vibesoc loop

Turn the request into one outcome-first packet with the fewest useful tool loops.

1. Resolve root and workspace, then run python3 <root>/.agents/scripts/loop_context.py <workspace> --format text. Request --mode merge only for delivery. Never lower returned mode or scope.
2. Treat paths, owner, rules, required_reads, checks, budget, cache, and next action as authoritative. Read listed rules and required_reads paths before owned edits (RTL full coding style injects this way). Use compact state; load full state/logs only on failure.
3. Dispatch one matching owner in dev. Use soc-pipeline only for multi-stage delivery. Parallelize only independent work — normally verif+syn once RTL evidence exists — within max_parallel_owners.
4. Before listed resource-heavy checks, perform one preflight for registered capability, required inputs, and tool/license availability. Record an unavailable lane once, skip only that blocked check, and continue independent available work. Never blind-retry a missing license.
5. Reuse every fresh stage. Keep full logs in artifact paths; return compact evidence: check status, run ID, source fingerprint, artifact paths, short failure tail. A changed fingerprint invalidates reuse.
6. Classify a failure before retrying. Respect the same-failure retry limit and one review run per unchanged snapshot. After the limit, stop with the blocker, evidence path, and concrete remediation instead of looping.
7. In dev, leave the owned stage open. In merge/signoff, close only stale stages, run the mapped reviewer once, then re-run router readiness.

Do not hand-write generated tops/CRG logic, bypass registered EDA tools, add a review/signoff pipeline stage, or narrate routine tool calls.