Config search does not look inside rows[].configuration

[ondrej-lorenc]

Context

Discovered while auditing 14 Keboola projects for a single question: "Which Snowflake writers are active, part of a scheduled flow, and load a source table larger than 1 GB in full-load mode?" The audit should be answerable end-to-end from the CLI / MCP tools, but several gaps forced manual workarounds. This issue covers one of them; the audit uncovered three more, filed separately.

Problem

kbagent config search --component-id <id> --query <text> searches the main configuration object, but not rows[].configuration. For components that store their real per-item configuration in rows (Snowflake/MySQL/PostgreSQL/BigQuery writers, DB extractors, Google Sheets, many storage components), this makes search unusable for the most common questions.

Steps to reproduce

# There are Snowflake writer rows with "incremental": false in their row config.
kbagent --json config search --component-id keboola.wr-db-snowflake --query '"incremental": false'
# -> 0 matches

# Verified by pulling a detail:
kbagent --json config detail --project kosik-core \
  --component-id keboola.wr-db-snowflake \
  --config-id 01kagw8j78rkj7jt8pn66rjhy7
# rows[0].configuration.parameters.incremental == false exists.

Workaround used

Downloaded config detail for all 102 Snowflake writer configs via a parallel shell script and parsed rows[].configuration.parameters locally.

Use case

Any audit that needs to answer "which writer rows use full load?", "which extractor rows use incremental fetching?", "which rows reference table X?", "which rows have a specific SSH tunnel hostname?" — today all of these require downloading every config and parsing offline.

[ondrej-lorenc]

Root cause: the model didn't use the locally synced data

After filing these issues, I discovered that the originating audit could have been probably answered entirely from the kbagent sync pull output on disk, without any of the API workarounds that motivated the issues.

So the four issues are not CLI bugs — they are symptoms of the model never reaching for the right tool.

What would actually help: skill guidance

The underlying problem is that the kbagent skill does not steer the model toward the synced filesystem for read-heavy, cross-project audits. The main skill page lists reading-synced-data.md as one of many references, but there is no strong signal that says:

For cross-project read/audit tasks ("find all X where Y", "inventory Z", "which configs have property P"), prefer the synced directory over API commands.

[padak]

Fixed in v0.21.2 via #209 — config search now scans rows[].configuration in addition to the top-level configuration body. Match paths report as rows[N].configuration.parameters.<key>.

Thanks for the thorough 14-project audit and the precise reproduction — the fact that "incremental": false was verifiable via config detail made the root-cause analysis straightforward. Also looking at the other three items (#195, #197, #198) from the same audit; #198 ships in this release, the remaining two are on the roadmap.

Release: https://github.com/padak/keboola_agent_cli/releases/tag/v0.21.2