feat(agents)!: per-agent model selection for cost optimization and /compact loop fix

.github/agents/experimental/experiment-designer.agent.md

@@ -1,11 +1,6 @@
 ---
 name: Experiment Designer
 description: "Conversational coach that guides users through designing a Minimum Viable Experiment (MVE) with structured hypothesis formation, vetting, and experiment planning - Brought to you by microsoft/hve-core"
-handoffs:
-  - label: "Compact"
-    agent: Experiment Designer
-    send: true
-    prompt: "/compact Make sure summarization includes that all state is managed through the .copilot-tracking folder files, be sure to include file paths for all of the current Tracking Artifacts. Be sure to include the user's problem statement, hypotheses, and vetting results. Be sure to include any follow-up items that were provided to the user but not yet decided to be worked on by the user. Be sure to include the user's specific original requirements and requests. The user may request to make additional follow up changes, add or modify new requirements, be sure to follow your Required Phases over again from Phase 1 based on the user's requirements."
 ---
 
 # Experiment Designer

.github/agents/experimental/pptx.agent.md

@@ -5,11 +5,6 @@ disable-model-invocation: true
 agents:
   - Researcher Subagent
   - PowerPoint Subagent
-handoffs:
-  - label: "Compact"
-    agent: PowerPoint Builder
-    send: true
-    prompt: "/compact Make sure summarization includes that all state is managed through the .copilot-tracking folder files, be sure to include file paths for all of the current Tracking Artifacts. Be sure to include any current analysis log artifacts. Be sure to include any follow-up items that were provided to the user but not yet decided to be worked on by the user. Be sure to include the user's specific requirements original requirements and requests. The user may request to make additional follow up changes, add or modify new requirements, be sure to follow your Required Phases over again from Phase 1 based on the user's requirements."
 ---
 
 # PowerPoint Builder

.github/agents/hve-core/prompt-builder.agent.md

@@ -8,10 +8,6 @@ agents:
   - Prompt Updater
   - Researcher Subagent
 handoffs:
-  - label: "Compact"
-    agent: Prompt Builder
-    send: true
-    prompt: "/compact Make sure summarization includes that all state is managed through the .copilot-tracking folder files, be sure to include file paths for all of the current Tracking Artifacts. Be sure to include any current analysis log artifacts. Be sure to include any follow-up items that were provided to the user but not yet decided to be worked on by the user. Be sure to include the user's specific requirements original requirements and requests. The user may request to make additional follow up changes, add or modify new requirements, be sure to follow your Required Phases over again from Phase 1 based on the user's requirements."
   - label: "💡 Update/Create"
     agent: Prompt Builder
     prompt: "/prompt-build"
@@ -62,6 +58,16 @@ Cross-run continuity: Subagents can read and reference files from prior sandbox
 * When using the `runSubagent` tool, select the named agent directly and provide the required inputs listed for that phase.
 * For all phases, avoid reading the prompt file(s) directly and instead have the subagents read the prompt file(s).
 
+### Model Selection for Subagents
+
+Apply cost-first model selection: use a fast model for tasks that do not write or design prompts.
+
+* Researcher Subagent: specify `model: "Claude Haiku 4.5 (copilot)"` (read-only research).
+* Prompt Evaluator: specify `model: "Claude Haiku 4.5 (copilot)"` (evaluation is pattern-matching against criteria, not authoring).
+* Prompt Tester: omit `model` (inherits session model) since literal execution of prompts needs full capability.
+* Prompt Updater: omit `model` (inherits session model) since prompt engineering is functionally code authoring.
+* When the cost tier constraint prevents downgrading, omit `model` and let the platform resolve it.
+
 ## Required Phases
 
 Repeat phases as often as needed based on *evaluation-log* findings.

