zy839971925-zyy/Agent--deep-research-Workflow-Skills / integrations/deepseek-harness/skills/reasoning-workflow
reasoning-workflow
VerifiedDefault silent workflow for everyday questions, research, writing, design and engineering. Match effort to the request, resolve costly ambiguity, find specialist resources and verify the requested outcome. Explicitly selectable with “调用我的 workflow skills”.
references/runtime-and-delegation.md
View on GitHub> Upgrade 2.1: read the updated [host contracts](host-adapter.md) and [migration rules](../UPGRADE.md) for executable admission, review and storage semantics. The shorter root SKILL is the active entry. Older examples are descriptive and must satisfy the current schemas.
# Runtime and delegation
**Trigger:** Tool-rich, cross-surface, private, delegated, sandboxed, capability-limited, or long-running tasks.
**Reads:** objective, required evidence/actions, available tools/surfaces, authorization, private-data constraints, current state.
**Updates:** capability plan, delegation branches, private/public routing, failure/blocked state.
**May invalidate:** plans that assumed unavailable capability or unauthorized access.
**Exit:** the task uses only capabilities actually available and authorized; delegated work has been integrated into the root state.
**Related:** [state model and invariants](state-model-and-invariants.md), [inquiry and research](inquiry-and-research.md), [evidence and provenance](evidence-and-provenance.md), [synchronization and recovery](synchronization-and-recovery.md).
**Return:** [root workflow](../SKILL.md) → Action / Project Lane or Inquiry loop.
## Capability preflight
Distinguish model capability, product/harness capability, connected tools, storage persistence, and user authorization. Web, files, private connectors, local repository, shell/code execution, media inspection, computer use, subagents, scheduling, and durable storage can vary by environment.
Do not assume a capability because another ChatGPT/Codex surface supports it. If the most direct surface is unavailable, use the best available route and narrow confidence or deliverable where necessary.
An attached ZIP supplies instructions but does not automatically install a persistent Skill or grant external tools.
## Private evidence routing
Classify material evidence where useful as public, user-provided, workspace/local-private, or confidential/restricted. Access for analysis does not imply authorization to publish or send the same content to a public search/tool.
When private evidence must be combined with public research, minimize disclosure. Generalize public queries, remove unnecessary identifiers, or perform joins locally when possible. Preserve the source visibility in provenance state.
## Delegation
Delegate only separable work that benefits from parallelism, independent discovery, another language/jurisdiction, reproduction, specialist method, or adversarial search. Root synthesis retains definitions, provenance deduplication, cross-branch conflict resolution, and final judgment.
### Hire by bottleneck
Do not start from an agent count. Start from the current bottleneck. A worker is justified when the parent is materially waiting on a branch such as research, verification, specialist execution, independent reproduction, or repeated rework that a scoped owner can resolve more effectively than the coordinator.
If the handoff cost, shared-state coupling, or reconciliation burden is likely to exceed the benefit, keep the work serial. More agents can reduce diversity or slow convergence when they share the same assumptions and evidence.
### Worker contract before handoff
Give a worker a compact contract, not an ambiguous role persona:
`job | allowed inputs/sources | judgment boundary | acceptance criteria | output contract | forbidden actions | uncertainty to preserve`
The contract should make clear what the worker owns and what remains a controller decision. Do not grant canonical-state authority, publishing authority, deletion authority, spending authority, or other external permission merely because the worker owns a branch.
### Handoff packet
Return state-bearing evidence rather than a long conversational transcript. A useful handoff packet is:
`task | status | material output | evidence/source locators | decisions made within scope | unresolved uncertainty | failed routes | state/items affected | next owner or next safe action`
When a shared workspace exists, use it as an internal bus for durable evidence, drafts, decisions, and handoffs; chat summaries are views of that state, not the authoritative state itself. This reduces copy-paste drift between workers.
Keep each handoff within the actual host's input/context limits: preserve the objective, constraints, named gap, essential source locators, and return contract; point to durable materials instead of copying long transcripts. If admission fails for size, treat the branch as not started, remove duplication or split only along separable evidence routes, then verify that the retry was admitted. Never truncate a material constraint or report scheduled work as completed; if a safe compact handoff still cannot fit, continue serially or state the concrete limit.
Do not ask subagents to write interdependent chapters in parallel when shared definitions or evidence state would drift. Multiple agents are not independent empirical sources.
## Root cancellation and late worker results
When the user or trusted controller cancels the root task, keep it terminal. A late worker report is evidence only: it cannot reopen the root, schedule another branch, or authorize a write or publication. Preserve a useful result for audit or a later explicit resumption, but do not continue stopped work merely because a worker finishes afterward.
Cancellation does not prove that an already-running worker stopped before an external side effect. If a side effect may have committed and its outcome matters, reconcile the target state through an authorized read-only route; do not blindly repeat it or claim it was canceled. Resume only after a new direct user request explicitly reopens the task, then reconcile the current scope and state. Instructions alone cannot enforce a cancellation fence on an already-running executor; only a capable host can do that.
## Specialist Skills
This Skill is the process/orchestration layer. Use specialist Skills when they own a better domain method or artifact capability. Preserve one governing objective and completion state while allowing the specialist to control implementation details.
Do not duplicate or override specialist instructions without a task-relevant reason. If two Skills conflict, higher-priority user/system instructions govern; otherwise prefer explicit ownership boundaries and reconcile outputs through the canonical state.
## GPT-6+ and frontier-model runtime adaptation
For GPT-6-generation and later models, including current Claude frontier models, adapt to actual runtime capabilities rather than hard-coding model versions. Use persisted reasoning when stable, compact when context becomes noisy, delegate when independent exploration or specialist work improves coverage, and move deterministic filtering or aggregation into code where available. Do not force fixed agent counts or uniform maximum reasoning effort across mechanical phases.
## Failures, triggers, and retries
When the runtime supports event-driven triggers, prefer a narrow event or state-change signal over constant polling. Poll only when no reliable trigger exists and the expected information value justifies the cost.
Retry only when a parameter change, alternate tool/surface, transient failure, new evidence, or different route creates a concrete prospect of success. Do not repeat unchanged failures. Record the consequence for the user outcome and state, not noisy tool logs.
When a capability is unavailable, leave a bounded gap or blocked item instead of inventing completion.
## Diagnose the failure before replanning
For material failure, distinguish where the chain broke before changing the plan:
`goal understanding → observation → interpretation → state carry → plan → tool selection → tool execution → dependency resolution → replan → verification → closure`
A later capability should not be blamed when the required state was never observed, was misread, or was lost upstream. Record the smallest useful failure class; do not create a heavy ontology for routine errors.
## Replan only at the affected level
Not every failed action invalidates the plan.
- **local execution failure:** repair parameters, retry with changed conditions, or use an alternate tool;
- **plan assumption invalidated:** reopen and replan the affected branch;
- **objective, constraint, or frame changed:** rebuild the relevant higher-level plan/model.
Trigger broad replanning from model-changing evidence, not from every transient tool error. Preserve unaffected plan semantics and completed work.
[← Return to root workflow](../SKILL.md)
## 2.1.1 ordering clarification
`parallel_policy=forbidden` cannot be bypassed by omitting `concurrency_group` or using different group names. In parallel/hybrid schedules a forbidden node must be explicitly ordered against every other assigned node through an acyclic dependency/`starts_after` chain; serial schedules retain ordered semantics. Group labels are descriptive, not host-enforced isolation or sequencing. Independent allowed nodes remain eligible for parallel scheduling.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation). | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK). | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |