vault-sync: 2026-10-02 07:10 (jimmy)
This commit is contained in:
+19
@@ -0,0 +1,19 @@
|
||||
---
|
||||
name: DR-20260917-ContextExpansion
|
||||
description: Documenting the move to 131K context and the specific VRAM trade-offs on the 3060.
|
||||
---
|
||||
# Decision Record: Context Window Expansion (131K)
|
||||
|
||||
## Context
|
||||
The gemma12 model was hitting its 65,536 token limit during a session, causing "Context Overflow" and interrupting the API call.
|
||||
|
||||
## Decision
|
||||
Increased the `n_ctx` (context size) for the gemma12 server on port :1923 to **131,072**.
|
||||
|
||||
## Rationale
|
||||
A smoke test on port :1924 confirmed that the RTX 3060 12GB VRAM could support the doubled KV cache, using ~11,754 MiB out of 12,288 MiB. This allows for much longer sessions without immediate overflow.
|
||||
|
||||
## Impact
|
||||
- **Positive**: Supports significantly longer conversations and larger document processing.
|
||||
- **Negative**: Decode speed dropped from ~39.6 t/s (at 65K) to ~26 t/s (at 131K).
|
||||
- **Risk**: High VRAM usage leaves very little headroom for concurrent processes.
|
||||
@@ -0,0 +1,19 @@
|
||||
---
|
||||
name: DR-20260917-ModelSwap
|
||||
description: Documenting the switch to route-flash-0731 as the default model.
|
||||
---
|
||||
# Decision Record: Default Model Swap (route-flash-0731)
|
||||
|
||||
## Context
|
||||
The gemma12 model (local) was frequently hitting its context limit, causing session interruptions. We needed a reliable fallback that could handle larger context windows without crashing the local VRAM.
|
||||
|
||||
## Decision
|
||||
Flipped the Hermes default model to `route-flash-0731` (DeepSeek-v4-flash-0731 via RouteAPI).
|
||||
|
||||
## Rationale
|
||||
DeepSeek-v4-flash-0731 provides high-reasoning capabilities with a large context window, while the local gemma12 server remains available on port :1923 for specific local-only needs. This prevents the system from defaulting to a 65K-limited model on every new session.
|
||||
|
||||
## Impact
|
||||
- **Positive**: More reliable session continuity; higher intelligence for complex tasks.
|
||||
- **Negative**: Increased API costs for RouteAPI usage.
|
||||
- **Risk**: Dependency on external API for daily-driver tasks.
|
||||
@@ -0,0 +1,19 @@
|
||||
---
|
||||
name: DR-20260917-VaultMap
|
||||
description: Documenting the adoption of the Meta/vault-map.md foundation.
|
||||
---
|
||||
# Decision Record: Vault Map Foundation
|
||||
|
||||
## Context
|
||||
The crew required a centralized way to resolve vault roles (e.g., `{{inbox}}`, `{{projects}}`) to ensure consistent behavior across different agents and skills.
|
||||
|
||||
## Decision
|
||||
Implemented a `Meta/vault-map.md` file as the authoritative source for vault path resolution.
|
||||
|
||||
## Rationale
|
||||
This allows the agent platform to adapt to any vault structure without hardcoding paths. By mapping roles to physical folders (e.g., `{{inbox}}` -> `00_Raw/Incoming`), we ensure that the crew remains portable and consistent.
|
||||
|
||||
## Impact
|
||||
- **Positive**: Enables high-agency delegation without pathing errors.
|
||||
- **Negative**: Requires maintenance of the map file when the vault structure changes.
|
||||
- **Risk**: If the map is not updated, agents will resolve to incorrect folders.
|
||||
Reference in New Issue
Block a user