.github/agents/hve-core/rpi-agent.agent.md

@@ -7,10 +7,6 @@ agents:
   - Researcher Subagent
   - Phase Implementor
 handoffs:
-  - label: Compact
-    agent: RPI Agent
-    prompt: "/compact Make sure summarization includes that all state is managed through the .copilot-tracking folder files, always include file paths for all of the Tracking Artifacts for this session. Indicate percent complete for each of the artifacts. Include the last Phase before compaction, steps of phase completed, in-progress step of phase, remaining steps of phase. Be sure to include executive details for each of the `Phase 4: Review` most recent findings. Must include all of the most recent `Phase 5: Discover` follow up work items and their order with complete and consistent details."
-    send: true
   - label: "1️⃣"
     agent: RPI Agent
     prompt: "/rpi continue=1"

.github/agents/hve-core/subagents/implementation-validator.agent.md

@@ -5,6 +5,9 @@ user-invocable: false
 tools:
   - read
   - search
+model:
+  - Claude Haiku 4.5 (copilot)
+  - GPT-5.4 mini (copilot)
 ---
 
 # Implementation Validator

.github/agents/hve-core/subagents/plan-validator.agent.md

@@ -2,6 +2,9 @@
 name: Plan Validator
 description: 'Validates implementation plans against research documents, updating the Planning Log Discrepancy Log section with severity-graded findings - Brought to you by microsoft/hve-core'
 user-invocable: false
+model:
+  - Claude Haiku 4.5 (copilot)
+  - GPT-5.4 mini (copilot)
 ---
 
 # Plan Validator

.github/agents/hve-core/subagents/prompt-evaluator.agent.md

@@ -2,6 +2,9 @@
 name: Prompt Evaluator
 description: 'Evaluates prompt execution results against Prompt Quality Criteria with severity-graded findings and categorized remediation guidance'
 user-invocable: false
+model:
+  - Claude Haiku 4.5 (copilot)
+  - GPT-5.4 mini (copilot)
 ---
 
 # Prompt Evaluator

.github/agents/hve-core/subagents/researcher-subagent.agent.md

@@ -2,6 +2,9 @@
 name: Researcher Subagent
 description: 'Research subagent using search tools, read tools, fetch web page, github repo, and mcp tools'
 user-invocable: false
+model:
+  - Claude Haiku 4.5 (copilot)
+  - GPT-5.4 mini (copilot)
 ---
 
 # Researcher Subagent

.github/agents/hve-core/subagents/rpi-validator.agent.md

@@ -2,6 +2,9 @@
 name: RPI Validator
 description: 'Validates a Changes Log against the Implementation Plan, Planning Log, and Research Documents for a specific plan phase - Brought to you by microsoft/hve-core'
 user-invocable: false
+model:
+  - Claude Haiku 4.5 (copilot)
+  - GPT-5.4 mini (copilot)
 ---
 
 # RPI Validator

.github/agents/hve-core/task-challenger.agent.md

@@ -4,10 +4,6 @@ description: 'Adversarial questioning agent that interrogates implementations wi
 disable-model-invocation: true
 tools: [read, search, edit/createFile, edit/editFiles, execute/runInTerminal, execute/getTerminalOutput]
 handoffs:
-  - label: "Compact"
-    agent: Task Challenger
-    send: true
-    prompt: "/compact Preserve the current challenge session state. Include the confirmed scope, all questions asked with the user's complete answers, any probe questions and responses, and all items marked unresolved. The challenge tracking document in .copilot-tracking/challenges/ contains the session record: reference the most recent document by date. When resuming, continue from the last question asked."
   - label: "🔬 Research Questions"
     agent: Task Researcher
     prompt: "/task-research Find and read the most recent challenge tracking document in .copilot-tracking/challenges/ (most recent by date prefix) for the Q&A log and unresolved items: these define the research scope."

.github/agents/hve-core/task-implementor.agent.md

@@ -6,10 +6,6 @@ agents:
   - Phase Implementor
   - Researcher Subagent
 handoffs:
-  - label: "Compact"
-    agent: Task Implementor
-    send: true
-    prompt: "/compact Make sure summarization includes that all state is managed through the .copilot-tracking folder files, and be sure to include that the next agent instructions will be Task Reviewer and the user will switch to it when they are done with Task Implementation"
   - label: "✅ Review"
     agent: Task Reviewer
     prompt: /task-review
@@ -56,6 +52,15 @@ The researcher-subagent returns deep research findings: subagent research docume
 
 Subagents can run in parallel when investigating independent topics or executing independent phases.
 
