feat: balance and simplify delivery arrangements - #3
Conversation
There was a problem hiding this comment.
🔵 Needs a closer look
Unresolved moderate findings remain in parent-only routing, with related documentation inconsistencies.
Pull request overview
Adds Balanced Delivery and simplifies Advanced and Lean arrangements through revised role catalogs, model allocations, concurrency caps, and routing documentation.
Changes:
- Adds the 12-role Balanced arrangement.
- Rebalances Advanced and Lean rosters and capacity limits.
- Updates routing, installation, rollback, validation, configuration, and changelog documentation.
File summaries
| File | Summary / final review status |
|---|---|
README.md |
Updates arrangement matrix and version status. |
docs/validation.md |
Documents revised caps and historical evidence. Nit (1 vote): Current status still presents 0.1.0 as current while documenting 0.2.0 caveats. |
docs/README.pt-BR.md |
Updates Portuguese arrangement guidance. Nit (1 vote): Lean role-model wording contradicts the allocation table and English guide. |
docs/orchestration.md |
Documents routing and responsibility guidance. Moderate (1 vote): Removing security-sweep leaves generic security coverage without an explicit role contract. Nit (1 vote): Lean’s lighter-model comparison is misleading for its Astra and Sol high roles. |
docs/installation.md |
Adds upgrade and rollback instructions. |
docs/configuration.md |
Records the 0.2.0 configuration changes. |
docs/CHANGELOG.md |
Summarizes the release changes and validation limits. |
arrangements/lean-delivery/README.md |
Documents Lean’s reduced roster and routing. |
arrangements/lean-delivery/config.toml |
Sets Lean’s three-thread cap. |
arrangements/lean-delivery/arrangement.toml |
Bumps Lean to 0.2.0. |
arrangements/lean-delivery/agents/test-engineer.toml |
Removes the obsolete test-author role. |
arrangements/lean-delivery/agents/security-sweep.toml |
Removes the obsolete security role. |
arrangements/lean-delivery/agents/researcher.toml |
Adds parent-mediated reporting guidance. |
arrangements/lean-delivery/agents/integrator-reviewer.toml |
Adds parent-only handoff guidance. Moderate (1 vote): High-risk certification still routes directly to critical-reviewer. |
arrangements/lean-delivery/agents/hard-task-specialist.toml |
Removes the obsolete coupled-work role. |
arrangements/lean-delivery/agents/frontend-worker.toml |
Routes escalation to the parent. |
arrangements/lean-delivery/agents/frontend-worker-light.toml |
Removes the obsolete light frontend role. |
arrangements/lean-delivery/agents/explorer.toml |
Routes diagnosis to the parent. Moderate (1 vote): Outside-source work still routes directly to researcher. |
arrangements/lean-delivery/agents/docs-writer.toml |
Removes the obsolete documentation role. |
arrangements/lean-delivery/agents/design-lead.toml |
Removes the obsolete design role. |
arrangements/lean-delivery/agents/debugger.toml |
Removes the obsolete debugger role. |
arrangements/lean-delivery/agents/critical-reviewer.toml |
Reallocates critical review to Luna. |
arrangements/lean-delivery/agents/code-reviewer.toml |
Adds parent-only handoff guidance. Moderate (1 vote): High-risk review needs still route directly instead of returning to the parent. |
arrangements/lean-delivery/agents/backend-worker.toml |
Consolidates small-work ownership. |
arrangements/lean-delivery/agents/backend-worker-light.toml |
Removes the obsolete light backend role. |
arrangements/lean-delivery/agents/advisor.toml |
Reallocates advice to Astra. |
arrangements/lean-delivery/AGENTS.with-skill.snippet.md |
Updates routing and capacity guidance. |
arrangements/lean-delivery/AGENTS.snippet.md |
Updates routing and capacity guidance. |
arrangements/balanced-delivery/README.md |
Adds Balanced arrangement documentation. |
arrangements/balanced-delivery/config.toml |
