跳转至

TRIZ Integration in EGM Plugin

  • Author: Brian Suchy (@briansuchy)
  • Status: Draft
  • Date: 2026-11-03
  • Issue: #37

Overview

This document specifies the integration of the TRIZ method (Theory of Inventive Problem Solving) into the open-source Elephant & Goldfish Model (EGM) framework (github.com/google/elephant-goldfish-model).

EGM separates high-context conversational drafting (The Elephant) from independent, zero-context critique (Goldfish personas). Existing Goldfish personas are predominantly evaluative (they audit designs to identify flaws, risks, cognitive friction, and unnecessary complexity).

TRIZ introduces an inventive and transformative dimension. Rather than merely flagging trade-offs or accepting compromises (e.g., higher reliability requires higher latency), TRIZ systematically identifies technical contradictions and applies recognized inventive principles (such as Inversion, Segmentation, and Self-Service) to resolve the conflict without premature surrender to false compromises.

This integration delivers:

  1. A new dedicated Goldfish persona: "The Inventive Catalyst" (agents/inventive_catalyst.md), featuring a self-contained inlined catalog of 10 software engineering patterns for zero-I/O execution.
  2. Integration of the persona into the upstream Goldfish shootdown swarm in skills/egm-brainstorm/SKILL.md.
  3. Integration of the persona into Step 5 (Initial Strategic Review) in skills/egm-review/SKILL.md, pairing with egm-zoom-out to audit foundational trade-offs before detailed reviews.

Problem Statement

Software system design is routinely constrained by perceived trade-offs (such as throughput versus memory, flexibility versus simplicity, or speed versus correctness). Engineers and AI agents frequently accept these compromises as inevitable laws of architecture.

Existing Goldfish personas audit designs against specific domain standards:

  • design_critic.md: Audits infrastructure scaling, bottlenecks, and costs.
  • zoom_out.md: Enforces simplicity and flags YAGNI (You Aren't Gonna Need It) violations.
  • codebase_details.md: Validates code reality and documentation drift.
  • comprehension.md: Enforces zero-assumption readability.
  • implementation_readiness.md: Verifies unambiguous file change steps.
  • structural.md and style.md: Verify template consistency and tone.

However, none of these personas actively challenge the premise of an architectural compromise or offer structural patterns to resolve it.

Furthermore, introducing inventive leaps without grounding risks generating impractical, overly clever abstractions. To be effective in software engineering, abstract TRIZ principles (such as "Pneumatics and Hydraulics" or "Thermal Expansion") must be translated into established, actionable software architecture patterns (such as Event Streams, Elastic Buffers, or Inversion of Control) spanning both single-process tools and distributed systems.


Goals & Non-Goals

Goals

  • Dedicated Inventive Persona: Define agents/inventive_catalyst.md (egm-inventive-catalyst) to identify technical contradictions and propose pattern-based resolutions without diluting the responsibility of existing evaluative personas.
  • Dual-Phase Integration:
    • Brainstorming: Add egm-inventive-catalyst to the parallel shootdown swarm coordinated by egm-synthesizer in /egm-brainstorm (the 7-lens swarm) to expand conceptual possibilities.
    • Review Pipeline: Add egm-inventive-catalyst to Step 5 (Initial Strategic Review) in /egm-review to audit high-altitude foundational compromises alongside egm-zoom-out before line-by-line review.
  • Self-Contained Inlined Pattern Knowledge: Inline 10 curated software architecture patterns directly into the persona prompt, eliminating external file I/O and cross-harness path resolution failures.
  • Harness-Agnostic Operation: Ensure all agent definitions and workflows operate across the harnesses in docs/harnesses.md using standard subagent tool conventions.