+### Model Selection for Subagents
+
+Apply cost-first model selection: use a fast model for tasks that do not write code, and inherit the session model for code generation.
+
+* Phase Implementor (writes code): omit the `model` parameter so it inherits the session model for maximum code quality.
+* Researcher Subagent (read-only research): specify `model: "Claude Haiku 4.5 (copilot)"` to reduce cost.
+* If a research task requires deep code-level analysis: omit `model` to inherit the session model.
+* When the cost tier constraint prevents downgrading below the session model, omit `model` and let the platform resolve it.
+
 ## Required Artifacts
 
 | Artifact               | Path Pattern                                                        | Required |

.github/agents/hve-core/task-planner.agent.md

@@ -6,10 +6,6 @@ agents:
   - Researcher Subagent
   - Plan Validator
 handoffs:
-  - label: "Compact"
-    agent: Task Planner
-    send: true
-    prompt: "/compact  make sure summarization includes that all state is managed through the .copilot-tracking folder files, and be sure to include that the next agent instructions will be Task Implementor and the user will switch to it when they are done with Task Planner"
   - label: "⚡ Implement"
     agent: Task Implementor
     prompt: /task-implement
@@ -58,6 +54,15 @@ Run `Plan Validator` using `runSubagent` or `task`, providing these inputs:
 
 Subagents can run in parallel when investigating independent topics or validating independent concerns.
 
+### Model Selection for Subagents
+
+Apply cost-first model selection: use a fast model for tasks that do not produce code or architectural decisions.
+
+* Researcher Subagent (read-only research): specify `model: "Claude Haiku 4.5 (copilot)"` to reduce cost.
+* Plan Validator (validation and comparison): specify `model: "Claude Haiku 4.5 (copilot)"` since validation is pattern-matching against documents, not code generation.
+* If a research or validation task involves complex architectural reasoning: omit the `model` parameter to inherit the session model.
+* When the cost tier constraint prevents downgrading, omit `model` and let the platform resolve it.
+
 ## File Locations
 
 Planning files reside in `.copilot-tracking/` at the workspace root unless the user specifies a different location.

.github/agents/hve-core/task-researcher.agent.md

@@ -5,10 +5,6 @@ disable-model-invocation: true
 agents:
   - Researcher Subagent
 handoffs:
-  - label: "Compact"
-    agent: Task Researcher
-    send: true
-    prompt: "/compact make sure summarization includes that all state is managed through the .copilot-tracking folder files, and be sure to include that the next agent instructions will be Task Planner and the user will switch to it when they are done with Task Researcher"
   - label: "📋 Create Plan"
     agent: Task Planner
     prompt: /task-plan
@@ -50,6 +46,14 @@ Run `Researcher Subagent` with `runSubagent` or `task`, and parallelize calls wh
 
 Subagents can run in parallel when investigating independent topics or sources.
 
+### Model Selection for Subagents
+
+Apply cost-first model selection when invoking subagents. Research tasks are read-heavy and do not generate code, so they benefit from a fast-tier model without sacrificing quality.
+
+* Research subagent calls: specify `model: "Claude Haiku 4.5 (copilot)"` on the `runSubagent` invocation to reduce cost.
+* If the research task involves complex code-level reasoning (tracing execution paths, analyzing architecture): omit the `model` parameter to inherit the session model.
+* When the fast model is unavailable or the cost tier constraint prevents downgrading, omit `model` and let the platform resolve it.
+
 ## File Locations
 
 Research files reside in `.copilot-tracking/` at the workspace root unless the user specifies a different location.

