Run reports#
--report path.json writes one JSON object when a headless run exits, on every outcome, including
an error before the first turn. Every field is always written: a zero is a measurement, and a
missing key would be a bug. The exceptions are the research fields quote_problems, claim_check and
claim_check_skipped, and a turn's failed_attempts, which are left out when there is nothing to record. schema_version changes only when a field changes meaning or is
removed; new fields are added without a bump.
{
"schema_version": 1,
"harness_build": "933da30",
"outcome": "verified",
"error": null,
"model_requested": "gemini-3.8-flash-high",
"model_id": "gemini/gemini-3.8-flash",
"effort": "high",
"context": "folded",
"gated": true,
"gate_source": "user",
"container": null,
"container_workdir": "/workspace",
"sandboxed": true,
"sandbox": "seatbelt",
"network": "on",
"allow_read": [],
"pass_env": [],
"withheld_env": ["GITHUB_TOKEN", "SSH_AUTH_SOCK"],
"duration_seconds": 186.4,
"turns": 38,
"empty_turns": 0,
"tool_calls": 51,
"gate_attempts": 1,
"rollbacks": 0,
"provider_retries": 0,
"cost_usd": 0.4566,
"max_cost_usd": null,
"cost_capped": false,
"fold_budget": null,
"turn_usage": [ { "turn": 1, "input": 14210, "cached": 0, "output": 312, "thinking": 540, "...": "..." } ],
"usage": { "input": 3120455, "cached": 2811020, "output": 18011, "thinking": 40210 },
"gate_ms": [41230],
"gate_turns": [38],
"gate_failures": [],
"tool_counts": { "bash": 22, "read": 17, "edit": 9, "grep": 3 },
"model_seconds": 121.7,
"tools_seconds": 20.1,
"gate_seconds": 41.2,
"final_message": "Changed parse() to return Err(ParseError::Empty) ... Deviations: none."
}
Fields#
The run
| field | meaning |
|---|---|
harness_build |
the commit blyte was built from |
outcome |
verified, finished (an ungated run that stopped by itself), budget_exhausted, gate_failed, blocked or error |
error |
the error message for an error or blocked outcome, else null |
model_requested / model_id |
the --model given, and what the provider actually called |
effort |
the effort level when the model supports it, else null |
context |
folded or conversational |
gated, gate_source |
whether a gate was given; user or none |
duration_seconds |
wall-clock time of the run |
final_message |
the model's last text: its own report of what it did (what blyte dispatch shows the orchestrator) |
Work
| field | meaning |
|---|---|
turns, empty_turns |
model calls, and calls that returned neither text nor a tool call |
tool_calls, tool_counts |
tool calls in total and by tool |
gate_attempts, gate_ms, gate_failures |
gate runs, the time of each, and the (trimmed) output of each failure |
gate_turns |
the turn on which each gate attempt ran, in order (one entry per gate attempt) |
rollbacks |
times the tree was reset after a failed gate |
provider_retries |
transient provider errors waited out |
gate_wait_ms |
time spent waiting for a machine-wide gate slot on each attempt, in order (0 when a slot was free) |
gate_left_out |
dispatch only: the gate steps blyte.toml records as left out, each with its reason |
truncated_turns |
turns whose reply hit the output token limit and was cut off |
big_read_turns |
turns whose tool output was more than half of the fold budget |
read_loop, peak_read_repeats |
whether the run stopped because it kept repeating read, grep and glob calls without changing the workspace, and the most repeats seen between two changes |
fold_thrash |
whether the run stopped because consecutive turns kept folding the previous turn's results |
stalled |
whether a fix run stopped because it made no change and kept calling tools |
Cost and time
| field | meaning |
|---|---|
usage |
tokens over the run: fresh input, cached input, output, thinking |
turn_usage |
the same per call (turn, input, cached, output, thinking), with the prompt's size (prompt_messages), whether it was the previous prompt plus new messages (appended) and how much was new (new_chars), the fold budget in force, and timings: model_ms (the model call, retries included), tools_ms (running the tool calls), first_data_ms and max_gap_ms (time to the first streamed data and the longest silence; null and 0 when the provider does not stream) |
cost_usd |
list-price cost, or null when the model's price is unknown |
max_cost_usd, cost_capped |
the --max-cost cap, and whether the run stopped at it |
fold_budget |
the --fold-budget given, if any |
model_seconds, tools_seconds, gate_seconds |
where the time went; the rest is start-up and the harness itself |
Security settings
| field | meaning |
|---|---|
sandboxed |
whether host commands ran sandboxed |
sandbox |
seatbelt, landlock, landlock-offline, container or off |
network |
on or off |
allow_read, pass_env |
the --allow-read grants and --pass-env names |
withheld_env |
names, never values, of variables withheld from commands (empty in a container) |
container, container_workdir |
the container, if the run used one |
Using reports#
Reports are designed to be compared. The evaluation harness that measures blyte reads exactly this
shape; an orchestrator can read outcome, gate_failures and cost_usd to decide whether to
merge, retry or escalate. jq is enough for most questions:
jq '{outcome, turns, cost_usd, model_s: .model_seconds, gate_s: .gate_seconds}' run.json