Ocean proposal + Leo field experience (2026-06-10 salience recalibration, 10k-row exports).
Add a result sink: `kkernel exec '' --save-file out.jsonl` and equivalently a `save_to="path"` arg on the MCP `request` tool. Behavior: results stream to the file; the caller gets back only `{path, rows, per-column null counts, schema fingerprint, checksum}`.
Why the manifest is the load-bearing part: today's null-content incident — an exported file silently had `content=null` for all 10,043 rows and a 5-agent judgment fleet graded blind before TBV caught it. A sink that self-reports "content: 10043 nulls" is caught in one second. Bulk data handoff between agents needs files that describe themselves.
This also collapses most of the CLI-vs-MCP split: agents get bulk exports via MCP without context cost, no presentation-flag proliferation.
Ocean proposal + Leo field experience (2026-06-10 salience recalibration, 10k-row exports).
Add a result sink: `kkernel exec '' --save-file out.jsonl` and equivalently a `save_to="path"` arg on the MCP `request` tool. Behavior: results stream to the file; the caller gets back only `{path, rows, per-column null counts, schema fingerprint, checksum}`.
Why the manifest is the load-bearing part: today's null-content incident — an exported file silently had `content=null` for all 10,043 rows and a 5-agent judgment fleet graded blind before TBV caught it. A sink that self-reports "content: 10043 nulls" is caught in one second. Bulk data handoff between agents needs files that describe themselves.
This also collapses most of the CLI-vs-MCP split: agents get bulk exports via MCP without context cost, no presentation-flag proliferation.