Non-Goals

  • Modifying Existing Personas: Modifying or augmenting existing Goldfish personas (design_critic, zoom_out, etc.) with TRIZ logic is out of scope. Their focused, single-responsibility rubrics remain intact.
  • Harness-Specific Modals: Introducing custom interactive question modals (such as ask_question) is out of scope. The persona runs as a standard peer in the automated shootdown swarm, adhering to the plain-text interaction constraints of skills/egm-brainstorm/SKILL.md.
  • Redundant Catalog File: Authoring a separate resources/triz_patterns.md document is out of scope. To avoid dual-maintenance debt and runtime drift, the pattern catalog lives strictly inside the persona definition.
  • Blocking Review Loop Gating in Step 7: Adding the Catalyst to the blocking repair loop (egm-review-loop) in Step 7 is out of scope. It runs upstream in Step 5 as an advisory high-altitude audit, avoiding cyclic conflict with Step 8 Zoom Out.

Proposed Architecture

1. Persona Specification: agents/inventive_catalyst.md

The new persona will be added to the repository's agents/ directory as a standalone Markdown file conforming to the standard Goldfish harness contract:

---
name: egm-inventive-catalyst
description: Inventive problem-solving Goldfish persona applying TRIZ principles to identify and eliminate system-level contradictions without compromise.
license: Apache-2.0
---

# Core Mandates (Universal Goldfish Context)

You are a specialized Goldfish subagent operating within the Elephant & Goldfish Model. You have been delegated a specific task by the root agent (the Elephant).

1. **Strict Read-Only Protocol**: You are a strictly read-only agent. You have NO tools to modify the workspace, write files, or compile code. Do not attempt to fix or refactor any code.
2. **Stateless Mindset (Goldfish Isolation)**: You evaluate plans with zero conversational history or external assumptions. You must read the document as a naive team member or fresh reader who has just joined the project.
3. **No Chitchat**: Avoid conversational filler. Focus exclusively on technical rationale and findings. Do not narrate your tool usage.

---

# Role

You are **Inventive Catalyst Goldfish**.
You are an expert in applying TRIZ (Theory of Inventive Problem Solving) principles to software architecture.
Your mission is to identify where proposed designs prematurely surrender to false compromises or technical contradictions (e.g., "higher throughput requires sacrificing durability", "extensibility requires complex hierarchies"). You uncover hidden contradictions and suggest established software architecture patterns to resolve them.

Crucial System Boundaries:
- Theoretical invariants are immutable system bounds, not technical contradictions to be eliminated.
- Every architectural pattern introduces secondary operational trade-offs; you MUST explicitly declare these trade-offs alongside every recommendation.

---

# Curated Inventive Software Patterns

Use this inlined catalog as your foundational pattern reference:

1. **Segmentation**: Divide monolithic structures into independent, decoupled units (modular packages, worker pools, sharded data).
    * *Secondary Trade-off*: Coordination overhead, interface boundary serialization.
2. **Extraction / Taking Out**: Separate cross-cutting or interfering concerns (interceptors, sidecars, out-of-process isolation).
    * *Secondary Trade-off*: Inter-process communication latency, additional operational artifacts.
3. **Preliminary Action**: Pre-compute or pre-allocate expensive resources (memoization, cache warming, connection pooling, ahead-of-time compilation).
    * *Secondary Trade-off*: Memory consumption during idle periods, initialization delay.
4. **Beforehand Cushioning**: Prepare fallback mitigations in advance (circuit breakers, stale-while-revalidate caches, bounded fallback queues).
    * *Secondary Trade-off*: Potential data staleness, fallback path maintenance.
5. **Inversion**: Invert control flow or data access (Inversion of Control, reactive streams, event-driven pull vs. push).
    * *Secondary Trade-off*: Execution indirection, harder stack-trace debuggability.
6. **Dynamics**: Adapt system characteristics dynamically based on workload (adaptive batching, autoscaling, dynamic rate limits).
    * *Secondary Trade-off*: System behavior unpredictability under boundary conditions.
7. **Blessing in Disguise**: Convert failure signals or liabilities into beneficial system actions (chaos engineering, backpressure signals, dead-letter re-routing).
    * *Secondary Trade-off*: Increased orchestration logic and queue management.
8. **Intermediary**: Decouple components using an intermediate coordinator (mediators, event buses, message queues, API gateways).
    * *Secondary Trade-off*: Extra execution hops, single point of failure without redundancy.
9. **Self-Service**: Make components autonomously self-heal or self-manage (reconciliation loops, watchdogs, self-tuning memory allocators).
    * *Secondary Trade-off*: Complex failure loops if health check logic is flawed.
10. **Cheap Dispensables**: Replace durable stateful instances with cheap, ephemeral components (stateless workers, ephemeral build containers, immutable runners).
    * *Secondary Trade-off*: Cold start latency, loss of persistent local disk state.

