Skip to content

A platform for community discussion. Free, open, simple.

FIX: Drop unsummarizable AI history safely (#44404)

Sam074b8462da98
diff --git a/plugins/discourse-ai/config/locales/client.en.yml b/plugins/discourse-ai/config/locales/client.en.yml
index 22ea3dbfb95..f8127136710 100644
--- a/plugins/discourse-ai/config/locales/client.en.yml
+++ b/plugins/discourse-ai/config/locales/client.en.yml
@@ -887,7 +887,7 @@ en:
no_llm_selected: "No language model selected"
use_parent_llm: "Use agent's language model"
max_turn_tokens: "Turn work allowance"
- max_turn_tokens_help: "Limits new output and tool results per reply, shared with child agents. Previous conversation does not count. This is not a spending limit. Leave empty to use the model’s context window."
+ max_turn_tokens_help: "Limits new output and tool results per reply, shared with child agents. Previous conversation does not count. This is not a spending limit. Leave empty to use 500,000 tokens."
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 uploads
diff --git a/plugins/discourse-ai/lib/agents/bot.rb b/plugins/discourse-ai/lib/agents/bot.rb
index 3ff5891ae93..0fcff1d6941 100644
--- a/plugins/discourse-ai/lib/agents/bot.rb
+++ b/plugins/discourse-ai/lib/agents/bot.rb
@@ -5,7 +5,7 @@ module DiscourseAi
class Bot
BOT_NOT_FOUND = Class.new(StandardError)
- FALLBACK_MAX_TURN_TOKENS = 32_000
+ DEFAULT_MAX_TURN_TOKENS = 500_000
MAX_DISCOVERED_IMAGE_REFERENCES = 20
MAX_TOOL_CALLS_PER_COMPLETION = 50
MAX_FINAL_ANSWER_TOKENS = DiscourseAi::Completions::TurnWorkBudget::MAX_FINAL_ANSWER_TOKENS
@@ -64,15 +64,8 @@ module DiscourseAi
inject_token_budget_final_answer_hint(prompt)
end
- def self.default_max_turn_tokens(llm, compression_threshold: 80)
- effective_max_turn_tokens(llm)
- end
-
- def self.effective_max_turn_tokens(llm, requested = nil, **_options)
- return requested.to_i if requested.to_i > 0
-
- context_window = llm&.max_prompt_tokens.to_i
- context_window > 0 ? context_window : FALLBACK_MAX_TURN_TOKENS
+ def self.effective_max_turn_tokens(requested = nil)
+ requested.to_i > 0 ? requested.to_i : DEFAULT_MAX_TURN_TOKENS
end
def self.as(bot_user, agent: DiscourseAi::Agents::General.new, model: nil)
@@ -175,7 +168,6 @@ module DiscourseAi
current_llm = llm
token_budget =
self.class.effective_max_turn_tokens(
- current_llm,
context.turn_token_budget.presence || agent.class.max_turn_tokens.presence,
)
execution_context ||=
diff --git a/plugins/discourse-ai/lib/agents/subagent_runner.rb b/plugins/discourse-ai/lib/agents/subagent_runner.rb
index ba1e2c9dd07..15642afa366 100644
--- a/plugins/discourse-ai/lib/agents/subagent_runner.rb
+++ b/plugins/discourse-ai/lib/agents/subagent_runner.rb
@@ -57,7 +57,7 @@ module DiscourseAi
yield(@prompt) if block_given?
child_agent = child_class.new
- child_context = build_child_context(child_record, model)
+ child_context = build_child_context(child_record)
response = collect_response(child_agent, child_context, model)
return error("completion_limit", child_record) if child_context.completion_limit_reached
@@ -111,7 +111,7 @@ module DiscourseAi
required_tools.empty? || (required_tools - Agent.all_available_tools).empty?
