Class: Phronomy::Agent::Base

Inherits:
Object
  • Object
show all
Includes:
Runnable
Defined in:
lib/phronomy/agent/base.rb

Overview

Base class for all Phronomy agents.

Subclass this to create a conversational agent powered by an LLM. DSL class methods configure the model, instructions, tools, memory, and retry behaviour. Instance methods handle invocation.

Examples:

Minimal agent

class GreetingAgent < Phronomy::Agent::Base
  model "gpt-4o-mini"
  instructions "You are a friendly greeter."
end
result = GreetingAgent.new.invoke("Hello!")
puts result[:output]

Agent with tools

class ResearchAgent < Phronomy::Agent::Base
  model "gpt-4o"
  instructions "You are a research assistant."
  tools WebSearchTool, CalculatorTool
  max_iterations 15
end

Direct Known Subclasses

Orchestrator, ReactAgent

Class Attribute Summary collapse

Instance Attribute Summary collapse

Class Method Summary collapse

Instance Method Summary collapse

Methods included from Runnable

#batch, #trace

Class Attribute Details

._retry_policy ⇒ Hash? (readonly)

Returns the configured retry policy, or nil when none is set.

Returns:

  • (Hash, nil) 


185
186
187
 
# File 'lib/phronomy/agent/base.rb', line 185

def _retry_policy
  @_retry_policy
end

._sleep_proc ⇒ #call

Injectable sleep callable for testing (shared with Tool::Base pattern).

Returns:

  • (#call) 


189
190
191
 
# File 'lib/phronomy/agent/base.rb', line 189

def _sleep_proc
  @_sleep_proc || method(:sleep)
end

Instance Attribute Details

#before_completion ⇒ #call?

Instance-level before_completion hook. When set, takes precedence over the class-level hook for this specific agent instance only.

Returns:

  • (#call, nil) 


384
385
386
 
# File 'lib/phronomy/agent/base.rb', line 384

def before_completion
  @before_completion
end

Class Method Details

._before_completion ⇒ #call?

Returns:

  • (#call, nil) 


376
377
378
 
# File 'lib/phronomy/agent/base.rb', line 376

def _before_completion
  @before_completion
end

._on_compact_callback ⇒ Proc?

Returns:

  • (Proc, nil) 


282
283
284
 
# File 'lib/phronomy/agent/base.rb', line 282

def _on_compact_callback
  @on_compact_callback
end

._on_compaction_trigger_callback ⇒ Proc?

Returns:

  • (Proc, nil) 


260
261
262
 
# File 'lib/phronomy/agent/base.rb', line 260

def _on_compaction_trigger_callback
  @on_compaction_trigger_callback
end

._on_trim_callback ⇒ Proc?

Returns:

  • (Proc, nil) 


237
238
239
 
# File 'lib/phronomy/agent/base.rb', line 237

def _on_trim_callback
  @on_trim_callback
end

.before_completion(callable = nil) ⇒ #call?

Sets or reads the class-level before_completion hook. The hook is called before every LLM request for instances of this class. Receives a Phronomy::Agent::BeforeCompletionContext; must return a Hash of params to merge into the LLM call, or nil to pass through unchanged.

Examples:

class MyAgent < Phronomy::Agent::Base
  before_completion ->(ctx) { { temperature: 0.2 } }
end

Parameters:

  • callable (#call, nil) (defaults to: nil) —

    lambda/proc to register, or nil to clear

Returns:

  • (#call, nil) 


367
368
369
370
371
372
373
 
# File 'lib/phronomy/agent/base.rb', line 367

def before_completion(callable = nil)
  if callable.nil? && !block_given?
    @before_completion
  else
    @before_completion = callable
  end
end

.cache_instructions(enabled = nil) ⇒ Object

When enabled, attaches Anthropic prompt-cache markers to the system message so that the fixed instructions are served from cache on subsequent turns, reducing input-token costs.

Only has an effect when the agent also declares provider :anthropic. The cache_control field is provider-specific (the format differs between Anthropic direct, Bedrock, etc.), so the agent must explicitly declare its provider via the DSL rather than having it inferred from the model name.

Examples:

class MyAgent < Phronomy::Agent::Base
  provider :anthropic
  cache_instructions true
end


301
302
303
304
305
306
307
 
# File 'lib/phronomy/agent/base.rb', line 301

def cache_instructions(enabled = nil)
  if enabled.nil?
    @cache_instructions
  else
    @cache_instructions = enabled
  end
end

.context_overhead(val = nil) ⇒ Object

Tokens reserved for the system prompt + tool definitions overhead. Subtract this from the context window before computing the memory budget.