---

# Instructions

Inspect the `Mode` parameter specified by the caller:

### IF Mode: BRAINSTORM (Upstream Concept Ideation)
Evaluate the feature concept from an inventive contradiction-resolution perspective:
0. **Verify Contradiction Existence**: Check whether a genuine technical contradiction exists. If the architecture is already balanced and no true compromise is being accepted, explicitly output "No false compromise detected" rather than manufacturing speculative transforms.
1. **Identify Technical Contradictions**: Detect points where improving one parameter degrades another.
2. **Suggest Inventive Possibilities**: Propose 2-3 concrete software engineering patterns from the catalog that could resolve the conflict.
3. **Formulate the Ideal Final Result (IFR)**: Ask: "How could this system deliver its core function with minimal added operational overhead?"

### IF Mode: REVIEW (Document Audit - Step 5 Initial Strategic Review)
Audit the target document at the strategic, high-altitude level:
0. **Verify Contradiction Existence**: Check whether a genuine technical contradiction exists in the document. If the foundational architecture is already balanced and no false compromise is being accepted, explicitly output "No false compromise detected" rather than manufacturing speculative transforms.
1. **Audit for False Compromises**: Inspect the foundational architecture to identify where the author accepted an unexamined trade-off or bottleneck as inevitable.
2. **Declare Secondary Liabilities**: For any pattern suggested, explicitly state its operational costs and trade-offs.

### Audit Vectors:
* **Unexamined Compromises**: Accepted performance, reliability, or maintainability bottlenecks presented as unavoidable laws.
* **Tight Architectural Coupling**: Subsystems coupled together when extraction or an intermediary would eliminate contention.
* **Unhandled Failure Potential**: Error handling treated as passive logging rather than active mitigation or backpressure.

---

# Output Format

Report your findings back to the orchestrator in plain, natural language.

#### IF Mode: BRAINSTORM
If no false compromise is detected, output: "No false compromise detected".
Otherwise, for each contradiction identified, include:
1. **Severity**: [Blocker | Major | Nit]
2. **Dimension**: Contradiction Elimination & Inventive Transforms
3. **Contradiction**: Parameter [X] versus Parameter [Y].
4. **Inventive Resolution Pattern**: Concrete software pattern recommendation from the catalog.
5. **Secondary Operational Trade-offs**: Anticipated operational overhead or liabilities.
6. **Ideal Final Result (IFR)**: What the self-sufficient architecture would look like.

#### IF Mode: REVIEW
If no false compromise is detected, output: "No false compromise detected".
Otherwise, for each finding, include:
1. **Severity**: [Major | Nit] (Advisory findings)
2. **Affected Scope**: Section header or architectural subsystem.
3. **False Compromise**: Description of the accepted trade-off in the text.
4. **Proposed Resolution**: Concrete architectural pattern to resolve the contradiction.
5. **Secondary Liabilities**: Operational overhead or complexity introduced by the pattern.

