Grayscale/nanobot
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(agent): make model presets session-scoped (#4866)

Browse Source
This commit is contained in:
chengyongru
2026-07-23 00:38:49 +08:00
committed by GitHub
parent 66690fdb0c
commit c22efb5f7a
41 changed files with 1140 additions and 155 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/my-tool.md
+16 -15
View File
@@ -27,7 +27,8 @@ To allow the agent to set its configuration (e.g. switch models, adjust paramete
Legacy `tools.myEnabled` / `tools.mySet` keys are auto-migrated on load, and rewritten in-place the next time `nanobot onboard` refreshes the config.
All modifications are held in memory only — restart restores defaults.
Most modifications are held in memory only. `model_preset` is the exception: it is
stored in the current session so the selection survives a restart.
---
@@ -77,20 +78,18 @@ my(action="check", key="web_config.enable")
## set — Runtime tuning
Changes take effect immediately, no restart required.
Changes do not require a restart. `model_preset` is saved for the current session and
applies to its next turn; other writable runtime tuning takes effect immediately.
Direct `model` and `context_window_tokens` writes are rejected during an active session
because those setters change the shared instance default. Configure a named preset for
model or context-window changes instead.
```text
my(action="set", key="max_iterations", value=80)
# → Bump iteration limit from 40 to 80
my(action="set", key="model_preset", value="fast")
# → Switch to a configured model preset
my(action="set", key="model", value="fast-model")
# → Switch to a raw model and clear the active preset
my(action="set", key="context_window_tokens", value=262144)
# → Expand context window for long documents
# → Use a configured model preset for this session's next turn
```
You can also store custom state in your scratchpad:
@@ -109,9 +108,9 @@ These parameters have type and range validation — invalid values are rejected:
| Parameter | Type | Range | Purpose |
|-----------|------|-------|---------|
| `max_iterations` | int | 1–100 | Max tool calls per conversation turn |
| `context_window_tokens` | int | 4,096–1,000,000 | Context window size |
| `model` | str | non-empty | LLM model to use |
| `model_preset` | str | configured preset name | Named preset to use |
| `context_window_tokens` | int | 4,096–1,000,000 | Instance default; during a session, select through a preset |
| `model` | str | non-empty | Instance default; during a session, select through a preset |
| `model_preset` | str | configured preset name | Current session's preset for its next turn |
Other parameters (e.g. `workspace`, `provider_retry_mode`, `max_tool_result_chars`) can be set freely, as long as the value is JSON-safe.
@@ -122,8 +121,8 @@ Other parameters (e.g. `workspace`, `provider_retry_mode`, `max_tool_result_char
### "This task is complex, I need more room"
```text
Agent: This codebase is large, let me expand my context window to handle it.
→ my(action="set", key="context_window_tokens", value=262144)
Agent: This codebase is large, let me switch this session to the configured deep preset.
→ my(action="set", key="model_preset", value="deep")
```
### "Simple question, don't waste compute"
@@ -180,7 +179,9 @@ Agent: The code review is progressing well. The test task hasn't started yet.
## Safety Mechanisms
Core design principle: **All modifications live in memory only. Restart restores defaults.** The agent cannot cause persistent damage.
Core design principle: **The tool does not rewrite `config.json`.** Instance-wide
changes live in memory only, while `model_preset` persists only as the current
session's selector.
### Off-limits (BLOCKED)