A platform for community discussion. Free, open, simple.
FIX: Agentic agents crash when `max_turn_tokens` is empty (#40952)
Rafael dos Santos Silva74693d5f7592
diff --git a/plugins/discourse-ai/config/locales/client.en.yml b/plugins/discourse-ai/config/locales/client.en.ymlindex b18664bd418..0efabc7dc96 100644--- a/plugins/discourse-ai/config/locales/client.en.yml+++ b/plugins/discourse-ai/config/locales/client.en.yml@@ -575,7 +575,7 @@ en:default: "Default (Fixed Limits)"agentic: "Agentic (Token Budget)"max_turn_tokens: "Max Turn Tokens"- max_turn_tokens_help: "Maximum total tokens (input + output) consumed across all LLM calls during a single bot reply. When set, replaces the fixed completion/tool limits with a token budget, enabling more agentic behavior. Leave empty for default limits."+ max_turn_tokens_help: "Maximum total tokens (input + output) consumed across all LLM calls during a single bot reply. When set, replaces the fixed completion/tool limits with a token budget, enabling more agentic behavior. Leave empty to default to half of the model's context window."compression_threshold: "Compression Threshold"compression_threshold_help: "Percentage of context usage (20–99) at which automatic compression kicks in. Lower values trigger compression earlier, preserving more headroom."vision_enabled: Include image uploadsdiff --git a/plugins/discourse-ai/lib/agents/bot.rb b/plugins/discourse-ai/lib/agents/bot.rbindex 47ffc7c1e58..bb9a80030d6 100644--- a/plugins/discourse-ai/lib/agents/bot.rb+++ b/plugins/discourse-ai/lib/agents/bot.rb@@ -22,6 +22,16 @@ module DiscourseAiprompt.push(type: :user, content: BUDGET_EXHAUSTED_HINT)end+ # When an agentic agent has no explicit max_turn_tokens, default the turn+ # budget to half of the LLM's context window (see the max_turn_tokens_help+ # string). Returns nil when the LLM exposes no usable context window, in+ # which case callers fall back to the fixed completion/tool limits.+ def self.default_max_turn_tokens(llm)+ max_prompt_tokens = llm&.max_prompt_tokens.to_i+ return if max_prompt_tokens <= 0+ max_prompt_tokens / 2+ end+def self.as(bot_user, agent: DiscourseAi::Agents::General.new, model: nil)new(bot_user, agent, model)end@@ -89,8 +99,17 @@ module DiscourseAineeds_newlines = falsetools_ran = 0- use_token_budget = agent.class.execution_mode == "agentic"token_budget = agent.class.max_turn_tokens+ use_token_budget = agent.class.execution_mode == "agentic"++ if use_token_budget && token_budget.blank?+ # max_turn_tokens is optional; default the turn budget to half of the+ # LLM context window. If the LLM exposes no context window, fall back+ # to the fixed completion/tool limits rather than crashing on a nil+ # budget comparison.+ token_budget = self.class.default_max_turn_tokens(current_llm)+ use_token_budget = token_budget.present?+ end# In token budget mode, compression manages context size instead of# the dialect's destructive trim_messages. Disabling trim preventsdiff --git a/plugins/discourse-ai/lib/ai_bot/stream_reply_custom_tools_session.rb b/plugins/discourse-ai/lib/ai_bot/stream_reply_custom_tools_session.rbindex d9c87a200b2..8d325bbaeeb 100644--- a/plugins/discourse-ai/lib/ai_bot/stream_reply_custom_tools_session.rb+++ b/plugins/discourse-ai/lib/ai_bot/stream_reply_custom_tools_session.rb@@ -178,6 +178,15 @@ module DiscourseAitoken_budget = resolve_token_budgetuse_token_budget = resolve_execution_mode == "agentic"+ if use_token_budget && token_budget.blank?+ # max_turn_tokens is optional; default the turn budget to half of the+ # LLM context window. If the LLM exposes no context window, fall back+ # to the fixed completion/tool limits rather than crashing on a nil+ # budget comparison.+ token_budget = DiscourseAi::Agents::Bot.default_max_turn_tokens(llm)+ use_token_budget = token_budget.present?