git:20260824.f7cdde8 to git:20261011.f93f7ea
16 added, 47 removed. Audit A to A.
---
name: sisyphus-junior-agent
- description: Developer reference for the Sisyphus-Junior category-spawned executor agent -- model variants and discipline.
+ description: Sisyphus-Junior factory, shared Native model core, and OpenCode role composition.
---
- # src/agents/sisyphus-junior/ -- Category-Spawned Executor
-
- **Generated:** 2026-05-18
-
- ## OVERVIEW
-
- 12 files. Sisyphus-Junior is a focused task executor spawned by `delegate-task` when category routing requires it. Runs in subagent mode with its own fallback chain. Does not delegate further; executes directly.
-
- ## FILES
-
- | File | Purpose |
- |------|---------|
- | `agent.ts` | `createSisyphusJuniorAgentWithOverrides()` factory, model-variant routing, `SISYPHUS_JUNIOR_DEFAULTS` |
- | `index.ts` | Barrel exports |
- | `default.ts` | Base/Claude prompt: todo discipline, verification, termination rules |
- | `gemini.ts` | Gemini-optimized prompt variant |
- | `gpt.ts` | Base GPT prompt variant |
- | `gpt-5-4.ts` | GPT-5.4-native prompt variant |
- | `gpt-5-5.ts` | Shared GPT-5.5/GPT-5.6 model-aware prompt variant |
- | `kimi-k2-6.ts` | Kimi K2.6 prompt variant |
- | `kimi-k2-7.ts` | Kimi K2.7-native prompt variant |
- | `kimi-k3.ts` | Kimi K3-native prompt variant (reasoning depth with built-in stop conditions) |
- | `glm-5-2.ts` | GLM-5.2-native prompt variant |
- | `index.test.ts` | Unit tests |
-
- ## VARIANT SELECTION
+ # Sisyphus-Junior
- Parent `agent.ts` selects prompt variant by model name:
- - Kimi K3 -> `kimi-k3.ts`
- - Kimi K2.7 -> `kimi-k2-7.ts`
- - Contains "kimi-k2" -> `kimi-k2-6.ts`
- - Contains "gpt-5.5" or "gpt-5.6" -> shared `gpt-5-5.ts` prompt family with model-aware identity
- - Contains "gpt-5.4" -> `gpt-5-4.ts`
- - Contains "gpt" -> `gpt.ts`
- - Contains "gemini" -> `gemini.ts`
- - Contains "glm" -> `glm-5-2.ts`
- - Default -> `default.ts` (Claude, etc.)
+ The category-spawned executor renders its model core through
+ `native-model-prompt.ts` and appends its OpenCode role through
+ `opencode-role-append.ts`. There are no local per-model prompt files or routing tables.
+ The public Senpi preset contract owns both selection and content.
- ## KEY BEHAVIORS
+ - `agent.ts` owns the factory, defaults, permission merging and user appends.
+ - `index.ts` exports the factory and prompt entrypoints.
+ - `index.test.ts` covers configuration and permission behavior.
+ - The parent `per-agent-prompt-parity.test.ts` compares rendered content over the catalog.
- - Mode: `subagent` (uses own fallback chain, ignores UI selection)
- - Default model: `claude-sonnet-5`
- - Default temperature: `0.1` (`SISYPHUS_JUNIOR_DEFAULTS`)
- - Fallback chain: claude-sonnet-5 -> kimi-k3 -> gpt-5.6-sol medium -> minimax-m3 -> MiniMax-M3 -> minimax-m2.7 -> big-pickle
- - Blocked tools: `task` (all models); `apply_patch` also blocked for GPT models
- - `call_omo_agent` explicitly allowed so subagents can spawn explore/librarian
- - Max tokens: 64000
- - Thinking enabled for non-GPT/non-GLM models (budgetTokens: 32000)
- - Reasoning effort "medium" for GPT models
+ Mode remains subagent, the default model is claude-sonnet-5, temperature is
+ 0.1 and maxTokens is 64000. The task tool is denied for every model;
+ call_omo_agent remains allowed for research. GPT uses medium reasoning by
+ default; GLM has no injected thinking, and the existing Claude thinking
+ configuration applies elsewhere. User file-backed prompt appends still
+ resolve through the existing URI boundary and remain last.