@apexops/cli 0.10.0-next.5 → 0.11.0-next.1

package/assets/customizations/.github/skills/apex-requirements/SKILL.md

@@ -1,46 +1,81 @@
 ---
 name: apex-requirements
-description: "Provides internal APEX requirements guidance for typed intake, scope, constraints, and budget."
+description: "Provides internal APEX requirements guidance for typed intake, scope, constraints, service preferences and Gate 1."
 user-invocable: false
 ---
 
+When calling any `apex/*` MCP tool, include the current session checkout or worktree as the required absolute
+`workspace` path.
+
 ## APEX Requirements
 
-Use this skill only for an active requirements task.
+Use this skill only when the kernel routes the active foreground `APEX` agent to a requirements input request, review
+or task.
 
 ## Prerequisites
 
-- `apex/taskContext` returns a requirements task envelope.
-- The interactive Requirements agent is active when user input may be needed.
+- Call `apex/status` first, then `apex/nextTask` only when status leaves runnable requirements work.
+- Task context is read only for `status=task` with a requirements task ID.
 - The kernel's returned input request is the authoritative requirements question catalog.
+- Apply the user's requested outcome and stop point before work begins. A prior handoff cannot expand scope. For
+  intake-only scope, stop after the matching intake context and do not submit artifacts, start review or ask for
+  approval.
 
 ## Workflow
 
-1. Call `apex/nextTask`. Validate that a returned `task` is owned by the requirements role before reading its context.
-2. When it returns `needs_input`, use its `intake` metadata and `questions` exactly as returned; do not maintain a
-   separate client-side question list.
-3. Ask the returned questions through the active client projection's question mechanism. Record each response as a
-   supplied value, typed unknown, or explicit deferral with its owner; never replace an unknown with an inferred value.
-4. Submit that request only through `apex/recordInput`, preserving its request ID, expected journal head, and owner
-   epoch.
-5. Call `apex/nextTask` again and repeat until it returns a requirements `task`; only then read `apex/taskContext`.
-6. Treat service questions in the workload panel as a preference boundary. Present the kernel recommendation, then
-   capture retained, prohibited, and preferred services, SKU preferences, and environment overrides without selecting
-   architecture, SKUs, or implementation details.
-7. Submit the task-context-defined output through `apex/requirementsComplete`, invoke the required Reviewer, and handle
-   `needs_review` through one native findings panel and `apex/reviewDecide`.
-8. After a clean or fully dispositioned review, ask for explicit Gate 1 approval. Call `apex/gateDecide` only after the
-   user chooses Proceed, then continue to Architecture in the same turn.
-
-The kernel catalog and its versioned input contracts are authoritative. Do not choose architecture, SKUs, or
-implementation details while gathering requirements.
+1. Reuse facts the user already supplied in the current request. Present matching supplied facts as recommended
+   confirmations when the kernel asks for them; do not make the user retype them, and do not record them until the user
+   confirms or corrects them. Ask only for missing values. Never invent a project value, target, environment, IaC tool,
+   service preference, budget posture, unknown, deferral, owner, risk acceptance or approval.
+2. For `status=needs_input`, ask every returned question through `ask_user` in chat. Use native single-select or
+   multi-select controls without adding or reordering kernel options. If native multi-select is unavailable, show the
+   exact kernel options numbered in kernel order, collect the numbers, resolve them back to option values, and confirm
+   the complete selection before recording. Invalid, duplicate, empty, ambiguous or out-of-range entries require
+   correction. Recommendations are proposals only; never submit them without confirmation.
+   When the request includes intake metadata, report the round as `request.intake.ordinal` of
+   `request.intake.total` and continue through each remaining intake round only up to the user's requested stop point;
+   each round has its own request ID and must be recorded separately.
+3. Submit accepted answers only through `apex/recordInput` with the exact request ID, expected head and owner epoch from
+   the request, plus one typed answer per question. Preserve arrays and typed values such as classifications,
+   compliance selections, explicit deferrals and unknowns. A chat answer is not kernel acceptance; wait for
+   `recorded: true` before advancing.
+4. After accepted input, call `apex/nextTask` again. Do not poll unresolved input or review. For `needs_review`, present
+   the findings in one decision panel and submit permitted decisions through `apex/reviewDecide`.
+5. For `status=task`, call `apex/taskContext` with the exact `task.taskId`. If the context is externalized, read it in
+   bounded chunks through `apex/readTaskInput`. Use `taskContext.recordedInput` and `outputTemplates` as the complete
+   contract; do not read repository schemas or session history.
+6. Build the requirements output with business context, measurable success criteria, non-functional requirements,
+   security/compliance posture, budget and operations posture, regional constraints and candidate-service rationale for
+   Architecture. Treat stated performance and scale values as later-validated goals. If none are provided, record
+   `Performance and scale: check later (validated at a later stage)` rather than asking supplemental questions.
+7. Treat Azure services as candidates. Recommend viable compute, data, integration, identity and observability options
+   with concise fit and trade-off rationale, but never record a service or SKU as an Architecture decision. Capture
+   retained, prohibited or preferred services, SKU constraints or an explicit no-preference position.
+8. Populate narrative recommendations for access, ingress, DNS, personal-data inventory, retention/deletion,
+   data-subject handling and telemetry minimization. Recommendations are proposed, not confirmed requirements,
+   assigned owners or compliance evidence.
+9. Submit the typed requirements artifact through `apex/requirementsComplete`. Report the materialized read-only Gate 1
+   review package under `agent-output/<project>/<run>/`, including requirements, recommendations, SKU preferences and
+   challenger findings.
+10. Immediately call `apex/nextTask` after submitting requirements. When it returns the `requirements-review` task,
+    run the rubber-duck review steps in `apex-next` with that exact `task.taskId`. If rubber-duck is unavailable or the
+    capture fails twice, report the pending review task and stop.
+11. When `needs_review` returns, do not request task context or run rubber-duck again. Record a disposition for every
+    finding: fix, accept or dismiss with a reason, or acknowledge an obligation with an owner. For accept-risk, show
+    `owner: <project risk owner>, expires in 90 days`, draft the rationale and ask only for confirmation. Existing
+    blocking findings require the normal correction and fresh review path; do not automatically acknowledge, dismiss or
+    accept risk.
+12. After `apex/reviewDecide`, call `apex/status`. If Gate 1 is pending, report it and stop. Do not call
+    `apex/nextTask` while a gate is pending. If the user's scope permits approval and the user explicitly approves or
+    rejects Gate 1, call `apex/gateDecide` with `confirm: true`.
 
