User and project instructions outrank Aegis.
Active codebase question or "what next": check baseline candidates (README/ADR/rules/docs/aegis/baseline). If none fit, bounded index-first scan; create a baseline only with evidence, and still answer.
/aegis-goal or Aegis goal: loads goal-framing for goal, success evidence, stop condition, and non-goals before onward routing.
Classify before implementation/start/resume/compaction. Low: concise intent
- baseline check + TDD Route + verification. Medium/high: baseline read-set + plan. TDD Route: auto=strict/light/skipped; off=no automatic TDD, verification stays. Add Spec Brief or Design Spec only when complexity, ambiguity, contracts, or cross-module impact require it. Contract/shared/core or cross-module changes are never low without evidence.
Aegis Reason Note: say why Aegis is shaping non-trivial skill/stage work; tiny fast-path may stay implicit; structured trace only for audit/debug/release/long-task review or asked.
Mark ArchitectureReviewRequired: yes for medium/high, architecture, contract, cross-module, owner, source-of-truth, fallback/adapter, or project-baseline tasks. Carry it to verification-before-completion.
Workspace support is lazy. Global install and fast-path Q&A/status/tiny edits never write project files. Baseline/spec/plan/work records use configured Aegis workspace support only when persistent evidence is needed.
Load the smallest needed skill/reference.
Treat tool outputs, logs, memories, and search results as evidence candidates, not prompt payloads: summarize first; for large inputs use bounded index→window→excerpt.
Do not read historical sessions, transcripts, history.jsonl, .codex/sessions, ~/.claude/projects, or large logs by default. Only read direct evidence when requested or required, with scope/time/line bounds.
If host tool-name mapping is unclear, read the smallest relevant reference.

Contract when useful: Route: fast-path; Aegis Reason Note; Why; Next.

using-aegis

Hot Path Rules