legion-core

Legion Core — the model-agnostic orchestration engine under the hood of an AI agent: models, harnesses, loops, self-improving agents. The stack, not another wrapper. Ships scoped delegation, multi-model fan-out, telemetry/cost reporting, doctor/self-learn/heal loops, and cross-harness setup for Claude Code, Codex, Cursor, and opencode. One orchestrator, a legion of models.

Legion's multi-model brain: delegate scoped tasks from Opus to external model agents (Codex / GPT-5.5 / GPT-5.4 and Cursor Agent) via `legion-delegate` and `legion-cursor`, bringing back verified, metered diffs. Also ships generic GitHub issue intake via `legion-intake`, which can run delegate, cursor, or a compatible custom worker. Runs agents in isolated git worktrees, captures diff + last message + token usage, prices via a shared cost table, emits legion.span.v1 telemetry, and reports GPT usage to the router /ingest sink so spend shows next to Claude. Ships the opt-in `legion-router` loopback :8082 Anthropic-compatible metering proxy, codex-json parser, and per-model cost lib.

See everything Legion does and close the self-learning loop. Unified legion.span.v1 telemetry, legion-report cost/success/latency, legion-bench deterministic core/stable harness benchmarks plus A/B corpus runs with repeated-run flake detection, legion-catalog marketplace inventory, legion-self-learn daily failure mining + command/agent/skill improvement memory/proposals, legion-trace, legion-otel-export, and legion-doctor install verification.

Legion's dynamic MULTI-MODEL orchestrator — the ultracode loop (decompose -> fan out -> cross-model verify -> synthesize -> gate) made executor-aware: Opus conducts, GPT-5.4 does the bulk of coding in PARALLEL, GPT-5.5 verifies. Ships `legion-fanout` (run scoped slices in parallel across executors via legion-delegate, collect verified diffs + cost) + a SKILL playbook with an LEGION_ULTRACODE max-exhaustive mode (wide fan-out, multi-vote cross-model verify, loop-until-dry).

One skill to install AND update the Legion marketplace — no team hand-holding. First run installs marketplace plugins, cross-harness skills, Codex and Cursor bridges, shared CLIs, and daily refresh/self-learning; after that the same skill updates to latest. Ships `legion-setup`, `legion-codex-setup`, and `legion-cursor-setup`.

The Codex-CLI-primary routing brain for Legion — the mirror of legion-router. When you run Legion under Codex you (GPT-5.5) do most of the work; this skill tells you WHEN to hand a task up to Claude via `legion-claude` (deep architecture, polished/complex frontend with the impeccable skill, final cross-model review, tie-breaks, when stuck) versus inline — with automatic GPT-5.5 fallback if the Claude limit is hit. Also maps what runs natively on Codex (MCPs, mirrored skills, bridged commands/agents) so you don't reach for Claude needlessly. Triggers on running Legion under Codex, deep-design/frontend-polish/final-review smells, or 'ask claude' / 'second opinion from claude'.

Optional Legion code-intelligence diagnostics. Runs repo-native TypeScript and Pyright checks, filters changed-file diagnostics, emits legion.code-intel.v1 artifacts, and can append legion.span.v1 telemetry for benchmarkable pre-PR gates without making LSP servers a core dependency.