pwn 0.5.639 → 0.5.643
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- checksums.yaml +4 -4
- data/Gemfile +3 -3
- data/README.md +1 -1
- data/documentation/Cron.md +10 -8
- data/documentation/Reinforcement-Learning.md +43 -21
- data/documentation/diagrams/cron-scheduling.svg +134 -93
- data/documentation/diagrams/dot/cron-scheduling.dot +9 -3
- data/documentation/diagrams/dot/pwn-ai-feedback-learning-loop.dot +2 -2
- data/documentation/diagrams/dot/reinforcement-learning.dot +9 -3
- data/documentation/diagrams/pwn-ai-feedback-learning-loop.svg +103 -101
- data/documentation/diagrams/reinforcement-learning.svg +211 -166
- data/documentation/pwn-ai-Agent.md +12 -0
- data/lib/pwn/ai/agent/curriculum.rb +647 -56
- data/lib/pwn/ai/agent/learning.rb +87 -15
- data/lib/pwn/ai/agent/loop.rb +164 -13
- data/lib/pwn/ai/agent/metrics.rb +32 -13
- data/lib/pwn/ai/agent/mistakes.rb +139 -8
- data/lib/pwn/ai/agent/prompt_builder.rb +2 -2
- data/lib/pwn/ai/agent/registry.rb +13 -3
- data/lib/pwn/ai/agent/result.rb +19 -2
- data/lib/pwn/ai/agent/reward.rb +415 -42
- data/lib/pwn/ai/agent/tools/curriculum.rb +51 -0
- data/lib/pwn/ai/agent/tools/reward.rb +15 -4
- data/lib/pwn/ai/anthropic.rb +265 -4
- data/lib/pwn/ai/ollama.rb +274 -10
- data/lib/pwn/ai/open_ai.rb +307 -4
- data/lib/pwn/config.rb +137 -12
- data/lib/pwn/cron.rb +24 -2
- data/lib/pwn/memory.rb +13 -1
- data/lib/pwn/version.rb +1 -1
- data/spec/integration/persistence_roundtrip_spec.rb +6 -1
- data/spec/integration/reinforced_feedback_loop_spec.rb +144 -8
- data/spec/lib/pwn/ai/agent/mistakes_spec.rb +27 -0
- data/spec/lib/pwn/ai/agent/result_spec.rb +11 -0
- data/spec/lib/pwn/ai/agent/reward_spec.rb +89 -0
- data/spec/lib/pwn/ai/ollama_spec.rb +251 -0
- data/third_party/pwn_rdoc.jsonl +58 -7
- metadata +8 -8
|
@@ -123,6 +123,18 @@ full `Loop.run` under a persona overlay) that share a JSONL bus. See
|
|
|
123
123
|
beats the old one on the current `Mistakes.top` set. See
|
|
124
124
|
[Reinforcement Learning](Reinforcement-Learning.md).
|
|
125
125
|
|
|
126
|
+
## RL feature flags (`PWN::Env[:ai][:agent]`)
|
|
127
|
+
|
|
128
|
+
| Flag | Default | Effect |
|
|
129
|
+
|---|---|---|
|
|
130
|
+
| `critic` / `counterfactual` / `red_team_plan` | `nil` (auto) | ON for remote engines, OFF for ollama |
|
|
131
|
+
| `hindsight` | `true` | C3 HER soft-relabel on failed turns |
|
|
132
|
+
| `reward_llm` | `nil` (auto) | ORM/PRM use LLM teacher on remote even when `module_reflection` is false |
|
|
133
|
+
| `verify_as_reward` | `nil` (auto) | E3 sample policy |
|
|
134
|
+
| `local_introspect` | `:failure_only` | ollama end-of-turn introspect policy |
|
|
135
|
+
|
|
136
|
+
Full tier table: [Reinforcement Learning](Reinforcement-Learning.md).
|
|
137
|
+
|
|
126
138
|
**See also:** [AI Integration](AI-Integration.md) ·
|
|
127
139
|
[Skills, Memory & Learning](Skills-Memory-Learning.md) ·
|
|
128
140
|
[Mistakes](Mistakes.md) · [Reinforcement Learning](Reinforcement-Learning.md) ·
|