Skip to main content

​
Reports & Metrics

Gump tracks structured metrics for every run. Use gump report to see them.

​
Basic report

gump report

Run: run_2026_03_24_1042
Workflow: tdd · Status: pass · Duration: 2m13s · Cost: $2.41

Step            Agent           Turns  Cost    Gate
───────────────────────────────────────────────────────
decompose       claude-opus       3   $0.31   schema ✓
build/1/tests   claude-haiku      8   $0.12   compile+test ✓
build/1/impl    claude-haiku     14   $0.18   compile+test ✓
build/2/tests   claude-haiku      6   $0.09   compile+test ✓
build/2/impl    claude-sonnet    12   $0.44   compile+test ✓ (escalated)
quality         —                 —   —       compile+lint+test ✓

​
Detailed step view

Drill into a specific step: 
gump report --detail build/2/impl
This shows the step’s attempts (including failed ones), state bag entries, files changed, tokens breakdown, and guard activity.

​
Comparing runs

gump report --last 5
Shows a summary table of your last 5 runs — workflow, status, duration, cost, retries. Useful for spotting trends: is a particular workflow getting more expensive? Are retries increasing?

​
What Gump measures

​
High confidence (100%)

These are deterministic and exact:
  • Git diff (files changed, lines added/removed)
  • Gate exit codes (pass/fail)
  • Wall-clock duration per step and per run
  • Number of retries and escalations
  • Guard triggers

​
Medium confidence (~80%)

Tracked live from the agent stream, but subject to provider accuracy:
  • Tokens (input, output, cache read, cache write)
  • Cost (estimated from token counts and provider pricing)
  • Timestamps from the agent’s native events
  • Context window usage per step

​
Reconstructed (~50%)

Inferred from heuristics, not directly reported by all providers:
  • Turn count (cognitive cycles)
  • Turn classification (coding, execution, exploration, planning)
  • Shell command classification for Codex

​
Key metrics

​
Cost

Tracked per step, per item, per run. Aggregated from agent token reports. The estimate is based on published provider pricing — actual bills may differ slightly.

​
TTFD (Time To First Diff)

The time between agent launch and the first file modification. A long TTFD suggests the agent spent time exploring or planning before writing code. A useful signal for prompt optimization.

​
Stall detection

Gump detects agents that spin in circles:
  • tool_error_count — repeated tool failures
  • correction_loops — edit → test → fail → edit cycles
  • fatal_loops — the agent hits the same error repeatedly
  • repeated_action_loops — identical tool calls in sequence
These counters appear in gump report --detail and help diagnose why a step needed many turns or failed.

​
Context window usage

How much of the agent’s context window was consumed per step. High usage suggests the prompt or accumulated context is too large, which can degrade agent performance.

​
Analytics model

Gump organizes data at four levels: Event → Turn → Step → Run Each event carries a timestamp and semantic labels. Turns are classified by their dominant activity (coding, exploration, execution). Steps aggregate turns. Runs aggregate steps. You can analyze at any level.