← Other Products

Span-level replay: fork from the failure

Fork-replay one recorded span from a failed agent run and get an attribution verdict, instead of re-running the whole run to reproduce it.

When an agent run fails, re-running the whole run to reproduce it is slow and often pointless — the failure lives in one step. This service fork-replays ONE recorded span (a JSON tool or model step from a tracer) out of the failed run and returns an attribution verdict: determinism, tool-nondeterminism, tool-drift, env-drift, world-state, or model-decision. It ingests normalized JSONL span records and OTel-shaped spans (including LangGraph/Langfuse-shaped foreign traces), verifies the recorded env_digest before any execution, and stubs side-effecting tools from the recording itself: a side-effecting tool NEVER executes — it is served its recorded result by canonical fingerprint, and a stub miss fails closed. Lab proof (LABREPORT-008): 37/37 tests; seeded attribution 24/24 vs 0.167 for the full-re-run baseline; tool-span replay uses 0 model tokens, 0.18–20.2 ms by span size.

Tool-span replay $0.02 per replay · Model-decision replay $0.05 per replay — pay per replay (x402, USDC on Base), no subscription, no free tier.

For agents

API endpoint (JSON): https://textforagents.com/products/span-level-replay-fork-from-failure/ · OpenAPI spec · Agent guide (llms.txt) · Full documentation (text)

Support: support@textforagents.com