build: Fixed some model settings

2026-04-09 16:41:09 -04:00
parent 6a78953691
commit fc9c730670
6 changed files with 6 additions and 166 deletions
@@ -8,7 +8,7 @@ description: >
 mode: subagent
 hidden: true
 temperature: 0.1
-model: anthropic/claude-haiku-4-6
+model: anthropic/claude-haiku-4-5
 color: info
 permission:
  edit: deny
@@ -1,160 +0,0 @@
-# New Escalation Architecture Proposal
-
-## Overview
-
-This proposal redesigns the escalation system to eliminate redundancy by leveraging the fact that when a model isn't specified, the agent inherits the model of its caller.
-
-## Current Problem
-
-We have massive duplication:
- `implementer-haiku.md`, `implementer-codex.md`, `implementer-sonnet.md`, `implementer-opus.md`
- `behave-tester-haiku.md`, `behave-tester-codex.md`, `behave-tester-sonnet.md`, `behave-tester-opus.md`
- `robot-tester-haiku.md`, `robot-tester-codex.md`, `robot-tester-sonnet.md`, `robot-tester-opus.md`
-
-Each contains 90% identical content with only the model and escalation context differing.
-
-## Proposed Solution
-
-### 1. Core Worker Agents (No Model Specified)
-
-Create model-agnostic worker agents that contain the actual implementation logic:
-
-```yaml
-# implementer.md
---
-description: >
-  Core implementation worker that writes code for subtasks. Model is inherited
-  from the calling tier agent, enabling progressive escalation without duplication.
-mode: subagent
-hidden: true
-temperature: 0.2
-# NO MODEL SPECIFIED - inherits from caller
-permission:
-  edit: allow
-  bash:
-    "*": allow
-    # Block ALL commands that could hit the label creation endpoints
-    "*api/v1/orgs/*/labels*": deny
-    "*api/v1/repos/*/labels*": deny
-    "*https://git.cleverthis.com/api/v1/repos/cleveragents/cleveragents-core/labels*": deny
-  task:
-    "*": deny
-    "ref-reader": allow
-  forgejo:
-    "*": allow
-    # CRITICAL: Label creation is COMPLETELY FORBIDDEN
-    "forgejo_create_label": deny
-    "forgejo_create_org_label": deny
-    "forgejo_create_repo_label": deny
-    "forgejo_add_issue_labels": deny
---
-
-# [Full implementation logic here - identical for all tiers]
-```
-
-### 2. Tier Selector Agents
-
-Create thin wrapper agents that only set the model and invoke the worker:
-
-```yaml
-# tier-haiku.md
---
-description: >
-  Haiku tier selector for progressive escalation. Sets model to Haiku and
-  invokes the requested worker agent.
-mode: subagent
-hidden: true
-temperature: 0.0
-model: anthropic/claude-haiku-4-6
-permission:
-  task:
-    "implementer": allow
-    "behave-tester": allow
-    "robot-tester": allow
-    "unit-test-runner": allow
-    "typecheck-fixer": allow
---
-
-# Tier Selector: Haiku
-
-You are a tier selector that enables Haiku model usage for worker agents.
-
-When invoked with a worker type and context, immediately invoke that worker
-with the provided context. Your Haiku model will be inherited by the worker.
-
-## Workers you can invoke:
- `implementer` - Code implementation
- `behave-tester` - Unit test writing
- `robot-tester` - Integration test writing
- `unit-test-runner` - Unit test execution and fixing
- `typecheck-fixer` - Type error resolution
-
-Pass ALL provided context directly to the worker without modification.
-```
-
-### 3. Updated Subtask Loop
-
-The subtask-loop agent would be updated to call tier selectors:
-
-```python
-# Instead of:
-invoke implementer-{tier}
-
-# Now:
-invoke tier-{tier} with:
-  worker_type: "implementer"
-  context: {all the implementation context}
-```
-
-### 4. Benefits
-
-1. **Eliminates redundancy** - One implementation per worker type
-2. **Easier maintenance** - Update logic in one place
-3. **Cleaner architecture** - Separation of concerns
-4. **Flexible escalation** - Easy to add new tiers or workers
-5. **Consistent behavior** - All tiers guaranteed to use identical logic
-
-### 5. Migration Path
-
-Phase 1: Create new architecture
- Create core worker agents (implementer.md, behave-tester.md, etc.)
- Create tier selector agents (tier-haiku.md, tier-codex.md, etc.)
- Update subtask-loop.md to use new pattern
-
-Phase 2: Testing
- Test with one worker type first (e.g., implementer)
- Verify escalation works correctly
- Ensure state persistence still functions
-
-Phase 3: Full rollout
- Migrate all worker types
- Delete old tier-specific agents
- Update any other agents that reference them
-
-### 6. State Persistence
-
-The escalation state tracking remains in subtask-loop and other orchestrators.
-The tier selectors are stateless - they just set the model and pass through.
-
-### 7. Example Usage
-
-```
-subtask-loop determines tier="haiku", worker="implementer"
-  → invokes tier-haiku with {worker_type: "implementer", context: {...}}
-    → tier-haiku (model: haiku) invokes implementer with context
-      → implementer (inherits haiku model) does the work
-
-If that fails, subtask-loop escalates to tier="codex"
-  → invokes tier-codex with {worker_type: "implementer", context: {...}}  
-    → tier-codex (model: codex) invokes implementer with context
-      → implementer (inherits codex model) does the work
-```
-
-## Implementation Order
-
-1. Start with implementer as proof of concept
-2. Add unit-test-runner and typecheck-fixer (as requested)
-3. Migrate behave-tester and robot-tester
-4. Consider other agents that could benefit (coverage-checker, test-fixer)
-
-This architecture provides maximum flexibility with minimal redundancy.
@@ -6,7 +6,7 @@ description: >
 mode: subagent
 hidden: true
 temperature: 0.0
-model: anthropic/claude-haiku-4-6
+model: anthropic/claude-haiku-4-5
 color: info
 permission:
  edit: deny
@@ -7,7 +7,7 @@ description: >
 mode: subagent
 hidden: true
 temperature: 0.3
-model: anthropic/claude-haiku-4-6
+model: anthropic/claude-haiku-4-5
 color: "#9B59B6"
 permission:
  edit: deny
@@ -6,7 +6,7 @@ description: >
 mode: subagent
 hidden: true
 temperature: 0.3
-model: anthropic/claude-haiku-4-6
+model: anthropic/claude-haiku-4-5
 color: secondary
 permission:
  edit: deny
@@ -5,7 +5,7 @@ description: >
 mode: subagent
 hidden: true
 temperature: 0.0
-model: anthropic/claude-haiku-4-6
+model: anthropic/claude-haiku-4-5
 permission:
  task:
    "implementer": allow
@@ -84,4 +84,4 @@ invoke implementer
  [all other fields exactly as provided]
 ```

-The implementer will then execute using the Haiku model inherited from you.
+The implementer will then execute using the Haiku model inherited from you.