Skip to content

161. Repo config limits what a run artifact publishes ​

Date: 2026-11-06

Amends: ADR 0021

Status ​

Accepted

Context ​

Every agent run uploads its whole output/ directory as the run artifact: fullsend-<agent> on GitHub Actions, and job artifacts on GitLab CI. Each run directory holds one iteration-<N>/ directory per iteration, with the raw conversation stream (output.jsonl), transcripts/ and the agent's output files. Next to them sit run-level files: metrics.json, run-telemetry.jsonl, eval-measurements.jsonl and security/findings.jsonl. The telemetry file carries conversation content when content capture is on (ADR 0050). Nothing skips the upload or narrows what it holds.

ADR 0021 keeps transcripts by default and relies on owner-scoped access. It lets an agent's own YAML opt out when the agent reads protected sources beyond the target repo. That opt-out was never built. Two facts change where the switch should live:

  • On a public repo, any signed-in user can download the artifact until it expires, so access is not owner-scoped.
  • Whether a run reads private data depends on the repo's sources and providers, not on the agent. A GitHub-triggered run with a read-only Jira provider attached can read private tracker text. A Jira-triggered run always does.

The output scan masks secret literals. It does not remove personal data, customer names or private business text (#9122).

This ADR aims to let a repo owner stop publishing conversation content in the run artifact for runs that read private data.

Decision ​

We add an artifacts setting to config.yaml that limits what a run artifact publishes, so a public repo can use private sources without publishing what its agents read.

yaml
artifacts:
  content: all          # all (default) | metadata | none
  by_source:
    jira: metadata      # key: the run's event.source.system
  by_provider:
    jira-ro: metadata   # key: an attached provider's name (example name)
LevelArtifact holds
allEverything in output/, as today. This is the default.
metadataOnly the run-level files: metrics.json, eval-measurements.jsonl, security/findings.jsonl, and run-telemetry.jsonl with conversation content removed. Every iteration-<N>/ directory and any other file is left out.
noneNo artifact is uploaded.
  • It lives in config.yaml, not in agent YAML. The setting applies the same way to every agent, so ADR 0080 puts it in config.yaml. This replaces ADR 0021's per-agent opt-out.
  • The strictest level wins. The runner picks one level per run: the strictest of content, the by_source entry for the run's event.source.system, and the by_provider entry for each attached provider. by_provider keys match a provider's configured name. When the source is missing or not listed, only content and the provider entries count. none is stricter than metadata, which is stricter than all.
  • Keys are open strings. A new source or provider name works as a key without a schema change.
  • Layers merge per key. Across config.base.yaml and config.yaml, a set content overrides, and by_source and by_provider merge per key. Lowering an effective level toward all is a relaxation under the ADR 0122 safety gate, so it needs an explicit declaration.
  • Metadata is an allowlist. metadata copies only the files named above. A file added to output/ later stays out until this list names it.
  • Filtering happens after local use, and fails closed. The runner leaves output/ untouched, so fullsend eval-measure and post-scripts read the full run. A step after them builds a staged copy, and only that copy is uploaded. If the level cannot be resolved, the staged copy is empty and the raw output/ is never uploaded.
  • Content capture is unchanged. Exporting over OTLP to a backend the repo owner controls still works under every level, as in ADR 0050.

The rest of ADR 0021 stands: full transcripts by default, and credential scanning as an invariant check. Artifact retention is out of scope.

Forge scope ​

  • GitHub: the existing upload step uploads the staged copy, and is skipped for none.
  • GitLab: the job artifact path points at the staged copy.
  • Forgejo: no run workflow yet; it uses the same staged copy when one lands.

All forges read the same config.yaml setting.

Consequences ​

  • A repo can withhold conversation content for Jira-triggered runs or runs with a private-data provider attached, and keep all elsewhere.
  • The retro agent, session resumption and follow-up workflows that download the artifact get no transcripts under metadata and no artifact under none, which the user guide must document.
  • When the setting is unset, the artifact matches today's.
  • Changing the level is a config.yaml change, so it goes through the repo's normal review and CODEOWNERS.