| name | deep-think |
| description | Mediator-first deep reasoning protocol for high-difficulty tasks. Triggers: deep-think, multi-hypothesis reasoning, structured refinement, mediator triage. |
deep-think
Usage
Use this skill when a task needs more than a single direct specialist pass: novel architecture, ambiguous tradeoffs, high-cost mistakes, or deliberate multi-hypothesis reasoning.
This is a repository-fit protocol surface for the active mediator stack. It is executed by @deep-think-mediator, not by the orchestrator or a leaf specialist directly.
Responsibility
This skill owns protocol-local reasoning flow only:
- Normalize the intake for
@deep-think-mediator.
- Triage whether the task should stay in deep-think or redirect to
@dt-council-mediator.
- Shape a bounded multi-hypothesis reasoning pass using the active agent fleet.
- Return a normalized payload back to the orchestrator.
Out of scope:
- Lifecycle writes, TODO ownership, and approval gates.
- User interaction and memory persistence.
- File editing or publication decisions.
- Direct vendor integration beyond the leaf-proxy boundary.
Active Repo Surfaces
Use the active repository surfaces instead of assuming the full reference stack:
- First protocol hop:
@deep-think-mediator
- Council redirect target:
@dt-council-mediator
- Optional Gemini runtime surface: host-provided native Gemini access outside the repo-owned runtime, when available
- Candidate reasoning participants:
@architect, @planner-gpt, @planner-gemini, @planner-claude, @research-gpt, @research-gemini, @research-claude, @validator
- Optional runtime helpers when configured:
session-gate and context-manager
Protocol
- Normalize intake
- Require: objective, constraints, budget hint, and expected output shape.
- If a protocol asset reference is provided, prefer that asset. Otherwise run this local repo-fit protocol.
- Run triage
- Classify difficulty as
lite, standard, or deep.
- Classify diversity need as
low, medium, or high.
- If diversity need is high enough that disagreement handling is the main value, return redirect metadata for
@dt-council-mediator instead of forcing deep-think.
- Build the reasoning set
- Choose 2 to 3 distinct reasoning participants from the active agent fleet.
- Prefer structural diversity over duplicated roles.
- Use a Gemini-family runtime path only if it is explicitly needed and the host environment provides native Gemini access outside the repo-owned runtime.
- Generate and critique
- Produce parallel hypotheses.
- Require each hypothesis to receive a self-critique or cross-check before synthesis.
- Keep the protocol bounded. If the budget hint is tight, reduce branch count before reducing verification.
- Synthesize
- Merge compatible findings.
- Preserve unresolved disagreements instead of flattening them.
- Return the best current answer plus explicit risks, weak points, and any redirect rationale.
Output Contract
Return one of these shapes:
deep_think_mediator_result:
status: complete
tier: lite | standard | deep
summary: <short outcome>
final_payload:
recommendation: <result>
risks: [<risk>]
open_questions: [<question>]
metadata:
protocol_asset: local-repo-fit | <path>
phases_completed: [triage, hypotheses, critique, synthesis]
participants: [<agent labels>]
deep_think_mediator_result:
status: redirect
suggested_redirect_target: dt-council-mediator
reason: <why council routing is preferred>
redirect_context: <handoff payload>
deep_think_mediator_result:
status: blocked | error
reason: <missing dependency or failure>
partial_payload: <optional>
Exit Criteria
- Intake is normalized.
- A bounded deep-think pass or a clean council redirect is produced.
- The result respects lifecycle ownership boundaries.
- No leaf vendor path is assumed without host-provided runtime access.
Failure Handling
- Missing mediator context or protocol inputs: return
blocked.
- Required Gemini leaf path unavailable: return
blocked or a degraded payload without vendor participation.
- Budget exhausted before critique and synthesis: return
error with the best partial payload.