mermaid

flowchart TD
    A([BEGIN]) --> B[Agent completes work on their branch]
    B --> C["Agent commits with [ACTION:submit] and pushes"]
    C --> D[Overseer detects submit commit]
    D --> E[Run code-review flow on submitted work]
    E --> F{Work approved?}
    F -->|No| G[Write feedback to .ai/reviews/]
    G --> H["Commit [ACTION:reject] with feedback summary"]
    H --> I[Agent reads feedback and revises]
    I --> J["Agent commits revised work with [ACTION:submit]"]
    J --> E
    F -->|Yes| K["Commit [ACTION:approve] for the task"]
    K --> L[Merge agent branch to pre-mortal]
    L --> M[Update task status to DONE in .ai/status.md]
    M --> N{Are there dependent tasks waiting?}
    N -->|Yes| O[Identify next unblocked task]
    O --> P[Determine which agent should handle it]
    P --> Q[Write instruction to .ai/instructions/]
    Q --> R["Commit [ACTION:delegate] for next task"]
    R --> S([END])
    N -->|No| T{Are there other pending tasks?}
    T -->|Yes| O
    T -->|No| U[All tasks complete - generate sprint report]
    U --> S

Flow Steps Detail

Step: Agent completes work

•Agent finishes implementation on their branch (<agent>/<task-id>)
•Agent runs build/tests locally to verify
•Agent reviews their own diff for boundary compliance

Step: Agent submits

•Commit message: [AGENT:cursor] [ACTION:submit] [TASK:TASK-P4-01] Description
•Push branch to remote
•Optionally write completion note to .ai/reports/

Step: Overseer detects submit

•Monitor for commits with [ACTION:submit] header
•Parse AGENT, ACTION, and TASK from commit message
•Identify which task and agent submitted

Step: Run code-review flow

•Invoke the code-review flow skill on the submitted work
•This checks boundaries, acceptance criteria, regressions, and format
•Produces a verdict: APPROVE, CHANGES REQUESTED, or REJECT

Step: Handle rejection

•Write detailed feedback to .ai/reviews/TASK-XXX-review.md
•Commit with [AGENT:kimi] [ACTION:reject] [TASK:X]
•Agent reads feedback, makes changes, and re-submits
•Loop back to code-review

Step: Handle approval

•Commit with [AGENT:kimi] [ACTION:approve] [TASK:X]
•Merge agent branch to pre-mortal: git merge <agent>/<task-id> --no-ff
•Commit merge with [AGENT:kimi] [ACTION:merge] [TASK:X]

Step: Update status

•Update .ai/status.md — move task to Recently Completed
•Update task brief status to DONE

Step: Check dependencies

•Read all task briefs in .ai/tasks/
•Find tasks with Depends on: TASK-XXX where XXX is the just-completed task
•These tasks are now unblocked and can be assigned

Step: Assign next task

•Pick highest-priority unblocked task
•Determine the appropriate agent based on file ownership
•Write instruction to .ai/instructions/<agent>-<task>.md
•Commit with [AGENT:kimi] [ACTION:delegate] [TASK:X]

Step: Sprint complete

•When no tasks remain (pending, in_progress, or blocked)
•Generate sprint summary in .ai/reports/sprint-X-summary.md
•Notify Human PM for pre-mortal → main review

Handoff Patterns

Sequential Handoff (most common)

code

Cursor completes TASK-P4-01 → Review → Approve → Merge
→ Lovable picks up TASK-LOVABLE-001 (depends on P4-01) → Review → Approve → Merge
→ Sprint complete

Parallel Handoff

code

Cursor works on TASK-P4-01 (no dependencies)
Lovable works on TASK-LOVABLE-001 (no dependencies)
Both submit → Both reviewed → Both merged → Next tasks assigned

Rejection Loop

code

Cursor submits TASK-P4-01 → Review → REJECT (missing error handling)
→ Cursor fixes → Re-submits → Review → APPROVE → Merge

Kimi Session Context in Handoffs

Session Continuity

Each sprint has a named Kimi session. During handoffs:

•The session retains context from previous reviews and decisions
•Use --continue to resume the session for the next task
•Key findings are always saved to .ai/ files (survive compaction)

Context-Aware Handoffs

After approve/merge, the post-commit hook:

•Runs context monitor in background (checks if compaction needed)
•Updates session metadata with task completion
•If all tasks DONE: triggers sprint evaluation and session archive

Handoff with Subagent Specialization

For complex handoffs, the overseer can create specialized subagents:

•Debugger: When a rejection requires deep investigation
•Performance analyzer: When a task involves optimization
•Documentation writer: When handoff requires docs update

Templates: .agents/subagents/ | Patterns: .ai/patterns/create-subagent-*.md