Defines the four-thread cap. |
arrangements/balanced-delivery/arrangement.toml |
Adds the Balanced manifest. |
arrangements/balanced-delivery/agents/test-runner.toml |
Adds parent-mediated test routing. Moderate (1 vote): Follow-up tasks can still bypass parent dispatch and slot management. |
arrangements/balanced-delivery/agents/researcher.toml |
Adds the researcher role. |
arrangements/balanced-delivery/agents/integrator-reviewer.toml |
Adds the Sol integration reviewer. Moderate (1 vote): High-risk work still routes directly to critical-reviewer. |
arrangements/balanced-delivery/agents/infra-sre.toml |
Adds the infrastructure role. Moderate (1 vote): Unknown failures can still route directly to hard-task-specialist. |
arrangements/balanced-delivery/agents/hard-task-specialist.toml |
Adds the Astra coupled-work role. |
arrangements/balanced-delivery/agents/frontend-worker.toml |
Adds the frontend role. Moderate (1 vote): Coupled UI state can still bypass the parent. |
arrangements/balanced-delivery/agents/explorer.toml |
Adds the explorer role. Moderate (1 vote): Child-to-child routing remains for coupled work and external research. |
arrangements/balanced-delivery/agents/database-engineer.toml |
Adds the database role. Moderate (1 vote): Migration risk still routes directly to critical-reviewer. |
arrangements/balanced-delivery/agents/critical-reviewer.toml |
Adds the Luna critical reviewer. |
arrangements/balanced-delivery/agents/code-reviewer.toml |
Adds the code reviewer. Moderate (1 vote): High-risk review can still bypass parent dispatch. |
arrangements/balanced-delivery/agents/backend-worker.toml |
Adds the Sol backend worker. Moderate (1 vote): Coupled work can still route directly to hard-task-specialist. |
arrangements/balanced-delivery/agents/advisor.toml |
Adds the Astra advisor. |
arrangements/balanced-delivery/AGENTS.with-skill.snippet.md |
Adds Balanced routing and capacity guidance. |
arrangements/balanced-delivery/AGENTS.snippet.md |
Adds Balanced standalone guidance. |
arrangements/advanced-delivery/README.md |
Documents the reduced Advanced roster. |
arrangements/advanced-delivery/config.toml |
Sets the five-thread cap. |
arrangements/advanced-delivery/arrangement.toml |
Bumps Advanced to 0.2.0. |
arrangements/advanced-delivery/agents/test-runner.toml |
Adds parent-only handoff guidance. Moderate (1 vote): Unknown failures and coverage work still dispatch directly to other children. |
arrangements/advanced-delivery/agents/test-engineer.toml |
Reallocates test authoring to Luna. Moderate (1 vote): Production fixes and reviewer selection can still be handed to another child. |
arrangements/advanced-delivery/agents/security-sweep.toml |
Removes the obsolete security role. |
arrangements/advanced-delivery/agents/researcher.toml |
Adds parent-mediated reporting. |
arrangements/advanced-delivery/agents/integrator-reviewer.toml |
Reallocates integration review to Sol. Moderate (1 vote): High-risk certification still routes directly to critical-reviewer. |
arrangements/advanced-delivery/agents/infra-sre.toml |
Reallocates infrastructure work to Luna. Moderate (1 vote): Unknown failures still route directly to debugger. |
arrangements/advanced-delivery/agents/hard-task-specialist.toml |
Preserves Astra coupled implementation. |
arrangements/advanced-delivery/agents/frontend-worker.toml |
Reallocates frontend work to Luna. Moderate (1 vote): Coupled state can still route directly to hard-task-specialist. |
arrangements/advanced-delivery/agents/frontend-worker-light.toml |
Removes the obsolete light frontend role. |
arrangements/advanced-delivery/agents/explorer.toml |
Adds parent-mediated reporting. Moderate (1 vote): Unknown-cause work and external research still route directly to children. |
arrangements/advanced-delivery/agents/docs-writer.toml |
Removes the obsolete documentation role. |
