feat(agents): optimize RPI agent context management with discipline rules

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

@@ -91,6 +91,29 @@ When a task requires subagents but neither `runSubagent` nor `task` tools are av
 
 Treat the phase guidance below as operating defaults rather than ceremony. Delegate only when it materially improves the outcome.
 
+## Context Discipline
+
+After any subagent returns, this turn must be lean:
+
+1. Emit one compact line per subagent (subagent name + one-line outcome + tracking file path).
+2. Update the relevant `.copilot-tracking/` file via a single edit if needed.
+3. Stop. Do not re-read large planning, research, or details files in the closing turn. Do not re-quote subagent payloads. Do not narrate the next phase plan.
+
+Choose the lightest response mode that satisfies the request:
+
+| Mode        | When to use                                                                                                                                                        |
+|-------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| Direct      | Answer from this turn's context only. No subagent, no file reads. Use for clarifications, status questions, or queries when the relevant file is already attached. |
+| Lightweight | Single subagent with a focused prompt. Skip re-reading prior phase tracking files. Use for summarizing findings or single-file edits.                              |
+| Standard    | Default behavior: subagent dispatch, tracking-file update, and handoff suggestion.                                                                                 |
+| Full        | Multiple parallel subagents and cross-phase synthesis. Use only when explicitly requested or when the phase contract requires it.                                  |
+
+Subagent result handling:
+
+* Treat the subagent's chat response as an index, not the full result.
+* When a decision (plan structure, phase ordering, accept/reject of an alternative, validation verdict) depends on detail beyond the summary bullets, re-read the subagent file directly and cite specific sections.
+* Do not re-read the file gratuitously: re-read only when the next action requires evidence the summary does not contain.
+
 ## Tracking Artifacts
 
 All persistent state, session notes, and workflow artifacts are tracked in `.copilot-tracking/` at the root of the workspace when the workflow needs durable records. For simple and most medium requests, the agent may keep research and planning in its own context and skip creating artifact files until task difficulty or workflow needs justify them.

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

@@ -199,13 +199,13 @@ During any validation scope, note issues that fall outside predefined categories
 
 ## Response Format
 
-Return implementation validation findings including:
-
-* Relative path to the implementation validation log.
-* Validation status: Passed, Partial, Failed, or Blocked.
-* Executive summary of findings with severity counts.
-* Count of findings by category.
-* Top critical and major findings with evidence.
-* Scopes completed, skipped, or blocked.
-* Suggested areas for additional investigation.
-* Clarifying questions requiring parent agent or user input.
+The subagent always writes complete validation findings to the review log before returning. The chat response is an executive summary only. Full fidelity lives on disk.
+
+Initial chat response, emit at most:
+* 1 line: review log file path (the parent re-reads this file when it needs detail).
+* 1 line: validation status (Pass / Pass with Warnings / Fail).
+* Up to 7 bullet-point findings (each ≤ 240 chars). Prioritize blocking issues and regressions.
+* Up to 3 clarifying questions, only when blocking.
+* 1 short "Full Detail" pointer line: "Re-read `<path>` for complete validation output, test results, and remediation guidance."
+
+Do not paste full test output, lint dumps, or complete file diffs into the chat response. The review log is the source of truth.

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

@@ -94,13 +94,14 @@ When prior planning logs are available, cross-run comparison notes reference res
 
 ## Response Format
 
-Return Plan Validation Findings including:
-
-* Relative path to the Planning Log file.
-* Validation status: Validated, Gaps Found, Incomplete, or Blocked.
-* Executive summary of identified discrepancies and unaddressed issues with severity and impact.
-* Count of DR- items added, updated, or removed in the Planning Log.
-* Count of DD- items added, updated, or removed in the Planning Log.
-* Coverage summary (internal analysis): counts of covered, partial, missing requirements.
-* Severity-ordered discrepancy findings with evidence references from research and plan files.
-* Clarifying questions requiring parent agent or user input.
+The subagent always writes complete validation findings to the Planning Log before returning. The chat response is an executive summary only. Full fidelity lives on disk.
+
+Initial chat response, emit at most:
+* 1 line: planning log file path (the parent re-reads this file when it needs detail).
+* 1 line: validation status (Pass / Fail - Critical / Fail - Major / Fail - Minor).
+* Up to 7 bullet-point severity-ordered findings (each ≤ 240 chars). Prioritize critical and major items.
+* 1 line: planning log deltas (DR- items added/updated/removed; DD- items added/updated/removed).
+* Up to 3 clarifying questions, only when blocking.
+* 1 short "Full Detail" pointer line: "Re-read `<path>` for complete discrepancy details, evidence, and recommended fixes."
+
+Do not paste full discrepancy tables, complete plan excerpts, or research quotes into the chat response. The planning log is the source of truth.