 ## Boundaries
 
-Do not read task context for `needs_input`, a task owned by another role, or a stale task ID. Treat `APEX_STALE` as a
-fresh-status requirement, and return kernel validation or authorization errors without fabricating a requirements
-result.
+Do not read task context for `needs_input`, `needs_review`, a stale task ID, or a task owned by another role. Write only
+through APEX MCP. ARM MCP access is read-only and only supports indicative pricing when the user asks for it. Generated
+review projections are derived from accepted state and are never editable authority.
 
 ## Output
 
-Return the kernel result plus any unresolved user-owned fields.
+Return the kernel result, review-package location, candidate-service rationale, challenger findings, unresolved
+user-owned fields and the reason for stopping.

package/assets/customizations/.github/skills/apex-terraform-import/SKILL.md

@@ -4,6 +4,9 @@ description: "Assess Terraform adoption of existing Azure resources in APEX. Use
 user-invocable: false
 ---
 
+When calling any `apex/*` MCP tool, include the current session checkout or worktree as the required absolute
+`workspace` path.
+
 # APEX Terraform Import Assessment
 
 Use this skill for an active Terraform adoption task concerning existing resources. It evaluates accepted inventory and

package/assets/customizations/.github/skills/apex-terraform-patterns/SKILL.md

@@ -4,6 +4,9 @@ description: "Apply approved Terraform architecture intent in APEX. Use for hub-
 user-invocable: false
 ---
 