arrangements/advanced-delivery/agents/design-lead.toml |
Preserves Astra design work. |
arrangements/advanced-delivery/agents/debugger.toml |
Reallocates diagnosis to Sol. Moderate (1 vote): Fixes and test execution still dispatch directly to other children. |
arrangements/advanced-delivery/agents/database-engineer.toml |
Reallocates database work to Luna. Moderate (1 vote): Migration risk still routes directly to critical-reviewer. |
arrangements/advanced-delivery/agents/critical-reviewer.toml |
Preserves Astra critical review. |
arrangements/advanced-delivery/agents/code-reviewer.toml |
Reallocates review to Sol. Moderate (1 vote): High-risk review still bypasses parent dispatch. |
arrangements/advanced-delivery/agents/backend-worker.toml |
Reallocates backend work to Sol. Moderate (1 vote): Coupled work still routes directly to hard-task-specialist. |
arrangements/advanced-delivery/agents/backend-worker-light.toml |
Removes the obsolete light backend role. |
arrangements/advanced-delivery/agents/advisor.toml |
Adds parent-mediated reporting. |
arrangements/advanced-delivery/AGENTS.with-skill.snippet.md |
Updates routing and capacity guidance. |
arrangements/advanced-delivery/AGENTS.snippet.md |
Updates standalone guidance. |
Review details
Suppressed comments (25)
arrangements/advanced-delivery/agents/backend-worker.toml:14
- The new parent-only instruction at line 14 is contradicted by this remaining direct handoff: a backend child can send coupled work straight to
hard-task-specialist, bypassing the parent’s dispatch and acceptance ownership. Return the recommendation to the parent instead.
Return every result, blocker, and role recommendation to the parent. References to other roles are suggestions for the parent, never child-to-child handoffs. Do not contact or wait for another child; the parent decides whether to act, spawn, or resume a role.
arrangements/advanced-delivery/agents/code-reviewer.toml:14
- The new parent-only instruction at line 14 is contradicted by this direct routing instruction. Return high-risk review needs to the parent instead of allowing a child reviewer to bypass the parent’s dispatch and briefing.
Return every result, blocker, and role recommendation to the parent. References to other roles are suggestions for the parent, never child-to-child handoffs. Do not contact or wait for another child; the parent decides whether to act, spawn, or resume a role.
arrangements/advanced-delivery/agents/database-engineer.toml:14
- The new parent-only instruction at line 14 is contradicted by this remaining direct handoff: the database child sends migration risk directly to
critical-reviewer, bypassing the parent’s review decision. Return the risk and recommendation to the parent instead.
Return every result, blocker, and role recommendation to the parent. References to other roles are suggestions for the parent, never child-to-child handoffs. Do not contact or wait for another child; the parent decides whether to act, spawn, or resume a role.
arrangements/advanced-delivery/agents/debugger.toml:14
- The new parent-only instruction at line 14 is contradicted by this remaining direct dispatch: the debugger sends fixes and test execution to other children. Return those recommendations to the parent so it can choose the next role and provide its brief.
Return every result, blocker, and role recommendation to the parent. References to other roles are suggestions for the parent, never child-to-child handoffs. Do not contact or wait for another child; the parent decides whether to act, spawn, or resume a role.
arrangements/advanced-delivery/agents/explorer.toml:14
- The new parent-only instruction at line 14 is contradicted by these direct child-to-child handoffs: discovery can send unknown-cause work to
debuggerand external research toresearcher. Return both recommendations to the parent, which owns dispatch.
