Module: Pikuri::Agent::Synthesizer
- Defined in:
- lib/pikuri/agent/synthesizer.rb
Overview
Prompt builder for the step-exhaustion rescue. When an Agent‘s Control::StepLimit trips with the :synthesize policy, Agent#run_loop runs this module’s prompt on a nested tools-free agent so the run still produces something useful — an assistant turn that answers the user’s question from whatever evidence the failed agent collected before running out of budget.
Why this exists
Without a rescue, a step-exhausted run just raises a stack trace past bin/pikuri-chat and the user gets nothing despite the agent having gathered useful information in the first N-1 steps. The observed failure mode is the “wait, but what about X?” death-loop: the agent collects sound evidence in the first few rounds, then spends the rest of the budget second-guessing. By the time the cap trips, the answer is largely in the messages — it just needs a tools-free pass to synthesize.
Salvage is the wrong move for some agents, which is why the policy lives on Control::StepLimit and defaults to :raise — a coding agent’s half-finished work can’t be completed by a tools-free pass, only described. See Control::StepLimit‘s class header.
Seam discipline
This module is pure prompt construction — no chat handling, no RubyLLM.chat call, no event wiring. The execution side (constructing the nested agent, sharing the parent’s listener stream and cancellable, capturing the answer) is Agent#run_synthesizer‘s job: the synth is a regular tools-free Agent, the same construction shape the agent tool from pikuri-subagents uses for sub-agents. The only RubyLLM::* surface read here is the value-type RubyLLM::Message / ToolCall passthrough (per the value-type rule in CLAUDE.md).
Constant Summary collapse
- SYSTEM_PROMPT =
The synthesizer’s system prompt. Strict and short: use the evidence, don’t apologize, admit gaps when present.
<<~PROMPT You are given evidence another agent collected before running out of steps. Answer the user's question using only this evidence. You have no tools. If the evidence is insufficient, state plainly what's missing and what partial answer you can give. Do not apologize or comment on the previous agent. PROMPT
Class Method Summary collapse
-
.build_prompt(parent_messages:, user_message:) ⇒ String
Render the user’s question plus an “Evidence gathered” section built from
parent_messagesas a single prompt string. -
.run_synthesizer(ctx, chat_messages, user_message) ⇒ String
The
:synthesizearm of the step-exhaustion policy (see the class header).
Class Method Details
.build_prompt(parent_messages:, user_message:) ⇒ String
Render the user’s question plus an “Evidence gathered” section built from parent_messages as a single prompt string. Pure function — no I/O, safe to test directly with fixture messages.
58 59 60 61 |
# File 'lib/pikuri/agent/synthesizer.rb', line 58 def self.build_prompt(parent_messages:, user_message:) transcript = format_evidence() "Question: #{}\n\nEvidence gathered:\n#{transcript}" end |
.run_synthesizer(ctx, chat_messages, user_message) ⇒ String
The :synthesize arm of the step-exhaustion policy (see the class header). Runs the Pikuri::Agent::Synthesizer prompt over the exhausted chat’s history on a nested tools-free Agent —the same construction shape the agent tool from pikuri-subagents uses for sub-agents, so the synth gets listener propagation, transport / context-window-cap / streaming inheritance, and teardown via close for free. The synth’s answer is returned.
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 |
# File 'lib/pikuri/agent/synthesizer.rb', line 119 def self.run_synthesizer(ctx, , ) # Check the cancel flag *before* constructing the synth: the # nested run_loop resets the shared cancellable at its turn # boundary, which would erase a cancel requested in this # window. The raise propagates without a parent-side # {Event::Cancelled} — a cancel *during* synthesis emits it # from the synth's own rescue (on the derived listener list) # instead, so either way the stream sees at most one. ctx.agent.cancellable&.check! ctx.emit_event(Event::FallbackNotice.new( reason: "agent exhausted #{ctx.agent.step_limit.max} steps; " \ 'synthesizing answer from gathered evidence' )) # Synth runs under this agent's identity but with a # different system prompt, so it gets a distinct # +_synthesizer+ suffix on the id — same +_+ separator the # sub-agent generator uses, so main becomes +"synthesizer"+ # and a sub-agent +"researcher 0"+ becomes # +"researcher 0_synthesizer"+. Any +TokenLog+ in the list # tags the synth's prompt under that bracket so it's # obvious from the log which turns were the rescue rather # than the original loop. synth_id = ctx.agent.id.empty? ? 'synthesizer' : "#{ctx.agent.id}_synthesizer" synth = Agent.new( # Carry the parent's resolved cap on the transport so the synth # reuses it without a re-probe — the cap rides {ChatTransport} # now, not an +Agent.new(context_window:)+ kwarg. transport: ctx.agent.transport.with(context_window: ctx.agent.context_window_cap), system_prompt: Synthesizer::SYSTEM_PROMPT, # Defensive budget with the default :raise policy: the # synth has no tools so it should never tick, but a buggy # provider that somehow returns a tool call must not loop # forever — and a synth that needs its own synth is a bug, # not a rescue. step_limit: Control::StepLimit.new(max: 1), cancellable: ctx.agent.cancellable, id: synth_id, streaming: ctx.agent.streaming ) { |c| c.add_listeners(ctx.sub_agent_listeners(id: synth_id)) } begin synth.run_loop(user_message: Synthesizer.build_prompt( parent_messages: , user_message: )) synth.last_assistant_content ensure synth.close end end |