Human Agent Live in Framework
HALF-Work is a lightweight Human–Agent collaboration framework for real software and hardware projects. It keeps the human's intent broad enough to describe naturally, then asks the Agent to refine scope, evidence, risks, and options before implementation begins.
Status: Personal Pilot — the first DeskNest engineering trials are complete; framework acceptance is still under revision.
An Agent can produce code quickly while still missing the actual product boundary, the board-level acceptance target, or the ownership of an existing dirty worktree. HALF-Work treats the task contract, evidence, and review as first-class work products.
These dimensions must not be collapsed into one overloaded “level” field:
| Dimension | Meaning |
|---|---|
| HALF-Work Task Class | L0 small fix, L1 standard task, L2 architecture task |
| Global Orchestration Level | The host project's Level 1/2/3 planning and Agent-dispatch rules |
| Model Thinking Level | 轻量, 标准, or 深入 reasoning depth requested by the user |
| Resource Mode | economy, balanced, or maximum task-local resource budget |
The task class describes the work. The other dimensions describe how the work is coordinated and verified.
- Human Draft — choose the task class and thinking level, then describe the goal, context, preferences, constraints, and unknowns in natural language.
- Agent Refinement / User Review — inspect the real repository, propose scope and options, define evidence and acceptance outcomes, and wait for user confirmation before implementation.
The reusable template is templates/ACTIVE_TASK_EXAMPLE.md.
Observe → Plan → Implement → Verify → Review → Close or Carry Risk
Small tasks can use a fast path. Architecture tasks must expose boundaries, sub-gates, ownership, and independent review evidence in proportion to risk.
Every gate records one explicit result:
Pass— the declared contract was met;Fail— the contract was not met;Not Run— the check was not executed;Accepted Limitation— a declared boundary was observed and accepted;Carried Risk— the item remains open with a named follow-up trigger.
An accepted limitation is evidence, not a disguised pass. A closed task must be archived or replaced before another task becomes current.
The first public seed intentionally contains only reusable workflow material. It does not contain:
- private pilot planning history;
- local control files or active task contracts;
- editor state or machine-local configuration;
- credentials, tokens, private URLs, or generated secrets;
- DeskNest or CNFontNest product code.
The initial private planning documents HALF-Work-ROADMAP.md and
design-draft.md remain local and are excluded by .gitignore.
The next real task should validate the revised template and record lightweight cost evidence: elapsed time, Agent calls, rework rounds, changed files, and human decisions. A public template should grow from another real project, not from publishing the entire private planning history at once.