Return every result, blocker, and role recommendation to the parent. References to other roles are suggestions for the parent, never child-to-child handoffs. Do not contact or wait for another child; the parent decides whether to act, spawn, or resume a role.
arrangements/advanced-delivery/agents/frontend-worker.toml:14
- The new parent-only instruction at line 14 is contradicted by this remaining direct handoff: a frontend child can send coupled state straight to
hard-task-specialist, bypassing the parent’s dispatch and brief. Change the later instruction to return the recommendation to the parent.
Return every result, blocker, and role recommendation to the parent. References to other roles are suggestions for the parent, never child-to-child handoffs. Do not contact or wait for another child; the parent decides whether to act, spawn, or resume a role.
arrangements/advanced-delivery/agents/infra-sre.toml:14
- The new parent-only instruction at line 14 is contradicted by this remaining direct handoff: the infrastructure child sends unknown failures straight to
debugger, bypassing parent routing and briefing.
Return every result, blocker, and role recommendation to the parent. References to other roles are suggestions for the parent, never child-to-child handoffs. Do not contact or wait for another child; the parent decides whether to act, spawn, or resume a role.
arrangements/advanced-delivery/agents/integrator-reviewer.toml:14
- The new parent-only instruction at line 14 is contradicted by this remaining direct handoff: the integration reviewer sends high-risk certification to
critical-reviewerinstead of returning the requirement to the parent.
Return every result, blocker, and role recommendation to the parent. References to other roles are suggestions for the parent, never child-to-child handoffs. Do not contact or wait for another child; the parent decides whether to act, spawn, or resume a role.
arrangements/advanced-delivery/agents/test-engineer.toml:14
- The new parent-only instruction at line 14 is contradicted by this remaining direct handoff: the test-engineer is told to hand production fixes to another child and identify a reviewer. Return those recommendations to the parent so it controls the next brief and slot.
Return every result, blocker, and role recommendation to the parent. References to other roles are suggestions for the parent, never child-to-child handoffs. Do not contact or wait for another child; the parent decides whether to act, spawn, or resume a role.
arrangements/advanced-delivery/agents/test-runner.toml:14
- The new parent-only instruction at line 14 is contradicted by this remaining direct dispatch: the test runner sends unknown failures and coverage work to other children. Return the classification and recommendations to the parent so it controls the next brief and slot.
Return every result, blocker, and role recommendation to the parent. References to other roles are suggestions for the parent, never child-to-child handoffs. Do not contact or wait for another child; the parent decides whether to act, spawn, or resume a role.
arrangements/balanced-delivery/agents/backend-worker.toml:18
- This direct handoff contradicts the parent-only contract added above: a child can bypass the parent’s brief and ownership by sending coupled work straight to
hard-task-specialist. Return the blocker and role recommendation to the parent, which should decide whether to dispatch the next child.
Prove the result at the lowest responsible layer and add focused regression coverage for changed executable behavior. Use real collaborators when claiming an integration works. Stop for an unresolved contract, authorization, billing, destructive-data, concurrency, or transaction-design decision; report the evidence and needed owner rather than guessing. Send coupled cross-component work to hard-task-specialist.
arrangements/balanced-delivery/agents/code-reviewer.toml:20
- This direct routing instruction conflicts with the parent-only contract added at line 14. A child reviewer should return the high-risk review recommendation to the parent; otherwise it can bypass the parent’s dispatch and complete brief.
arrangements/balanced-delivery/agents/database-engineer.toml:18 - This direct handoff contradicts the parent-only contract: the database child is instructed to send a risk review directly to
critical-reviewer, bypassing the parent’s routing and acceptance decision. Return the risk and review recommendation to the parent instead.
arrangements/balanced-delivery/agents/explorer.toml:18 - This still routes work directly from one child to two other children, contradicting the parent-only dispatch contract added above. Return both recommendations to the parent; it must decide whether to dispatch
hard-task-specialistorresearcher.
arrangements/balanced-delivery/agents/frontend-worker.toml:20 - This direct handoff contradicts the parent-only contract added above: a child can bypass the parent by sending coupled UI state to another child. Return the blocker and role recommendation to the parent, which should decide whether to dispatch
hard-task-specialist.
