Add JSON tags to backend.GenerationParams and embed it directly in
AgentConfig. The 14 duplicated sampling fields and the field-by-field
copy method disappear. Callers use cfg.GenerationParams directly.
Replace the monolithic preamble string with Preamble struct — an
ordered list of named sections joined on demand. Updates are simple
Set() calls; no marker-based string splicing.
Changes:
- Delete prompts/tool_discovery.md (content already in system_prompt.md)
- Reduce prompts/ package to just the embed + ResolveSystemPrompt
- Replace buildPreamble() with direct section construction
- Replace refreshToolListing() 90-line string surgery with one Set()
- Replace SetCWD string.ReplaceAll with env section rebuild
- Remove RenamePreamble (was a no-op; session ID is a template var)
- Export EnvironmentBlock() for session setup
- All callers use PreambleString() instead of raw string field
The agent always talks to a *Conn (never a *Server directly).
Replace the Runner interface with the concrete type throughout:
- Runtime.ToolServer is now *toolsrv.Conn
- BuildRuntime takes *toolsrv.Conn
- session.Config, SessionInfra, InfraConfig use *toolsrv.Conn
- All inline type assertions in agent.go deleted — methods called
directly on *Conn (SetEnv, SetCWD, Close, Detach, ListDetachedRaw,
SignalDetached, GetDetachedOutput, DismissDetached, SetAllowTools,
SetOnToolsChanged, Ping, ToolRegistryRevision)
- Added Conn.ToolRegistryRevision() that returns 0 (remote connections
don't support revision tracking; use push-based OnToolsChanged instead)
- Deleted the Runner interface and its test
- Renamed LoadToolOnRunner → LoadToolOnConn
The entire classifier indirection (interfaces, type assertions, RPC
methods for classification, closure wrappers in BuildRuntime) is replaced
by a simple map[string]ToolInfo on Runtime.
The .meta file is read once into ToolInfo. The loop reads fields directly:
rt.ToolMeta[name].ReadOnly (was: rt.CanParallelize(name))
rt.ToolMeta[name].Tier (was: rt.MemoryTier(name, args))
rt.ToolMeta[name].OutputFormat (was: rt.OutputFormat(name))
rt.ToolMeta[name].ResetsCounter (was: rt.ResetsCounter(name))
Removed: ParallelClassifier, MemoryTierClassifier, OutputFormatClassifier
interfaces, toolClassifier type, and all closure-wrapping in BuildRuntime.
The RPC methods on Server/Conn stay (used by ollie-remote) but the agent
no longer routes through them.
- Kill agentConfig struct. run() takes (*Runtime, TurnCtx, state) directly.
TurnCtx holds only 7 per-turn closures; stable config reads come from Runtime.
- Lazy tool refresh: toolsNeedRefresh() checks ToolRegistryRevision() on the
tool server. Skips ListTools IPC when registry hasn't changed. Remote Conn
falls back to always-refetch (same as before).
- AgentConfig.GenParams() method replaces the 16-line manual field copy in
BuildRuntime.
- Remove CfgBackend/CfgModel from Runtime. Callers read cfg.Backend/cfg.Model
directly since they already have the AgentConfig in scope.
Tools like file_write, file_edit, and shell now reset the agent's step
counter when called successfully. This allows the agent to continue
working without hitting the soft step-budget guardrail as long as it's
making active progress (writing files, running commands) rather than
looping on research/reading.
Changes:
- Add ResetsCounter bool to ToolInfo and MetaFile structs
- Parse resetsCounter from .meta JSON files
- Reset step counter in agent loop when ResetsCounter tool succeeds
- Add resetsCounter: true to file_write, file_edit, shell
- Reduce maxSteps from 50 to 25 (resetsCounter makes this safe)
Ollie Agent
Levi Neely