Menu
Compatibility-only helper for legacy or ad-hoc DS delegation.
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/edwinhu/workflows --skill ds-delegateThe command stays on one line. Scroll horizontally to inspect it before copying.
Prefer a local copy? Download the files currently available to SkillsMP.
Showing SKILL.md
| name | ds-delegate |
| description | Compatibility-only helper for legacy or ad-hoc DS delegation. |
| user-invocable | false |
| disable-model-invocation | true |
| hooks | {"PostToolUse":[{"matcher":"Agent","hooks":[{"type":"command","command":"bun ${CLAUDE_PLUGIN_ROOT}/hooks/ds-post-subagent-guard.ts"}]}],"PreToolUse":[{"matcher":"Agent","hooks":[{"type":"command","command":"bun ${CLAUDE_PLUGIN_ROOT}/hooks/ds-pre-subagent-clear.ts"}]},{"matcher":"Read","hooks":[{"type":"command","command":"bun ${CLAUDE_PLUGIN_ROOT}/hooks/ds-read-after-subagent-guard.ts"}]},{"matcher":"Grep","hooks":[{"type":"command","command":"bun ${CLAUDE_PLUGIN_ROOT}/hooks/ds-read-after-subagent-guard.ts"}]},{"matcher":"Glob","hooks":[{"type":"command","command":"bun ${CLAUDE_PLUGIN_ROOT}/hooks/ds-read-after-subagent-guard.ts"}]},{"matcher":"Write","hooks":[{"type":"command","command":"bun ${CLAUDE_PLUGIN_ROOT}/hooks/orchestrator-mutation-guard.ts --workflow ds"}]},{"matcher":"Edit|MultiEdit|NotebookEdit","hooks":[{"type":"command","command":"bun ${CLAUDE_PLUGIN_ROOT}/hooks/orchestrator-mutation-guard.ts --workflow ds"}]},{"matcher":"Bash","hooks":[{"type":"command","command":"bun ${CLAUDE_PLUGIN_ROOT}/hooks/orchestrator-mutation-guard.ts --workflow ds"}]}]} |
ds-delegate is retained only for legacy or explicitly ad-hoc work. It is not the normal DS
implementation path: approved native-plan work must use and its shared sequential
ready-wave runner plus independent verifier loop.
ds-implementA shortcut that bypasses the agent's output evidence and technical review is not helpful: it leaves the user with an unverified analysis result.
Never call this from ds-implement for an approved native-plan task. Use it only when an explicit
legacy/ad-hoc compatibility request cannot use the native runner.
Before dispatching, read:
TaskList — the live task ID, status, dependencies, and any task comments.The receipt-selected native generated plan, identified by its exact {planFile, planHash}, is the approved specification; TaskList is live workflow state; and project auto-memory is reusable project knowledge. The main orchestrator supplies exact relevant auto-memory topic-file paths; do not discover, create, or write auto-memory files. Do not create or use SPEC.md, STATE.md, LEARNINGS.md, or agent-specific memory files.
This flowchart is authoritative.
Read receipt-selected `{planFile, planHash}` generated plan + TaskList + relevant project auto-memory
|
v
Dispatch appropriate implementer
|
needs clarification? -- yes --> Record in TaskList; route scope changes through planning
|
no
v
Read returned evidence and reusableFacts
|
v
Technical PASS inside ds-implement? -- no --> Fix-and-verify loop (max 3)
| |
yes still failing
v v
Update TaskList / offer reusableFacts Escalate to user
|
v
Continue ds-implement; human ds-accept only after all implementation tasks finish
Use the full task text from the receipt-selected {planFile, planHash} generated plan, never a summary.
| Task type | Dispatch |
|---|---|
engineering — ETL, schema, joins, pipelines, cleaning, transformations | workflows:ds-engineer |
analysis — estimation, modeling, tests, visualization, summaries | workflows:ds-analyst |
| Missing or ambiguous | Default to workflows:ds-analyst; record the ambiguity in the task report. |
Match the model tier to the task: cheapest capable for mechanical one-file work, standard for integration, and the most capable available for methodology design or technical review.
