prime-agent #

fork of PrimeIntellect-ai/prime-agent for changes that are not upstreamable (and some fixes). rebased onto upstream 0.9.3. this fork lives at next.tangled.org/ptr.pet/prime-agent.

what's different #

how the agent runs #

• extensions can read and change messages sent between agents, and can tell apart "the human asked" from "the agent woke itself up"
• a python api to list, watch, message, stop, and restart running sessions from outside the agent
• the agent's python can move to another machine over ssh: it sets itself up there, caches what it needs, copies your skills over, and can run subagents on that machine
• the bash tool pushes work toward python: it refuses git/jj/grep/find/ls and shell pipes (there are skills and plain python for that), and it can feed a running command input while it runs
• ssh and scp to other machines skip the local repo safety checks; huge command output is saved to a file instead of flooding the screen
• background commands report when they finish, even if they were killed or the machine ran out of memory, without spamming the chat
• . keeps working on the last task; python cells show a short useful preview instead of a wall of code
• useful python skills ship with the agent: git, jj, nix, repo, search, ssh, atproto, tangled, display, websearch, edit
• queued messages go out as one batch: ten messages typed while the agent works land in a single turn instead of ten turns with the model thinking in between. steeringMode and followUpMode in settings bring back the one-at-a-time behavior

subagents #

• discover available categorized models with await rlm.models(), which returns a dict-like object (e.g. rlm.models().scouts[idx], rlm.models().coders)
• categories and model lists live in ~/.prime/agent/rlm-models.toml, loaded as-is
• calling await rlm.models() at least once per compaction window is required before calling await rlm.spawn(prompt, name="...", model="...")
• model= specifies the child's model; optional thinking= sets reasoning level; optional host= routes execution to a remote host via SSH control master

accounts and providers #

• stay logged into several accounts per provider: requests rotate between them, a session can stick to one account, and a failed login or rate limit quietly moves to the next account. /logout lists each account, /usage shows what you spent
• /usage reports each provider's real quota from a per-provider feed, and prime-agent usage --json prints the same report for anything else to read: one file per provider, and a provider without a feed is named instead of looking unlimited
• that query is pure — a request shape in, a report out, no socket or clock — so the same rules run outside javascript: the zig core lives in its own repository (usage-core, checked out at native/usage/ as a submodule) behind a C ABI for native hosts (a macOS app links it and keeps NSURLSession for the request). a parity test computes both implementations from the same recorded bodies, so neither drifts quietly

models #

• the model list refreshes itself from models.dev about once an hour; /reload models forces it. refresh errors no longer wreck the screen
• you can ask a model for json that matches a schema — works on google and openai-style providers
• fixed tool results getting lost when a user message interrupts a turn
• every tool call shows how long it took, and the model sees that too

remote work #

• --remote-python-host runs the agent's python and its subagents on another machine; the banner, tray, and agents view show which one
• remote skills mirror the local ones (rem.bash, rem.os, ...)
• the ssh skill can pipe text into commands and reconnects itself when its master connection dies

looks and speed #

• context and cache numbers update live while the model works; the tray and agents view show a cache timer and how much of the context is cached
• the agents view and session switching got much faster: parallel disk scans, cached tool paths, reused ui services
• pin sessions to keep them on top (ctrl+u)
• diffs in the ui are syntax highlighted; the display skill shows rich output (markdown, diffs, tables)
• a failed update now says it failed instead of claiming success

extensions #

separate repos that plug into prime agent:

• prime-antigravity — sign in with google antigravity: oauth login and streaming for google's cloud code assist, ported from oh-my-pi
• prime-commitment-observer — remembers unfinished requests across turns; a small model files each one as open, active, waiting, blocked, or done
• prime-snapcompact — compaction that keeps a plaintext and image archive of old turns instead of squashing them to text, so code and tool output survive
• prime-meat — turns big diffs into short readable summaries
• prime-narrate-jjk — narrates agent turns in jujutsu kaisen narrator voice
• vibecoding-quota-tray — provider quota awareness in the tray
• omp-codex-bridge — presents omp as codex for use with chatgpt remote

install #

stable release (upstream builds):

curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh

from source (this fork):

npm install        # from the repo root
./prime-agent.sh   # run via tsx, or ./prime-agent.sh --dist for the built bundle

npm run build builds all packages. start in a project directory, /login on first launch.