📜 Reading a run page
Every run has a report page (the Chronicle lists them all). It replays the full transcript — prompts, responses, commands, diffs — live while the run happens and forever after. The row types:
State rows and their basis badges
A staterow records the machine moving from one stage to the next. Each one must answer “what earned this?” with a badge:
Real recorded evidence earned the transition — the badge names its kind (an ai verdict, a command run, an agent diff, a check) and the detail points at the evidence.
A step the pipeline advances by policy because there is no evidence to have (for example, your ask itself is the sensing evidence). Declared, not disguised.
An honest admission: the stage was promoted despite a failed gate so the walk could finish. The claim is explicitly not proven.
The transcript predates evidence-bound transitions — old runs admit they carry no proof rather than borrowing any.
Check rows
check ✓ / check ✗ rows are the turn contracts firing: every agent input and output is validated (empty input, malformed output, and model refusals all fail). A ✗ is a recorded failure — and a wipe. Only an explicit pass renders as one.
Command rows and test counts
Command rows show the actual command, its exit code, and — for test runs — how many tests executed. That count is load-bearing: a run with exit 0 but 0 tests ran is not green, and the report flags it. Output (stdout/stderr) is expandable under each row.
Everything expands
- Model-call rows open to the verbatim prompt and response, with token counts.
- Agent rows show the coding agent’s duration and its real diff.
- The stage ladder across the top tracks progress left to right — click any reached stage to jump to its evidence.