kubefleet-dev/kubefleet/iterative-retrieval

Iterative Retrieval Skill

Squad agents frequently spawn sub-agents to complete scoped work. Without structure, these handoffs become vague, cycles multiply, and outputs land without being checked. The Iterative Retrieval Pattern caps cycles at 3, mandates WHY context in every spawn, and requires the coordinator to validate agent output before closing an issue.

Spawn Prompt Template

Every agent spawn must include the following four sections. Copy and fill in the template:

## Task
{What you need done — concrete and bounded}
 
## WHY this matters
{The motivation and context. What system or user goal does this serve? What breaks if skipped?}
 
## Success criteria
{How you will know the output is correct. Be explicit — list acceptance criteria, not vibes.}
Example:
- [ ] File X exists and contains Y
- [ ] No regressions in existing tests
- [ ] PR is open targeting main with description matching the issue
 
## Escalation path
{What the agent should do if uncertain or stuck. "Stop and ask me" is valid.}
Example:
- If requirements are ambiguous → stop, comment on the issue, set label status:needs-decision
- If blocked by a dependency → label status:blocked, explain in a comment
- If 3 cycles exhausted without resolution → write a summary to inbox and surface to coordinator

3-Cycle Protocol

Rules

1. After each cycle, the coordinator evaluates the output against the success criteria

before accepting it or spawning the next cycle.

1. Objective context forward: each subsequent spawn includes a summary of what was tried

and what is still missing — not just a repeat of the original task.

1. Cycle 3 exhausted → escalate: write a summary to .squad/decisions/inbox/, label the

issue status:needs-decision, and notify the user.

Coordinator Validation Checklist

Before accepting agent output and closing an issue, the coordinator must check:

• [ ] All success criteria from the spawn prompt are met
• [ ] PR exists and description matches the issue (if code work)
• [ ] No obvious regressions (grep for TODO/FIXME introduced, build passes)
• [ ] Agent did not silently skip parts of the task
• [ ] If the agent reported uncertainty — was it resolved or escalated?

If any item fails → do not accept. Spawn cycle N+1 (up to cycle 3) with specific deltas.

When to Escalate vs Retry

Retry (cycle N+1) when:

• Output is structurally correct but missing specific items
• Agent misunderstood scope (provide more context and re-run)
• Partial success — clearly identified remaining delta

Escalate when:

• Requirements are fundamentally unclear (decision needed)
• 3 cycles complete without convergence
• Agent returned conflicting results across cycles
• Task requires elevated permissions or external action
• The work depends on another issue that isn't done yet

Issue Dedup Check (Mandatory)

Before any agent creates a GitHub issue, it must search for existing open issues to avoid duplicates.

# Check for existing open issues before creating a new one
gh issue list --search "<keywords from your issue title>" --state open

• If an open issue already covers the same problem → comment on it instead of creating a new one.
• If no duplicate → proceed to create the issue.
• Use 2–3 representative keywords from the planned issue title as the search query.

Mandatory Output Requirement (Research-Then-Execute)

Every research or analysis task completed under this protocol MUST end with at least one concrete action before the cycle is closed. Acceptable follow-up actions:

• GitHub issue created documenting the findings and next steps
• PR opened implementing a recommendation
• Decision recorded in .squad/decisions/inbox/
• Documented recommendation with a named assignee and due date

Pure analysis reports without actionable follow-up will be rejected during triage. If no action is warranted, the agent must explicitly state why and get coordinator sign-off.

Anti-Patterns

• Spawning without WHY — agents can't prioritise trade-offs without motivation context.
• Accepting output without validating — one failed check avoids merging broken work.
• Cycle 4+ — if 3 cycles haven't converged, the problem is in the requirements, not the agent.
• Vague success criteria — "looks good" is not a criterion. Use checkboxes.
• Forwarding WHAT without delta — cycle 2+ prompts must include what cycle 1 got wrong.
• Creating issues without dedup check — always search before creating.
• Research without action — delivering analysis with no issue, PR, decision, or assignee is incomplete work.

Examples

Good spawn prompt

## Task
Add an "Iterative Retrieval Protocol" section to `.squad/agents/coordinator/charter.md` explaining
the 3-cycle rule, WHY format, and validation checklist.
 
## WHY this matters
The coordinator spawns sub-agents on every round. Without a documented protocol, agents run unbounded
cycles and outputs go unvalidated — leading to stale issues and silent failures.
 
## Success criteria
- [ ] Section "Iterative Retrieval Protocol" exists in charter.md
- [ ] Section documents max-3-cycles rule
- [ ] Section documents WHY format requirement
- [ ] Section contains validation checklist (at least 4 items)
- [ ] No other sections of charter.md are modified
 
## Escalation path
If the charter.md format is unclear, check another agent charter as a reference.
If uncertain about content, stop and surface to coordinator.

Bad spawn prompt (don't do this)

Update the coordinator charter with the iterative retrieval stuff.