+When calling any `apex/*` MCP tool, include the current session checkout or worktree as the required absolute
+`workspace` path.
+
 # APEX Terraform Patterns
 
 Use this skill for an active Terraform-bound planning or CodeGen task. It records approved architecture intent and

package/assets/customizations/.github/skills/apex-terraform-test/SKILL.md

@@ -4,6 +4,9 @@ description: "Design receipt-gated Terraform test intent in APEX. Use for unit a
 user-invocable: false
 ---
 
+When calling any `apex/*` MCP tool, include the current session checkout or worktree as the required absolute
+`workspace` path.
+
 # APEX Terraform Test Design
 
 Use this skill for an active Terraform test-design or validation task. It specifies coverage and assertion intent;

package/assets/customizations/.github/skills/apex-unslop/LICENSE.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Lauren Tan
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/assets/customizations/.github/skills/apex-unslop/SKILL.md

@@ -0,0 +1,94 @@
+---
+name: apex-unslop
+user-invocable: true
+disable-model-invocation: true
+argument-hint: "text or document path, audience, and review or edit scope"
+description: '**UTILITY SKILL** — Manually polish authorized prose for clarity and concrete language while preserving technical facts and APEX contracts. WHEN: explicitly invoked as /apex-unslop for prose review or cleanup. DO NOT USE FOR: code changes, technical validation, or automatic workflow finalization.'
+license: MIT
+metadata:
+  author: jonathan-vella
+  version: "1.0"
+  upstream: cursor/plugins/pstack/skills/unslop
+  upstream_commit: e8d856f0273b42ebafe0ec3546bd645709e7c1b0
+---
+
+# APEX prose cleanup
+
+Improve clarity when the user explicitly invokes this skill. Do not run automatically,
+load it for every agent turn, or add it as a mandatory workflow step. Manual invocation
+does not change the caller's model, tools, role or edit permissions. This is prose editing,
+not an AI-authorship detector or a substitute for technical review.
+
+## Prerequisites
+
+- Identify the requested text or file, audience, tone and whether the user wants review or edits.
+- Read the current selected text and applicable file instructions. Ask only for unresolved scope.
+- For files, establish ownership and review status before editing. Preserve unrelated user changes.
+
+## Scope and protected content
+
+Edit only authorized natural-language prose. Review requests produce suggestions without file changes.
+An ambiguous request does not authorize a repository-wide cleanup. The default for code, commands, identifiers,
+configuration, structured data, quotations, license notices and generated content is leave unchanged.
+
+Preserve facts, numbers, units, currencies, dates, SKU names, source URLs, citations, negation and requirement strength.
+Preserve uncertainty, evidence limits, accepted risks and qualifications. Never turn "may" into "will", an estimate
+into a guarantee, or accepted risk into remediation. Flag missing sources; never invent citations or delete an
+unsupported claim's caveat just to make the prose sound confident. Do not invent an actor to force active voice.
+
+APEX file contracts outrank style preferences. Preserve required H2 text/order, emoji-bearing template headings,
+anchors, badge structures, links, tables and code fences. Use sentence case only for
+unconstrained headings; do not globally replace punctuation or Unicode characters.
+
+Do not edit kernel-managed APEX artifacts under `.apex/`, approved or hash-reviewed artifacts, or generated files in
+place as a style pass. Propose the edits instead; the APEX agent stages artifact changes through the kernel before the
+review, and later changes need a fresh review, never hash restamping. This skill cannot grant approvals, resolve
+findings, change APEX run state, or waive a required check.
+
+## Workflow
+
+1. Establish scope and protected content. Reuse current context rather than loading unrelated files or all skills.
+2. Identify specific clarity problems. If the text is already clear, leave it alone.
+3. Make the smallest authorized edits, preserving meaning and the intended tone. Prefer one pass and one self-check.
+4. Compare the result with the original for factual changes, lost caveats, broken structure and over-compression.
+5. For file edits, run the relevant existing checks. Kernel validation and reviews own APEX artifacts; do not run them
+   as part of a style pass. Report any unverified requirement rather than claiming technical correctness.
+
+## Editing guidance
+
+- Remove filler, generic conclusions, flattery and conversational boilerplate when they add no information.
+- State mechanisms and concrete outcomes instead of praise or decorative metaphors. Do not invent measurements.
+- Prefer familiar words and consistent terminology. "Use" often replaces "utilize"; technical terms such as
+  vector, API surface, harness, policy and managed identity remain when precise.
+- Replace repetitive framing with a direct statement. Do not force every explanation into three items.
+- Split dense sentences when it improves comprehension. Keep necessary articles, verbs and qualifications;
+  fewer words are not automatically clearer.
+- Prefer active voice when the actor is known and relevant. Reduce stacked hedges, not justified uncertainty.
+- Reduce gratuitous bold, punctuation and decorative emojis only where the document contract allows it.
+- Preserve useful lists, labels, parentheses and colons. Treat vocabulary patterns as review cues, not a blacklist.
+- Keep general security or safety requirements even when the same sentence could apply to another project.
+
+## Regression examples
+
+| Original | Safe treatment |
+| --- | --- |
+| In order to validate the file, run the existing check. | To validate the file, run the existing check. |
+| The estimate may exceed EUR 200 if usage increases. | Leave unchanged: uncertainty, currency and threshold matter. |
+| Deployment MUST remain blocked until approval. | Preserve MUST, condition and meaning; flag technical concerns separately. |
+| `P0v4`, `Standard_LRS`, `apex-recall`, `--verify-cache` | Preserve these identifiers exactly. |
+| `## 🔒 Approval Gate` | Preserve the required heading and its anchor. |
+| The accepted recovery risk has not been remediated. | Preserve the distinction between acceptance and remediation. |
+| Industry reports suggest this is faster. | Request a source or propose a qualified rewrite; do not fabricate evidence. |
+
+## Output
+
+For supplied text, return the revised text or the requested review. For file edits, name the changed files and
+briefly summarize meaningful edits and checks. Separate technical concerns from style suggestions. Do not add
+generic reassurance, an "AI score", or a claim that the result is human-authored.
+
+## Attribution
+
+Adapted from Lauren Tan's [Unslop skill](https://github.com/cursor/plugins/tree/e8d856f0273b42ebafe0ec3546bd645709e7c1b0/pstack/skills/unslop),
+part of Cursor's pstack plugin, under the [MIT license](LICENSE.txt).
+APEX scopes the original style guidance to explicit requests and adds technical-content and review safeguards.
+Upstream rule numbers are not retained as APEX rule identifiers. Updates require a reviewed adaptation, not a live import.

