feat(intel): always-on code intelligence for every repository (universal gate, prebuild priority, reconciler, full-set rebuild)
- MUSEHUB_INTEL_ALL_REPOS=1 (exactly) makes every repository eligible in a deployment whose identity is known and matches MUSE_ENV; local still needs a local DB. Off by default; never a wildcard in the allowlist. Persisted for hosted environments in deploy/secrets.sh. - Prebuild priority in the shared scheduler: non-intel work and prebuilds are claimed before intel; a running intel job yields to a waiting prebuild (bounded: at most MAX_INTEL_YIELDS, not an attempt); an aged intel job gets one turn per slice so it cannot starve. - Reconciler (event-driven path stays primary): enqueues exactly the missing/stale/skipped/retry-eligible providers of the registry per repository, bounded (backlog, per-pass cap, resumable cursor, bounded retries with back-off, dead-index holds dependents, exhausted pending jobs failed). Also the historical backfill. - The rebuild route enqueues the index AND every provider; provider state gains queued and not_applicable; aggregate fleet endpoint /api/intel/fleet. - Full symbol history load shares repeated strings (memory); hosted child limits sized 1024 MiB / 1800 s (a cold 1.7k-commit index died at 768 MiB on production). Tests: tests/test_intel_universal.py (56); updated rebuild-route, serial-barrier and health-status tests for the deliberate semantic changes. Focused regression: 840+ intel/job tests pass.
Semantic Changes
151 symbols
0 comments
muse hub commit comment sha256:545266da3f09ffbfe65a21bfc233bb91fe37f83ab714dd2d07db1fee98c3988b --body "your comment"
No comments yet. Be the first to start the discussion.