Highlights
This release lands kid agents (sandboxed children in containers), Codex as a 7th LLM provider, the official Polymarket integration, full Pump.fun livestream stack (video + voice + chat + captions), the plan-review trio, an ego layer for evaluative self-image, swarm projects for cross-spawn continuation, and a complete onboarding rewrite.
Kid agents — sandboxed children in containers
Spawn disposable child EloPhanto instances inside hardened Docker containers to run dangerous shell commands (rm -rf, fork bombs, untrusted package installs) without touching the host. Hardening baked into core/kid_runtime.py at the API surface — no override knobs at the call site. --cap-drop=ALL, read-only rootfs, non-root uid 10001, no host bind-mounts (named volume only), no docker-socket, no --privileged. Vault scoping via new Vault.subset(keys) — default empty, kids get zero secrets unless explicitly granted. Registry filter strips kid_*/payment_*/wallet_*/polymarket_* from kid environments. Synchronous kid_exec request/response via gateway hook so kid responses don't pollute the parent's main conversation. Five tools (kid_spawn/kid_exec/kid_list/kid_status/kid_destroy) plus elophanto kid build / list / destroy admin CLI. See docs/66-KID-AGENTS.md.
Codex (ChatGPT Plus/Pro subscription) as 7th LLM provider
Use your existing ChatGPT subscription as an LLM backend via the Codex CLI's OAuth credentials (~/.codex/auth.json). Responses API, streaming, auto-refreshes tokens, per-model reasoning effort clamping. Auto-detects on startup — run codex login once and it's wired in. ⚠️ ToS grey area (ChatGPT sold as UI, not API). See docs/63-CODEX-PROVIDER.md.
Polymarket — prediction market trading
Installed the official Polymarket/agent-skills bundle. Skill-only (no native tool group, by design): py-clob-client reads orderbooks, streams WebSocket updates, places GTC/GTD/FOK/FAK orders on Polygon CLOB. All order placement gated behind owner approval. Vault stores polymarket_private_key / polymarket_funder_address. Reliability: auto-detects which proxy wallet (EOA / POLY_PROXY / GNOSIS_SAFE) holds collateral before placing orders, fetches tick_size/neg_risk per market, hard rule against browser fallback. See docs/64-POLYMARKET.md.
Pump.fun livestream — full multi-modal autonomous channel
End-to-end pump.fun streaming from chat. Auth signs frontend-api-v3.pump.fun/auth/login with the agent's existing Solana wallet. Publishing through pump.fun's WHIP/RTMP ingress on LiveKit Cloud via ffmpeg.
pump_livestream— start/stop/status, supports video file loopingpump_say— voice mode: queues lines, OpenAI TTS → PCM → named FIFO → ffmpeg, swaps video for static idle.pngpump_chat— posts to pump.fun's livechat Socket.IO server (wss://livechat.pump.fun)pump_caption— drawtext overlay (~33ms updates) when libfreetype is present, Pillow-bake fallback when not- ffmpeg supervisor — re-fetches credentials per retry, exponential backoff, IPv6→RTMP failover when LiveKit hands back unreachable v6 ICE candidates, distinct exit code on JWT expiry, bitrate caps + 720p downscale, stereo audio (WHIP rejects mono),
-loglevel warningto kill log bloat
See docs/65-PUMPFUN-LIVESTREAM.md.
Swarm projects — continuation across spawns
swarm_spawn(project=<slug>) reuses an existing worktree so the next spawn SEES and UPDATES prior code instead of starting from git init. Auto-derives a slug from the task if no project name is passed (so the agent never accidentally orphans a project). Local-only projects (repo='new') first-class — no remote ops on continuation. GitHub-backed projects fetch + reset to origin/main so updates start from latest merged code. Worktree-missing recovery rebuilds from scratch. New swarm_list_projects (planner calls before spawning) and swarm_archive_project (hides dead projects). Planner gets a <continuation_protocol> block forcing "list before new" on update/extend/fix requests.
Ego layer — evaluative self-image computed from outcomes
Identity is descriptive (who I claim to be); ego is evaluative (how reality has graded that claim). New EgoManager records per-capability outcomes after each task, applies asymmetric exponential smoothing (failures hit harder), maintains a humbling-events log capped at 5, and on every 25 outcomes recomputes a one-paragraph self-image + one-line self-critique via the LLM. The LLM only writes the prose; never the numbers. New <self_perception> block prepended to the system prompt so the planner sees both claim and measurement. learned/ego.md markdown mirror written automatically on every humbling event and recompute.