Examples:

class MyAgent < Phronomy::Agent::Base
  context_overhead 500
end


348
349
350
351
352
353
354
 
# File 'lib/phronomy/agent/base.rb', line 348

def context_overhead(val = nil)
  if val.nil?
    @context_overhead || 0
  else
    @context_overhead = val.to_i
  end
end

.context_window(val = nil) ⇒ Object

Overrides the context window size used for token budget calculations. When set, this value takes precedence over the RubyLLM model registry, which is useful for locally-hosted models (e.g. LM Studio) where the actually-loaded context length may differ from the catalogue value.

Examples:

class MyAgent < Phronomy::Agent::Base
  context_window 4096
end


333
334
335
336
337
338
339
 
# File 'lib/phronomy/agent/base.rb', line 333

def context_window(val = nil)
  if val.nil?
    @context_window
  else
    @context_window = val.to_i
  end
end

.instructions(text = nil) { ... } ⇒ String, ...

Sets or reads the system instructions for this agent. Accepts a String, a PromptTemplate, or a block (Proc). When used as a reader (no argument, no block), returns the stored value.

Examples:

String instructions

class MyAgent < Phronomy::Agent::Base
  instructions "You are a helpful assistant."
end

Block instructions

class MyAgent < Phronomy::Agent::Base
  instructions { |input| "Answer in #{input[:lang]}." }
end

Parameters:

Yields:

  • optionally provide instructions as a block

Returns:



65
66
67
68
69
70
71
72
 
# File 'lib/phronomy/agent/base.rb', line 65

def instructions(text = nil, &block)
  if text || block_given?
    @instructions = text || block
  else
    return @instructions if instance_variable_defined?(:@instructions)
    superclass.respond_to?(:instructions) ? superclass.instructions : nil
  end
end

.max_iterations(val = nil) ⇒ Integer

Sets or reads the maximum number of LLM call cycles for ReAct agents. Each tool call and follow-up counts as one iteration. Defaults to 10.

Examples:

class MyAgent < Phronomy::Agent::Base
  max_iterations 5
end

Parameters:

  • val (Integer, nil) (defaults to: nil)

Returns:

  • (Integer) 


160
161
162
163
164
165
166
 
# File 'lib/phronomy/agent/base.rb', line 160

def max_iterations(val = nil)
  if val
    @max_iterations = val
  else
    @max_iterations || 10
  end
end

.max_output_tokens(val = nil) ⇒ Object

Tokens to reserve for the model's output. When nil, the model's max_output_tokens from the registry is used.

Examples:

class MyAgent < Phronomy::Agent::Base
  max_output_tokens 4096
end


316
317
318
319
320
321
322
 
# File 'lib/phronomy/agent/base.rb', line 316

def max_output_tokens(val = nil)
  if val.nil?
    @max_output_tokens
  else
    @max_output_tokens = val.to_i
  end
end

.model(name = nil) ⇒ String?

Sets or reads the LLM model identifier for this agent. When called without an argument, returns the stored model or the global default from Phronomy.configuration.

Examples:

class MyAgent < Phronomy::Agent::Base
  model "gpt-4o"
end

Parameters:

  • name (String, nil) (defaults to: nil) —

    model identifier (e.g. "gpt-4o", "claude-3-5-sonnet")

Returns:

  • (String, nil) —

    the model name when used as a reader



42
43
44
45
46
47
48
 
# File 'lib/phronomy/agent/base.rb', line 42

def model(name = nil)
  if name
    @model = name
  else
    @model || Phronomy.configuration.default_model
  end
end

.on_compact {|ctx| ... } ⇒ Object

Registers a callback that performs the actual compaction when the +on_compaction_trigger+ callback fires. The block receives a Context::CompactionContext and should call +ctx.compact+ to specify which messages to summarise.

Examples:

Replace the first 4 messages with a short summary

on_compact do |ctx|
  ctx.compact(0..3) do |elements|
    texts = elements.map { |e| e[:message].content }.join(" | ")
    "Earlier conversation summary: #{texts}"
  end
end

Yields:

  • (ctx) —

    Phronomy::Context::CompactionContext



277
278
279
 
# File 'lib/phronomy/agent/base.rb', line 277

def on_compact(&block)
  @on_compact_callback = block
end

.on_compaction_trigger {|ctx| ... } ⇒ Boolean

Registers a callback that decides whether compaction should run. Evaluated before every LLM call (after on_trim). If the block returns truthy AND an +on_compact+ callback is also registered, the compact pipeline is executed.