arrangements/balanced-delivery/agents/infra-sre.toml:20 - This direct handoff contradicts the parent-only contract: the infrastructure child can route an unknown failure straight to
hard-task-specialist. Return the failure mechanism and recommendation to the parent, which should own the next dispatch.
arrangements/balanced-delivery/agents/integrator-reviewer.toml:20 - This direct handoff contradicts the parent-only routing contract: the integration reviewer is told to send high-risk work straight to
critical-reviewer. Return the certification requirement to the parent so it can choose the reviewer and provide the complete brief.
arrangements/balanced-delivery/agents/test-runner.toml:20 - This tells a child to dispatch both follow-up tasks directly, despite the new rule that role recommendations return only to the parent. That can bypass the parent’s independent brief and slot management; return the classification and recommendation to the parent instead.
arrangements/lean-delivery/agents/code-reviewer.toml:14 - The new parent-only instruction at line 14 is contradicted by this direct routing instruction. Return high-risk review needs to the parent instead of allowing a child reviewer to bypass the parent’s dispatch and briefing.
arrangements/lean-delivery/agents/explorer.toml:18 - This remaining direct handoff contradicts the parent-only routing contract: the Lean explorer sends outside-source work directly to
researcher. Return that role recommendation to the parent, which must decide whether to dispatch it.
arrangements/lean-delivery/agents/integrator-reviewer.toml:14 - The new parent-only instruction at line 14 is contradicted by this remaining direct handoff: the Lean integrator sends high-risk certification to
critical-reviewerrather than returning the requirement to the parent.
docs/README.pt-BR.md:19 - This Portuguese sentence says that the other Lean roles also use Luna, but the immediately preceding allocation assigns the advisor to Astra and the integrator to Sol. That makes the translated routing guide contradict both the role table and the English guide; narrow the sentence to frontend/backend (and the Luna roles) instead of all remaining roles.
docs/orchestration.md:44 - Assigning the removed
security-sweeptocritical-reviewerleaves a scope gap: the retained critical-reviewer contract names authorization, billing, public APIs, destructive migrations, and core invariants, while the deleted sweep covered secrets, injection/traversal/SSRF, and dependency risk. A routine security sweep therefore has no role with an explicit contract; expand critical-reviewer or document that generic sweeps require separately qualified help.
docs/orchestration.md:22 - This still describes Lean as using lighter models, but its
advisorisgpt-6-astraathigh, the same allocation as Advanced's advisor, and its integrator is also Solhigh. That makes the comparison misleading for users choosing the lower-cost topology; qualify the claim as a narrower roster with lighter models for most roles.
docs/validation.md:53 - These new 0.2.0 caveats are correct, but the
Current measured statussection above still begins with “Version 0.1.0 is...”, so this page presents the current release as 0.1.0 while documenting 0.2.0 here. Update that status text or explicitly label it as historical so readers do not mistake the old evidence for the current release.
- Files reviewed: 69/69 changed files
- Comments generated: 0
- Review effort level: Lite
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
Arrangement or component and purpose
Add a middle-ground delivery option and reduce role-selection and coordination overhead. Advanced, Balanced, and Lean now provide 15, 12, and 8 roles with at most 4, 3, and 2 simultaneous children. Model allocation follows responsibility; Luna implements established patterns, not just narrow discovery. References #1; does not close its unverified runtime acceptance.
Changes
Functional evidence
Catalog validation loads all three packages and all 35 role TOMLs. The existing 37 validator/component/hook behavioral tests pass on Windows using the standard-library tooling. Component preflight reports two skills and zero plugin bundles.
The target remains Codex CLI 0.155.0-alpha.3.10. No new authenticated native test was run for this revision. Existing native writer/hook results are historical, not proof of the new allocation; complete role coverage, saturation and slot release remain unverified. No permission, authentication, MCP or hook behavior is changed.
Checks