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

@@ -58,10 +58,14 @@ External URLs may still use markdown link syntax.
 
 ## Response Format
 
-Return Subagent Research Executive Details and include the following requirements:
+The subagent always writes complete findings to its subagent file before returning. The chat response is an executive summary only. Full fidelity lives on disk.
 
-* The relative path to the subagent research document.
-* The status of the subagent research: Complete, In-Progress, Blocked, etc.
-* The important details from the subagent research document based on your interpretation.
-* A checklist of recommended next research not completed during this session.
-* Any clarifying questions that require more information or input from the user.
+Initial chat response, emit at most:
+* 1 line: subagent file path (the parent re-reads this file when it needs detail).
+* 1 line: status (Complete / Blocked / Needs Clarification).
+* Up to 7 bullet-point key findings (each ≤ 240 chars). Prioritize findings the parent cannot act on without reading the file.
+* A checklist of up to 5 recommended next research items not completed during this session.
+* Up to 3 clarifying questions, only when blocking.
+* 1 short "Full Detail" pointer line: "Re-read `<path>` for complete evidence, code blocks, file/line citations, and rejected alternatives."
+
+Do not paste file contents, code blocks, long quotes, or full evidence tables into the chat response. The subagent file is the source of truth.

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

@@ -64,10 +64,14 @@ Create and update the validation document progressively documenting:
 
 ## Response Format
 
-Return RPI Validation Executive Details and include the following requirements:
+The subagent always writes complete validation findings to the review log before returning. The chat response is an executive summary only. Full fidelity lives on disk.
 
-* The relative path to the validation document.
-* The status of the validation: Passed, Partial, Failed, or Blocked.
-* The important details from the validation document based on your interpretation.
-* A checklist of recommended next validations not completed during this session.
-* Any clarifying questions that require more information or input from the user.
+Initial chat response, emit at most:
+* 1 line: review log file path (the parent re-reads this file when it needs detail).
+* 1 line: validation status (Pass / Fail).
+* Up to 7 bullet-point findings (each ≤ 240 chars). Prioritize schema violations and missing required sections.
+* A checklist of up to 5 recommended next validations not completed during this session.
+* Up to 3 clarifying questions, only when blocking.
+* 1 short "Full Detail" pointer line: "Re-read `<path>` for complete RPI artifact validation details."
+
+Do not paste full artifact contents, schema dumps, or long quotes into the chat response. The review log is the source of truth.

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

@@ -56,6 +56,29 @@ The researcher-subagent returns deep research findings: subagent research docume
 
 Subagents can run in parallel when investigating independent topics or executing independent phases.
 
+## Context Discipline
+
+After any subagent returns, this turn must be lean:
+
+1. Emit one compact line per subagent (subagent name + one-line outcome + tracking file path).
+2. Update the relevant `.copilot-tracking/` file via a single edit if needed.
+3. Stop. Do not re-read large planning, research, or details files in the closing turn. Do not re-quote subagent payloads. Do not narrate the next phase plan.
+
+Choose the lightest response mode that satisfies the request:
+
+| Mode        | When to use                                                                                                                                                        |
+|-------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| Direct      | Answer from this turn's context only. No subagent, no file reads. Use for clarifications, status questions, or queries when the relevant file is already attached. |
+| Lightweight | Single subagent with a focused prompt. Skip re-reading prior phase tracking files. Use for summarizing findings or single-file edits.                              |
+| Standard    | Default behavior: subagent dispatch, tracking-file update, and handoff suggestion.                                                                                 |
+| Full        | Multiple parallel subagents and cross-phase synthesis. Use only when explicitly requested or when the phase contract requires it.                                  |
+
+Subagent result handling:
+
+* Treat the subagent's chat response as an index, not the full result.
+* When a decision (plan structure, phase ordering, accept/reject of an alternative, validation verdict) depends on detail beyond the summary bullets, re-read the subagent file directly and cite specific sections.
+* Do not re-read the file gratuitously: re-read only when the next action requires evidence the summary does not contain.
+
 ## Required Artifacts
 
 | Artifact               | Path Pattern                                                        | Required |

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

@@ -58,6 +58,29 @@ Run `Plan Validator` using `runSubagent` or `task`, providing these inputs:
 
 Subagents can run in parallel when investigating independent topics or validating independent concerns.
 