Dispatch one fresh implementation agent with the applicable constraints and this information:
# Approved Task
[Paste the complete task text and acceptance criteria from the receipt-selected generated plan]
# Live Task
TaskList ID: [id]
Status and dependencies: [live TaskList details]
# Project Context
[Only relevant, curated facts from project auto-memory]
[Relevant data/source paths and prior verified outputs]
# Required Work
- Implement only this approved task.
- Follow output-first verification: show the relevant before/after state, checks, and output locations.
- Load the applicable DS constraints and data-quality checks.
- Do not invent requirements. If the plan is ambiguous, stop and ask a precise question.
- Operational clarification may be recorded in TaskList. If it changes scope, acceptance criteria, methodology, deliverables, or data treatment, block the task and route it through planning for a newly approved receipt-selected generated plan; TaskList is not a substitute specification.
- Do not modify the receipt-selected generated plan or create SPEC.md, STATE.md, LEARNINGS.md, or agent-memory.
# Return exactly
1. status: COMPLETE | BLOCKED | NEEDS_CLARIFICATION
2. evidence: commands/checks run, their salient results, and output file paths
3. changedFiles: files created or changed
4. blockersOrQuestions: [] or precise blockers/questions
5. reusableFacts: [] or short, durable project facts worth the main orchestrator considering for project auto-memory
For every compatibility task, require the agent to read all three aggregate indexes:
${CLAUDE_SKILL_DIR}/../../references/constraints/ds-common-conventions.md${CLAUDE_SKILL_DIR}/../../references/constraints/ds-analysis-constraints.md${CLAUDE_SKILL_DIR}/../../references/constraints/ds-engineering-constraints.md${CLAUDE_SKILL_DIR}/../../skills/ds-verify/references/ds-checks.mdAdd the task-specific atomic constraint files named by those indexes. Require deterministic schema and join checks for engineering work; require assumption, specification, robustness, and standard-error checks when applicable for analysis work.
After an agent returns, read its returned report and evidence. Main chat may check that named outputs exist and may update TaskList, but it must not inspect source/data, run implementation code, query data stores, or repair the work. A failure or doubt goes back to a fresh implementation agent.
ds-implement owns technical PASS. For each attempt:
PASS or actionable ISSUES with file references.PASS permits the task to be marked complete in TaskList.If evidence is absent, incomplete, or the verifier returns ISSUES, redispatch a fresh implementation agent with the exact findings. There are at most three total fix-and-verify cycles per task. After the third unresolved cycle, mark or retain the task as blocked in TaskList and escalate the specific evidence and blocker to the user.
Do not dispatch the human ds-accept during this loop. ds-accept is the post-implementation human review phase, after all implementation tasks have technical PASS.
On technical PASS:
reusableFacts unchanged to the main orchestrator. The main orchestrator alone decides whether a fact is durable enough to curate into project auto-memory.| About to | STOP because | Do instead |
|---|---|---|
| Edit or infer a new requirement from an implementation result | The receipt-selected {planFile, planHash} generated plan is the approved scope | Escalate the ambiguity or create a new TaskList task only through the approved planning path. |
| Mark a task complete from an agent's prose alone | A completion claim is not evidence | Read evidence and obtain technical PASS. |
Write LEARNINGS.md, STATE.md, SPEC.md, or agent memory | These conflict with the receipt-selected generated-plan/TaskList/project-memory architecture | Return reusableFacts for the main orchestrator to curate. |
Send work to ds-accept before technical PASS | Human review cannot replace implementation verification | Finish the bounded technical loop inside ds-implement. |
| Fix an agent's work in main chat | It defeats delegated, evidence-based implementation | Redispatch a fresh task agent with precise findings. |