zuar-portal-mcp: Command for Claude Code

Run ONE bounded portal-improvement pass. Writes are gated and provable — this is the loop the safety primitives were built for.

Confirm a portal is connected and writable: check_connection. If connected:false or read-only, relay the reason + fix and stop — an improvement pass with no write access is a no-op.

Parse $ARGUMENTS for limit (blocks per pass, default 5), min_score (only blocks below this are touched, default 90), tier (fast/standard/max, default standard). State the scope before starting.

Run the .claude/workflows/portal-improve.js Workflow with those args. What it does — and what you verify in its report:

1. Baseline — snapshot_portal (the rollback point), score_portal update_baseline=true (pins the pre-state), worst-N selection with named deductions.
2. Improve — one portal-block-debugger agent per block, fixing ONLY the named deductions: reads first, writes with expected_updated_at, content tools only, scratch named TMP · …, product decisions reported instead of guessed.
3. Verify — the read-only portal-block-adversary re-checks each changed block; a regression verdict blocks the baseline update and triggers restore_resource from the snapshot.
4. Close — re-score (the delta vs baseline is the proof), cleanup_scratch confirm=true, baseline locked in only when nothing regressed and the delta is non-negative.

Report to the user: baseline → final score, per-block improvements, any regressions (and that they were restored), anything flagged needs_human. The pass is designed to be safe to schedule nightly (/schedule): bounded blast radius, snapshot-first, provable delta, self-cleaning.

Scores measure the mechanically checkable — never "improve" a block's design intent from a score alone. Design work goes through /portal-build's full pipeline.