+## Context Discipline
+
+After any subagent returns, this turn must be lean:
+
+1. Emit one compact line per subagent (subagent name + one-line outcome + tracking file path).
+2. Update the relevant `.copilot-tracking/` file via a single edit if needed.
+3. Stop. Do not re-read large planning, research, or details files in the closing turn. Do not re-quote subagent payloads. Do not narrate the next phase plan.
+
+Choose the lightest response mode that satisfies the request:
+
+| Mode        | When to use                                                                                                                                                        |
+|-------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| Direct      | Answer from this turn's context only. No subagent, no file reads. Use for clarifications, status questions, or queries when the relevant file is already attached. |
+| Lightweight | Single subagent with a focused prompt. Skip re-reading prior phase tracking files. Use for summarizing findings or single-file edits.                              |
+| Standard    | Default behavior: subagent dispatch, tracking-file update, and handoff suggestion.                                                                                 |
+| Full        | Multiple parallel subagents and cross-phase synthesis. Use only when explicitly requested or when the phase contract requires it.                                  |
+
+Subagent result handling:
+
+* Treat the subagent's chat response as an index, not the full result.
+* When a decision (plan structure, phase ordering, accept/reject of an alternative, validation verdict) depends on detail beyond the summary bullets, re-read the subagent file directly and cite specific sections.
+* Do not re-read the file gratuitously: re-read only when the next action requires evidence the summary does not contain.
+
 ## 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

@@ -50,6 +50,29 @@ Run `Researcher Subagent` with `runSubagent` or `task`, and parallelize calls wh
 
 Subagents can run in parallel when investigating independent topics or sources.
 
+## Context Discipline
+
+After any subagent returns, this turn must be lean:
+
+1. Emit one compact line per subagent (subagent name + one-line outcome + tracking file path).
+2. Update the relevant `.copilot-tracking/` file via a single edit if needed.
+3. Stop. Do not re-read large planning, research, or details files in the closing turn. Do not re-quote subagent payloads. Do not narrate the next phase plan.
+
+Choose the lightest response mode that satisfies the request:
+
+| Mode        | When to use                                                                                                                                                        |
+|-------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| Direct      | Answer from this turn's context only. No subagent, no file reads. Use for clarifications, status questions, or queries when the relevant file is already attached. |
+| Lightweight | Single subagent with a focused prompt. Skip re-reading prior phase tracking files. Use for summarizing findings or single-file edits.                              |
+| Standard    | Default behavior: subagent dispatch, tracking-file update, and handoff suggestion.                                                                                 |
+| Full        | Multiple parallel subagents and cross-phase synthesis. Use only when explicitly requested or when the phase contract requires it.                                  |
+
+Subagent result handling:
+
+* Treat the subagent's chat response as an index, not the full result.
+* When a decision (plan structure, phase ordering, accept/reject of an alternative, validation verdict) depends on detail beyond the summary bullets, re-read the subagent file directly and cite specific sections.
+* Do not re-read the file gratuitously: re-read only when the next action requires evidence the summary does not contain.
+
 ## 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

@@ -37,6 +37,29 @@ Reviews completed implementation work from `.copilot-tracking/` artifacts. Valid
 * Match `applyTo` patterns from `.github/instructions/` files against changed file types to identify applicable conventions.
 * Subagents return structured, evidence-based responses with severity levels and can ask clarifying questions rather than guessing.
 
+## Context Discipline
+
+After any subagent returns, this turn must be lean:
+
+1. Emit one compact line per subagent (subagent name + one-line outcome + tracking file path).
+2. Update the relevant `.copilot-tracking/` file via a single edit if needed.
+3. Stop. Do not re-read large planning, research, or details files in the closing turn. Do not re-quote subagent payloads. Do not narrate the next phase plan.
+
+Choose the lightest response mode that satisfies the request:
+
+| Mode        | When to use                                                                                                                                                        |
+|-------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| Direct      | Answer from this turn's context only. No subagent, no file reads. Use for clarifications, status questions, or queries when the relevant file is already attached. |
+| Lightweight | Single subagent with a focused prompt. Skip re-reading prior phase tracking files. Use for summarizing findings or single-file edits.                              |
+| Standard    | Default behavior: subagent dispatch, tracking-file update, and handoff suggestion.                                                                                 |
+| Full        | Multiple parallel subagents and cross-phase synthesis. Use only when explicitly requested or when the phase contract requires it.                                  |
+
+Subagent result handling:
+
+* Treat the subagent's chat response as an index, not the full result.
+* When a decision (plan structure, phase ordering, accept/reject of an alternative, validation verdict) depends on detail beyond the summary bullets, re-read the subagent file directly and cite specific sections.
+* Do not re-read the file gratuitously: re-read only when the next action requires evidence the summary does not contain.
+
 ## Review Artifacts
 
 | Artifact            | Path Pattern                                                        | Required |