A Novaprospect fork of Waveterm that makes Claude the ambient intelligence layer of the entire workspace — native Anthropic provider, NOVAICOM operator dispatch, Jira-aware tabs, and role-based presets for Code, Arch, Research, Review, and Ops.
Waveterm treats AI as a chat sidebar. RogueWave makes Claude the ambient intelligence layer of the entire workspace — aware of every block, capable of acting on anything, integrated with the NOVAICOM operator dispatch system.
Where Waveterm asks "which AI provider do you want?" RogueWave asks "what are you trying to build, and how can I help right now?"
simulated transcript
[ rogue v0.0.0-dev · workspace=novaprospect/ncentcom · preset=Code ] director@rogue ~/repos/ncentcom $ wsh rogue dispatch --role code --issue NCC-481 ⤷ novaicom dispatch · session=rogue-NCC-481-a91c · pid=48213 ⤷ rogue-ops block opened in split[2] (tail+status live) ⤷ rogue-jira block bound to tab — NCC-481 context piped to system prompt director@rogue ~/repos/ncentcom $ wsh rogue context --show context buffer · 18,422 tokens · cache hit 71% pinned src/lib/server/auth.ts 3,104 ✓ pinned docs/NCC-481-PROMPT.md 1,266 ✓ ambient terminal scrollback (last 400ln) 8,710 ambient jira NCC-481 (summary+comments) 2,118 system novaicom-ops.md 3,224 cached director@rogue ~/repos/ncentcom $ wsh rogue claude "why is the test for guardClause failing?" ⠋ thinking · 2.1s · 1,402 thinking tokens (collapsed) ⠿ response streaming → ai panel split[3] director@rogue ~/repos/ncentcom $ █
Every wsh rogue command is a GUI-aware projection of novaicom — calls the daemon, then drives WSH RPC to open blocks, badge tabs, post notifications.
Upstream shoehorns Anthropic through an OpenAI shim — losing extended thinking, native tool use, prompt caching, streaming thinking tokens.
novaicom dispatches operators (Code, Arch, Research, QA, DevOps) as scoped subprocesses. Director has no visual into what they're doing in real time.
Waveterm sees blocks only when you attach them. The workspace state should BE the context.
Every dev task at Novaprospect maps to a Jira issue. Claude should know which one without prompting.
Tab templates are planned upstream but not shipped. RogueWave can deliver opinionated presets that match the operator roles.
wsh bridges terminal scripts to the GUI. Extend it as a GUI-aware projection of the entire novaicom CLI.
serializable layouts. commit .roguewave/workspace.json to share with the team.
All RogueWave code lives in namespaced locations so upstream Waveterm rebases stay clean. Do not fork the core block protocol, WSH RPC wire format, or build system.
{
"default": "claude-opus",
"modes": {
"claude-opus": {
"provider": "anthropic",
"model": "claude-opus-4-7",
"extended_thinking": true,
"cache_system_prompt": true
}
}
}ordered by value over effort
Anthropic SDK streaming + extended thinking. Wire to existing waveai block. Highest leverage — unblocks everything else.
Establish the command surface. Start with claude, status, context — thin wrapper over NOVAICOM and the new client.
Background goroutine in the Go backend. Feeds terminal scrollback + open files into the system prompt.
rogue-ops view type. Connect to novaicom session list --json. Stream stdout of active sessions into the block.
rogue-jira view type. Atlassian REST API. Read-only first.
Serialize current split layout to JSON. Restore on load. Ship the five role presets.
resolved during the architecture phase