Class: Insika::ChatBuilder

Inherits:
Object
  • Object
show all
Defined in:
lib/insika/chat_builder.rb

Overview

Assembles the turn's chat (pipeline stages 5-7): instructions, eager/deferred tool partition, system tools (tool_search/load_skill/remember), history and callbacks. Extracted from the Executor so it can coordinate the pipeline without also carrying the RubyLLM glue.

The Executor creates the chat (stage 6, the RubyLLM boundary) and passes it here to be configured; event numbering (monotonic seq per task) stays in the Executor, injected as the emit callable.

Instance Method Summary collapse

Constructor Details

#initialize(tool_registry:, skill_catalog:, checkpoint_store:, event_stream:, hooks:, tool_catalog: nil, memory_store: nil, subagent_runner: nil, tool_trace_store: nil) ⇒ ChatBuilder

Returns a new instance of ChatBuilder.



15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
 
# File 'lib/insika/chat_builder.rb', line 15

def initialize(tool_registry:, skill_catalog:, checkpoint_store:, event_stream:,
               hooks:, tool_catalog: nil, memory_store: nil, subagent_runner: nil,
               tool_trace_store: nil)
  @tool_registry = tool_registry
  @skill_catalog = skill_catalog
  @checkpoint_store = checkpoint_store
  @event_stream = event_stream
  @hooks = hooks
  @tool_catalog = tool_catalog
  @memory_store = memory_store
  # only to trace load_skill, which is not enveloped — nil = no trace (parity).
  @tool_trace_store = tool_trace_store
  # the object exposing #run_subagent (the Executor). nil = the
  # spawn_subagent system tool is never wired (parity for a builder used
  # without delegation, e.g. some unit stubs).
  @subagent_runner = subagent_runner
end

Instance Method Details

#anthropic_provider?(chat) ⇒ Boolean

The RESOLVED provider (chat.model.provider is the slug string, e.g. "anthropic"), authoritative even when the agent left provider nil and RubyLLM inferred it from the model id. Any surface without a model (fakes, a provider that raises) -> false: caching silently stays off.

Returns:

  • (Boolean) 


200
201
202
203
204
 
# File 'lib/insika/chat_builder.rb', line 200

def anthropic_provider?(chat)
  chat.respond_to?(:model) && chat.model && chat.model.provider.to_s == "anthropic"
rescue StandardError
  false
end

#apply_instructions(chat, system, state) ⇒ Object

R3: opt-in Anthropic prompt caching. When the agent enables prompt_caching AND the resolved provider is Anthropic, wrap the system in the provider's native Content helper with cache: true — ONE breakpoint at the END of the system block. By Anthropic's prefix order (tools -> system -> messages), a breakpoint on the last system block caches tools + system together and is immune to history eviction (messages come after it). RubyLLM::Content::Raw is Anthropic-specific: build_system_content emits its blocks verbatim, so the cache_control rides along.

Any other case (caching off, or a non-Anthropic provider) uses the plain string — OpenAI caches its prefix on its own; the Raw shape would confuse non-Anthropic providers. The gem only supports MANUAL caching, and only for Anthropic.

PRE-AUDIT (why this is opt-in): the system must be BYTE-STABLE between turns for a read hit. A context provider that injects volatile content into :system (timestamps, per-turn data) makes every turn a paid cache WRITE with no hit — worse than off. Enable only for stable-system agents.



188
189
190
191
192
193
194
 
# File 'lib/insika/chat_builder.rb', line 188

def apply_instructions(chat, system, state)
  if state.profile.prompt_caching && anthropic_provider?(chat)
    chat.with_instructions(RubyLLM::Providers::Anthropic::Content.new(system, cache: true))
  else
    chat.with_instructions(system)
  end
end

#assemble(chat, state, emit:) ⇒ Object

Configures an already-created chat with the context (stage 2) and the Resolution (stage 3), seeds the history and wires the callbacks. emit is the Executor's emitter (seq+task correlation), called as emit.call(type, data). The system tools (Tools::ToolSearch/LoadSkill/Remember) were already lazy-loaded by Executor#create_chat before reaching here.