package/assets/customizations/.github/skills/apex-workflow/SKILL.md

@@ -4,6 +4,9 @@ description: "Provides internal APEX workflow routing for status, resume, projec
 user-invocable: false
 ---
 
+When calling any `apex/*` MCP tool, include the current session checkout or worktree as the required absolute
+`workspace` path.
+
 ## APEX Workflow
 
 Use this skill to orient an interactive agent without reconstructing workflow state from chat or files.
@@ -16,21 +19,28 @@ Use this skill to orient an interactive agent without reconstructing workflow st
 ## Workflow
 
 1. Call `apex/status` for the selected project.
+   For a status-only request, report that result and stop without calling `apex/nextTask`.
+   If status returns `runtime_mismatch`, report both versions and its `nextAction`, then stop; every other APEX tool
+   fails with `APEX_RUNTIME_MISMATCH` until the user fixes the runtime or workspace version.
 2. If the selected project is wrong or absent, use the active client question mechanism to select a project, then repeat
    `apex/status`.
 3. If status identifies blockers, gates, or a terminal run, report that kernel state and do not infer a task.
-4. Otherwise call `apex/nextTask`. For `needs_input`, return the exact request to its owning interactive role. For
-   `task`, route only to the returned task owner through the active client projection.
+4. Otherwise route the next step with the `apex-next` skill (`.github/skills/apex-next/SKILL.md`). It maps
+   `needs_input`, `needs_review` and `status=task` results to their owner and preserves the exact request, review hash
+   or `task.taskId` for that owner task only. It must not add later-stage prerequisites to an earlier task, and it
+   continues directly when the active agent already owns the next result. Read task context only for `status=task`, as
+   its owner. Do not poll unresolved input or review results; obtain a fresh result after the owning role records
+   answers or dispositions.
 5. For resume, fetch fresh status instead of relying on an earlier conversation. On an `APEX_STALE` error, refresh
    status and do not reuse a task ID, request ID, journal head, or owner epoch.
 