end
- def build_child_context(child_record, model)
+ def build_child_context(child_record)
child_depth = @context.subagent_depth.to_i + 1
feature_context =
@context.feature_context.to_h.merge(
@@ -119,11 +119,7 @@ module DiscourseAi
"subagent_agent_id" => child_record.id,
"subagent_depth" => child_depth,
)
- child_budget =
- Bot.effective_max_turn_tokens(
- DiscourseAi::Completions::Llm.proxy(model),
- child_record.max_turn_tokens,
- )
+ child_budget = Bot.effective_max_turn_tokens(child_record.max_turn_tokens)
effective_budget = [child_budget, @context.subagent_execution_state.remaining_tokens].min
child_execution = @context.execution_context.dup
diff --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.rb
index f0371ac4c6c..0f014073ee0 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
@@ -308,7 +308,7 @@ module DiscourseAi
@work_budget ||=
DiscourseAi::Completions::TurnWorkBudget.new(
- limit: DiscourseAi::Agents::Bot.effective_max_turn_tokens(llm, resolve_token_budget),
+ limit: DiscourseAi::Agents::Bot.effective_max_turn_tokens(resolve_token_budget),
)
token_usage_tracker =
DiscourseAi::Completions::TokenUsageTracker.new(
diff --git a/plugins/discourse-ai/lib/completions/context_preparation.rb b/plugins/discourse-ai/lib/completions/context_preparation.rb
index e8b7bc79b52..31478f83284 100644
--- a/plugins/discourse-ai/lib/completions/context_preparation.rb
+++ b/plugins/discourse-ai/lib/completions/context_preparation.rb
@@ -7,6 +7,14 @@ module DiscourseAi
MAX_OUTPUT_TOKENS = 2048
MAX_SECONDS = 90
MAX_ALLOWANCE_TOKENS = 1_000_000
+ RETAINED_ATTACHMENTS_ACK = "The preceding attachments are retained historical evidence."
+ HISTORY_DROPPED_NOTICE = <<~TEXT.squish
+ Part of the earlier conversation was omitted because it could not be summarized to fit
+ the context window. Tool actions in the omitted part may already have been completed;
+ do not repeat them unless the user asks.
+ TEXT
+ PREVIOUS_SUMMARY_LABEL =
+ "\nSummary from an earlier checkpoint (it does not cover messages omitted after it):\n"
class Budget
attr_reader :calls, :spent_tokens, :elapsed_seconds
@@ -118,6 +126,10 @@ module DiscourseAi
# when retaining it would prevent hard fit.
source = prompt.messages[1...protected_user_index]
tail = prompt.messages[protected_user_index..]
+ # Captured before completed current-turn evidence can move into the summary
+ # source: dropping may only cut earlier turns, never actions that already ran.
+ history = source.reject { |message| transient_hint?(message) }
+ current_turn = tail
if size > capacity && !fits?(prompt, [prompt.messages.first, *tail], **options)
pending_ids =
prompt.messages.filter_map { |message| message[:id] if message[:type] == :tool_call }
@@ -144,13 +156,11 @@ module DiscourseAi
message[:type] == :user &&
Array(message[:content]).first.to_s.start_with?(
"Retained historical attachments",
- ) &&
- acknowledgement&.dig(:content) ==
- "The preceding attachments are retained historical evidence."
+ ) && acknowledgement&.dig(:content) == RETAINED_ATTACHMENTS_ACK
end
if source.empty? || checkpoint_only
- raise Error.new("hard_overflow") if size > capacity
- return :not_needed
+ return :not_needed if size <= capacity
+ return drop_oldest_history!(prompt, history, current_turn, "hard_overflow", **options)
end
begin
@@ -167,21 +177,8 @@ module DiscourseAi
)
messages = [
prompt.messages.first,
- {
- type: :user,
- content:
- "#{PromptMessagesBuilder::COMPRESSED_CONTEXT_PREFIX}#{summary}#{PromptMessagesBuilder::COMPRESSED_CONTEXT_SUFFIX}",
- },
- { type: :model, content: PromptMessagesBuilder::COMPRESSED_CONTEXT_ACK },
- *attachments.flat_map do |message|
- [
- message,
- {
- type: :model,
- content: "The preceding attachments are retained historical evidence.",
- },
- ]
- end,
+ *checkpoint_messages(summary),
+ *attachment_messages(attachments),
*tail,
]
candidate = copy_prompt(prompt, messages)
@@ -198,16 +195,99 @@ module DiscourseAi
rescue => error
reason = error.is_a?(Error) ? error.reason : "summary_failed"
if !error.is_a?(Error)
- Rails.logger.warn("DiscourseAi: context_preparation reason=#{reason}")
+ Rails.logger.warn(
+ "DiscourseAi: context_preparation reason=#{reason} error=#{error.class}: #{error.message}",
+ )
end
return :cancelled if cancel_manager&.cancelled?
