v0.2.0 — Parallel fanout, effort control, capable tier

Parallel multi-agent fan-out, first-class reasoning-effort control, and a capability-vs-price tier model. No breaking changes.

Added

• fanout — run N delegations in parallel across any provider (--agents DIR, --tasks FILE, or inline -- "t1" "t2"), concurrency-capped with --max, then fanout status|wait|collect|list. The supported way to run a multi-agent gauntlet (torture-room, review sweeps) through the Outsourcerer: N fast headless jobs, not 16 interactive sessions.
• --effort minimal|low|medium|high|xhigh|max (alias --reasoning) — universal reasoning-effort control. Native on codex lanes (model_reasoning_effort) and Claude lanes (MAX_THINKING_TOKENS), advisory on Gemini. Never silently dropped; forwards through bg and fanout.

Changed

• Capability tiers decoupled from price — new capable tier for GLM-5.2 / Hy3 / DeepSeek-v4-pro (~Opus-4.8 class): the thin, high-autonomy prompt scaffold and generous stall windows of a frontier model, not the strict budget "worker-drone" wrapper. Cheap is not dumb.

Fixed

• OpenRouter alias resolution on the cc/codex lanes (-m glm → z-ai/glm-5.2; was a 400).
• Cross-lane self-heal: a codex→OpenRouter tool-type 400 auto-reroutes to the cc lane.
• Job classification honors the last OSRC:: marker (no more false blocked on completed agents).
• Portable run-or-model.sh PATH (no pinned Node version).

Install: /plugin marketplace add alexgreensh/outsourcerer then /plugin install outsourcerer@outsourcerer · update with /plugin update.

v0.1.0 - Delegate the grind, keep the glory

First public release. Outsourcerer lets your main AI agent delegate grunt work to the cheapest engine you already pay for, pull a panel of stronger models in as advisors, mirror your whole setup onto whichever engine runs the job, and keep a real cost receipt. Cross-harness: Claude Code, Codex, Antigravity, Devin.

Highlights

• Delegation lanes: OpenRouter (via Claude Code or Codex), native subscription lanes (Sol/Terra/Luna on ChatGPT, Fable/Opus/Sonnet/Haiku on Claude), keyless Gemini/Antigravity, plus GPT-image.
• Interactive sessions (tmux) for every provider: the mode that brings the full workshop (tool exec + skills + MCP) and is watchable. Headless one-shot stays for small tasks.
• Verified claude-native lane: reads the model actually used and prints a receipt, so it can never mislabel (e.g. an Opus run as Fable).
• Self-healing codex lane (auto-handles a missing code_mode_host binary) and per-lane doctor preflight.
• Real cost Tab: per-generation OpenRouter cost read back from the provider; honest cash vs plan-limit columns.
• Live model discovery (suggest): the cheap and free models available on each platform right now.
• Advisor panels with consensus-gated second opinions; per-model prompting.

Security: repo-forensics audited, 0 critical / 18 high / 1 medium, all triaged (see SECURITY.md).

Install (Claude Code): /plugin marketplace add alexgreensh/outsourcerer then /plugin install outsourcerer@outsourcerer