Install
$ agentstack add skill-samrom3-claude-hyper-plugs-hyperteam ✓ scanned · ✓ verified — works with Claude Code, Cursor, and more.
Security review
✓ PassedNo issues found. Passed automated security review. · v0.1.0 How review works →
- ✓ Prompt-injection patterns
- ✓ Secret / credential exfiltration
- ✓ Dangerous shell & filesystem operations
- ✓ Untrusted network calls
- ✓ Known-malicious package signatures
What it can access
- ✓ Network access No
- ✓ Filesystem access No
- ✓ Shell / process execution No
- ✓ Environment & secrets No
- ✓ Dynamic code execution No
From automated source analysis of v0.1.0. “Used” means the capability is present in the source — more access means more to trust, not that it’s unsafe.
About
Hyperteam
Converts session-spec into autonomous agent team. Executes full task DAG, tracks state in plans/-team-state.json, coordinates via native task list, offers PR when gate passes.
______________________________________________________________________
Phase 0: Pre-Flight
> Prerequisites: Requires Agent Teams feature + gh CLI. Set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Verify gh CLI installed and authenticated for PR creation.
Run checks in order. Stop and surface each issue as encountered.
Step 1 — Scan plans/ and select a session-spec
- Legacy check: List files matching
*-prd.mdinplans/. Any found → stop:
> Legacy -prd.md plans found: `. Rename to -session-spec.md to proceed, or create a new spec with /session-spec`. Do not silently process old-format files.
- List all files in
plans/matching*-session-spec.md. - For each
plans/-session-spec.md, determine state:
- No
plans/-team-state.json→ unstarted. metadata.status = "running"→ in-progress.metadata.status = "complete"→ complete.- Any other value → in-progress.
- Exclude complete specs from selection.
- No incomplete specs → stop:
> No incomplete session-specs found in plans/. Create a session-spec first with /session-spec.
- Build ordered selection list: unstarted first, then in-progress. Within each group, sort by file modification time (most recent first). Format each entry:
. plans/-session-spec.md. Append for in-progress entries:⚠ This spec may have an in-flight hyperteam run. Ensure no other session is working on it before proceeding. - Single spec: Use
AskUserQuestion:
> Only one incomplete session-spec found: > > plans/-session-spec.md [warning if in-progress] > > Proceed with this spec?
User confirms → select it. Otherwise stop.
- Multiple specs: Use
AskUserQuestion:
> Multiple session-specs found. Choose one to run: > >
Wait for user's choice.
- Derive `
from selected filename: stripplans/prefix and-session-spec.md` suffix. - Derive `
fromby stripping leadingfeat-prefix if present. Ifdoes not start withfeat-, useas` unchanged.
Step 2 — Checkout git branch
- Run
git branch --show-current. - Result matches `` → proceed to Step 3.
- Mismatch:
a. Run git branch --list . b. Branch exists locally → git checkout . c. Branch absent → git fetch origin main && git checkout -b origin/main. d. Verify git branch --show-current equals `. Mismatch → AskUserQuestion` and stop.
Step 3 — Verify symlink
- Run
test -L plans/. - Absent or not symlink → create:
`` mkdir -p plans && ln -sf ~/.claude/tasks/ plans/ ``
- Verify:
readlink plans/must return path ending in.claude/tasks/. Fails →AskUserQuestionand stop.
> Note: Task list scoping handled automatically by TeamCreate in Phase 2, Step 2. TeamCreate with team_name: "" creates task list at ~/.claude/tasks// and sets CLAUDE_CODE_TEAM_NAME on all teammates. No manual export of CLAUDE_CODE_TASK_LIST_ID needed.
Step 4 — Detect fresh start vs. resume
Check plans/-team-state.json:
- Absent → Read
references/phase-1-fresh-start.mdand follow in full. Return here, proceed to Step 5. - Present → Read
references/phase-1-resume.mdand follow in full. Return here and proceed to Step 5 (or stop if user declines).
Step 5 — Gate discretion
After spec parsed and task DAG proposed, assess whether back-pressure gate task is needed.
Gate recommended when any of:
- Total step count ≥ 4
- Cross-step integration required (one step's output is another's input)
- Output not easily human-verifiable in-session (e.g., data pipelines, infra changes)
Gate optional when ALL of:
- Total step count ≤ 3
- Each step has independent verify criteria (no cross-step deps)
- Human already in verification loop (e.g., UI changes requiring manual review, or user will run
claude plugin validate .themselves) - Unit/integration tests alone sufficient to confirm correctness
Present recommendation via AskUserQuestion:
> Gate task recommended / not recommended because [reason]. > > Proceed with gate? [Yes / No / Override]
User can override either way. If Yes (or override to Yes) → include GATE task in DAG and Phase 2 team creation. If No → omit GATE task.
Proceed to Phase 2.
______________________________________________________________________
Phase 2: Team Creation and Coordination
Step 1 — Count parallel-eligible tasks
- Read
plans/-team-state.json. - Count tasks where
status: pendingANDblocked_byis empty (or all listed blockers already terminal). Call this countN. - Clamp to
M:M = min(max(N, 1), 4).
Step 2 — Create the team
Call TeamCreate with:
- Team name: ``
- Teammates: 1
hyperteam-lead,Mhyperteam-workerinstances, 1hyperteam-reviewer - Prompt includes: branch name, paths to
plans/-team-state.json,plans/-progress.txt,plans/-session-spec.md
Step 3 — Seed the native task list
For every task in team-state.json with status: pending:
- Call
TaskCreatewith YAML front-matter block + full step text asdescription:
``` --- id: type: skills:
blocked_by:
---
```
Step 4 — Broadcast kickoff
Send broadcast SendMessage to team:
> Hyperteam ` is starting. > State file: plans/-team-state.json > Progress log: plans/-progress.txt > > All workers: claim tasks from native task list. Parse YAML front-matter in each task's description for type (FEAT or DOC) and blocked_by. Load skills: listed in task front-matter via Skill tool before beginning work. Resolve blockers via team-state.json (blocker terminal when status validated or completed). > > Reviewer: begin scanning team-state.json for completed FEAT tasks with reviewed: false` immediately.
Step 5 — Monitor
Main thread monitors run. Lead agent (dispatched in Step 2) handles coordination: review outcomes, failure resets, blocker broadcasts, GATE readiness detection.
React to events:
SendMessagefrom lead signalling GATE PASS → proceed to Phase 4.SendMessagefrom any teammate requiring main-thread intervention → address and resume.
Main thread does not dispatch individual workers or validators. Teammates self-claim.
______________________________________________________________________
Phase 3: Back-Pressure Gate
> Runs inside reviewer agent — not main thread. Reviewer claims GATE native task when lead broadcasts GATE OPEN. See references/gate-task-template.md for full gate agent instructions.
Lead notifies main thread only after GATE passes. Proceed to Phase 4.
______________________________________________________________________
Phase 4: Completion and PR Offer
Read references/phase-4-completion.md and follow in full.
______________________________________________________________________
Phase 5: Team Cleanup
After Phase 4 completes (summary written, PR offered/created/declined), call TeamDelete for team ``. Removes all shared team resources. Must be done after Phase 4 so all teammates fully idle before cleanup.
Source & license
This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.
- Author: samrom3
- Source: samrom3/claude-hyper-plugs
- License: MIT
Install and usage instructions live in the source repository linked above.
Reviews
No reviews yet — be the first.
Write a review
Versions
- v0.1.0 Imported from the upstream source.