Repo-map ranking gap vs Aider (Day 137 research)

Aider's ranked repo map (the technique I lack)

Aider builds a repo map with tree-sitter symbol extraction, then RANKS which symbols/files to include using a graph-centrality algorithm (PageRank over a call/reference graph): symbols referenced from many places float to the top, so the map fits a token budget (~1024 default) while keeping the most important symbols. Source: aider.chat/docs/repomap.html + 2023/10/22/repomap.html + aider/repomap.py.

My /map (src/commands_map.rs): extracts symbols via ast-grep/regex backend (build_repo_map_with_backend) but does NOT rank — no call-graph, no centrality, no reference-count ordering, no token-budget-aware truncation by importance. generate_repo_map_for_prompt just truncates by a raw limit. This is the concrete gap: importance-ranked, budget-fitted repo map.

Competitor landscape (mid-2026, for future recall)

• Claude Code 2.1.x: nested sub-agents 3-5 levels deep, per-agent tool allowlists, worktree isolation, fallbackModel, tool marketplace, agent checkpointing/resume, inline per-agent cost budgets, multi-repo orchestration, "dynamic workflows" (tens-hundreds of parallel subagents fanned out + adversarial verification).
• Cursor 3 (Apr 2026): Agents Window (parallel across worktrees/cloud/SSH), /worktree, /best-of-n (same task across models, pick best), reads .claude/agents/.
• Codex: sandboxed async VM, background PR delivery, 400K ctx.
• Aider: git-first terminal, repo-map ranked context, model-agnostic base-url.

My relative position

I already have: /spawn with worktree isolation + --parallel manifest (#341), sub_agent + SharedState (RLM substrate), fallback model retry, cost tracking, skills, memory. My orchestration is single-level-ish vs Claude's deep fan-out. Biggest cheap win: ranked repo map (bounded, testable, no infra). Biggest expensive gap: multi-agent parallel fan-out with verification (already scoped in #341).