38
39
40
41
42
43
 
# File 'lib/insika/chat_builder.rb', line 38

def assemble(chat, state, emit:)
  configure_chat(chat, state)
  seed_history(chat, Array(state.context.history))
  wire_callbacks(chat, state, emit)
  chat
end

#configure_chat(chat, state) ⇒ Object

Assembles the chat with the context (stage 2) and the Resolution's tools (stage 3).



46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
 
# File 'lib/insika/chat_builder.rb', line 46

def configure_chat(chat, state)
  system = state.context.system.to_s
  apply_instructions(chat, system, state) unless system.empty?

  tools = Array(state.allowed_tools).dup

  # Tool Search: the partition only runs with @tool_catalog present (parity
  # when nil — the `&&` short-circuits before reading `.name`).
  # `deferred_allowed` = allowed_tools ∩ tools_deferred. The <available_tools>
  # catalog comes from Context::Providers::ToolSearch (stage 2); here we only
  # decide chat.tools.
  deferred_allowed = if @tool_catalog
                       Array(state.profile.tools_deferred).map(&:to_s) &
                         tools.map { |t| t.name.to_s }
                     else
                       []
                     end

  unless deferred_allowed.empty?
    tools.reject! { |t| deferred_allowed.include?(t.name.to_s) }
    # a system tool (outside the allowlist), like load_skill — never enveloped.
    tools << Tools::ToolSearch.new(@tool_catalog, deferred_allowed, chat,
                                   tool_registry: @tool_registry,
                                   checkpoint_store: @checkpoint_store,
                                   event_stream: @event_stream, state: state)
  end

  # load_skill is a system default (outside the allowlist), otherwise
  # progressive disclosure breaks. allowed_skills comes from the Resolution
  # (policy), minus the EAGER ones: their bodies are already in the prompt, so a
  # call could only pay for a duplicate. Nothing lazy left -> the tool is not
  # wired at all. Keeping it for the discretionary skills is deliberate: that
  # call is the only record of which skill the model actually reached for.
  skill_names = lazy_skill_names(state)
  unless skill_names.empty?
    tools << Tools::LoadSkill.new(@skill_catalog, skill_names,
                                  trace_recorder: @tool_trace_store, state: state,
                                  agent: state.profile.id)
  end

  # remember is the memory-write system tool — wired only with
  # @memory_store present AND profile.memory (a double gate). Never enveloped.
  if @memory_store && state.profile.memory
    tools << Tools::Remember.new(@memory_store, state.tenant,
                                 event_stream: @event_stream, state: state)
  end

  # signal_stuck is the "I cannot proceed" system tool (WS5) — wired only
  # when the agent opted in (`profile.stuck_signal`), never enveloped. It is a
  # deterministic signal; the consumer decides what "stuck" means. Defensive
  # read: a minimal profile double without the reader means off (nil = parity).
  if state.profile.respond_to?(:stuck_signal) && state.profile.stuck_signal
    tools << Tools::StuckSignal.new(state: state)
  end

  # spawn_subagent is the delegation system tool — wired only with a
  # runner present AND profile.subagents non-empty (a double gate, like
  # remember). Never enveloped: in the synchronous mode the child lives in the
  # parent's envelope. The runtime gate on WHICH agent is spawnable is the
  # parent's subagents allowlist, enforced in Executor#run_subagent.
  if @subagent_runner && !Array(state.profile.subagents).empty?
    tools << Tools::Subagent.new(runner: @subagent_runner, state: state)
    # and its parallel sibling: fan-out N children at once.
    tools << Tools::Subagents.new(runner: @subagent_runner, state: state)
  end

