Set Up Issue Tracker
This Kit initializes one existing Space, not a Workspace or a repository. The user supplies the GitHub repository and Kit path; clone the requested revision into an unused directory, record its commit, and locate this file. No built-in discovery, package service, desktop integration, canvas ledger Note, or Huabu source-checkout working directory is required. All paths below are relative to this file.
Prepare
Load the authenticated Huabu root guide, including live Skills, using node "$HUABU_RFS_ENTRY" get-skill (or native HTTP if HUABU_RFS_ENTRY is absent). Load the live agents guide and capability details for the operations this installer uses: GET_SPACE_OUTLINE, INSPECT_NODES, INSPECT_EDGES, CREATE_NODES, MERGE_NODE_DATA, SET_FRAME_LAYOUT, and CONNECT_NODES. Follow current guidance; a rejected operation is not permission to guess a replacement.
Read the installer before executing it. It is a dependency-free Node script, not a document to paste into a Prompt. Use an existing Node 24.16+ runtime; never install a runtime merely for Huabu access. If that runtime is unavailable, stop and explain the requirement. The script resolves its assets relative to itself, independently of cwd, and uses HUABU_RFS_ENTRY when present (falling back to native HTTP via HUABU_RFS_URL and AGENTLET_TOKEN).
Confirm installation authorization and the destination Space. The installer automatically resolves the current installer Agent's Node ID from HUABU_THREAD_ID when present; pass --installer-node NODE_ID only if HUABU_THREAD_ID is absent. You do not need a Profile ID or repository paths before laying out the Space.
Step 1. Lay out Space content
Run a read-only preview first without --apply:
node ./scripts/install.mjs
The script reads complete bounded inventory, lists selectable external Agent Profiles, and prints an installation plan without writes. On initial layout, blank means only this identified installer Agent and no edges; unrelated content or truncated inventory blocks overlay installation.
After confirming the preview, execute with --apply to lay out the Space content:
node ./scripts/install.mjs --apply
This creates native Frames/Notes/Text and the empty Issues area without leaving any machine-state ledger Note on the canvas. Instruction Frames have their final recognized titles from creation: prompt: issue tracking common, prompt: issue tracking coordinator, skill: issue fixing, and the amber prompt: issue tracking configuration. Do not rename them during onboarding. Retain the returned nodes map in your conversation context so Step 3 can address those exact created node IDs directly (and if resumed without --nodes, the script resolves the unique canonical Frames on the Space). It copies installed Markdown bytes verbatim from assets; do not regenerate or summarize them.
Install the current single-Agent-node layout, naming, and asynchronous-dispatch contracts unchanged: each issue is represented by a single Fixing Agent node inside the Issues Frame (no per-issue Frame and no Execution Note node), and Issue Frame Layout owns its compact <status emoji> <issue identity> · <stable issue title> label. The installed Coordinator passes execution identities in the initial prompt and returns after prompt acceptance/Thread confirmation rather than waiting for a Fixing Agent's final answer; the Fixing Agent reports findings/plans directly in chat and updates its own Agent node label. These rules apply to future issue units, not the setup Coordinator's issue-free label. Do not copy older Fixing Agent/Issue Frame label suffixes, Execution Note contracts, or synchronous waiting instructions from a previous installation.
When run before configuration is filled in, Step 1 stops cleanly in state: "needs_configuration" with coordinator: null. The recognized instruction titles are already in place; only Coordinator creation waits for configuration. --profile PROFILE_ID can override the Coordinator selection but never bypasses valid confirmed configuration.
Step 2. Fill in configuration on the Space
The amber prompt: issue tracking configuration Frame contains only the editable Issue Tracker Configuration Note, using the Markdown table template. It is the single source of configuration: no user-maintained JSON copy. As a direct Note child of a Prompt Frame its body is included in the Agent's initial Prompt. A separate root-level Text node, outside every Prompt/Skill Frame, explains required fields, defaults, confirmation, and the available external Profile names/IDs. Do not put onboarding instructions inside an instruction Frame. Missing, invalid, or unverified values are reported by field, along with availableProfiles from the installer output.