- raise Error.new(reason) if size > capacity
- :skipped
+ return :skipped if size <= capacity
+ drop_oldest_history!(prompt, history, current_turn, reason, **options)
end
end
private
+ # Last resort when summarizing cannot make the prompt fit. The notice is
+ # written as a checkpoint so later replies resume after the dropped history.
+ # Priority: recent turns, then the previous summary, then dropped attachments.
+ def drop_oldest_history!(prompt, history, tail, reason, **options)
+ raise Error.new(reason) if history.empty?
+
+ build =
+ lambda do |notice, start, attachments = []|
+ [
+ prompt.messages.first,
+ *checkpoint_messages(notice),
+ *attachment_messages(attachments),
+ *history[start..],
+ *tail,
+ ]
+ end
+ # Agent examples precede a restored checkpoint, so only cut after it; any
+ # user message there keeps tool calls with their results.
+ checkpoint_index = PromptMessagesBuilder.compression_checkpoint_index(history)
+ first_cut = checkpoint_index ? checkpoint_index + 2 : 0
+ starts =
+ (first_cut...history.length).select do |index|
+ history[index][:type] == :user && !transient_hint?(history[index])
+ end
+ starts << history.length
+ start =
+ starts.bsearch do |candidate|
+ fits?(prompt, build.call(HISTORY_DROPPED_NOTICE, candidate), **options)
+ end
+ raise Error.new(reason) if start.nil?
+
+ notice = HISTORY_DROPPED_NOTICE
+ if checkpoint_index
+ # Unwrap an earlier drop notice so repeated drops don't nest notices.
+ previous_summary =
+ PromptMessagesBuilder
+ .message_text(history[checkpoint_index])
+ .delete_prefix(PromptMessagesBuilder::COMPRESSED_CONTEXT_PREFIX)
+ .delete_suffix(PromptMessagesBuilder::COMPRESSED_CONTEXT_SUFFIX)
+ .delete_prefix(HISTORY_DROPPED_NOTICE)
+ .delete_prefix(PREVIOUS_SUMMARY_LABEL)
+ if previous_summary.present?
+ with_summary = "#{HISTORY_DROPPED_NOTICE}#{PREVIOUS_SUMMARY_LABEL}#{previous_summary}"
+ notice = with_summary if fits?(prompt, build.call(with_summary, start), **options)
+ end
+ end
+
+ retained = []
+ @llm
+ .context_attachments(history[0...start])
+ .reverse_each do |attachment|
+ candidate = [attachment, *retained]
+ retained = candidate if fits?(prompt, build.call(notice, start, candidate), **options)
+ end
+
+ Rails.logger.warn(
+ "DiscourseAi: context_preparation reason=history_dropped original_reason=#{reason} dropped_history_messages=#{start}",
+ )
+ messages = build.call(notice, start, retained)
+ @protected_user_index = messages.length - tail.length
+ prompt.messages.replace(messages)
+ :compressed
+ end
+
+ def checkpoint_messages(summary)
+ [
+ {
+ type: :user,
+ content:
+ "#{PromptMessagesBuilder::COMPRESSED_CONTEXT_PREFIX}#{summary}#{PromptMessagesBuilder::COMPRESSED_CONTEXT_SUFFIX}",
+ },
+ { type: :model, content: PromptMessagesBuilder::COMPRESSED_CONTEXT_ACK },
+ ]
+ end
+
+ def attachment_messages(attachments)
+ attachments.flat_map do |message|
+ [message, { type: :model, content: RETAINED_ATTACHMENTS_ACK }]
+ end
+ end
+
def fits?(prompt, messages, **options)
size, capacity = @llm.prompt_capacity(copy_prompt(prompt, messages), **options)
size <= capacity
diff --git a/plugins/discourse-ai/lib/summarization/fold_content.rb b/plugins/discourse-ai/lib/summarization/fold_content.rb
index 146ae48f05f..6d1992d2253 100644
--- a/plugins/discourse-ai/lib/summarization/fold_content.rb
