Menu
Use when you need persistent looping until a task is fully complete.
The review-first prompt is selected by default. You can switch to a direct command or download a local copy.
Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
npx skills add https://github.com/seokan-jeong/team-shinchan --skill team-shinchan-ralphThe command stays on one line. Scroll horizontally to inspect it before copying.
Prefer a local copy? Download the files currently available to SkillsMP.
Use when you want to start a new task with the integrated workflow.
Use when you want autonomous completion from requirements to verification without intervention.
Use when you need backend development for APIs, databases, servers, or endpoints.
Based on SOC occupation classification
Showing SKILL.md
| name | team-shinchan:ralph |
| description | Use when you need persistent looping until a task is fully complete. |
| user-invocable | false |
Output immediately before executing:
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
๐ฉ [Kazama] Persistent mode โ won't stop until done! ๐ช
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
If args length > 2000 characters:
Truncate to 2000 characters
Warn user: "Request was truncated to 2000 characters"
This skill (Kazama's narrated boulder loop via Task โ cheap, delegatable, auto-recovering) is Tier 1 and the default. For high-stakes, genuinely long-running work where a stalled or prematurely-stopped loop is costly, escalate to Tier 2 โ team-shinchan:fierce-ralph (a main-loop Workflow whose loop condition is owned by the script: workerโverifier iterations bounded by an iteration cap, a token budget, and a stagnation limit, closed by an Action-Kamen gate).
| Stay on Tier 1 (this skill) | Escalate to fierce-ralph |
|---|---|
| Routine persistence, quick "keep going till done" | High-stakes long task; a stalled/early-stopped loop is costly |
| A narrated loop in one context is acceptable | You want the loop + stop bounds ENFORCED deterministically |
| Auto-triggered or delegated, cheaper | Explicit user opt-in only (Workflow can't fire from a subagent; set a budget) |
Never silently jump to Tier 2 โ on a high-stakes long task, offer the opt-in; the user launches /team-shinchan:fierce-ralph.
If invoked standalone (not as Stage 3 of an active /start/autopilot workflow),
run the START โ In Progress transition per agents/_shared/linear-sync.md before
executing. When driven inside a workflow, skip โ the workflow owns the issue lifecycle.
No-op if no Linear issue / Linear MCP unavailable.
Do not read further. Execute this Task NOW:
Task(
subagent_type="team-shinchan:kazama",
model="opus",
prompt=`/team-shinchan:ralph has been invoked.
## Persistent Execution Until Complete Mode (Boulder v2)
Don't stop until complete. Idle detection enabled.
### Core Loop
1. Check TODO list
2. Execute next task (delegate to appropriate agent)
3. Verify results
4. **Measurable progress check** (after each iteration):
- Compare current state to previous iteration:
a. File changes: \`git diff --stat\` output differs from last check
b. TODO count: unchecked items \`- [ ]\` in PROGRESS.md decreased
c. Phase completion: checked items \`- [x]\` in PROGRESS.md increased
- If ANY of (a, b, c) is true โ progress detected โ reset idle counter to 0
- If ALL of (a, b, c) are false โ no progress โ increment idle counter
5. **Stagnation detection** (after each idle-detection check) โ a STOP condition, not just narration:
Run: \`node src/stagnation-detector.js --jsonl .shinchan-docs/work-tracker.jsonl --doc-id {DOC_ID} --window 20\`
\`stagnation: true\` means a hard pattern (REPEAT_ERROR / OSCILLATION / AC_STALL) RECURRED in the
window โ the loop is stuck, not merely slow (the detector already requires repeated occurrences, so
it is not a single transient).
- **First** \`stagnation: true\`: narrate the pattern and try ONE corrective approach โ
"Stagnation detected: {pattern names}. Adjusting approach to address: {evidence}."
- **Second consecutive** \`stagnation: true\` (the correction did not break the pattern): **STOP the
loop NOW and escalate to the user** with the pattern + evidence โ do NOT continue toward the
15-iteration cap. A confirmed-recurring stagnation pattern will not resolve by looping; burning
the remaining iterations is waste. Record \`event: stagnation_stop\` to boulder-log.jsonl.
6. **Idle detection**:
- If idle counter >= 3: "Idle detected โ 3 consecutive iterations with no measurable progress"
- Auto-generate continuation prompt with new approach:
"Previous approach stalled on [last task]. Trying alternative: [different strategy or next task]."
- Apply exponential backoff before retry:
- Iteration 1-2: 0 seconds (immediate)
- Iteration 3-4: 2 seconds wait
- Iteration 5-6: 4 seconds wait
- Iteration 7+: min(8 * 2^(iter-7), 60) seconds (max 60s)
- Record to boulder-log.jsonl (if .shinchan-docs/{DOC_ID}/ exists):
\`{"ts":"ISO8601","iteration":N,"event":"idle_detected","reason":"...","backoff_ms":N}\`
7. On failure โ Analyze cause โ Retry (max 3 retries per task)
8. On success โ Next task
9. All tasks complete โ Action Kamen final verification
10. Verification fails โ Fix and re-verify
### Safety Limits
- **Max iterations**: 15 (hard limit โ "Boulder limit reached โ manual intervention required")
- **Stagnation stop**: 2 consecutive `stagnation: true` detections โ stop + escalate before the cap (step 5)
- **Max retries per task**: 3 (report failure if exceeded)
- **Progress narration**: Every iteration, show current Phase, task, and progress indicator
- **Log file**: .shinchan-docs/{DOC_ID}/boulder-log.jsonl (append-only JSONL, skip if no active workflow)
Format: \`{"ts":"ISO8601","iteration":N,"event":"idle_detected|retry|success|limit_reached|stagnation_stop","reason":"...","backoff_ms":N}\`
### Completion Conditions
Complete only when all conditions met:
- All TODO list items completed
- Build/tests pass
- Action Kamen review approved
**If conditions not met, automatically continue (respecting safety limits)!**
User request: ${args || '(Task to complete)'}
`
)
When the loop completes (all conditions met, Action Kamen approved) and this was a
standalone invocation, run the FINISH โ In Review transition per
agents/_shared/linear-sync.md for the same issue (no-op if not Linear-based / driven
inside a workflow), then STOP.
STOP HERE. The above Task handles everything.