.github/agents/hve-core/task-reviewer.agent.md

@@ -7,10 +7,6 @@ agents:
   - Researcher Subagent
   - Implementation Validator
 handoffs:
-  - label: "Compact"
-    agent: Task Reviewer
-    send: true
-    prompt: "/compact Make sure summarization includes that all state is managed through the .copilot-tracking folder files, be sure to include file paths to the review documents and executive details about each individual finding. Be sure to include that the next agent instructions will be one-of Task Researcher for deeper research on the chosen findings to address, Task Planner to go right into planning based off of the chosen findings from the review document, or right back into implementation addressing the chosen findings from the review document. The user will switch to the agent instructions when they are done with Task Review."
   - label: "🔬 Research More"
     agent: Task Researcher
     prompt: /task-research
@@ -106,6 +102,15 @@ Read the validation files produced by each `RPI Validator` run. Synthesize findi
 
 When findings require deeper investigation, run additional `RPI Validator` calls for specific phases. Run `Researcher Subagent` when context is missing, providing research topics and a subagent research document path.
 
+#### Model Selection for Subagents
+
+Apply cost-first model selection when spawning validation and research subagents.
+
+* RPI Validator and Implementation Validator: specify `model: "Claude Haiku 4.5 (copilot)"` since validation compares artifacts without generating code.
+* Researcher Subagent: specify `model: "Claude Haiku 4.5 (copilot)"` for read-only research.
+* If validation requires complex code reasoning or architectural judgment: omit `model` to inherit the session model.
+* When the cost tier constraint prevents downgrading, omit `model` and let the platform resolve it.
+
 Proceed to Phase 3 when RPI validation is complete.
 
 ### Phase 3: Quality Validation

.github/agents/rai-planning/rai-planner.agent.md

@@ -8,10 +8,6 @@ handoffs:
     agent: Security Planner
     prompt: /security-capture
     send: true
-  - label: "Compact"
-    agent: RAI Planner
-    send: true
-    prompt: "/compact Make sure summarization includes that all state is managed through .copilot-tracking/rai-plans/ folder files, and be sure to include the current phase, entry mode, and project slug"
 tools:
   - read
   - edit/createFile

.github/agents/security/security-planner.agent.md

@@ -14,10 +14,6 @@ tools:
   - web
   - agent
 handoffs:
-  - label: "Compact"
-    agent: Security Planner
-    send: true
-    prompt: "/compact  make sure summarization includes that all state is managed through the .copilot-tracking folder files, and be sure to include the current security planning phase and project slug"
   - label: "RAI Planner"
     agent: RAI Planner
     prompt: /rai-plan-from-security-plan

.github/agents/security/security-reviewer.agent.md

@@ -107,6 +107,16 @@ Skill resolution: Read the applicable security skill (e.g., `owasp-top-10`, `owa
 | Report Generator      | `.github/agents/**/report-generator.agent.md`      | Collates all verified findings and generates the final vulnerability report.       |
 | Skill Assessor        | `.github/agents/**/skill-assessor.agent.md`        | Assesses a single skill against the codebase, returning structured findings.       |
 
+### Model Selection for Subagents
+
+Apply cost-first model selection when invoking subagents. Security scanning subagents compare code against reference patterns rather than generating code.
+
+* Codebase Profiler: specify `model: "Claude Haiku 4.5 (copilot)"` (read-only scanning and classification).
+* Skill Assessor: specify `model: "Claude Haiku 4.5 (copilot)"` (pattern matching against vulnerability references).
+* Finding Deep Verifier: omit `model` (inherits session model) since adversarial verification requires deeper reasoning.
+* Report Generator: specify `model: "Claude Haiku 4.5 (copilot)"` (collation and formatting, not analysis).
+* When the cost tier constraint prevents downgrading, omit `model` and let the platform resolve it.
+
 ### Available Skills
 
 * owasp-agentic

.github/agents/security/sssc-planner.agent.md

@@ -8,10 +8,6 @@ description: >-
 agents:
   - Researcher Subagent
 handoffs:
-  - label: "Compact"
-    agent: SSSC Planner
-    send: true
-    prompt: "/compact Make sure summarization includes that all state is managed through .copilot-tracking/sssc-plans/ folder files, and be sure to include the current phase, entry mode, and project slug"
   - label: "Security Planner"
     agent: Security Planner
     prompt: /security-capture

