval_fallback_search
    -> Avoid fallback search -- use primary retrieval
  [HIGH] tool_timeout @event 5
    -> 1 tool call(s) timed out
  [MEDIUM] reliability_errors @event 3
    -> Review 3 error(s) at event indices [3, 4, 8]
  [MEDIUM] context_compression
    -> Context compression triggered 1 time(s)

Compare

trace-eval compare examples/before.jsonl examples/after.jsonl
COMPARISON: before vs after
=======================================================
  Total score:   67.5 ->   99.3
  Change:      +31.9 (improved)

  reliability            45.0 ->  100.0  ^ +55.0
  efficiency             93.5 ->   96.8  ^ +3.2
  retrieval              50.0 ->  100.0  ^ +50.0
  tool_discipline        90.0 ->  100.0  ^ +10.0
  context                95.0 ->  100.0  ^ +5.0

  FLAG CHANGES:
    [RESOLVED] reliability_errors
    [RESOLVED] reliability_terminal_partial
    [RESOLVED] retrieval_no_entrypoint
    [RESOLVED] tool_retries

Quick Start

# Install
pip install -e .
# Or with uv:
uv sync --all-extras

# Validate a trace
trace-eval validate examples/hermes_good.jsonl

# Run a scorecard
trace-eval run examples/hermes_good.jsonl

# Machine-readable output (for agents)
trace-eval run examples/hermes_bad.jsonl --format json

# Compare before/after
trace-eval compare examples/before.jsonl examples/after.jsonl

# CI gate
trace-eval ci examples/hermes_good.jsonl --min-score 80

Scoring Dimensions

Dimension	Weight	What It Measures
Reliability	35%	Did it succeed? Errors, timeouts, partial results
Efficiency	20%	Token usage, cost, tool call density
Retrieval	20%	Canonical entrypoint, deprecated files, fallback search
Tool Discipline	15%	Retries, redundant calls, timeouts
Context	10%	Context pressure, warnings, compression events

Weights are configurable. Unscorable dimensions redistribute proportionally.

Examples

Trace	Source	Events	Score	File
Good run (synthetic)	Modeled after Hermes	8	98.9	examples/hermes_good.jsonl
Bad run (synthetic)	Modeled after Hermes	11	32.4	examples/hermes_bad.jsonl
Real Claude Code	Stillness project session	3,694	60.3	examples/claude_code_real.jsonl
OpenClaw before	Real OpenClaw session	158	44.3	examples/openclaw_before.jsonl
OpenClaw after	Simulated fixes	158	57.3	examples/openclaw_after.jsonl
Compare demo	before vs after	—	+66.5	examples/case_study.md

See examples/case_study.md for a complete walkthrough of bad run → diagnosis → fix → compare.

Adapters

Format	Adapter	Capability
JSONL (.jsonl)	Generic JSONL	Full — all fields available if present in file
Hermes SQLite (.db)	Hermes	Honest/lossy — populates what exists, nulls what doesn't

Adding your own adapter? The adapter interface is simple: implement load(path) -> Trace and capability_report(trace) -> dict.

Agent Integration (--format json)

The --format json flag produces stable, machine-readable output designed for agent consumption. An AI agent that just completed a task can pipe its trace through trace-eval and use the results to self-diagnose and guide remediation.

How an agent uses it

{
  "total_score": 32.43,
  "dimension_scores": {
    "reliability": 5.0,
    "efficiency": 77.42,
    "retrieval": 0.0,
    "tool_discipline": 80.0,
    "context": 32.0
  },
  "friction_flags": [
    {
      "id": "retrieval_no_entrypoint",
      "severity": "critical",
      "dimension": "retrieval",
      "event_index": null,
      "suggestion": "Use canonical retrieval entrypoint"
    }
  ],
  "likely_causes": [
    "Use canonical retrieval entrypoint",
    "Stop accessing deprecated files"
  ],
  "suggestions": [
    "Use canonical retrieval entrypoint",
    "Stop accessing deprecated files"
  ],
  "scorable_dimensions": ["reliability", "efficiency", "retrieval", "tool_discipline", "context"],
  "unscorable_dimensions": [],
  "judge_coverage": { "...": "per-judge scorable + confidence" },
  "adapter_capability_report": { "...": "field availability from adapter" },
  "failed_thresholds": []
}

Agent remediation pattern

1. Run trace-eval run trace.jsonl --format json
2. Parse likely_causes — these are the root-cause hypotheses
3. Parse suggestions — each one maps to a concrete fix
4. Apply fixes, re-run the agent, compare the new trace
5. Use trace-eval compare old.jsonl new.jsonl --format json to quantify improvement

Fields are stable and typed. suggestions is a plain string array designed for direct use in agent prompts or remediation logic.

What's Coming

• More adapters (OpenAI traces, LangSmith, LangGraph, custom formats)
• Score profiles (balanced, reliability_first, cost_conscious)
• Baseline comparison (cost vs similar tasks)
• Parallelization analysis in Tool Discipline

Known Limitations

Hermes SQLite Adapter (lossy by design)

The Hermes adapter maps the real Hermes DB schema (sessions + messages tables) to the canonical trace format. It is honest and lossy — it populates what exists and nulls what doesn't. It does not synthesize span IDs, fabricate relationships, or guess at missing data.

Fields the Hermes schema does not provide:

Missing field	Impact on scoring
error_type	Reliability judge counts errors but can't classify them
retrieval_entrypoint, retrieval_steps	Retrieval judge always scores 50.0 (no entrypoint)
context_pressure_pct, context_tokens	Context judge returns unscorable — weights redistributed
latency_ms	Efficiency can't penalize slow tool calls
span_id, parent_span_id	No trace-level deduplication or call tree analysis
cost_estimate (Hermes stores in sessions, not events)	Efficiency cost sub-score unavailable

What this means for Hermes users:

• Retrieval score will always be 50.0 (the no-entrypoint baseline) — this is correct behavior, not a bug
• Context judge will show N/A (low) — the weight (10%) redistributes to other judges
• Errors embedded in tool result content (e.g., "error": "File not found") are not parsed by the adapter — you must add status: "error" to the JSONL if you want them scored

Generic JSONL Adapter

• Requires the canonical field names (event_type, status, tool_name, etc.)
• Fields not present are simply absent — no synthetic values
• Does not parse proprietary trace formats (OpenAI, LangSmith) — use those formats' native tools to export to canonical JSONL first

Scoring

• Scores are relative to the trace content, not to a baseline of "similar tasks"
• Tool Discipline does not yet analyze parallel tool calls
• Weight redistribution is proportional — some users may want custom profiles

What This Is NOT

• A dashboard — CLI-first, local-first
• An LLM judge — all scoring is deterministic
• An auto-fix tool — it tells you what to fix, not how
• A broad observability platform — focused on agent trace evaluation

License

MIT