2. Workflow & Review Pipeline Integration

A. Upstream Brainstorming Integration (/egm-brainstorm)

  1. Skill Table Update: Add Inventive Catalyst to the table of Goldfish Brainstorming Lenses in skills/egm-brainstorm/SKILL.md:

    Lens / Persona Core Objective Key Questions Challenged
    7. Inventive Catalyst Contradiction elimination & inventive transforms What fundamental trade-off is this design accepting as inevitable? How can we resolve this conflict using proven software transforms to achieve the Ideal Final Result without compromise?
  2. Synthesizer Dispatch Update: In agents/synthesizer.md and skills/egm-brainstorm/SKILL.md, update the subagent dispatch list for Mode: BRAINSTORM:

    • Update Reviewers to include egm-inventive-catalyst.
  3. Generic Multi-Lens Synthesis: egm-synthesizer aggregates findings across all lenses symmetrically into the ## Analysis section of concepts_brief.md. Extend the failed-reviewer rule in agents/synthesizer.md from Mode: REVIEW to Mode: BRAINSTORM: a lens that times out or returns nothing is retried once, then reported as not run in concepts_brief.md, never as a lens with no findings.

B. Review Loop Integration (/egm-review)

  1. Step 5 (Initial Strategic Review) Pairing: In skills/egm-review/SKILL.md, update Subflow: Zoom Out Review (Step 5) to spawn two parallel high-altitude auditors:
    • egm-zoom-out: Audits for YAGNI, over-engineering, and root-problem alignment against zoom-out-standard.md.
    • egm-inventive-catalyst: Audits for foundational false compromises and inventive opportunities across full file without a document standard (relying on its self-contained inlined catalog, analogous to egm-comprehension).
  2. Advisory Checkpoint Gate: Step 5 is an interactive manual checkpoint gate. Findings from the Catalyst are reported in zoom_out_report.md under an Inventive Opportunities section. The author can choose to pivot their architecture early before proceeding. In Subflow: Uninterrupted Full Flow, these advisory opportunities in zoom_out_report.md are excluded from the mandatory defect repair traceability matrix.
  3. Unidirectional Review Pipeline: Because foundational compromises are evaluated in Step 5:
    • Step 6 (egm-comprehension) audits the doc for clarity.
    • Step 7 (egm-design, egm-structural, egm-style) audits implementation mechanics without cyclic whiplash.
    • Step 8 (egm-zoom-out) acts as the final pruning shears before Step 9 readiness testing, evaluating adopted inventive resolutions in the context of resolved trade-offs so legitimate conflict resolutions are not treated as premature complexity.

Risks & Limits

1. Hallucinating Contradictions on Hard Theoretical Invariants

  • Risk: An inventive agent may misinterpret fundamental mathematical, physical, or theoretical limits as solvable compromises, generating unfeasible proposals.
  • Mitigation: The persona prompt explicitly bounds contradiction analysis: theoretical invariants are immutable system bounds, not technical contradictions to be eliminated.

2. Secondary Operational Liabilities & YAGNI Bloat

  • Risk: Applying architectural patterns (e.g., event buses, sidecars, caching layers) to resolve minor friction introduces operational and maintenance overhead.
  • Mitigation:
    1. The Catalyst must explicitly output Secondary Liabilities for every pattern.
    2. In /egm-review, Step 5 findings are strictly advisory, and Step 8 (Intermediate Zoom Out) prunes any speculative complexity before implementation planning.

3. Swarm Execution Resilience

  • Risk: Adding a 7th subagent to the parallel shootdown swarm increases request concurrency, risking timeout failures in latency-sensitive environments.
  • Mitigation: Item 3 extends the failed-reviewer rule in agents/synthesizer.md from Mode: REVIEW to Mode: BRAINSTORM: a lens that times out or returns nothing is retried once, then reported as not run in concepts_brief.md, never as a lens with no findings.

Alternatives Considered

Alternative A: Augmenting Existing Lenses with TRIZ Logic

  • Description: Inject TRIZ prompts into design_critic.md or zoom_out.md rather than creating a new persona.
  • Pros: Avoids adding another persona file to agents/.
  • Cons: Dilutes the core mission of existing personas. design_critic is specifically designed to be skeptical and pessimistic; forcing it to also be an inventive brainstormer creates conflicting directives and lowers finding quality.
  • Reason for Rejection: EGM relies on isolated, single-responsibility critics. A dedicated persona keeps evaluative and inventive mindsets cleanly separated.

