Files
APAW/.kilo/rules/session-persistence.md
NW f01e2064fb feat(evolution): Kilo Code release sync & APAW system hardening (v2026-05-07)
Security & Permissions:
- All 30 agents: task[*]=deny, task[subagent]=deny (cascade prevention)
- orchestrator & release-manager: bash=ask (hardening)
- New .kilo/rules/subagent-security.md with audit rules
- Updated .kilo/rules/global.md with Security & Permissions section
- Updated .kilo/agents/orchestrator.md with Security Enforcement block

Session Management:
- New .kilo/rules/session-persistence.md (checkpoint format, worktree isolation)
- Updated .kilo/rules/branch-strategy.md (worktree per agent)
- pipeline-runner.ts: Checkpoint interface + save/load/resume methods

Plan Persistence:
- Updated .kilo/rules/lead-developer.md (plan handover section)

Per-Agent Reasoning:
- capability-index.yaml: reasoning_effort for all 30 agents (xhigh/high/medium/low)

MCP Cleanup:
- New .kilo/skills/docker-security/SKILL.md (--rm, orphaned process cleanup)

Config Validation:
- Updated .kilo/rules/docker.md (startup checks, commit scoping, location awareness)

Docs:
- README.md: v2026-05-07 evolution badges
- .kilo/EVOLUTION_LOG.md: Entry #6 with full metrics
- .gitignore: ignore dist/ + bun.lock

Gitea: Milestone #66, Issues #91-#98
Architect: 9/9 sections fresh (express project type)
2026-05-08 18:54:08 +01:00

1.3 KiB

Session Persistence & Checkpoint Rules

Session State Preservation

  1. After each pipeline phase completes, the orchestrator MUST write a checkpoint to .kilo/logs/checkpoints/{issue}-{phase}.json.
  2. The checkpoint JSON must contain:
    • issue_number
    • phase (e.g., designing, testing, implementing)
    • agent_name
    • files_modified
    • status (success, fail, blocked)
    • timestamp (ISO 8601)
    • next_agent
  3. If the pipeline is interrupted (orchestrator restart), the orchestrator MUST read the latest checkpoint and resume from the next phase instead of restarting.

Session Fork

  1. Before any agent that performs destructive edits (lead-developer, the-fixer, frontend-developer), the orchestrator SHOULD create a git stash or worktree to allow rollback.
  2. The stash/worktree name format: checkpoint/{issue}-{agent}-{timestamp}.

Diff Viewer

  1. After any edit tool execution, the orchestrator MUST log the diff to .kilo/logs/diffs/{issue}-{agent}.patch.
  2. Diff files are available for per-file revert decisions.

Worktree Isolation

  1. Each agent task SHOULD run in its own git worktree when the agent performs file edits.
  2. Worktrees are created under .kilo/worktrees/{issue}/{agent}/.
  3. After task completion, worktree is merged back or discarded based on checkpoint status.