-Do not infer progress from generated files, conversation history, or handoff completion. If selection is ambiguous, the
+Do not infer progress from generated files, conversation history, or routing. If selection is ambiguous, the
 interactive agent may use its active client question mechanism before repeating `apex/status`.
 
 ## Boundaries
 
-`apex/nextTask` returns only `needs_input` or `task`. Terminal, blocked, missing-project, authorization, and stale
-conditions are represented by status or stable kernel errors; do not invent additional tool results or transitions.
+`apex/nextTask` returns `needs_input`, `needs_review`, or `task`. Terminal, blocked, missing-project, authorization,
+and stale conditions are represented by status or stable kernel errors; do not invent additional results or transitions.
 
 ## Output
 

package/assets/customizations/.github/workflows/governance-policy-baseline.yml

@@ -0,0 +1,101 @@
+name: Governance Policy Baseline
+
+on:
+  schedule:
+    - cron: "0 5 * * 1"
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+
+jobs:
+  collect-baseline:
+    if: vars.GOVERNANCE_BASELINE_ENABLED == 'true'
+    runs-on: ubuntu-latest
+    environment: governance
+    permissions:
+      contents: write
+      pull-requests: write
+      id-token: write
+    env:
+      AZURE_CLIENT_ID: ${{ vars.AZURE_CLIENT_ID || secrets.AZURE_CLIENT_ID }}
+      AZURE_TENANT_ID: ${{ vars.AZURE_TENANT_ID || secrets.AZURE_TENANT_ID }}
+      AZURE_SUBSCRIPTION_ID: ${{ vars.AZURE_SUBSCRIPTION_ID || secrets.AZURE_SUBSCRIPTION_ID }}
+      GOVERNANCE_MG_ID: ${{ vars.GOVERNANCE_MG_ID }}
+      GOVERNANCE_SUBSCRIPTION_ID: ${{ vars.GOVERNANCE_SUBSCRIPTION_ID }}
+      GOVERNANCE_MAX_SUBSCRIPTIONS: ${{ vars.GOVERNANCE_MAX_SUBSCRIPTIONS || '100' }}
+    steps:
+      - name: Validate collection configuration
+        shell: pwsh
+        run: |
+          $ErrorActionPreference = 'Stop'
+          $hasManagementGroup = -not [string]::IsNullOrEmpty($env:GOVERNANCE_MG_ID)
+          $hasSubscription = -not [string]::IsNullOrEmpty($env:GOVERNANCE_SUBSCRIPTION_ID)
+          if ($hasManagementGroup -eq $hasSubscription) {
+            throw 'Configure exactly one governance management group or subscription.'
+          }
+          if ($hasManagementGroup -and $env:GOVERNANCE_MG_ID -cnotmatch '\A[a-zA-Z0-9](?:[a-zA-Z0-9_().-]{0,88}[a-zA-Z0-9_()-])?\z') {
+            throw 'Invalid governance management group ID.'
+          }
+          $guidPattern = '\A[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}\z'
+          if ($hasSubscription -and $env:GOVERNANCE_SUBSCRIPTION_ID -cnotmatch $guidPattern) {
+            throw 'Invalid governance subscription ID.'
+          }
+          foreach ($value in @($env:AZURE_CLIENT_ID, $env:AZURE_TENANT_ID, $env:AZURE_SUBSCRIPTION_ID)) {
+            if ([string]::IsNullOrEmpty($value) -or $value -cnotmatch $guidPattern) {
+              throw 'Configure valid Azure OIDC client, tenant, and login subscription IDs.'
+            }
+          }
+          $maximum = 0
+          if ($env:GOVERNANCE_MAX_SUBSCRIPTIONS -cnotmatch '\A[0-9]+\z' -or
+              -not [int]::TryParse($env:GOVERNANCE_MAX_SUBSCRIPTIONS, [ref]$maximum) -or $maximum -lt 1) {
+            throw 'Maximum subscriptions must be an integer from 1 through 2147483647.'
+          }
+
+      - name: Checkout repository
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+
+      - name: Azure Login (OIDC)
+        uses: azure/login@a641126d1b8aa4d1fa005f4f92df94a3a4c4c906 # v3.1.0
+        with:
+          client-id: ${{ env.AZURE_CLIENT_ID }}
+          tenant-id: ${{ env.AZURE_TENANT_ID }}
+          subscription-id: ${{ env.AZURE_SUBSCRIPTION_ID }}
+
+      - name: Collect governance baseline
+        shell: pwsh
+        run: |
+          $ErrorActionPreference = 'Stop'
+          $parameters = @{ MaxSubscriptions = [int]$env:GOVERNANCE_MAX_SUBSCRIPTIONS; IncludeDefenderAuto = $true; IncludeDescendants = $true }
+          if ($env:GOVERNANCE_MG_ID) {
+            $parameters.ManagementGroupId = $env:GOVERNANCE_MG_ID
+          } else {
+            $parameters.SubscriptionId = $env:GOVERNANCE_SUBSCRIPTION_ID
+          }
+          try {
+            & ./tools/scripts/collect-governance-baseline.ps1 @parameters *> $null
+            $baseline = Get-Content -LiteralPath '.github/data/governance-policy-baseline.json' -Raw
+            if (-not (Test-Json -Json $baseline -SchemaFile './tools/schemas/governance-baseline.schema.json' -ErrorAction Stop)) {
+              throw 'Invalid baseline.'
+            }
+          } catch {
+            throw 'Governance collection or schema validation failed; no baseline PR will be created.'
+          }
+
+      - name: Create pull request
+        uses: peter-evans/create-pull-request@5f6978faf089d4d20b00c7766989d076bb2fc7f1 # v8.1.1
+        with:
+          token: ${{ secrets.GITHUB_TOKEN }}
+          add-paths: .github/data/governance-policy-baseline.json
+          commit-message: "chore(data): Update governance policy baseline"
+          branch: "chore/governance-baseline-update"
+          delete-branch: true
+          title: "chore(data): Update governance policy baseline"
+          body: |
+            Updates only `.github/data/governance-policy-baseline.json`.
+            Human review and manual merge are required. Review scope, coverage,
+            exclusions, and policy changes before approving this baseline.

