University of Pittsburgh
Open source · Rust · Apache-2.0 · v0.5.22
A toolchain that treats the seven recurring failure modes of multi-agent coding as coordination problems, not prompt problems.
GitHub → · Blog: Why I built AiPlus
v0.5.22 closes the agent-key zero-touch UX (K1→K7). Install once, paste your first API key into a native OS dialog once, and every future agent session in every project picks it up from the OS keyring automatically — no more re-handing API keys to every fresh agent. Works across Claude Code & Codex CLI; sandbox-aware via the NEEDS_ELEVATED protocol status.
Agent-driven coding fails in predictable ways across Claude Code, Codex, and OpenCode. Each failure looks like a model problem — “the agent forgot,” “the agent ignored the plan,” “the agent hallucinated” — but the root cause is usually structural: missing memory across sessions, lost context after compaction, blurred roles, untracked time, plan-time blind spots, uncoordinated parallel work, and the agent re-asking for the same API keys every fresh session.
AiPlus addresses each one with a small Rust module that runs project-local — no telemetry, no global config edits, no runtime network fetches.
01
Agents start every session blank. Context window resets erase project conventions, prior decisions, and active threads — forcing re-prompting and re-orientation.
aiplus-agent-memory — project-local memory store with typed records (user, feedback, project, reference) the agent re-reads at session start.
02
When the host compacts mid-task, the agent loses the working set: open files, partial diffs, the reason behind the current step. The post-compact agent often re-does or contradicts pre-compact work.
aiplus-auto-compact + aiplus-compact-reminder — checkpoint before compact, validate state, resume best-effort after.
03
A single agent juggling planner / builder / reviewer roles drifts mid-conversation, second-guessing its own plan and shipping half-reviewed code.
aiplus-auto-team-consultant — explicit Advisor / CEO / Reviewer / Builder routing so each turn has a single role.
04
Token and wall-clock costs are real but invisible. Sessions burn budget on retries, context refresh, and idle polling that the user never sees attributed.
aiplus-agent-velocity — per-session velocity and cost accounting so the agent and user can see where time goes.
05
Planning happens at the wrong altitude — too vague to execute, too detailed to be wrong cheaply. The agent commits to a path before the cheap-to-discover unknowns are surfaced.
aiplus-auto-team-consultant Advisor routing — surface unknowns before commitment; cheap doubts before expensive work.
06
Multiple agents (or one agent across windows) editing the same files produce silent conflicts, wasted retries, and lost intent.
aiplus-agent-team — permanent virtual SWE team with explicit role separation (Advisor / CEO / Architect / PM / Engineer×2 / Reviewer / QA) and execution boundaries.
07
Every fresh agent session re-asks the user for OpenAI / Anthropic / cloud keys, or the user manually re-pastes them into env vars. Plaintext keys leak into shell history; agents that ought to escalate via a broker fall back to asking the Owner directly.
aiplus-agent-key (broker) — set keys once into the OS keyring via a native password dialog; agents read them via aiplus secret-broker need <alias> per a documented protocol. Cross-project share built-in. v0.5.22 (K1→K7) closes the loop: install auto-wires a cd-hook, version-skew checks prevent silent protocol failures, and a NEEDS_ELEVATED status handles sandboxed agent runtimes.
Install the aiplus command:
curl -fsSL https://raw.githubusercontent.com/izhiwen/aiplus/main/install.sh | shThen install AiPlus into a project:
aiplus install claude-code # or: codex | opencode | allThe installer writes only project-local files (.aiplus/, .codex/compact/, .claude/, .opencode/, and the AiPlus managed block in AGENTS.md). No sudo, no global config edits, no telemetry.
From v0.5.22 onward, aiplus install also offers (interactive [Y/n], default yes) to wire a cd hook into your shell rc so API keys are auto-injected as env vars when you enter a project — set the key once with the OS-native password dialog, and every future agent session in that project (or any other with the same alias) starts with the key already loaded. Append-only; opt out via AIPLUS_SKIP_SHELL_INIT=1.