qwen-sci 0.2.2__py3-none-any.whl

src/agents/idea_agent/agent/skills/edit_operator_skills/surgical-modularity/SKILL.md

@@ -0,0 +1,39 @@
+---
+name: surgical-modularity
+description: Perform a localized block-level intervention with replace/rewire and prove gain with focused ablation.
+---
+
+## defect_tags
+- feature_dumping
+- monolithic_design
+- harder_to_ablate
+
+## guardrails
+- Touch only one weak block.
+- Keep interfaces explicit after rewiring.
+- Include a mandatory block-level ablation.
+
+## structural_mode
+- local_refinement
+
+## scope_preference
+- existing_component
+
+## requires_control_centered_parent
+- false
+
+## atomic_blueprint
+- REPLACE_COMPONENT(weak_block -> modular_block)
+- REWIRE(modular_block -> downstream_interface)
+- ADD_PROTOCOL(ablation)
+
+## required_protocols
+- ablation
+
+## avoid_combinations
+- ADD_COMPONENT(extra_auxiliary_stack) in the same plan
+
+## execution_logic
+1. Identify the defect from `defect_tags`.
+2. Generate the solution by **instantiating** the `atomic_blueprint`.
+3. Format: `FUNCTION_NAME(ARGUMENTS) -> REASONING`

src/agents/idea_agent/agent/skills/edit_operator_skills/surgical-modularity/references/modularity_patterns.md

@@ -0,0 +1,22 @@
+# Surgical-Modularity Patterns
+
+Use these cards as soft mechanism patterns and failure-mode checks for `surgical-modularity`.
+Reuse the pattern, not the literal module names.
+
+## Weak-Block Swap
+
+- Best for defects: `feature_dumping`, `monolithic_design`, `harder_to_ablate`
+- Core move: replace one weak block with one modular block that exposes a clearer input-output interface and a narrower causal claim
+- Injection point: an existing scoring, aggregation, or update block already blamed for failure
+- Keep fixed: upstream and downstream modules, training setup, and evaluation suite wherever possible
+- Minimal validation: swap only this block back to the baseline while keeping interfaces and instrumentation unchanged
+- Failure boundary: the "localized" edit silently changes multiple behaviors such as data flow, objective, and scheduler at once
+
+## Interface-First Refactor
+
+- Best for defects: `monolithic_design`, `harder_to_ablate`
+- Core move: rewrite a tangled block around one explicit downstream contract so the module can be studied or removed without rewriting the rest of the pipeline
+- Injection point: immediately before an existing downstream interface or consumer module
+- Keep fixed: the semantics expected by the downstream consumer unless the compiled plan explicitly changes them
+- Minimal validation: report whether the module can be bypassed or replaced without collapsing the surrounding system
+- Failure boundary: the new module is only a wrapper around the old behavior and does not isolate a real mechanism

src/agents/idea_agent/agent/skills/edit_operator_skills/theory-transfer-injection/SKILL.md

@@ -0,0 +1,40 @@
+---
+name: theory-transfer-injection
+description: Inject a theory-backed mechanism or invariant from another domain and wire it into the core objective or mechanism rule.
+---
+
+## defect_tags
+- stagnant_novelty
+- theory_gap
+- weak_generalization
+
+## guardrails
+- Name the transferred principle and where it enters the core path.
+- Limit transfer to one main mechanism instead of a bundle of unrelated ideas.
+- Add the minimum stress coverage needed to detect negative transfer.
+
+## structural_mode
+- objective_injection
+
+## scope_preference
+- core_objective
+
+## requires_control_centered_parent
+- false
+
+## atomic_blueprint
+- ADD_COMPONENT(theory_transfer_module)
+- REWIRE(theory_transfer_module -> core_objective)
+- ADD_PROTOCOL(ablation,stress)
+
+## required_protocols
+- ablation
+- stress
+
+## avoid_combinations
+- ADD_COMPONENT(unrelated_parallel_module) in the same plan
+
+## execution_logic
+1. Identify the defect from `defect_tags`.
+2. Generate the solution by **instantiating** the `atomic_blueprint`.
+3. Format: `FUNCTION_NAME(ARGUMENTS) -> REASONING`

src/agents/idea_agent/agent/skills/edit_operator_skills/theory-transfer-injection/references/transfer_patterns.md

@@ -0,0 +1,24 @@
+# Theory-Transfer Injection Patterns
+
+Use these cards as soft transfer patterns and negative-transfer checks for `theory-transfer-injection`.
+Reuse the imported principle, not the literal names from the source domain.
+
+## Lagrangian Constraint Coupling
+
+- Source principle: constrained optimization with a dual variable
+- Imported invariant: performance improves only while a resource or quality constraint stays satisfied
+- Injection point: one scalar rule such as retrieval scope, allocation weight, or facet allocation
+- Concrete rewrite: maintain a multiplier that reweights the target variable when the observed quality or consistency constraint is violated
+- Keep fixed: backbone modules, memory schema, and retrieval stack unless the compiled plan explicitly edits them
+- Minimal validation: compare against a static rule and a heuristic reweighting scheme with the same monitoring setup
+- Negative transfer signal: the dual variable reacts faster than the environment changes and starts chasing noise instead of the true constraint
+
+## Asymmetric Persistence Rule
+
+- Source principle: control systems avoid switch chatter by using separate enter/exit boundaries
+- Imported invariant: state transitions should require stronger evidence to flip back than to stay in the current regime
+- Injection point: mode switch, fallback criterion, or multi-path interface
+- Concrete rewrite: replace one symmetric trigger rule with asymmetric entry and exit conditions plus a short persistence window
+- Keep fixed: the scorer that produces the switching signal; only alter the switching law
+- Minimal validation: compare against a single symmetric trigger rule under drift and bursty noise
+- Negative transfer signal: hysteresis masks real rapid changes and delays necessary adaptation