package/assets/customizations/README.md

@@ -1,16 +1,15 @@
 ## Managed Workspace Source
 
-This directory is the versioned source bundle installed into consumer repositories by `apex init` and updated by
-`apex update`. Its payload uses supported workspace discovery paths:
+This directory is the versioned source for two outputs:
 
-- `.github/agents/`
-- `.github/skills/`
-- `.github/copilot-instructions.md`
-- `.vscode/mcp.json`
+- The thin workspace projection that `apex init` installs and `apex update` updates: `.github/copilot/settings.json`,
+  `.github/copilot-instructions.md`, `.github/instructions/`, and the governance workflow and scripts.
+- The `apex` plugin built by `tools/scripts/build-plugin.mjs`: agents from `.github/agents/` and skills from
+  `.github/skills/`. The plugin owns these paths, so they are never copied into a consumer workspace.
 
-`manifest.json` records the bundle version, managed files, agent roles, supported targets, invocation edges, interaction
-types, recommended models, and cost tiers. A client projection omits roles that do not declare its target; unsupported
-roles are never installed or added to the managed lock.
+`manifest.json` records the bundle version, managed workspace files, the plugin declaration and the files it owns,
+agent roles, supported targets, invocation edges and interaction types. The plugin renders only roles that declare the
+`github-copilot` target.
 
 ## Editing Policy