What capture stores

[capture] in the sweep file turns on OpenTelemetry capture. Two of its four fields decide what ends up in the store.

content is false by default, and under that default prompts and completions never leave the process. An attribute whose key is gen_ai.input.messages or gen_ai.output.messages, or which is one of gen_ai.prompt, gen_ai.completion, traceloop.entity.input, traceloop.entity.output, llm.input_messages, llm.output_messages or anything under them, is dropped in the child process, before the span is serialised — so it reaches neither the pipe nor the Parquet file. content = true keeps all of it.

Tool arguments and results are kept whichever way content is set. gen_ai.tool.call.arguments and gen_ai.tool.call.result are what the agent did rather than what the model was told, and a run that cannot see them cannot tell a wrong tool call from a right one. If they carry data you will not store, do not instrument that span.

One gap, named rather than left to be discovered: input.value and output.value on an llm span are prompt content under OpenInference’s convention, and the default keeps them. The strip runs on the attribute key alone, before a span’s kind is known, so closing that is a later change.

max_observation_bytes is required, and bounds one observation value. Every [capture] table declares it and the tool supplies no number of its own: a wrong number fails silently, where a missing declaration fails loudly, at parse, before the run exists. A value whose byte form is larger than the declared limit is stored as its leading limit - limit // 4 bytes, its trailing limit // 4 bytes and its original byte length — together never more than the limit, at any limit. Lowering it is the supported way to measure how large observations really are without storing them: every value’s true length is still recorded, though the values themselves are not kept.

max_observation_bytes = "none" keeps every value whole. It is the one spelling of no bound, and it is a declaration like any other: the manifest records none, and report, the receipt, the viewer page and inspect read it back as none — which is a different fact from a run that recorded no limit at all, and prints in different words.

A sweep file with no [capture] table declares nothing and is not refused. It did not ask for capture, and what reaches the store on that path is a rollout’s own record() trace, which is stored whole. Nothing is backfilled: a run stored before this change keeps what it recorded, under the spelling it recorded it in.

Migrating a file written before this change: add max_observation_bytes = 8192. That is exactly what such a file stored, so the run reproduces — and it is what --resume of a run begun before the change requires, since a resume whose declaration differs from what the run recorded is refused rather than mixed into one run directory.

A cut is not silent. poisson-prove inspect prints an observation: cut (...) line above the span lines of any row whose trace has a cut span, naming each cut span by the index it prints under, its values’ original size in bytes, and the limit the run recorded. report and serve print, for each run, the byte limit it recorded, how many spans its stored traces hold and how many of them were stamped as cut, in the provenance list of report.md, receipt.html and the viewer page. The count is over spans, read from the stored envelopes and nothing else, and a span with two cut values counts once. A run written before the per-value cut recorded its limit under another name and stamped nothing, so its counts read not counted and never 0; a run with no capture record reads not recorded for the limit.

A sweep file may declare what the run expects. A [declare] table names the steps the target is expected to execute (steps), the outcome labels it can emit (outcomes) and which of those steps are gates (gates, a subset of steps). Any of the three may be absent; all three empty is refused, and so is a duplicate — a repeated name in a hand-written list is a typo, and collapsing it would be guessing at what was meant. The declaration is recorded on the run’s manifest as declared. It is not part of the sweep’s identity, so a declaring arm and a non-declaring arm derive the same seeds and still pair under diff. On the way to the store, a span whose step_name is a declared gate and which carries no gate of its own is stamped gate = "declared"; a span the target itself marked "emitted" keeps that.

What never reached means, and what it does not. For a run that declared, report, the receipt, serve and inspect print, against the declaration: the declared steps, the declared steps no span in this run carried (never reached), the observed step names the declaration does not name, the declared outcomes, the observed outcomes it does not name, the declared gates, the declared gates with no span (gates never reached), and the gates the target emitted that the declaration does not name. A declared step with no span means exactly that: this run’s spans did not show it. It does not mean the target cannot reach the step; it does not mean the step was not instrumented; it does not mean anything is wrong — a sweep whose configs never take the verify branch reports never reached: verify correctly, and that is the finding; and it says nothing at all about a run with no declaration, for which the lines do not exist and the phrase appears nowhere. A declaration that matches nothing, as on a run whose capture was off for the whole run, prints every declared step and is not refused. diff refuses two runs that declare different, non-empty outcome sets, since two competing sets are two quantities wearing one name; --resume refuses a sweep whose declaration differs from what the run recorded, while --pair-with does not.

A misspelled table name is ignored. The sweep file refuses an unknown field inside [declare], as it does inside [capture] and [chaos], but not an unknown top-level table: a [declair] table is read by nothing and refused by nothing, exactly as it was before this change. A run that prints no declaration block declared nothing as far as the tool can see; check the spelling.