  unless tools.empty?
    # the ONLY place the gem is told to run tool calls in parallel.
    # `:fibers` is not a preference but the only admissible mode — `:threads`
    # breaks ToolEnvelope's `Async::Task.current.with_timeout`, the SQLite
    # store's fiber semaphore, and the turn's own durability (mailbox,
    # approvals, cancellation are all expressed in fiber terms). The number
    # the operator configured is OUR cap (ToolAssembly#install_tool_gate);
    # the gem has none.
    if tool_concurrency_for(state)
      chat.with_tools(*tools, concurrency: :fibers)
    else
      chat.with_tools(*tools)
    end
  end

  chat
end

#lazy_skill_names(state) ⇒ Object

The skills the model still has to ask for: the Resolution's set minus the eager ones. Intersected by NAME against the catalog's own verdict (SkillCatalog#eager_for) so the tool and the level-1 list can never disagree about who is eager. A catalog without the reader (a unit stub) falls back to the whole set — parity.



134
135
136
137
138
139
140
 
# File 'lib/insika/chat_builder.rb', line 134

def lazy_skill_names(state)
  names = Array(state.allowed_skills).map { |s| s.respond_to?(:name) ? s.name : s.to_s }
  return names unless @skill_catalog.respond_to?(:eager_for)

  eager = @skill_catalog.eager_for(state.profile).map(&:name)
  names - eager
end

#rehydrate_tool_calls(list) ⇒ Object

[id,name,arguments] (string|symbol keys) -> => RubyLLM::ToolCall, the shape RubyLLM seeds an assistant message with. Called only when the gem is already loaded (create_chat required it before assemble).



228
229
230
231
232
233
234
235
236
 
# File 'lib/insika/chat_builder.rb', line 228

def rehydrate_tool_calls(list)
  Array(list).each_with_object({}) do |tc, acc|
    id = (tc[:id] || tc["id"]).to_s
    acc[id] = RubyLLM::ToolCall.new(
      id: id, name: (tc[:name] || tc["name"]).to_s,
      arguments: tc[:arguments] || tc["arguments"] || {}
    )
  end
end

#seed_history(chat, messages) ⇒ Object

History comes from the context/checkpoint. The content: shape tolerates string keys (JSON from the stores). flatten(1) dissolves the Session provider's "eviction units" (an assistant+tool_results cycle grouped as one Array, R1) back into a flat message stream.

tool_calls / tool_call_id are rehydrated ONLY when present, so a message without them keeps the 2-arg shape the specs' FakeChat expects (no unknown keyword). This is what lets the model SEE the tools it already called.



214
215
216
217
218
219
220
221
222
223
 
# File 'lib/insika/chat_builder.rb', line 214

def seed_history(chat, messages)
  Array(messages).flatten(1).each do |m|
    attrs = { role: (m[:role] || m["role"]).to_sym, content: m[:content] || m["content"] }
    tool_calls = m[:tool_calls] || m["tool_calls"]
    tool_call_id = m[:tool_call_id] || m["tool_call_id"]
    attrs[:tool_calls] = rehydrate_tool_calls(tool_calls) if tool_calls && !Array(tool_calls).empty?
    attrs[:tool_call_id] = tool_call_id if tool_call_id
    chat.add_message(**attrs)
  end
end

#tool_concurrency_for(state) ⇒ Object

The turn's effective tool concurrency (nil = serial), plus the ONE thing the gate owes the operator: when the profile asked for parallel tool calls and this turn silently cannot have them (an approval-required tool would deadlock two fibers on the single per-task mailbox), say so once. Otherwise the speedup just vanishes with no reason given. The rule itself lives in TurnState; a state predating those readers (a unit stub) means off.



