reasoning-sentinels

Check causal reasoning in debugging, root cause analysis, or system analysis. Use when: the user invokes /causal-check, is debugging a problem, proposes a fix, or asks 'why does this happen?'. Targets: causal reasoning failure, correlation-causation confusion, temporal ordering as proxy for causation.

Challenge the current approach or reasoning with genuine counter-arguments. Use when: the user invokes /challenge, suspects sycophancy, wants a devil's advocate perspective, or asks 'am I wrong about this?'. Targets: sycophancy, anchoring bias, overconfidence, framing effects.

Decompose a complex task into atomic, independently verifiable steps before execution. Use when: the user invokes /decompose, a task has many dependent steps, the task is complex enough to risk compositional failure, or reasoning chains exceed 5 dependent steps. Targets: compositional failure, logic collapse, long reasoning chain drift.

Check for context drift in long conversations or multi-step tasks. Use when: the user invokes /drift-check, a conversation has been going for many turns, work seems to have drifted from the original goal, or the user asks 'where are we?'. Targets: context drift, memory breaks, entropy collapse, lost-in-the-middle.

Test understanding by reversing the question or relationship. Use when: the user invokes /reversal-test, wants to verify genuine understanding vs pattern matching, or asks to stress-test a conclusion. Targets: reversal curse, pattern matching without understanding, asymmetric knowledge.

Verify structural validity of formal/symbolic content like SQL, regex, math, schemas, configs, or code logic. Use when: the user invokes /symbolic-guard, generates structured content that must be formally correct, works with complex queries or expressions, or asks to verify a formal structure. Targets: symbolic collapse, constraint hallucination, prose-structure mismatch.

Detect opposing or contradictory instructions within a prompt, system prompt, or set of instructions. Use when: the user invokes /tension-check, is writing or editing an LLM prompt, has a prompt that produces inconsistent results, or asks 'why does this prompt sometimes ignore X?'. Targets: prompt tension from competing objectives, silent instruction dropping, semantic stress between opposing requirements, inconsistent LLM behavior from contradictory directives.

Trace and verify the sources behind claims, code references, and recommendations. Use when: the user invokes /trace-sources, wants to audit where information came from, needs to verify code references exist, or asks 'how do you know that?'. Targets: hallucination, black-box debugging, retrieval failures.

Validate reasoning quality after generating a response, code solution, or analysis. Use when: the user invokes /validate-reasoning, asks to check reasoning, wants to verify claims are grounded, or suspects hallucination. Targets: hallucination, unfaithful chain-of-thought, overconfidence, self-contradiction.