+++ b/plugins/discourse-ai/lib/summarization/fold_content.rb
@@ -136,10 +136,7 @@ module DiscourseAi
work_budget =
DiscourseAi::Completions::TurnWorkBudget.new(
limit:
- DiscourseAi::Agents::Bot.effective_max_turn_tokens(
- llm,
- bot.agent.class.max_turn_tokens,
- ),
+ DiscourseAi::Agents::Bot.effective_max_turn_tokens(bot.agent.class.max_turn_tokens),
)
output_limit =
work_budget.generation_options(
diff --git a/plugins/discourse-ai/lib/translation/base_translator.rb b/plugins/discourse-ai/lib/translation/base_translator.rb
index 3e5fb8b2aa8..4e6c02af721 100644
--- a/plugins/discourse-ai/lib/translation/base_translator.rb
+++ b/plugins/discourse-ai/lib/translation/base_translator.rb
@@ -77,8 +77,7 @@ module DiscourseAi
available_tokens = capacity * threshold / 100 - size - PROMPT_TOKEN_RESERVE
work_budget =
DiscourseAi::Completions::TurnWorkBudget.new(
- limit:
- DiscourseAi::Agents::Bot.effective_max_turn_tokens(llm, agent.class.max_turn_tokens),
+ limit: DiscourseAi::Agents::Bot.effective_max_turn_tokens(agent.class.max_turn_tokens),
)
output_limit =
work_budget.generation_options(
diff --git a/plugins/discourse-ai/spec/lib/agents/bot_spec.rb b/plugins/discourse-ai/spec/lib/agents/bot_spec.rb
index f07bb333966..4e47276fa1c 100644
--- a/plugins/discourse-ai/spec/lib/agents/bot_spec.rb
+++ b/plugins/discourse-ai/spec/lib/agents/bot_spec.rb
@@ -25,27 +25,11 @@ RSpec.describe DiscourseAi::Agents::Bot do
let(:llm_responses) { [function_call, response] }
describe ".effective_max_turn_tokens" do
- it "defaults to the model context window while keeping explicit allowances exact" do
- [4096, 200_000, nil].each do |context_window|
- model = instance_double(DiscourseAi::Completions::Llm, max_prompt_tokens: context_window)
- default = context_window || described_class::FALLBACK_MAX_TURN_TOKENS
- expect(described_class.effective_max_turn_tokens(model, 8000)).to eq(8000)
- expect(described_class.effective_max_turn_tokens(model, 60_000)).to eq(60_000)
- expect(described_class.effective_max_turn_tokens(model, 10)).to eq(10)
- expect(described_class.default_max_turn_tokens(model)).to eq(default)
- expect(
- described_class.effective_max_turn_tokens(
- model,
- 4000,
- compression_threshold: 50,
- max_tokens: 1,
- thinking_effort: "high",
- ),
- ).to eq(4000)
- expect(described_class.default_max_turn_tokens(model, compression_threshold: 50)).to eq(
- default,
- )
- end
+ it "defaults to a fixed allowance while keeping explicit allowances exact" do
+ expect(described_class.effective_max_turn_tokens(8000)).to eq(8000)
+ expect(described_class.effective_max_turn_tokens(10)).to eq(10)
+ expect(described_class.effective_max_turn_tokens(nil)).to eq(500_000)
+ expect(described_class.effective_max_turn_tokens(0)).to eq(500_000)
end
end
@@ -630,7 +614,7 @@ RSpec.describe DiscourseAi::Agents::Bot do
it "allows a final response when the tracker starts at the token budget" do
tracker =
DiscourseAi::Completions::TokenUsageTracker.new(
- base_request: described_class.effective_max_turn_tokens(bot.llm, 5000),
+ base_request: described_class.effective_max_turn_tokens(5000),
base_response: 0,
)
execution_context =
@@ -665,7 +649,7 @@ RSpec.describe DiscourseAi::Agents::Bot do
)
end
- it "defaults the work allowance to the model context window" do
+ it "defaults the work allowance to DEFAULT_MAX_TURN_TOKENS" do
no_budget_agent =
Fabricate(