The block receives a read-only Context::TriggerContext.

Examples:

Trigger when messages exceed 70% of token budget

on_compaction_trigger do |ctx|
  limit = ctx.budget&.available(used: 0) || Float::INFINITY
  ctx.total_tokens > limit * 0.7
end

Yields:

  • (ctx) —

    Phronomy::Context::TriggerContext

Returns:

  • (Boolean) —

    truthy → run on_compact; falsy → skip



255
256
257
 
# File 'lib/phronomy/agent/base.rb', line 255

def on_compaction_trigger(&block)
  @on_compaction_trigger_callback = block
end

.on_trim {|ctx| ... } ⇒ Object

Registers a callback that is invoked before every LLM call so the application can remove stale or irrelevant messages from the conversation history.

The block receives a Context::TrimContext and may call +ctx.remove(seqs)+ to drop messages by seq number. Changes affect only the current invocation; the underlying memory store is unchanged.

Examples:

Drop the oldest message when over 80% of budget is used

on_trim do |ctx|
  limit = ctx.budget&.available(used: 0) || Float::INFINITY
  ctx.remove(ctx.message_elements.first[:seq]) if ctx.total_tokens > limit * 0.8
end

Yields:

  • (ctx) —

    Phronomy::Context::TrimContext



232
233
234
 
# File 'lib/phronomy/agent/base.rb', line 232

def on_trim(&block)
  @on_trim_callback = block
end

.provider(name = nil) ⇒ Symbol?

Sets or reads the LLM provider for this agent. Required when using a model not registered in RubyLLM's model registry (e.g. locally-hosted models via LM Studio or Ollama).

Examples:

class MyAgent < Phronomy::Agent::Base
  model "openai/gpt-oss-20b"
  provider :openai
end

Parameters:

  • name (Symbol, nil) (defaults to: nil) —

    e.g. +:openai+, +:anthropic+, +:ollama+

Returns:

  • (Symbol, nil) 


125
126
127
128
129
130
131
132
 
# File 'lib/phronomy/agent/base.rb', line 125

def provider(name = nil)
  if name
    @provider = name
  else
    return @provider if instance_variable_defined?(:@provider)
    superclass.respond_to?(:provider) ? superclass.provider : nil
  end
end

.retry_policy(times: 0, wait: 0, base: 1.0) ⇒ Object

Configures a retry policy that wraps the full #invoke call. GuardrailError is never retried regardless of this setting.

Examples:

class MyAgent < Phronomy::Agent::Base
  retry_policy times: 2, wait: :exponential, base: 1.0
end

Parameters:

  • times (Integer) (defaults to: 0) —

    maximum retry attempts (default: 0)

  • wait (Symbol, Numeric) (defaults to: 0) —

    :exponential, :linear, or a fixed Float

  • base (Float) (defaults to: 1.0) —

    base wait time in seconds (default: 1.0)



179
180
181
 
# File 'lib/phronomy/agent/base.rb', line 179

def retry_policy(times: 0, wait: 0, base: 1.0)
  @_retry_policy = {times: times, wait: wait, base: base}
end

.static_knowledge(*sources) ⇒ Object

Registers one or more static knowledge sources on the agent class. Static sources are fetched once per agent instance and their content is cached in ContextVersionCache keyed by a fingerprint of the instruction text + source content. The cache is invalidated automatically when the fingerprint changes (e.g. because a source was updated).

Examples:

class PolicyAgent < Phronomy::Agent::Base
  static_knowledge Phronomy::KnowledgeSource::StaticKnowledge.new(POLICY_TEXT)
end

Parameters:



208
209
210
 
# File 'lib/phronomy/agent/base.rb', line 208

def static_knowledge(*sources)
  @static_knowledge_sources = sources.flatten
end

.static_knowledge_sources ⇒ Array<Phronomy::KnowledgeSource::Base>

Returns the registered static knowledge sources.

Returns:



214
215
216
 
# File 'lib/phronomy/agent/base.rb', line 214

def static_knowledge_sources
  @static_knowledge_sources || []
end

.temperature(val = nil) ⇒ Float?

