#833 docs: bundled swamp agent skill lacks datastore-namespace guidance (giga-swamp)

Summary

The giga-swamp datastore-namespace feature and its user-facing web docs shipped (#542, #547, #548, #551), but the bundled swamp agent skill (the SKILL.md reference tree the CLI ships, which drives coding agents) was never updated to cover datastore namespaces. An agent operating purely from the skill cannot discover or safely run the feature — in a real session I had to be prompted from the web manual to learn swamp datastore namespace.

Evidence

Packaged skill, swamp 20260626.005336.0-sha.f302d18e: "namespace" appears 27x but exclusively in unrelated senses — the data.*/env/webhook CEL accessors, HashiCorp Vault namespaces (enterprise vault config), the @swamp/* package namespace, the TS namespace-export lint rule, and test keyPrefixes. Zero mentions of datastore namespaces. The skill documents datastore setup/sync and S3/GCS as storage backends (references/repo, ci-integration, troubleshooting/tracing, extension/datastore authoring) but nothing on the namespace/multi-repo dimension.

What the skill should add (all shipped CLI surface)

• swamp datastore namespace set | migrate | list | unset command surface, and the set -> migrate (preview) -> migrate --confirm -> sync --push flow.
• Multi-repo shared datastores: same bucket+prefix + a DISTINCT namespace per repo; namespaces are repoId-owned via a {namespace}/.namespace.json manifest with check-then-write conflict detection.
• CEL point lookups (data.latest/data.version) implicitly scope to the current namespace; cross-namespace reads go through the foreign-catalog methods (exportCatalog/pullForeignCatalogs/fetchForeignContent).
• The footgun: two repos sharing one prefix UN-namespaced both write <prefix>/.datastore-index.json and clobber each other — so namespacing is required, not optional, before a second repo joins.
• A pointer from the repo/datastore guide and the SKILL routing table so the guidance is actually discovered.

Why it matters

Multi-repo / shared-datastore splits (e.g. separating infra provisioning and app deployment across two repos on one GCS datastore) are exactly where namespaces are load-bearing. Without skill coverage an agent either won't use the feature or hits the un-namespaced index-clobber footgun. The web docs (#551) exist, but in-repo agents are driven by the bundled skill, not the manual.

Environment

swamp 20260626.005336.0-sha.f302d18e; @swamp/gcs-datastore datastore.