:ai_agent,
@@ -676,8 +660,6 @@ RSpec.describe DiscourseAi::Agents::Bot do
klass = no_budget_agent.class_instance
- expect(described_class.default_max_turn_tokens(bot.llm)).to eq(gpt_4.max_prompt_tokens)
-
tool_call =
DiscourseAi::Completions::ToolCall.new(id: "call_1", name: "categories", parameters: {})
@@ -698,7 +680,7 @@ RSpec.describe DiscourseAi::Agents::Bot do
if (tracker = kwargs[:execution_context]&.token_usage_tracker)
tracker.add_effective(request: 90_000, response: 30_000)
kwargs[:execution_context].work_budget.debit(
- gpt_4.max_prompt_tokens,
+ described_class::DEFAULT_MAX_TURN_TOKENS,
event_id: SecureRandom.uuid,
)
end
@@ -726,9 +708,6 @@ RSpec.describe DiscourseAi::Agents::Bot do
)
expect(canned.completions).to eq(0)
end
- expect(described_class.default_max_turn_tokens(nil)).to eq(
- described_class::FALLBACK_MAX_TURN_TOKENS,
- )
end
it "retains a tool-heavy previous turn despite a small work allowance" do
@@ -821,7 +800,7 @@ RSpec.describe DiscourseAi::Agents::Bot do
agent_bot.reply(context)
llm = agent_bot.llm
input, _, output = llm.prompt_capacity(prompts.last)
- baseline = described_class.effective_max_turn_tokens(llm, 8000)
+ baseline = described_class.effective_max_turn_tokens(8000)
expect(options.map { |option| option[:feature_name] }).to eq(%w[context_compression bot])
expect(prompts.last.tool_choice).not_to eq(:none)
diff --git a/plugins/discourse-ai/spec/lib/completions/context_preparation_spec.rb b/plugins/discourse-ai/spec/lib/completions/context_preparation_spec.rb
index e01fd3d0804..6bac250e2db 100644
--- a/plugins/discourse-ai/spec/lib/completions/context_preparation_spec.rb
+++ b/plugins/discourse-ai/spec/lib/completions/context_preparation_spec.rb
@@ -24,6 +24,17 @@ RSpec.describe DiscourseAi::Completions::ContextPreparation do
)
end
+ def expect_history_dropped(prompt, retained: [])
+ expect(prompt.messages.map { |message| message[:type] }).to eq(
+ [:system, :user, :model, *retained.map { |message| message[:type] }, :user],
+ )
+ expect(prompt.messages[1][:content]).to start_with(
+ "<compressed_context>#{described_class::HISTORY_DROPPED_NOTICE}",
+ )
+ expect(prompt.messages[3...-1]).to eq(retained)
+ expect(prompt.messages.last[:content]).to eq("What codeword?")
+ end
+
describe "#prepare!" do
it "shares maintenance bounds across contexts forked before their first preparation" do
stub_const(described_class, :MAX_CALLS, 2) do
@@ -98,20 +109,166 @@ RSpec.describe DiscourseAi::Completions::ContextPreparation do
end
end
- it "fails hard overflow without publishing a partial checkpoint or answering" do
+ it "drops the oldest history when an unusable summary leaves the prompt over capacity" do
prompt = history_prompt(repetitions: 4000)
- original = prompt.messages.deep_dup
DiscourseAi::Completions::Llm.with_prepared_responses([""]) do |canned|
llm = DiscourseAi::Completions::Llm.proxy(model)
+ preparation = described_class.new(llm)
+
+ expect(preparation.prepare!(prompt)).to eq(:compressed)
+ expect(canned.completions).to eq(1)
+ expect_history_dropped(prompt)
+ expect(preparation.protected_user_index).to eq(prompt.messages.length - 1)
+ end
+ end
+
+ it "keeps the newest complete earlier turns that fit when dropping history" do
+ prompt = history_prompt(repetitions: 4000)
+ # Merged consecutive user messages lose their id but are still turn boundaries.