148
149
150
151
152
153
154
155
156
157
 
# File 'lib/insika/chat_builder.rb', line 148

def tool_concurrency_for(state)
  return nil unless state.respond_to?(:tool_concurrency)

  effective = state.tool_concurrency
  return effective if effective

  requested = state.requested_tool_concurrency
  warn_tool_concurrency_gated(state, requested) if requested
  nil
end

#warn_tool_concurrency_gated(state, requested) ⇒ Object 



159
160
161
162
163
164
165
166
167
168
 
# File 'lib/insika/chat_builder.rb', line 159

def warn_tool_concurrency_gated(state, requested)
  gated = Array(state.requires_approval)
  @event_stream.emit(Insika::Event.new(
                       type: :provider_warning,
                       data: { provider: "tool_concurrency",
                               message: "parallel tool calls (#{requested}) disabled for this turn: " \
                                        "#{gated.size} tool(s) require approval (#{gated.join(', ')})" },
                       meta: { task_id: state.task&.id, at: Time.now.utc.iso8601 }
                     ))
end

#wire_callbacks(chat, state, emit) ⇒ Object

RubyLLM's additive callbacks become events. load_skill becomes :skill_activated. Adds the max_tool_calls counter: the loop is RubyLLM's; here we only count and abort.



241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
 
# File 'lib/insika/chat_builder.rb', line 241

def wire_callbacks(chat, state, emit)
  # Per-TURN counter: safe as a closure local even under concurrent tool calls
  # (MRI fibers do not preempt between the read and the write). The per-CALL
  # correlation is NOT — it lives in fiber storage behind TurnState, because
  # each call gets its own fiber once tool concurrency is on.
  tool_calls = 0
  max_tool_calls = state.profile.limits[:max_tool_calls] || 50

  # the loop detector. Needs #after_message + #add_message for the
  # batch-boundary intervention; a chat without them (smoke shim, minimal
  # double) stays bounded by max_tool_calls alone — never half-wired.
  detector = if %i[after_message add_message].all? { |m| chat.respond_to?(m) }
               repeat = state.profile.limits[:max_tool_repeat] || Insika::AgentProfile::DEFAULT_LIMITS[:max_tool_repeat]
               Insika::LoopDetector.new(chat: chat, limit: repeat, emit: emit) if repeat >= 2
             end

  chat.before_tool_call do |tool_call|
    # call<->decorator correlation (side-effects/skip) — 1st line.
    state.current_tool_call = tool_call
    # max_tool_calls guard-rail: stays inline (not as a registered hook)
    # because Hooks is shared across turns and has no unregister.
    tool_calls += 1
    if tool_calls > max_tool_calls
      raise Insika::TimeoutError.new("tool call limit exceeded (#{max_tool_calls})",
                                       stage: :tool_limit)
    end

    # AFTER the count, BEFORE the call runs — a post-warning
    # repeat raises here, so the stubborn loop pays for no extra call.
    detector&.tool_call(tool_call.name, tool_call.arguments)

    # :tool pair: RubyLLM's callbacks are additive — the altered subject
    # feeds later hooks and the events, but does not rewrite the call the
    # model executes. A hook exception here aborts the turn.
    tool_call = @hooks.run_before(:tool, tool_call)

    # The name of the (possibly hook-altered) subject, for the :tool_result
    # label. Also fiber-scoped: as a closure local it belonged to the TURN, so
    # under concurrency `after_tool_result` labelled every result with whichever
    # call started last.
    state.current_tool_name = tool_call.name.to_s
    if state.current_tool_name == "load_skill"
      args = tool_call.arguments || {}
      emit.call(:skill_activated, { name: args["name"] || args[:name] })
    else
      emit.call(:tool_call, { name: tool_call.name, arguments: tool_call.arguments })
    end
  end

  chat.after_tool_result do |result|
    # the RAW result — the only place a Tool::Halt (halt_when) is
    # still recognizable, and a halted batch must receive no intervention.
    detector&.tool_result(result)
    result = @hooks.run_after(:tool, result)
    emit.call(:tool_result, { name: state.current_tool_name, result: result.to_s })
  end

  # the intervention appends at the batch boundary (the Nth tool
  # result closing) — never between two tool results of one batch.
  chat.after_message { |message| detector.message_ended(message) } if detector
end