القائمة
Portable "loop kit" — inject durable state, terminal-condition, auto-resume and auto-escalation into any session. Makes the loop reliable without a daemon.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Lookup reference for executing a task already decided on — code style per language, test/mock boundaries, API conventions, the local CI sequence before push, merge discipline, how to resolve review comments, repo settings and git hooks. Read it when you are about to push, merge, review, or set up a repo; not before.
Evidence-first done-gate. Refuses to mark work complete without concrete artifacts (commit SHA, PR link, file path, test output) verified empirically. The cardinal principle.
Platform-agnostic ship workflow — local CI gate, commit per phase, push, PR with Summary + Test plan, watch CI, merge-commit only. Honors human gates on main.
Helps figure out WHICH problems are worth solving, not just how to solve them. Orchestrates: problem discovery/validation with users (focus-group) → prioritization (severity×frequency×reachability×fit) → solution stress-test (premortem). TRIGGER: 'is it worth solving X', 'vale la pena risolvere X', 'is this a real problem', 'è un problema vero', 'which problem should I attack', 'quale problema attacco', 'should I build this', 'dovrei costruire questo', 'validate before building', 'validiamo prima di costruire', 'prioritize these problems', 'prioritizza questi problemi', 'is this worth building'. YES upstream of any non-trivial build/investment.
Operate the way Roberto D'Angelo works — total autonomy scoped by evidence-first verification, artifact-per-phase discipline, human confirmation before irreversible actions, and his communication voice for drafting. Use whenever asked to operate "in roberto-mode", when starting a complex multi-step task, when drafting a document/email/reply in his voice, when reasoning through a real decision, or when asked "do you know who I am" / "are you in my context".
Pre-landing code review — correctness, security, patterns, reuse/simplification. Severity-classified findings with file:line evidence. Read-only; the owner applies fixes.
| name | auto-checkpoint |
| description | Portable "loop kit" — inject durable state, terminal-condition, auto-resume and auto-escalation into any session. Makes the loop reliable without a daemon. |
| providers | ["claude","copilot","codex"] |
A kit injectable into any session to make it loop-reliable without a daemon.
Implements the contract in loop/loop-protocol.md:
durable state, terminal-condition, auto-resume, auto-escalation.
state.db at a known path + .agent-state/<task>.jsonl (append-only cursor).done step.state store: ~/.convergio/v3/state.db (if present)
~/.roberdan-os/state.db (fallback)
cursor: .agent-state/<task>.jsonl (gitignored, 1 record/step + evidence)
timestamp: RFC3339
Convergio, if active, reads the same state — optional observer, never a dependency.
on start:
state = read(state.db, cursor) # idempotent resume
step = last_done(state) + 1
loop:
result = execute(step)
append(cursor, {step, result, evidence}) # checkpoint = 1 commit/phase, evidence-first
if terminal_condition(): break # empirical verification (thor / job-specific check)
if failed_twice(step): escalate(opus); log(reason)
if no_progress(2 passes): STOP; report_wedged(); break
step += 1
on each phase end:
post_task_sync() # vault + cvg + repo
/loop + ScheduleWakeup for external waits (CI/deploy/embed) —
submit → wakeup +Nmin → check terminal-condition → done | re-arm.launchd/cron re-read the cursor and relaunch until the terminal-condition.Every checkpoint = an evidence-first update: [phase N/M ✓] commit <sha> · <check> · next: ….
Never "still working."