+ recent_turn = [
+ { type: :user, content: "Also remember AMBER-FOX-42" },
+ { type: :tool_call, id: "recent", name: "read", content: '{"arguments":{}}' },
+ { type: :tool, id: "recent", name: "read", content: "AMBER-FOX-42 stored" },
+ { type: :model, content: "Noted." },
+ ]
+ prompt.messages.insert(-2, *recent_turn.deep_dup)
+ DiscourseAi::Completions::Llm.with_prepared_responses([""]) do
+ llm = DiscourseAi::Completions::Llm.proxy(model)
+
+ expect(described_class.new(llm).prepare!(prompt)).to eq(:compressed)
+ expect_history_dropped(prompt, retained: recent_turn)
+ end
+ end
+
+ it "carries a previous checkpoint summary restored behind agent examples" do
+ prompt = history_prompt(repetitions: 4000)
+ prompt.messages.insert(
+ 1,
+ { type: :user, content: "Example question" },
+ { type: :model, content: "Example answer" },
+ { type: :user, content: "<compressed_context>Earlier summary</compressed_context>" },
+ {
+ type: :model,
+ content: DiscourseAi::Completions::PromptMessagesBuilder::COMPRESSED_CONTEXT_ACK,
+ },
+ )
+ DiscourseAi::Completions::Llm.with_prepared_responses([""]) do
+ llm = DiscourseAi::Completions::Llm.proxy(model)
+
+ expect(described_class.new(llm).prepare!(prompt)).to eq(:compressed)
+ expect_history_dropped(prompt)
+ expect(prompt.messages[1][:content]).to include("Earlier summary")
+ expect(prompt.messages.to_s).not_to include("Example question")
+ expect(
+ DiscourseAi::Completions::PromptMessagesBuilder.compression_checkpoint_index(
+ prompt.messages,
+ ),
+ ).to eq(1)
+ end
+ end
+
+ it "does not nest notices when dropping history again" do
+ prompt = history_prompt(repetitions: 4000)
+ previous =
+ "#{described_class::HISTORY_DROPPED_NOTICE}#{described_class::PREVIOUS_SUMMARY_LABEL}Earlier summary"
+ prompt.messages.insert(
+ 1,
+ { type: :user, content: "<compressed_context>#{previous}</compressed_context>" },
+ {
+ type: :model,
+ content: DiscourseAi::Completions::PromptMessagesBuilder::COMPRESSED_CONTEXT_ACK,
+ },
+ )
+ DiscourseAi::Completions::Llm.with_prepared_responses([""]) do
+ llm = DiscourseAi::Completions::Llm.proxy(model)
+
+ expect(described_class.new(llm).prepare!(prompt)).to eq(:compressed)
+ expect(prompt.messages[1][:content]).to eq(
+ "<compressed_context>#{previous}</compressed_context>",
+ )
+ end
+ end
+
+ it "prefers recent turns over a previous summary when both do not fit" do
+ prompt = history_prompt(repetitions: 4000)
+ prompt.messages.insert(
+ 1,
+ {
+ type: :user,
+ content: "<compressed_context>#{"summary detail " * 4000}</compressed_context>",
+ },
+ {
+ type: :model,
+ content: DiscourseAi::Completions::PromptMessagesBuilder::COMPRESSED_CONTEXT_ACK,
+ },
+ )
+ recent_turn = [
+ { type: :user, content: "recent evidence " * 4000, id: "alice" },
+ { type: :model, content: "Noted." },
+ ]
+ prompt.messages.insert(-2, *recent_turn.deep_dup)
+ DiscourseAi::Completions::Llm.with_prepared_responses([""]) do
+ llm = DiscourseAi::Completions::Llm.proxy(model)
+
+ expect(described_class.new(llm).prepare!(prompt)).to eq(:compressed)
+ expect_history_dropped(prompt, retained: recent_turn)
+ expect(prompt.messages[1][:content]).not_to include("summary detail")
+ end
+ end
+
+ it "keeps the current turn's completed tool batch while dropping earlier history" do
+ prompt = history_prompt(repetitions: 4000)
+ current_batch = [
+ { type: :tool_call, id: "now", name: "read", content: '{"arguments":{}}' },
+ { type: :tool, id: "now", name: "read", content: "Side effect completed" },
+ ]
+ prompt.messages.concat(current_batch.deep_dup)
+ DiscourseAi::Completions::Llm.with_prepared_responses([""]) do
+ llm = DiscourseAi::Completions::Llm.proxy(model)