+ end+execution_context = niltoken_usage_tracker = nilif use_token_budgetdiff --git a/plugins/discourse-ai/spec/lib/agents/bot_spec.rb b/plugins/discourse-ai/spec/lib/agents/bot_spec.rbindex 2c998a9c08f..60da5352f18 100644--- a/plugins/discourse-ai/spec/lib/agents/bot_spec.rb+++ b/plugins/discourse-ai/spec/lib/agents/bot_spec.rb@@ -255,6 +255,56 @@ RSpec.describe DiscourseAi::Agents::Bot doexpect(call_count).to eq(described_class::MAX_COMPLETIONS + 1)end+ it "defaults the turn budget to half the context window when agentic without max_turn_tokens" do+ # The editor lets an agentic agent be saved without a max_turn_tokens+ # budget ("Leave empty for default limits"). This previously crashed+ # with "comparison of Integer with nil failed". Instead the agent stays+ # agentic and the budget defaults to half the LLM context window.+ agentic_no_budget_agent =+ Fabricate(+ :ai_agent,+ execution_mode: "agentic",+ max_turn_tokens: nil,+ compression_threshold: 80,+ tools: [["ListCategories", nil, false]],+ )++ klass = agentic_no_budget_agent.class_instance++ # gpt_4 has max_prompt_tokens 131_072, so the default budget is 65_536.+ expect(DiscourseAi::Agents::Bot.default_max_turn_tokens(bot.send(:llm))).to eq(65_536)++ tool_call =+ DiscourseAi::Completions::ToolCall.new(id: "call_1", name: "categories", parameters: {})++ responses = [tool_call, "Final answer"]+ call_count = 0++ DiscourseAi::Completions::Llm.with_prepared_responses(responses) do+ agent_bot = described_class.as(bot_user, agent: klass.new)+ context =+ DiscourseAi::Agents::BotContext.new(messages: [{ type: :user, content: "test" }])++ allow_any_instance_of(DiscourseAi::Completions::Llm).to receive(+ :generate,+ ).and_wrap_original do |original, *args, **kwargs, &blk|+ call_count += 1+ result = original.call(*args, **kwargs, &blk)+ # 70_000 tokens per call exceeds the 65_536 default budget after the+ # first call, so the loop stops on the token budget (not the legacy+ # MAX_COMPLETIONS limit).+ if (tracker = kwargs[:execution_context]&.token_usage_tracker)+ tracker.add_effective(request: 40_000, response: 30_000)+ end+ result+ end++ agent_bot.reply(context) { |_partial| }+ end++ expect(call_count).to eq(2)+ end+it "forces a final text-only call with budget hint when budget exhausted after tool execution" do# budget=2000, first call adds 3000 tokens → tool runs → budget exceeded# but prompt ends with :tool, so model gets one more tool_choice=:none calldiff --git a/plugins/discourse-ai/spec/lib/modules/ai_bot/stream_reply_custom_tools_session_spec.rb b/plugins/discourse-ai/spec/lib/modules/ai_bot/stream_reply_custom_tools_session_spec.rbindex 1bf5db98dec..ca676010cc8 100644--- a/plugins/discourse-ai/spec/lib/modules/ai_bot/stream_reply_custom_tools_session_spec.rb+++ b/plugins/discourse-ai/spec/lib/modules/ai_bot/stream_reply_custom_tools_session_spec.rb@@ -141,6 +141,47 @@ RSpec.describe DiscourseAi::AiBot::StreamReplyCustomToolsSession doendend+ it "defaults the budget to half the context window when agentic without max_turn_tokens" do+ # "Leave empty for default limits": an agentic agent with no+ # max_turn_tokens must not crash, and the budget defaults to half the LLM+ # context window (fake_llm has max_prompt_tokens 131_072 → 65_536).+ ai_agent.update!(max_turn_tokens: nil)++ tool_call =+ DiscourseAi::Completions::ToolCall.new(+ name: "client_tool",+ parameters: {+ input: "test",+ },+ id: "tool_4",+ )++ DiscourseAi::Completions::Llm.with_prepared_responses(+ [tool_call, "Summary after budget hit."],+ ) do+ session = build_session++ allow_any_instance_of(DiscourseAi::Completions::Llm).to receive(+ :generate,+ ).and_wrap_original do |original, *args, **kwargs, &blk|+ result = original.call(*args, **kwargs, &blk)+ # 70_000 tokens exceeds the 65_536 default budget after the first call.+ if (tracker = kwargs[:execution_context]&.token_usage_tracker)+ tracker.add_effective(request: 40_000, response: 30_000)+ end+ result+ end++ events = collect_events(session)++ partials = events.select { |type, _| type == :partial }.map { |_, data| data }+ expect(partials.join).to eq("Summary after budget hit.")++ tool_events = events.select { |type, _| type == :tool_calls }+ expect(tool_events).to be_empty+ end+ end+it "persists request and response token counters in resume state" dotool_call =DiscourseAi::Completions::ToolCall.new(