.github/agents/security/subagents/codebase-profiler.agent.md

@@ -8,6 +8,9 @@ tools:
   - search/textSearch
   - read/readFile
 user-invocable: false
+model:
+  - Claude Haiku 4.5 (copilot)
+  - GPT-5.4 mini (copilot)
 ---
 
 # Codebase Profiler

.github/agents/security/subagents/report-generator.agent.md

@@ -7,6 +7,9 @@ tools:
   - search/fileSearch
   - read/readFile
 user-invocable: false
+model:
+  - Claude Haiku 4.5 (copilot)
+  - GPT-5.4 mini (copilot)
 ---
 
 # Report Generator

.github/copilot-instructions.md

@@ -52,7 +52,7 @@ Scripts are organized by function:
 
 * Collections (`scripts/collections/`) - Collection validation and shared helper modules.
 * Extension (`scripts/extension/`) - Extension packaging and preparation.
-* Linting (`scripts/linting/`) - Markdown validation, link checking, frontmatter validation, and PowerShell analysis.
+* Linting (`scripts/linting/`) - Markdown validation, link checking, frontmatter validation, model reference validation, and PowerShell analysis.
 * Security (`scripts/security/`) - Dependency pinning validation, SHA staleness checks, and action version consistency.
 * Library (`scripts/lib/`) - Shared utilities such as verified downloads.
 * Plugins (`scripts/plugins/`) - Plugin generation and marketplace validation.
@@ -207,7 +207,9 @@ Agents should use npm scripts for all validation:
 * `npm run lint:version-consistency` - Action version consistency
 * `npm run lint:marketplace` - Marketplace validation
 * `npm run lint:py` - Python linting via ruff
-* `npm run lint:all` - Run all linters (chains `format:tables`, `lint:md`, `lint:ps`, `lint:yaml`, `lint:links`, `lint:frontmatter`, `lint:collections-metadata`, `lint:marketplace`, `lint:version-consistency`, `lint:permissions`, `lint:dependency-pinning`, `lint:py`, and `validate:skills`)
+* `npm run lint:models` - Model reference validation against catalog
+* `npm run lint:models:refresh` - Refresh model catalog from upstream documentation
+* `npm run lint:all` - Run all linters (chains `format:tables`, `lint:md`, `lint:ps`, `lint:yaml`, `lint:links`, `lint:frontmatter`, `lint:collections-metadata`, `lint:marketplace`, `lint:version-consistency`, `lint:permissions`, `lint:dependency-pinning`, `lint:py`, `validate:skills`, `lint:ai-artifacts`, and `lint:models`)
 * `npm run validate:copyright` - Copyright header validation
 * `npm run validate:skills` - Skill structure validation
 * `npm run spell-check` - Spelling validation

.github/instructions/hve-core/prompt-builder.instructions.md

@@ -412,7 +412,7 @@ Optional fields available by file type:
 * `disable-model-invocation:` - Boolean. Set to `true` to prevent Copilot from automatically invoking the agent. Use for agents that run subagents, agents that cause side effects (git operations, backlog management, deployments), or agents that should only run when explicitly requested. Defaults to `false` when omitted.
 * `agent:` - Agent delegation for prompt files and handoffs. Use the human-readable name from the agent's `name:` frontmatter (for example, `Prompt Builder`).
 * `argument-hint:` - Hint text for prompt picker display.
-* `model:` - Model specification. Accepts any valid model identifier string (for example, `gpt-4o`, `claude-sonnet-4`). When omitted, the default model is used.
+* `model:` - Model specification. For **agent files**, accepts a single model identifier string (for example, `claude-sonnet-4`) or a prioritized array of model identifiers; when an array is specified, the system tries each model in order until an available one is found. For **prompt files**, accepts a single model identifier string only. When omitted, the currently selected model in the model picker is used.
 * `license:` - SPDX license identifier for skill content (for example, `MIT`, `CC-BY-SA-4.0`). Defaults to the repository license when omitted. Use for skills that incorporate third-party content under a specific license.
 * `metadata:` - Object containing provenance and versioning metadata for skills. Recognized fields include `authors`, `spec_version`, `framework_revision`, `last_updated`, `skill_based_on`, and `content_based_on`.