Updated: 2026-06-08 ยท Migration plan for carrying the useful OpenClaw operating model into the new Paperclip agent company.
The review covered the current OpenClaw runtime instructions, heartbeat/autonomy policy, local queue and human ToDo files, daily and long-term memory notes, work-session ledger, dream question ledger, custom workflow skills, cron/job state, task-run ledger, and the existing Paperclip COO/PM/Development/Testing agent contracts.
Sensitive config was treated only as policy shape. Credentials, auth profiles, tokens, and provider keys should not be copied into Paperclip prompts, skills, tasks, or public docs.
| OpenClaw surface | Migrate? | Paperclip destination |
|---|---|---|
| Role split: orchestrator, PM/Design, Development, Testing. | Yes. | COO, PM, Development, and Testing agent instructions. |
| Work-session supervision and liveness evidence. | Yes. | Paperclip task templates, run comments, checkpoint fields, and COO status rules. |
| Codex work-cycle launch fields. | Yes. | Required PM-to-Dev and Dev-to-Test handoff sections. |
| Portfolio PM audit. | Yes. | PM routine and Paperclip company skill/template. |
| Human ToDo delegation gate. | Yes. | PM decision gate and COO notification policy. |
| GitHub PR flow. | Yes. | Development implementation contract and Testing verification contract. |
| Telegram briefing and interview rules. | Yes. | COO Telegram/OpenClaw bridge contract. |
| Memory layers. | Partially. | GitHub for public memory, Paperclip for local operational memory, private OpenClaw files only through sanitized summaries. |
OpenClaw isolated agentTurn cron runner. | No. | Retire after Paperclip routines prove reliable; do not reproduce its failure mode. |
OpenClaw's strongest pattern is the split between orchestration, PM, Development, and independent Testing. Paperclip should keep that small-company shape:
OpenClaw learned not to call work active without evidence. Paperclip should encode this as a task contract: every PM, Development, or Testing task needs expected first checkpoint, evidence type, timeout, and recovery owner.
Before Development or Testing starts, PM or COO must define the target issue/PR, role, repository, branch/worktree strategy, allowed files, acceptance criteria, checks, evidence, checkpoint, non-goals, and recovery plan. Development should refuse tasks that omit fields that could change behavior or scope.
The PM routine should continue to audit the full obot-claw portfolio before selecting Development work. It should inspect Projects, Requirements, Tasks, PRs, Hub status, human ToDos, dirty worktree risk, parent/sub-issue links, and stale blockers. The local AUTONOMOUS_QUEUE.md file should become transitional backup once Paperclip/GitHub task sync is active.
Before creating any Jeremy ToDo, PM must ask whether PM, Development, Testing, or a new specialist agent can do the work instead. If Jeremy truly owns the decision, the linked issue or PR must contain an explicit @jwildfire question or instruction with the decision criteria and evidence.
Development should inherit the OpenClaw PR discipline: write only inside approved obot-claw repositories, capture repo freshness before editing, use branches and PRs for repo changes, assign Obot where appropriate, request Jeremy review when ready, and monitor checks through durable task/routine state instead of long silent polling.
OpenClaw should remain the human front door, but it should address the COO rather than direct agent execution. Telegram output must stay short, plain text, and artifact-backed. The bridge should expose allowlisted actions such as health, latest status, review next queue item, start approved task, check failures, summarize latest artifact, and pause/stop where supported.
| Memory type | Current role | Paperclip migration rule |
|---|---|---|
| GitHub issues and PRs | Durable public memory. | Keep as source of truth for project decisions, work evidence, PRs, reviews, and human asks. |
| Paperclip tasks, comments, runs | New local operational memory. | Use for task state, assignments, run evidence, liveness, and local workflow state. |
| Daily memory notes | Private operational diary. | Migrate only sanitized summaries into Paperclip when needed for task continuity. |
| Long-term memory | Personal and operating preferences. | Do not expose wholesale. Convert only relevant policy preferences into scoped instructions. |
| Dream notes and question ledger | Private synthesis and question dedupe. | Keep private until a question is promoted into a Paperclip task or GitHub issue comment. |
| Human ToDo file | Local decision queue. | Move public items into GitHub/Paperclip issue state; keep private/config items direct-only. |
| Work-session ledger | Liveness and recovery history. | Translate the evidence model into Paperclip run and checkpoint discipline. |
Do not copy every OpenClaw cron job into Paperclip immediately. The old isolated cron path repeatedly produced startup stalls and main-session reminder fallbacks. Paperclip routines should be promoted in stages.
agentTurn cron as the production executor.| Agent | Updates to add |
|---|---|
| COO | Telegram output contract, liveness rules, schedule promotion gates, Human ToDo notification gate, artifact-backed status summaries, and child-task creation rules. |
| PM | Portfolio audit loop, Human ToDo gate, launch-field completeness checklist, GitHub-native memory rule, queue migration rule, and Project/Requirement/Task hygiene checklist. |
| Development | GitHub PR flow, freshness capsule, branch/PR/check evidence, long-running watch rule, and stop conditions for ambiguous scope or dirty unrelated worktrees. |
| Testing | Independent acceptance evidence, browser QA rule, PR review watch behavior, no-edit compliance checks, artifact-placement verification, and disposition verification. |
| Before enabling | Required evidence |
|---|---|
| Scheduled Paperclip autonomy | Health OK, backup exists, contracts versioned, PM no-edit routine posts durable evidence, COO status is artifact-backed, and Telegram remains allowlisted. |
| Development scheduling | PM routine passes repeatedly, GitHub path is scoped and documented, no-edit and issue-comment tasks pass without operator repair, and artifact placement is enforced. |
| OpenClaw schedule retirement | Paperclip equivalent succeeds on the same cadence, failure alerts work, anti-spam behavior works, public/private boundaries are verified, and rollback is documented. |
Then run a no-edit PM task against the rollout queue. The PM should review the Chapter 8/9 rollout tasks and produce a readiness note without changing repositories.