Skip to content

163. Repo config limits what the live progress log shows ​

Date: 2026-11-09

Status ​

Accepted

Context ​

While an agent runs, fullsend run prints its progress to the CI job log. The log shows:

  • the model's reasoning and replies, in full;
  • one line per tool call: the tool name and a short summary, such as the first 120 characters of a command;
  • task and status progress, whose labels and details can carry text the agent wrote;
  • error messages from the agent's transcript, and the output of the validation script, which can quote the agent's output files;
  • fullsend's own lines: heartbeat, token counts, cost and the result.

On GitHub Actions, tool lines and progress lines also become notice annotations, and transcript errors become error annotations, on the run summary page.

Tool results are not printed, but reasoning and replies often quote them. On a public repo, anyone can read the job log and its annotations until log retention expires. The output scan masks secret literals only. It does not remove personal data, customer names or private business text (#9123).

Suppressing an agent's extracted transcripts (ADR 0021) and limiting the run artifact per run source (ADR 0161) do not cover the job log. A run started from an issue tracker that is not public still prints the issue's text, and anything the agent quotes from it, to a public job log.

This ADR aims to let a repo owner stop the live log from showing model- and tool-written text for runs started from a source that is not public.

Decision ​

We add a progress setting to config.yaml that limits what the live progress log shows for each run source, so a run's log on a public repo does not echo text from a private source.

yaml
progress:
  verbosity: full       # every run: full (default) | tools | quiet
  by_source:
    jira: quiet         # example: runs started from a Jira issue
LevelThe log shows
fullEverything, as today. This is the default.
toolsTool lines, progress, token counts and the result. Reasoning and replies are left out. Tool lines and task labels can still carry text from the source, so use quiet for a source that is not public.
quietOnly lines fullsend writes from its own fields: heartbeat, event kinds and states, iteration and token counts, cost, error types, validation passed or failed, and the result. Leaves out reasoning, replies, tool lines, task labels, status details, error message text, validation output, and the annotations that carry them.
  • by_source keys name where the run started. Each key is the value Common Expression Language (CEL) triggers see as event.source.system, such as github, gitlab, jira, manual or schedule. Keys are open strings, so an adapter for another tracker adds its source name with no schema change.
  • by_source looks only at where the run started, not at every private system the agent can read. A source level protects private data only when access to that data is attached per source, as in the shipped agents.
  • The stricter level wins. The runner uses the stricter of verbosity and the by_source entry for the run's source. A source with no entry gets verbosity. quiet is stricter than tools, which is stricter than full.
  • Layers merge per key. Across config.base.yaml and config.yaml, an explicit verbosity overrides, and by_source merges per key. Lowering an effective level toward full is a relaxation under the ADR 0122 safety gate, so it needs an explicit declaration. These rules match the artifacts setting in ADR 0161.
  • One setting for every runtime. Claude Code, Codex, OpenCode and pi runs print through the same renderer and runner loop, so the level applies to all of them.
  • Only the printed log changes. Telemetry, content capture (ADR 0050), validation feedback to the next iteration and the run artifact still receive everything.
  • It fails closed. If the level cannot be resolved, the run logs at quiet.
  • quiet keeps a stuck or failed run visible. The heartbeat, error types, the validation result and the result line still print.
  • Per-agent levels are out of scope. This ADR decides the repo-level setting only.
  • Provider-keyed levels (by_provider) are out of scope. For the shipped agents, a provider such as jira-ro is attached only on runs from the matching source, so by_source selects the same runs. A repo whose agent reads a private system on runs from every source should set the repo-wide verbosity to quiet.
  • Pre-scripts and post-scripts are out of scope. They are the repo owner's own code.

Forge scope ​

  • GitHub: the level applies to the job log and to its annotations.
  • GitLab: the level applies to the job log. GitLab has no annotations.
  • Forgejo: no run workflow yet; the same level applies when one lands.

All forges read the same config.yaml setting.

Consequences ​

  • A repo can keep text from a source that is not public out of its public job logs and keep full for other sources.
  • Under quiet, someone debugging a run sees no reasoning, tool lines or error text in the log; they need OpenTelemetry (OTLP) export to a backend the repo owner controls, or the run artifact if ADR 0161 allows it.
  • This ADR adds no private log for the hidden lines, and none is planned. To keep reasoning and tool lines privately, export them over OTLP to a backend the repo owner controls (ADR 0050).
  • When the setting is unset, the log is the same as today.
  • Changing the level is a config.yaml change, so it goes through the repo's normal review and CODEOWNERS.