kf-g-agent-worker-common

Common worker rules

Common to subagent workers in an orchestrator-workers configuration. For the procedure and output format of each worker type, the definition of each agent takes precedence.

Skill reference

• Delegation instruction skills Reference only the skills listed in
• Read targets: those stated in the delegation instruction path の SKILL.md, and reference files within the same skill directory
• If a skill outside the list is needed, do not start work; needs-escalation do

Pre-start check

1. In the delegation instruction, skills If there is no section → immediately needs-escalation(confirm with orchestrator)
2. skills: none → Can work without reading skills (exception)
3. none Other than kf-g-agent-worker-common is not listed → immediately needs-escalation(worker skill unspecified)
4. In the listed skills, path If there is no → immediately needs-escalation
5. After reading the listed skills, proceed to the procedure of the relevant worker agent

Additional skill needed → do not start work; needs-escalation(orchestrator reselects)

Common reporting rules

• Delegation instruction report_budget (default 600 tokens, max 1000)
• State conclusions, rationale, and verification results concisely
• Follow the "output format" of the relevant worker agent

Budget and checkpoint

Delegation instruction budget If there is, stop work at one of the following and needs-escalation return a partial result / checkpoint with.

Stop condition	Judgment
Budget reached	time_budget(soft budget) or max_tool_calls reached
Same failure repeated	Failures of the same hypothesis max_retries_per_hypothesis over
No progress	checkpoint_interval No change in phase / deliverables per unit

• time_budget is a soft budget (recommended upper limit). Not a hard timeout
• checkpoint_interval Update the checkpoint every time, and return the latest when stopping
• Minimum checkpoint items: phase, completed, next, blocked_on, tool_calls

status

Value	Meaning
completed	Acceptance was satisfied
blocked	Stopped due to external factors such as insufficient access, tool failure, or waiting for user input
needs-escalation	Ambiguous specification, excessive scope, authentication/security boundary, design decision needed, additional skill needed, worker skill unspecified, budget reached, same failure repeated, no progress (partial result / checkpoint included)