Guide the user to edit only the Value column in the existing Issue Tracker Configuration Note. Keep the Field keys unchanged; blank optional cells mean use defaults/derivation, not the literal string null. Only two fields are required upfront (codebasePath and fixingProfileId):
codebasePath(required): absolute path to the local Git repository on the intended execution device.fixingProfileId(required): explicit selectable external Profile ID (neverhuabuor an inferred alias). Also used for the Coordinator unlesscoordinatorProfileIdis set.coordinatorProfileId(optional): defaults tofixingProfileIdwhennull.repository(optional):owner/repooverride; whennull, the Coordinator derives it fromcodebasePathGit remotes (and asks for confirmation if remotes such as forkoriginvsupstreamare ambiguous).worktreeRoot(optional): absolute parent directory for issue worktrees; whennull, defaults to<codebasePath>-worktrees.developmentBaseRefandpullRequestBase(optional): default to"origin/main"and"main". Override them when the project uses a custom branch workflow (such asorigin/alpha+alphaororigin/develop+develop); once configured, the Coordinator enforces those exact refs and fails closed rather than silently falling back.confirmedByUser: change the table cell fromfalsetotrueafter the user confirms the configuration.
Do not copy the Kit repository's settings, guess a default branch, insert host paths, or treat a Server-local check as evidence about another machine. Read the target repository's own instructions; do not impose Huabu's architecture directories, package manager, checks, or Canary policy. Edit the existing configuration Note, not another copy. Validation checks shape, selected Profiles and confirmation; the Coordinator must still verify Git/path/device facts on the intended device before creating an execution environment.
Step 3. Finalize installation and create Coordinator
Once the configuration Note is updated, preview validation without --apply (passing --nodes "$NODES_JSON" from Step 1's output when available in your conversation history):
node ./scripts/install.mjs --nodes "$NODES_JSON"
When configurationState is valid, run with --apply to finalize the installation:
node ./scripts/install.mjs --nodes "$NODES_JSON" --apply
Using the validated coordinatorProfileId (or fixingProfileId when the override is blank, or --profile PROFILE_ID), the script creates the fresh idle Coordinator with noStart (X-Huabu-Agent-Start: false), no prompt, no parent lineage, and no host-thread correlation; labels it Issue Tracker Coordinator; connects its specific Prompt Frame; turns the configuration Frame green; and returns state: "ready" with the created Coordinator identity. Frame titles never change. Common and configuration instructions are global, and the fixing contract is a lazy Skill Note. No prompt, warm-up, mode/model/config control, or implementation authorization is sent.
Creation sets workingDirPath to the confirmed codebasePath, overriding the Profile's default working directory on its execution device. This does not create an issue worktree or authorize repository changes. Rerunning setup reuses an existing Coordinator without changing its launch overrides; if an older Coordinator has the wrong working directory, agree on a replacement with the user rather than silently recreating it.
Require state: "ready" with valid configuration and expected direct Prompt adjacency. New installations keep the Coordinator directly below its dedicated Prompt with a measured 64px gap; the fixing Skill is below the common Prompt, and configuration stays to the right rather than between the Coordinator and its Prompt. Final geometry is checked after creation; reruns preserve existing Agent positions. Review the visible Space as well: browser text measurement may refine content-driven heights after headless installation. Report the installed Coordinator Node/Thread, configuration state and limitations. Only then invite the user to send the Coordinator its first message.
Space Prompt is captured at first realization, including external mode/model/config controls, not node creation. Do not use the already-realized installer as Coordinator. Color communicates setup status; it is not a technical launch lock. The Coordinator must download and verify the current Issue Tracker Configuration Note before new work, not rely on its initial Prompt snapshot. If the user already interacted with the Coordinator before configuration/topology completion, stop: editing the Space does not replace its frozen Prompt; agree on a replacement Agent rather than silently continuing.
Failure and rerun boundaries
Setup state stays in this setup Agent's conversation context and the Space's own instruction Frames—no ledger Note is written to the Space, and user notes or edits after Step 1 do not block Step 3. Rerunning after completion reuses the existing Coordinator and performs no writes when unchanged.
RFS is not transactional. Accepted commands and Coordinator creation are separate durable boundaries. If a command fails or a response drops, never blindly retry: re-query actual Space state (GET_SPACE_OUTLINE / INSPECT_NODES), reconcile against the node IDs in your conversation context, explain what already exists, and obtain an explicit recovery decision before creating or deleting nodes.
Installing this Kit is not approval to fix an issue. The installed Coordinator owns runtime coordination; the safety contract owns Fixing Agent authorization. For requests such as "fix issue 12" or "approve that implementation", use the existing execution conversation, not this installation Skill.