Plan-review trio + plan_autoplan pipeline
Three new SKILL.md files (plan-review-ceo, plan-review-eng, plan-review-design) plus a tool that runs them sequentially with auto-decisions and escalations.
- CEO: 4 modes (SCOPE EXPANSION / SELECTIVE / HOLD / REDUCTION), 6-dimension scoring rubric
- Eng: 6-dimension scoring + required additions (module map, schema diff, failure modes, rollback)
- Design: 6-dimension scoring + required wireframe / state table / accessibility floor (WCAG AA)
plan_autoplantool: threads each stage's revised plan into the next, aggregatesdecisions[]+escalations[], returnsready_to_implementso heartbeat / autonomous mind can detect "needs human input" without parsing free text
Six baked-in decision principles (ship-over-perfect, reversibility-wins, existing-pattern-wins, agent-leverage, user-facing-wins, escalate-irreversible-public) so reviewers don't contradict each other.
Onboarding — doctor preflight + provider-first init wizard
Major usability fix.
elophanto doctor— green/yellow/red preflight that catches placeholder API keys (YOUR_*,<TODO>), missing Chrome profile paths, uninitialised vault, missing bootstrap docs.start.shruns it before chat; override withSKIP_DOCTOR=1- Init wizard rewrite — Codex auto-detected from
~/.codex/auth.json; OpenRouter mandatory unless Codex is present; Z.ai / Kimi / OpenAI collapsed under one optional "Add more providers?" prompt. Models per task come fromconfig.demo.yamlverbatim — wizard no longer asks about model selection, fixing the "init defaulted to Claude Sonnet and blew per-session budget on first call" feedback - Placeholder detection — providers with
YOUR_*/<TODO>keys auto-disabled at config load elophanto bootstrap— regeneratesknowledge/system/{identity,capabilities,styleguide}.mdso the planner has grounding from day one
Three money-making skills
Aimed at the agent's actual edge (operating capability, not directional prediction):
indie-saas-shipper— validate willingness-to-pay BEFORE writing code, kill at day 14 if no first dollar, geography-neutral via Stripe/LemonSqueezysmart-contract-audit— per-phase checklist (reentrancy / oracle / access control / integer / token-specific / Solana-Anchor / economic-MEV / upgradeability), severity rubric, full report template, direct-to-protocol cold outreachcrypto-launch-ops— productize the pumpfun + X engagement + livestream stack as a paid service for crypto projects, with hard refusal rules on rug-prone projects
Deep-research skill
Seven-phase rubric: falsifiable claim, tiered sources, steel-manned dissent, cross-validation, confidence interval — replaces "I'll do a quick web search" with structured investigation. Iron rule: never state a conclusion without (a) the source tier, (b) the strongest counter-argument and what would refute it, (c) what you don't know.
AlphaScala broker skill
Documents the POST /api/match broker-matching API (3 scored picks + reasoning), URL templates for reviews / comparisons / curated lists / affiliate signup, plus stock research (Alpha Score), 13F clusters, insider activity, and TradingView indicators. Operator is ROGA AI — same company that runs EloPhanto — so the skill prefers AlphaScala over generic web search for trading-platform questions.
Reliability fixes
- Flaky scheduler tests — replaced
asyncio.sleep(0.05)synchronization with deterministicasyncio.Eventsignals so tests no longer race against slower CI runners - Deferred tool activation — registry rebuilds
_toolseach agent loop iteration so newly-discovered tools become callable immediately - Codex adapter parity — actually forwards tool schemas and parses tool calls (was broken)
- Dashboard mouse SGR leak — strips ANSI mouse-tracking bytes from Input field, disables mouse tracking on mount
replicate_generate— uses absolute paths + configured workspace, auto-registers outputs in knowledge
Stats
- 1243 tests passing (was 1053 in v2026.04.05, +190)
- 168+ built-in tools (was 163+)
- 170+ skills (added: deep-research, plan-review-{ceo,eng,design}, alphascala, pumpfun-livestream, indie-saas-shipper, smart-contract-audit, crypto-launch-ops, kid-agents)
- 66 docs (added: 63-CODEX-PROVIDER, 64-POLYMARKET, 65-PUMPFUN-LIVESTREAM, 66-KID-AGENTS)
Full changelog: v2026.04.05...v2026.05.02