Alternative B: Interactive Multi-Select Gates via Custom Modal Tools

  • Description: Add an interactive multi-select UI dialog (e.g., ask_question) in /egm-brainstorm allowing users to selectively toggle individual Goldfish personas on or off before each shootdown.
  • Pros: Allows granular per-session customization.
  • Cons: Violates skills/egm-brainstorm/SKILL.md Constraint 2 (no structured question tools during ideation), depends on a structured question tool that only some harnesses have, and adds user friction.
  • Reason for Rejection: The automated parallel swarm handles reviews cleanly without manual toggling.

Alternative C: Synthesizer-Embedded Prompting (Zero-Spawn Alternative)

  • Description: Instruct egm-synthesizer to perform TRIZ analysis directly during synthesis rather than launching a 7th subagent.
  • Pros: Saves token expenditure and eliminates cold-start latency for one subagent.
  • Cons: Violates EGM's core invariant of zero-context Goldfish isolation. The Synthesizer's role is objective aggregation; forcing it to generate inventive patterns biases its synthesis and dilutes independent critical findings.
  • Reason for Rejection: Independent Goldfish isolation produces higher quality, unpolluted critique.

Alternative D: In-Prompt Pattern Knowledge vs. External Reference File

  • Description: Author a separate resources/triz_patterns.md document for the agent to read at runtime.
  • Pros: Keeps the persona prompt shorter.
  • Cons: Introduces file I/O latency, creates dual-maintenance drift between documentation and prompt, and fails with FileNotFound across diverse harnesses in client repositories.
  • Reason for Rejection: Inlining the 10 core pattern summaries into the system prompt guarantees zero-I/O, deterministic execution in all harnesses without dead artifacts.

Alternative E: Step 5 Strategic Audit vs. Step 7 Blocking Review Loop

  • Description: Add egm-inventive-catalyst to the Step 7 (egm-review-loop) blocking repair loop.
  • Pros: Runs alongside design critics on FeatureRange.
  • Cons: Causes severe cyclic whiplash with Step 8 (egm-zoom-out), which immediately prunes complexity added in Step 7. Furthermore, mapping to feature-spec-standard.md creates a phantom contract.
  • Reason for Rejection: Placing the Catalyst in Step 5 provides the earliest strategic trade-off audit before line-by-line review, creating a unidirectional pipeline without blocking loops.

Detailed Implementation Plan

The implementation is partitioned into four sequential phases, strictly enforcing the Zero-Decision Rule and establishing deterministic verification checkpoints at each stage.

Phase 1: Agent Persona Creation & Manifest Registration

1. agents/inventive_catalyst.md (Create)

  • Type / Scope: Agent Persona Specification
  • Rationale: Define the system prompt, mandates, role, inlined pattern catalog, and output format for egm-inventive-catalyst.
  • Proposed Changes: Author the file following the exact schema specified in Section 1 of the Proposed Architecture.

2. Host packages (Modify)

  • Type / Scope: Plugin manifests and host adapters
  • Rationale: Every host that ships the reviewers needs the new persona.
  • Proposed Changes: Add "./.claude-plugin/agents/inventive_catalyst.md" to the agents array in .claude-plugin/plugin.json. Add the persona to .claude-plugin/agents/, .cursor-plugin/agents/, com.github.copilot/agents/egm-inventive-catalyst.agent.md, the reviewer list in adapters/opencode/install.py, and the persona list in scripts/pack_plugin.py. Update reviewer and adapter counts in docs/harnesses.md from nine to ten.

Phase 1 Checkpoint: Validate YAML frontmatter across all agent files using pytest tests/unit/test_frontmatter.py.


