docs(repo-arch): teacher7b v2 experiment and adapter

.repo-arch/adapters/distill/adapter_config.json

@@ -0,0 +1,41 @@
+{
+    "adapter_path": "/Users/pavel/repos/fiale-plus/pi/.repo-arch/adapters/distill",
+    "batch_size": 4,
+    "clear_cache_threshold": 0,
+    "config": null,
+    "data": "/Users/pavel/repos/fiale-plus/pi/.repo-arch/training-data/distill",
+    "fine_tune_type": "lora",
+    "grad_accumulation_steps": 1,
+    "grad_checkpoint": false,
+    "iters": 300,
+    "learning_rate": 1e-05,
+    "lora_parameters": {
+        "rank": 8,
+        "dropout": 0.0,
+        "scale": 20.0
+    },
+    "lr_schedule": null,
+    "mask_prompt": false,
+    "max_seq_length": 2048,
+    "model": "Qwen/Qwen2.5-Coder-1.5B-Instruct",
+    "num_layers": 4,
+    "optimizer": "adam",
+    "optimizer_config": {
+        "adam": {},
+        "adamw": {},
+        "muon": {},
+        "sgd": {},
+        "adafactor": {}
+    },
+    "project_name": null,
+    "report_to": null,
+    "resume_adapter_file": null,
+    "save_every": 100,
+    "seed": 0,
+    "steps_per_eval": 10,
+    "steps_per_report": 10,
+    "test": false,
+    "test_batches": 500,
+    "train": true,
+    "val_batches": 10
+}

.repo-arch/adapters/hybrid-lora/adapter_config.json

@@ -0,0 +1,41 @@
+{
+    "adapter_path": "/Users/pavel/repos/fiale-plus/pi/.repo-arch/adapters/hybrid-lora",
+    "batch_size": 4,
+    "clear_cache_threshold": 0,
+    "config": null,
+    "data": "/Users/pavel/repos/fiale-plus/pi/.repo-arch/training-data/hybrid",
+    "fine_tune_type": "lora",
+    "grad_accumulation_steps": 1,
+    "grad_checkpoint": false,
+    "iters": 300,
+    "learning_rate": 1e-05,
+    "lora_parameters": {
+        "rank": 8,
+        "dropout": 0.0,
+        "scale": 20.0
+    },
+    "lr_schedule": null,
+    "mask_prompt": false,
+    "max_seq_length": 2048,
+    "model": "Qwen/Qwen2.5-Coder-1.5B-Instruct",
+    "num_layers": 4,
+    "optimizer": "adam",
+    "optimizer_config": {
+        "adam": {},
+        "adamw": {},
+        "muon": {},
+        "sgd": {},
+        "adafactor": {}
+    },
+    "project_name": null,
+    "report_to": null,
+    "resume_adapter_file": null,
+    "save_every": 100,
+    "seed": 0,
+    "steps_per_eval": 10,
+    "steps_per_report": 10,
+    "test": false,
+    "test_batches": 500,
+    "train": true,
+    "val_batches": 10
+}

.repo-arch/adapters/repo-arch-b8125c4/adapter_config.json

@@ -0,0 +1,41 @@
+{
+    "adapter_path": "/Users/pavel/repos/fiale-plus/pi/.repo-arch/adapters/repo-arch-b8125c4",
+    "batch_size": 4,
+    "clear_cache_threshold": 0,
+    "config": null,
+    "data": "/Users/pavel/repos/fiale-plus/pi/.repo-arch/training-data",
+    "fine_tune_type": "lora",
+    "grad_accumulation_steps": 1,
+    "grad_checkpoint": false,
+    "iters": 200,
+    "learning_rate": 1e-05,
+    "lora_parameters": {
+        "rank": 8,
+        "dropout": 0.0,
+        "scale": 20.0
+    },
+    "lr_schedule": null,
+    "mask_prompt": false,
+    "max_seq_length": 2048,
+    "model": "Qwen/Qwen2.5-Coder-1.5B-Instruct",
+    "num_layers": 4,
+    "optimizer": "adam",
+    "optimizer_config": {
+        "adam": {},
+        "adamw": {},
+        "muon": {},
+        "sgd": {},
+        "adafactor": {}
+    },
+    "project_name": null,
+    "report_to": null,
+    "resume_adapter_file": null,
+    "save_every": 100,
+    "seed": 0,
+    "steps_per_eval": 10,
+    "steps_per_report": 10,
+    "test": false,
+    "test_batches": 500,
+    "train": true,
+    "val_batches": 10
+}

