Loki Mode

Added

• Hybrid codebase search: loki code search "<query>". Combines ripgrep keyword
  matching with ChromaDB semantic search via reciprocal-rank fusion (deduped by
  file:line), with token-budget-aware output. Flags: --grep-only, --semantic-only,
  --budget N, --top N, --json. When ChromaDB is unavailable it degrades to a
  grep-only result instead of failing, so search always returns something.
• Incremental code-index freshness. tools/index-codebase.py --changed
  re-indexes only files that changed (mtime or content hash) and deletes orphan
  chunks left by renamed or removed functions, fixing stale hits after refactors.
  A manifest at .loki/state/code-index-manifest.json tracks per-file state. The
  MCP code-search tools now report stale / stale_files; opt-in
  LOKI_CODE_INDEX_AUTOREINDEX=1 reindexes on demand (default is warn-if-stale).
• Dynamic resource-aware session concurrency. Opt-in LOKI_DYNAMIC_CONCURRENCY=1
  makes the parallel-session cap adapt to machine load: it reads
  .loki/state/resources.json and scales the cap down under high CPU or memory,
  always clamped to [1, ceiling], never above the configured maximum. Knobs:
  LOKI_MAX_PARALLEL_SESSIONS_CEILING, LOKI_CONCURRENCY_CPU_THRESHOLD (85),
  LOKI_CONCURRENCY_MEM_THRESHOLD (85), LOKI_CONCURRENCY_CRITICAL_THRESHOLD
  (95). Default off: behavior is identical to before unless explicitly enabled. A
  supervisor/judge pattern (CONTINUE / COMPLETE / ESCALATE / PIVOT) is documented
  in skills/parallel-workflows.md with an honest scaling ceiling (dozens of
  adaptive concurrent sessions on a local machine, not thousands).

Fixed

• CI reliability: the CLI command test (tests/test-cli-commands.sh) no longer
  false-fails on a loaded runner. The pattern check piped output into grep -q,
  which exits on first match and kills the upstream echo with SIGPIPE ("write
  error: Broken pipe"); on a slow CI cell that broken-pipe exit was misread as a
  no-match. It now uses a race-free in-shell case-insensitive match. This is what
  turned the v7.22.0 Tests workflow red on one matrix cell (the product artifacts
  were correct; only the test harness raced).