Configuration¶

Every field has a working default. Nothing needs to be set for the package to work.

Precedence¶

Defaults, then environment variables, then keyword arguments to start().

aftersight.start(keep_runs=200, quiet=True)

export AFTERSIGHT_KEEP_RUNS=200
export AFTERSIGHT_QUIET=1

Fields¶

Field	Environment variable	Default
`enabled`	`AFTERSIGHT`	`1`	`0` disables everything. No folders are created and every call is inert.
`root`	`AFTERSIGHT_ROOT`	`.runs`	Where runs are written.
`keep_runs`	`AFTERSIGHT_KEEP_RUNS`	`50`	Oldest run folders are pruned past this at the start of a run. `0` keeps everything.
`blob_max_bytes`	`AFTERSIGHT_BLOB_MAX`	`8192`	Payloads larger than this move to `blobs/`.
`blob_preview_lines`		`12`	Lines of a spilled payload kept inline in the transcript.
`capture_otel`	`AFTERSIGHT_OTEL`	`1`	Install the OpenTelemetry span processor.
`capture_logging`	`AFTERSIGHT_LOGGING`	`1`	Install the stdlib `logging` bridge.
`log_level`		`WARNING`	Threshold for the logging bridge.
`redact`	`AFTERSIGHT_REDACT`	`1`	Scrub credentials before writing.
`quiet`	`AFTERSIGHT_QUIET`	`0`	Suppress the one-line startup notice.

Boolean environment variables accept 1/0, true/false, yes/no, on/off.

Turning it off¶

AFTERSIGHT=0 is a hard off switch. start() returns an inert object, no directory is created, and span(), trace and log() do nothing. This is the intended way to run the same code in an environment where you do not want traces on disk.

Bounding disk use¶

Two knobs. keep_runs bounds how many run folders survive, and blob_max_bytes decides when a large payload stops being inlined into the transcript. Neither is a hard size cap, because nothing is ever truncated. If you record a run with a 40 MB prompt, that 40 MB is on disk.

Metadata versus configuration¶

Keyword arguments to start() that are not configuration fields are recorded as run metadata:

aftersight.start(
    keep_runs=200,               # configuration
    framework="langgraph",       # metadata, ends up in meta.json and outline.md
    model="claude-sonnet-5",
)

framework, framework_version, model and task are rendered specially in outline.md and index.jsonl. Any other key is still recorded, just not formatted.

Redaction¶

Applied at the sink, so it covers every writer at once. It matches sk- prefixed keys, bearer tokens, api_key / secret / token / password assignments, GitHub tokens and AWS access key ids, replacing the value with [REDACTED].

It is a filter, not a guarantee. Treat run folders as containing whatever your prompts contained.