+ preparation = described_class.new(llm)
+
+ expect(preparation.prepare!(prompt)).to eq(:compressed)
+ expect(prompt.messages.map { |message| message[:type] }).to eq(
+ %i[system user model user tool_call tool],
+ )
+ expect(prompt.messages.last(2)).to eq(current_batch)
+ expect(preparation.protected_user_index).to eq(3)
+ end
+ end
+
+ it "fails rather than dropping current-turn evidence that cannot fit without history" do
+ prompt = history_prompt(repetitions: 10)
+ prompt.push(type: :tool_call, id: "large", name: "read", content: '{"arguments":{}}')
+ prompt.push(type: :tool, id: "large", name: "read", content: "large result " * 20_000)
+ original = prompt.messages.deep_dup
+ DiscourseAi::Completions::Llm.with_prepared_responses([""]) do
+ llm = DiscourseAi::Completions::Llm.proxy(model)
+
expect { described_class.new(llm).prepare!(prompt) }.to raise_error(
described_class::Error,
/unusable_summary/,
)
- expect(canned.completions).to eq(1)
expect(prompt.messages).to eq(original)
end
end
+ it "retains attachments from dropped history when they fit" do
+ prompt = history_prompt(repetitions: 4000)
+ image = { encoded_upload: { kind: :image, mime_type: "image/png", base64: "image-data" } }
+ prompt.messages[1][:content] = ["Remember this image", image]
+ DiscourseAi::Completions::Llm.with_prepared_responses([""]) do
+ llm = DiscourseAi::Completions::Llm.proxy(model)
+
+ expect(described_class.new(llm).prepare!(prompt)).to eq(:compressed)
+ expect(prompt.messages.map { |message| message[:type] }).to eq(
+ %i[system user model user model user],
+ )
+ expect(prompt.messages[3][:content]).to include(image)
+ expect(prompt.messages[4][:content]).to eq(described_class::RETAINED_ATTACHMENTS_ACK)
+ expect(prompt.messages.last[:content]).to eq("What codeword?")
+ end
+ end
+
it "preserves every fragment of source larger than the model window through bounded merges" do
prompt = history_prompt(repetitions: 5000)
original_result = prompt.messages[3][:content]
@@ -163,16 +320,12 @@ RSpec.describe DiscourseAi::Completions::ContextPreparation do
expect(logs.last.prev_log_id).to eq(logs[-2].id)
end
- it "bounds maintenance calls and leaves hard-overflow history intact on exhaustion" do
+ it "bounds maintenance calls and drops hard-overflow history on exhaustion" do
prompt = history_prompt(repetitions: 5000)
- original = prompt.messages.deep_dup
DiscourseAi::Completions::Llm.with_prepared_responses(["summary"]) do |canned|
llm = DiscourseAi::Completions::Llm.proxy(model)
- expect { described_class.new(llm, calls: 8).prepare!(prompt) }.to raise_error(
- described_class::Error,
- /preparation_budget_exhausted/,
- )
- expect(prompt.messages).to eq(original)
+ expect(described_class.new(llm, calls: 8).prepare!(prompt)).to eq(:compressed)
+ expect_history_dropped(prompt)
expect(canned.completions).to eq(0)
end
end
@@ -256,7 +409,7 @@ RSpec.describe DiscourseAi::Completions::ContextPreparation do
end
end
- it "falls back intact when completion reservation fails softly and rejects hard overflow" do
+ it "falls back intact when completion reservation fails softly and drops history on hard overflow" do
tracker = DiscourseAi::Completions::TokenUsageTracker.new
execution = DiscourseAi::Completions::ExecutionContext.new(token_usage_tracker: tracker)
state =
@@ -277,12 +430,11 @@ RSpec.describe DiscourseAi::Completions::ContextPreparation do
).to eq(:skipped)
expect(soft_prompt.messages).to eq(original)
expect(soft_prompt.skip_trim).to eq(true)
- expect {
- described_class.new(llm).prepare!(
- history_prompt(repetitions: 4000),
- subagent_execution_state: state,
- )