Sets or reads the sampling temperature sent to the LLM. When nil, the provider's default is used.

Examples:

class MyAgent < Phronomy::Agent::Base
  temperature 0.2
end

Parameters:

  • val (Float, nil) (defaults to: nil) —

    temperature (0.0 to 2.0 depending on provider)

Returns:

  • (Float, nil) 


143
144
145
146
147
148
149
 
# File 'lib/phronomy/agent/base.rb', line 143

def temperature(val = nil)
  if val
    @temperature = val
  else
    @temperature
  end
end

.tool_aliases ⇒ Hash{Class => String}

Returns the alias map registered via the hash form of .tools.

Returns:

  • (Hash{Class => String}) 


110
111
112
 
# File 'lib/phronomy/agent/base.rb', line 110

def tool_aliases
  @tool_aliases ||= {}
end

.tools(*args) ⇒ Object

Registers tool classes for this agent.

Accepts either a splat of classes (backward-compatible) or a Hash mapping each class to an explicit alias name (String) or nil (use tool's own name). The alias form is useful when two tools share the same auto-generated name (e.g. two SearchTool classes from different modules).

Examples:

Splat form (no alias)

tools WeatherTool, TimeTool

Hash form (with optional per-tool alias)

tools(
  Weather::SearchTool => "weather_search",
  Places::SearchTool  => "places_search",
  CurrentTimeTool     => nil
)


90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
 
# File 'lib/phronomy/agent/base.rb', line 90

def tools(*args)
  if args.empty?
    if instance_variable_defined?(:@tools)
      return @tools
    end
    return superclass.respond_to?(:tools) ? superclass.tools : []
  end

  if args.length == 1 && args.first.is_a?(Hash)
    hash = args.first
    @tools = hash.keys
    @tool_aliases = hash.transform_values { |v| v&.to_s }.reject { |_, v| v.nil? }
  else
    @tools = args
    @tool_aliases = {}
  end
end

Instance Method Details

#_add_handoff_tool(tool_class) ⇒ self

Registers an anonymous handoff tool class on this agent instance. Called by Runner during construction when routes are configured.

Parameters:

Returns:

  • (self) 


390
391
392
393
394
 
# File 'lib/phronomy/agent/base.rb', line 390

def _add_handoff_tool(tool_class)
  @_handoff_tools ||= []
  @_handoff_tools << tool_class
  self
end

#_handoff_tools ⇒ Array<Class>

Returns handoff tool classes registered on this instance by Runner.

Returns:

  • (Array<Class>) 


398
399
400
 
# File 'lib/phronomy/agent/base.rb', line 398

def _handoff_tools
  @_handoff_tools || []
end

#add_input_guardrail(guardrail) ⇒ Object

Attach a guardrail that validates input before every #invoke call.

Parameters:



525
526
527
528
529
 
# File 'lib/phronomy/agent/base.rb', line 525

def add_input_guardrail(guardrail)
  @input_guardrails ||= []
  @input_guardrails << guardrail
  self
end

#add_output_guardrail(guardrail) ⇒ Object

Attach a guardrail that validates output before it is returned.

Parameters:



533
534
535
536
537
 
# File 'lib/phronomy/agent/base.rb', line 533

def add_output_guardrail(guardrail)
  @output_guardrails ||= []
  @output_guardrails << guardrail
  self
end

#context_version_cache ⇒ Object

This method is part of a private API. You should avoid using this method if possible, as it may be removed or be changed in the future.

Returns the Context::ContextVersionCache for the current thread.



541
542
543
 
# File 'lib/phronomy/agent/base.rb', line 541

def context_version_cache
  (Thread.current[:phronomy_context_version_caches] ||= {})[object_id]
end

#invoke(input, config: {}) ⇒ Hash

Invokes the agent with the given input and returns a result Hash. Applies the retry policy configured via retry_policy when transient errors occur. GuardrailError is never retried.

Examples:

Normal invocation

result = MyAgent.new.invoke("What is Ruby?")
puts result[:output]

Suspend / resume flow

result = agent.invoke("Perform task X")
if result[:suspended]
  result = agent.resume(result[:checkpoint], approved: true)
end
puts result[:output]

Parameters:

  • input (String, Hash) —

    the user message; a Hash may supply +:message+, +:query+, or +:user+ as the text key, plus any template variables consumed by the configured instructions template.

