Module run_log 

Per-run NDJSON output capture. Agent and aftercare stdout/stderr are untrusted evidence: they are stored as ordered chunks, never parsed as lines and never routed through the dispatcher.

Structs

AgentOutput

OutputPage

OutputRecord

PageQuery

One read of the captured log: everything after a cursor, optionally narrowed to a stage and trimmed to a trailing window.

RunLogWriter

Append-only writer shared by the stdout and stderr reader threads of one run. Sequence numbers are capture order across both pipes.

StageFilter

Selects the records belonging to one flow stage. Aftercare records carry their stage name; agent records captured before stages were tagged carry none, so agent_fallback lets the flow’s agent stage claim them.

Enums

OutputChunk

One captured chunk. UTF-8 chunks stay readable; anything else round-trips through base64 rather than being lossily converted.

OutputSource

OutputStream

Functions

read_agent_output

Reassembles each agent stream from every complete captured record. Test and merge output is deliberately excluded from vendor classification.

read_filtered_page

Reads one page under a filter. next_cursor advances past every record the read consumed, not just the ones it returned: a filtered-out record is still evidence that has been examined, so a follower resuming from the cursor neither replays it nor stalls behind it.

read_page

Reads a finite page of records with sequence > after, in order. A missing file is an empty page: the run may exist before any output does.

visit_agent_output

Visits decoded agent chunks in capture order without retaining the whole log. A malformed or crash-truncated record does not hide earlier evidence.