ds-delegate

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 Compatibility Path

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 ds-implement and its shared sequential ready-wave runner plus independent verifier loop.

The Iron Law of Delegation

**NO IMPLEMENTATION OUTSIDE A TASK AGENT.** The orchestrator delegates each approved task; it does not write analysis code, edit notebooks, run quick data checks, or make a one-off plot.

A 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.

Authoritative Context

Before dispatching, read:

1. the receipt-selected generated plan — the exact, approved task text and acceptance criteria. Do not edit it.
2. TaskList — the live task ID, status, dependencies, and any task comments.
3. The project's auto-memory topic files — reusable, curated technical facts relevant to this task.

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.

Flowchart: Per-Task Implementation

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

Select the Implementer

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 Template

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.md

Add 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.

Post-Subagent Boundary

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.

Technical Verification Loop

ds-implement owns technical PASS. For each attempt:

1. IDENTIFY — map the returned evidence to the task's receipt-selected generated-plan acceptance criteria.
2. READ — read the actual returned evidence, not a bare completion claim.
3. VERIFY — dispatch a fresh, read-only technical verifier for every task. For mechanical tasks its review may be limited to the receipt-selected generated plan's named acceptance checks and output evidence; for complex work it checks implementation and outputs in depth. It returns explicit PASS or actionable ISSUES with file references.
4. CLAIM — only a technical 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.

Close the Task

On technical PASS:

1. Update the corresponding TaskList item to completed, with concise evidence and output locations.
2. Return reusableFacts unchanged to the main orchestrator. The main orchestrator alone decides whether a fact is durable enough to curate into project auto-memory.
3. Immediately proceed to the next unblocked TaskList task. Do not pause between approved tasks.

Red Flags — STOP

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.

Failure Handling

• Needs clarification: record the precise question and status in TaskList. Resolve an operational question before redispatching; if its answer 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.
• Missing outputs or failed checks: redispatch with the missing evidence or verifier findings; do not investigate independently.
• Task is no longer suitable for the approved scope: do not mutate the receipt-selected generated plan. Escalate or route the proposed change through the planning workflow.
• Protocol violation: if implementation was performed in main chat, delete that unverified work and restart the task through a fresh agent.