  • config (Hash) (defaults to: {}) —

    runtime options: +:messages+ (ArrayRubyLLM::Message) — conversation history from a previous invocation +:thread_id+ (+String+) — conversation thread identifier +:user_id+ (+String+, optional) — caller identity forwarded to the tracer +:session_id+ (+String+, optional) — session identity forwarded to the tracer

Returns:

  • (Hash) —

    +{ output: String, messages: Array, usage: Phronomy::TokenUsage }+, or +{ output: nil, suspended: true, checkpoint: Phronomy::Agent::Checkpoint, messages: Array }+ when the invocation was suspended awaiting tool approval.

Raises:



427
428
429
 
# File 'lib/phronomy/agent/base.rb', line 427

def invoke(input, config: {})
  _invoke_impl(input, config: config)
end

#on_approval_required(&block) ⇒ self

Registers a callback that is invoked before executing any tool that has +requires_approval true+ set. The block receives the tool name (String) and the arguments Hash, and must return a truthy value to allow execution. Returning a falsy value causes the tool to return a denial message instead of executing.

When no handler is registered and a tool with +requires_approval+ is called, #invoke returns a suspended result hash containing a Checkpoint. Call #resume to continue execution after obtaining an approval decision from the user or an external system.

Examples:

Synchronous handler

agent = MyAgent.new
agent.on_approval_required { |tool_name, args| prompt_user(tool_name, args) }

Returns:

  • (self) 


518
519
520
521
 
# File 'lib/phronomy/agent/base.rb', line 518

def on_approval_required(&block)
  @approval_handler = block
  self
end

#resume(checkpoint, approved:, config: {}) ⇒ Hash

Resumes a previously suspended invocation from a Checkpoint.

This method reconstructs the conversation state captured at suspension time, injects the tool result (executed or denied), and continues the LLM loop until it produces a final answer.

Parameters:

  • checkpoint (Phronomy::Agent::Checkpoint) —

    the checkpoint returned by the suspended #invoke call

  • approved (Boolean) —

    +true+ to execute the pending tool; +false+ to inject a denial message and let the LLM handle it gracefully

  • config (Hash) (defaults to: {}) —

    same runtime options as #invoke

Returns:

  • (Hash) —

    +{ output: String, suspended: false, messages: Array, usage: Phronomy::TokenUsage }+

Raises:



467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
 
# File 'lib/phronomy/agent/base.rb', line 467

def resume(checkpoint, approved:, config: {})
  checkpoint.thread_id

  # Build a fresh chat with all tools registered.
  chat = build_chat

  # Restore the full conversation (system + history + user + assistant).
  checkpoint.messages.each { |msg| chat.messages << msg }

  # Determine the tool result: execute it or inject a denial string.
  tool_result =
    if approved
      tool_instance = chat.tools[checkpoint.pending_tool_name.to_sym]
      tool_instance ? tool_instance.call(checkpoint.pending_tool_args) : "Tool not found."
    else
      "Tool execution denied."
    end

  # Inject the tool result so the LLM can continue.
  chat.add_message(
    role: :tool,
    content: tool_result.to_s,
    tool_call_id: checkpoint.pending_tool_call_id
  )

  # Continue the React loop.
  response = chat.complete

  output = response.content
  usage = Phronomy::TokenUsage.from_tokens(response.tokens)

  run_output_guardrails!(output)

  {output: output, suspended: false, messages: chat.messages, usage: usage}
end

#stream(input, config: {}) {|Phronomy::Agent::StreamEvent| ... } ⇒ Hash

Streaming version of #invoke. Yields StreamEvent objects as they are produced by the underlying LLM.

Events emitted (in order): :token — each content delta from the LLM :tool_call — when the LLM requests a tool (ReactAgent subclasses only) :tool_result — after a tool completes (ReactAgent subclasses only) :done — final event carrying output, messages, and usage :error — if an unrecoverable error occurs

Parameters:

  • input (String, Hash) —

    same as #invoke

  • config (Hash) (defaults to: {}) —

    same as #invoke

Yields:

Returns:

  • (Hash) —

    { output:, messages:, usage: } — same as #invoke



445
446
447
448
449
450
451
452
 
# File 'lib/phronomy/agent/base.rb', line 445

def stream(input, config: {}, &block)
  return invoke(input, config: config) unless block

  _stream_impl(input, config: config, &block)
rescue => e
  block&.call(StreamEvent.new(type: :error, payload: {error: e}))
  raise
end