.repo-arch/adapters/teacher7b-v2/adapter_config.json

@@ -0,0 +1,41 @@
+{
+    "adapter_path": "/Users/pavel/repos/fiale-plus/pi/.repo-arch/adapters/teacher7b-v2",
+    "batch_size": 4,
+    "clear_cache_threshold": 0,
+    "config": null,
+    "data": "/Users/pavel/repos/fiale-plus/pi/.repo-arch/training-data/teacher7b",
+    "fine_tune_type": "lora",
+    "grad_accumulation_steps": 1,
+    "grad_checkpoint": false,
+    "iters": 300,
+    "learning_rate": 1e-05,
+    "lora_parameters": {
+        "rank": 8,
+        "dropout": 0.0,
+        "scale": 20.0
+    },
+    "lr_schedule": null,
+    "mask_prompt": false,
+    "max_seq_length": 2048,
+    "model": "Qwen/Qwen2.5-Coder-1.5B-Instruct",
+    "num_layers": 4,
+    "optimizer": "adam",
+    "optimizer_config": {
+        "adam": {},
+        "adamw": {},
+        "muon": {},
+        "sgd": {},
+        "adafactor": {}
+    },
+    "project_name": null,
+    "report_to": null,
+    "resume_adapter_file": null,
+    "save_every": 100,
+    "seed": 0,
+    "steps_per_eval": 10,
+    "steps_per_report": 10,
+    "test": false,
+    "test_batches": 500,
+    "train": true,
+    "val_batches": 10
+}

.repo-arch/adapters/teacher7b/adapter_config.json

@@ -0,0 +1,41 @@
+{
+    "adapter_path": "/Users/pavel/repos/fiale-plus/pi/.repo-arch/adapters/teacher7b",
+    "batch_size": 4,
+    "clear_cache_threshold": 0,
+    "config": null,
+    "data": "/Users/pavel/repos/fiale-plus/pi/.repo-arch/training-data/teacher7b",
+    "fine_tune_type": "lora",
+    "grad_accumulation_steps": 1,
+    "grad_checkpoint": false,
+    "iters": 300,
+    "learning_rate": 1e-05,
+    "lora_parameters": {
+        "rank": 8,
+        "dropout": 0.0,
+        "scale": 20.0
+    },
+    "lr_schedule": null,
+    "mask_prompt": false,
+    "max_seq_length": 2048,
+    "model": "Qwen/Qwen2.5-Coder-1.5B-Instruct",
+    "num_layers": 4,
+    "optimizer": "adam",
+    "optimizer_config": {
+        "adam": {},
+        "adamw": {},
+        "muon": {},
+        "sgd": {},
+        "adafactor": {}
+    },
+    "project_name": null,
+    "report_to": null,
+    "resume_adapter_file": null,
+    "save_every": 100,
+    "seed": 0,
+    "steps_per_eval": 10,
+    "steps_per_report": 10,
+    "test": false,
+    "test_batches": 500,
+    "train": true,
+    "val_batches": 5
+}

.repo-arch/eval/pi-behavioral-questions.md

@@ -0,0 +1,75 @@
+# pi Behavioral Eval Questions
+
+Fixed question set for comparing base model, retrieval-only, LoRA-only, and retrieval+LoRA.
+
+## Package Boundaries
+
+1. Which packages tend to co-change together, and what does that imply about their coupling?
+2. When modifying `packages/ai/src/models.generated.ts`, which other files should you inspect?
+3. What architectural decisions are visible around the unified LLM provider layer?
+4. What is the risk profile of changing `packages/coding-agent/src/core/agent-session.ts`?
+5. Which package boundaries are the most fragile based on historical breakages?
+
+## Risky Files
+
+6. Which files have been reverted most often, and why?
+7. What are the riskiest files to modify in the coding agent package?
+8. Which files in the TUI package have required repeated fixes?
+9. What files in the AI package have no test coverage despite high change frequency?
+10. Which configuration files change without corresponding test updates?
+
+## Historical Patterns
+
+11. What historical breakages should I know before changing the provider routing?
+12. Where has CLI state/config handling been fragile?
+13. What changed repeatedly in the TUI package and why?
+14. What web UI changes historically required backend or agent changes?
+15. What is the reversion history of the coding agent README?
+16. How has the changelog management pattern evolved across packages?
+
+## Test Coverage
+
+17. Which packages have the most untested high-churn files?
+18. What would you tell a new contributor about test expectations before touching package boundaries?
+19. Which files have been fixed most often without corresponding test additions?
+20. Are there test gaps in the AI provider layer specifically?
+
+## Coding Agent Internals
+
+21. What should a coding agent inspect before modifying provider-tool-calling behavior?
+22. What is the relationship between interactive mode and theme configuration historically?
+23. How fragile is the agent session lifecycle based on commit patterns?
+24. What modes have been added or removed from the coding agent?
+25. What is the history of model resolver changes?
+
+## AI Provider Layer
+
+26. How many LLM providers are supported and how are they registered?
+27. What is the architectural pattern for adding a new provider?
+28. Which provider has the most complex auth requirements historically?
+29. How have streaming options evolved across providers?
+30. What is the relationship between model generation scripts and the type system?
+
+## TUI/UI
+
+31. What are the most frequently changed TUI components?
+32. How has the keybinding system evolved?
+33. What theme-related changes have been made and reverted?
+34. What is the history of the prompt input handling?
+35. How has error display evolved in the TUI?
+
+## Root/Infrastructure
+
+36. What npm scripts and config are essential for the build pipeline?
+37. How has the CI configuration evolved?
+38. What is the history of dependency management (package-lock churn)?
+39. Which root-level config files are most frequently modified and why?
+40. What release process conventions are visible in the commit history?
+
+## Cross-Cutting
+
+41. What is the relationship between CHANGELOG updates and actual code changes across packages?
+42. How do model updates cascade through the system?
+43. What is the most common cause of bugs based on fix commit patterns?
+44. How does the system handle provider API changes over time?
+45. What architectural debt is visible from the commit history?

