MCP Shrimp Execution Loop

mcp-shrimp-execution-loop

Runs the execute → verify → update cycle against Shrimp's persistent task store. Planning new epics belongs in task-planning-shrimp; same-session-only work belongs in Plan Mode.

Activation gate

Activate when:

• Shrimp tasks already exist (list_tasks shows in_progress or pending), or
• split_tasks just created a graph and user wants execution, or
• /next selected a Shrimp-backed next task.

Skip when:

• No Shrimp MCP available — use Agent + Plan Mode instead.
• User is in ASK mode (read-only; no execute_task / file edits).
• Task is a one-shot change with no Shrimp taskId.

Resume protocol (/next, new chat)

1. list_tasks(status: "in_progress") — if any, resume first (get_task_detail → execute_task).
2. Else list_tasks(status: "pending") — pick first task whose dependencies are all completed (check names against completed list).
3. If global volume has unrelated tasks, filter by epic keywords in task name / globalAnalysisResult (see task-planning-shrimp global-volume warning).

Execution loop (one task at a time)

Shrimp expects one task per cycle unless user requests continuous mode.

┌─────────────────────────────────────┐
│ 1. get_task_detail(taskId)          │
│ 2. execute_task(taskId)             │  ← follow returned guidance; NOT "done" yet
│ 3. Implement in repo (minimal diff) │
│ 4. Verify (tests/lint/run)          │
│ 5. verify_task(taskId, score, summary) │  score ≥80 → auto-completes
│ 6. If score <80: fix → re-verify    │
└─────────────────────────────────────┘
         │
         ▼
   Next pending with satisfied deps, or stop

Step details

execute_task — returns instructional guidance only. You must still edit files and run verification. Never mark done immediately after the tool call.

verify_task — requires score (0–100) and summary (≥30 chars). Use verificationCriteria from get_task_detail. Score ≥80 completes the task in Shrimp.

update_task — use mid-flight for notes, relatedFiles, or dependency fixes; completed tasks allow only summary + relatedFiles.

Blocked dependencies — do not skip silently. Either complete prerequisites first or ask user to reprioritize / update_task dependencies.

Pairing with other tooling

After epic completion: propose /save_memory with outcome — do not auto-save.

Continuous mode

Only if user explicitly asks ("continuous", "ciągły", "all remaining tasks"):

• Loop pending tasks in dependency order.
• Stop on: failing verify_task (<80), ambiguity, destructive ops needing confirmation, or blocker outside scope.

Global volume discipline

Same as planning skill: one shrimp_data volume for all projects. When executing, confirm tasks belong to this repo/epic before changing code. If mixed backlog, ask user which task ID to run.

Output (each cycle)

Shrimp: <task name> (<taskId>) — <status after verify>
Done: <1 line>
Next: <task name + taskId | "none — epic complete">
Verify: <command run + result>

Final epic summary: counts completed/skipped, link to memory save prompt if useful.

Anti-patterns

• Using Shrimp to re-plan what Plan Mode already approved in the same session.
• clearAllTasks mid-execution without user consent.
• Parallel execute_task on dependent tasks.
• Claiming task complete without verify_task or repo verification.