Phase 2: Workflow Skill Orchestration

3. agents/synthesizer.md (Modify)

  • Type / Scope: Intermediate Orchestrator & Aggregator
  • Rationale: Instruct egm-synthesizer to launch and aggregate findings from inventive_catalyst during brainstorming.
  • Proposed Changes:
    • In IF Mode: BRAINSTORM Step 1: Update the subagent dispatch list to launch the 7-lens swarm in parallel, adding inventive_catalyst with the role title Inventive Catalyst Brainstormer.
    • In IF Mode: BRAINSTORM Step 1: Add the Failed Reviewers rule from Mode: REVIEW Step 1: retry a lens once, then report it as not run in concepts_brief.md, never as a lens with zero findings.
    • In IF Mode: BRAINSTORM Step 2: Update aggregation instructions: "Collect findings from all 7 Goldfish, deduplicate overlapping points, synthesize inventive contradictions and pattern resolutions alongside risk dimensions, and compile them into concepts_brief.md."
    • In Step 5 of egm-review: egm-inventive-catalyst is dispatched alongside egm-zoom-out; findings under ## Inventive Opportunities are recorded directly into zoom_out_report.md without altering goldfish_review_report.md's strict Blocker/Major/Nit defect schema.

4. skills/egm-brainstorm/SKILL.md (Modify)

  • Type / Scope: Core Skill Workflow
  • Rationale: Register egm-inventive-catalyst in the brainstorming shootdown swarm.
  • Proposed Changes:
    • Add the 7th lens to the table: | **7. Inventive Catalyst** | Contradiction elimination & inventive transforms | *What fundamental trade-off is this design accepting as inevitable? How can we resolve this conflict using proven software transforms to achieve the Ideal Final Result without compromise?* |
    • Update Step 2.1 lead-in to include egm-inventive-catalyst in the Reviewers array parameter: Reviewers: [egm-design, egm-codebase-details, egm-zoom-out, egm-comprehension, egm-implementation-readiness, egm-structural, egm-inventive-catalyst].
    • In the embedded concepts_brief.md template, extend the ## Analysis placeholder to cover inventive contradictions, the catalog pattern suggested for each, its secondary trade-offs, and the Ideal Final Result.

5. skills/egm-review/SKILL.md (Modify)

  • Type / Scope: Review Orchestrator
  • Rationale: Integrate egm-inventive-catalyst into Step 5 (Initial Strategic Review) without colliding with Step 8 post-critic pruning, while staying strictly under the 24,000-byte skill load limit enforced by tests/unit/test_skill_load_limits.py.
  • Skill Size Budget: The egm-review body is 23,743 bytes at main, leaving 257 bytes of headroom. To accommodate Step 5 without exceeding the limit:
    1. The detailed report schema for Inventive Opportunities is defined strictly inside agents/inventive_catalyst.md's # Output Format, avoiding a duplicate multiline schema block in SKILL.md.
    2. Subflow: Zoom Out Review stays shared. Its Step 5 case and the Step 5 bullet in Subflow: Uninterrupted Full Flow add egm-inventive-catalyst. This brings the body to about 23,885 bytes, so nothing else needs cutting.
  • Proposed Changes:
    • In Subflow: Zoom Out Review step 1, after the Step 8 clause, add: "When running as Step 5, also spawn egm-inventive-catalyst in parallel and write both into zoom_out_report.md." Step 8 still runs only egm-zoom-out.
    • In the Step 5 bullet of Subflow: Uninterrupted Full Flow, which the Consolidated report preset runs, change Reviewers: [egm-zoom-out] to Reviewers: [egm-zoom-out, egm-inventive-catalyst].

Phase 2 Checkpoint: Validate skill frontmatter and subflow parsing via pytest tests/unit/test_frontmatter.py.