.repo-arch/eval/questions.jsonl

@@ -0,0 +1,45 @@
+{"id": 1, "question": "Which packages tend to co-change together, and what does that imply about their coupling?"}
+{"id": 2, "question": "When modifying `packages/ai/src/models.generated.ts`, which other files should you inspect?"}
+{"id": 3, "question": "What architectural decisions are visible around the unified LLM provider layer?"}
+{"id": 4, "question": "What is the risk profile of changing `packages/coding-agent/src/core/agent-session.ts`?"}
+{"id": 5, "question": "Which package boundaries are the most fragile based on historical breakages?"}
+{"id": 6, "question": "Which files have been reverted most often, and why?"}
+{"id": 7, "question": "What are the riskiest files to modify in the coding agent package?"}
+{"id": 8, "question": "Which files in the TUI package have required repeated fixes?"}
+{"id": 9, "question": "What files in the AI package have no test coverage despite high change frequency?"}
+{"id": 10, "question": "Which configuration files change without corresponding test updates?"}
+{"id": 11, "question": "What historical breakages should I know before changing the provider routing?"}
+{"id": 12, "question": "Where has CLI state/config handling been fragile?"}
+{"id": 13, "question": "What changed repeatedly in the TUI package and why?"}
+{"id": 14, "question": "What web UI changes historically required backend or agent changes?"}
+{"id": 15, "question": "What is the reversion history of the coding agent README?"}
+{"id": 16, "question": "How has the changelog management pattern evolved across packages?"}
+{"id": 17, "question": "Which packages have the most untested high-churn files?"}
+{"id": 18, "question": "What would you tell a new contributor about test expectations before touching package boundaries?"}
+{"id": 19, "question": "Which files have been fixed most often without corresponding test additions?"}
+{"id": 20, "question": "Are there test gaps in the AI provider layer specifically?"}
+{"id": 21, "question": "What should a coding agent inspect before modifying provider-tool-calling behavior?"}
+{"id": 22, "question": "What is the relationship between interactive mode and theme configuration historically?"}
+{"id": 23, "question": "How fragile is the agent session lifecycle based on commit patterns?"}
+{"id": 24, "question": "What modes have been added or removed from the coding agent?"}
+{"id": 25, "question": "What is the history of model resolver changes?"}
+{"id": 26, "question": "How many LLM providers are supported and how are they registered?"}
+{"id": 27, "question": "What is the architectural pattern for adding a new provider?"}
+{"id": 28, "question": "Which provider has the most complex auth requirements historically?"}
+{"id": 29, "question": "How have streaming options evolved across providers?"}
+{"id": 30, "question": "What is the relationship between model generation scripts and the type system?"}
+{"id": 31, "question": "What are the most frequently changed TUI components?"}
+{"id": 32, "question": "How has the keybinding system evolved?"}
+{"id": 33, "question": "What theme-related changes have been made and reverted?"}
+{"id": 34, "question": "What is the history of the prompt input handling?"}
+{"id": 35, "question": "How has error display evolved in the TUI?"}
+{"id": 36, "question": "What npm scripts and config are essential for the build pipeline?"}
+{"id": 37, "question": "How has the CI configuration evolved?"}
+{"id": 38, "question": "What is the history of dependency management (package-lock churn)?"}
+{"id": 39, "question": "Which root-level config files are most frequently modified and why?"}
+{"id": 40, "question": "What release process conventions are visible in the commit history?"}
+{"id": 41, "question": "What is the relationship between CHANGELOG updates and actual code changes across packages?"}
+{"id": 42, "question": "How do model updates cascade through the system?"}
+{"id": 43, "question": "What is the most common cause of bugs based on fix commit patterns?"}
+{"id": 44, "question": "How does the system handle provider API changes over time?"}
+{"id": 45, "question": "What architectural debt is visible from the commit history?"}