agkan-icebox

Overview

A workflow to review icebox tasks and decide whether to promote each one to backlog or close it.

Icebox definition: Tasks that are ideas or candidates not yet ready for planning. They have unclear requirements, are on hold pending external factors, or are low-priority items that may be addressed in the future.

When to Run

• When icebox tasks have accumulated and need periodic review
• When external conditions change and previously deferred tasks may now be relevant
• When starting a new planning cycle and wanting to surface hidden candidates

Workflow

0. Fetch Config

CONFIG=$(agkan config get --json 2>/dev/null || echo '{}')
ICEBOX_MODEL=$(echo "$CONFIG" | jq -r '.config.models.icebox.model // "sonnet"')
ICEBOX_EFFORT=$(echo "$CONFIG" | jq -r '.config.models.icebox.effort // "high"')

1. Retrieve Icebox Tasks

agkan task list --status icebox --json

2. Review Tasks One by One with Sub-agents

For each task, use the Task tool (general-purpose sub-agent) to review. Do not use Skill("agkan-icebox-subtask"). Instead, invoke it by having the sub-agent read the SKILL.md file:

Task(
  subagent_type="general-purpose",
  model="<ICEBOX_MODEL>",
  description="Review icebox task #<id>",
  prompt="""
Please review the following icebox task.

## Task Information
- ID: <id>
- Title: <title>
- Body: <body>

## Procedure
Read .claude/skills/agkan-icebox-subtask/SKILL.md and follow its procedures to review.

## Effort / Thoroughness
Effort level: <ICEBOX_EFFORT>
- low: Quick assessment. Focus on obvious gaps or blockers. Minimal research.
- medium: Standard review. Check requirements clarity and promote/close decision.
- high: Thorough review. Deep analysis of relevance, dependencies, and edge cases. Research codebase as needed.
- xhigh: Recommended default for planning/agentic work; maximize coverage of dependencies and edge cases.
- max: Reserve for the highest-stakes or most complex tasks.
"""
)

If a task ID is specified by the user, retrieve and review only that target task:

agkan task get <id> --json

Then delegate only that single task to a sub-agent using the same Task call format above.

3. Re-fetch Icebox Tasks and Continue or End Session

After the sub-agent completes, re-fetch the icebox task list to pick up any tasks added during processing:

agkan task list --status icebox --json

If there is no instruction to end from the user and icebox tasks remain, select the next task and repeat from step 2.

If no icebox tasks remain, end the session.

Decision Flow

Retrieve icebox tasks
    ↓
Delegate one task to sub-agent
    ↓
Sub-agent reviews and makes decision
  - Requirements now clear? → Promote to backlog
  - No longer needed / superseded → Close
  - Still unclear / waiting → Keep in icebox
    ↓
Move to next task (repeat until all are done)

Important Notes

• Icebox is not a permanent graveyard — review it periodically
• When promoting to backlog, the task will be picked up by agkan-planning next
• Do not move icebox tasks directly to ready; always go through backlog → planning first