Phase 3: Global Documentation & Lens Count Synchronization

6. docs/designs/review.md (Modify)

  • Type / Scope: Architectural Specification
  • Rationale: Reconcile the review pipeline specification with Step 5 dual-auditor integration.
  • Proposed Changes:
    • Update the Mermaid flowchart node Step5 to Step 5: Initial Strategic Review (Zoom Out + Inventive Catalyst).
    • Update the Step 5 Stage Breakdown section to document the parallel execution of zoom_out and inventive_catalyst.
    • Add agents/inventive_catalyst.md to the Reviewer Persona Catalog table.

7. Design System Documentation (Modify)

  • Type / Scope: Architectural Specifications
  • Rationale: Synchronize component inventories and lens counts across design specs.
  • Proposed Changes:
    • docs/designs/architecture.md: Update 6-lens shootdown swarm references to 7-lens, and add inventive_catalyst to the Personas Layer listing.
    • docs/designs/brainstorm.md: Update Mermaid diagrams and lens descriptions from 6 to 7 lenses, and map inventive_catalyst.md to the Inventive Catalyst lens in the file mapping table.
    • docs/designs/synthesizer.md: Update the 6-lens swarm section to 7-lens, and add inventive_catalyst to the brainstorm perspectives list.
    • docs/designs/README.md: Update the existing entry for triz_integration.md to remove reference to rejected "dynamic reviewer selection" and reflect integration into /egm-brainstorm and Step 5 of /egm-review.

8. User-Facing Documentation (Modify)

  • Type / Scope: Documentation
  • Rationale: Ensure getting started documentation reflects multi-lens brainstorming without hardcoding brittle lens counts.
  • Proposed Changes:
    • docs/getting_started.md: Change the "6-lens Goldfish" wording in the brainstorming overview to "multi-lens Goldfish".

Phase 3 Checkpoint: Validate link integrity across all documentation files using pytest tests/unit/test_link_integrity.py.


Phase 4: Test Suite Updates & Behavioral Verification

9. Test Specifications & Behavioral Scenarios (Modify)

  • Type / Scope: Test Suite & Evaluation Scenarios
  • Rationale: Ensure automated test coverage for the 7-lens swarm and Step 5 Catalyst integration across all harnesses.
  • Proposed Changes:
    • skills/egm-brainstorm/TEST.md: Update test expectations and lens enumeration to include Inventive Catalyst (7 lenses).
    • skills/egm-review/TEST.md: Update Step 5 test expectation to verify Initial Strategic Review with parallel Zoom Out and Catalyst execution.
    • skills/egm-brainstorm/test_scenarios.yaml: Add behavioral test scenarios asserting that the ## Analysis section names an inventive contradiction, the catalog pattern suggested for it, and its secondary trade-offs, plus a scenario validating the "No false compromise detected" early-exit guardrail when the input is already optimal.
    • skills/egm-review/test_scenarios.yaml: Add scenarios validating detection of false architectural compromises in Step 5 and the early-exit guardrail.
    • tests/unit/test_frontmatter.py: Add validation asserting frontmatter conformance for all markdown files under agents/.
    • adapters/opencode/install.py: Update agents["egm"]["prompt"] lead-in string from "complete all six brainstorm reviewers" to "all seven".
    • tests/unit/test_opencode_adapter.py: Update agent count assertion from 8 to 9 (assert len(config["agent"]) == 9).
    • tests/unit/test_harnesses.py: Add "inventive_catalyst" to the central REVIEWERS tuple.
    • scripts/pack_plugin.py: Add agents/inventive_catalyst.md to the REQUIRED_SHARED list to ensure Claude distribution packaging passes tests/unit/test_packaging.py.

Phase 4 Checkpoint: Run the unit suite. test_skill_load_limits.py and test_scenario_coverage.py cover the skill changes:

pytest tests/unit

Verification Plan

  • Markdown Formatting & Linting: Run mdformat across all created and modified documentation files:

    mdformat agents/ skills/ docs/ README.md
    
  • Static Test Suite Execution: Run the test suite from the repository root:

    pytest tests/unit
    

本页原文 Markdown:在 AtomGit 查看·内容源自开源项目 el/elephant-goldfish-model