- }.to raise_error(described_class::Error, /completion_limit/)
+ hard_prompt = history_prompt(repetitions: 4000)
+ expect(
+ described_class.new(llm).prepare!(hard_prompt, subagent_execution_state: state),
+ ).to eq(:compressed)
+ expect_history_dropped(hard_prompt)
expect(canned.completions).to eq(0)
expect(state.reserve_root_final_completion).to eq(true)
end
@@ -417,14 +569,10 @@ RSpec.describe DiscourseAi::Completions::ContextPreparation do
child_preparation = described_class.new(llm, threshold: 20)
prompt = history_prompt(repetitions: 4000)
- original = prompt.messages.deep_dup
- expect { child_preparation.prepare!(prompt, execution_context: execution) }.to raise_error(
- described_class::Error,
- /preparation_budget_exhausted/,
- )
+ expect(child_preparation.prepare!(prompt, execution_context: execution)).to eq(:compressed)
expect(child_preparation.calls).to eq(8)
expect(child_preparation.spent_tokens).to eq(preparation.spent_tokens)
- expect(prompt.messages).to eq(original)
+ expect_history_dropped(prompt)
expect(canned.completions).to eq(1)
end
end
diff --git a/plugins/discourse-ai/spec/lib/modules/ai_bot/playground_spec.rb b/plugins/discourse-ai/spec/lib/modules/ai_bot/playground_spec.rb
index 38e59511447..6cb237258a7 100644
--- a/plugins/discourse-ai/spec/lib/modules/ai_bot/playground_spec.rb
+++ b/plugins/discourse-ai/spec/lib/modules/ai_bot/playground_spec.rb
@@ -175,6 +175,112 @@ RSpec.describe DiscourseAi::AiBot::Playground do
)
end
end
+
+ it "drops unsummarizable history, then resumes from that checkpoint on the next turn" do
+ first_post.update!(raw: "Remember QUARTZ-OTTER-731")
+ large_result = "Record: amber inventory checked. " * 4000
+ PostCustomPrompt.create!(
+ post_id: second_post.id,
+ custom_prompt: [
+ %w[{"arguments":{}} read tool_call read],
+ [large_result, "read", "tool", "read"],
+ ["DATA READ", bot_user.username],
+ ],
+ )
+ claude_2.update!(max_prompt_tokens: 16_000)
+ agent_record =
+ Fabricate(:ai_agent, max_turn_tokens: 4000, compression_threshold: 50, tools: [])
+ dropping_playground =
+ described_class.new(
+ DiscourseAi::Agents::Bot.as(
+ bot_user,
+ agent: agent_record.class_instance.new,
+ model: claude_2,
+ ),
+ )
+
+ DiscourseAi::Completions::Llm.with_prepared_responses(
+ ["", "Answer without the old read", "Follow-up answer"],
+ ) do |_, _, prompts, options|
+ reply = dropping_playground.reply_to(third_post, stream_reply: false, auto_set_title: false)
+
+ expect(reply.raw).to eq("Answer without the old read")
+ checkpoint = reply.post_custom_prompt.custom_prompt.first
+ expect(checkpoint[0]).to include(
+ DiscourseAi::Completions::ContextPreparation::HISTORY_DROPPED_NOTICE,
+ )
+ expect(checkpoint[6]).to include("source_id" => third_post.id)
+
+ followup = Fabricate(:post, topic: pm, user: user, raw: "Repeat the codeword")
+ dropping_playground.reply_to(followup, stream_reply: false, auto_set_title: false)
+
+ expect(options.map { |option| option[:feature_name] }).to eq(
+ %w[context_compression bot bot],
+ )
+ [prompts[1], prompts.last].each do |prompt|
+ contents = prompt.messages.map { |message| message[:content].to_s }
+ expect(contents).to include(checkpoint[0])
+ expect(contents.join).not_to include(large_result)
+ end
+ expect(prompts.last.messages.map { |message| message[:content] }).to include(followup.raw)
+ end
+ end
+
+ it "persists a partially retained earlier turn once across the next turn" do
+ large_result = "Record: amber inventory checked. " * 4000
+ PostCustomPrompt.create!(
+ post_id: second_post.id,
+ custom_prompt: [
+ %w[{"arguments":{}} read tool_call read],
+ [large_result, "read", "tool", "read"],
+ ["DATA READ", bot_user.username],
+ ],
+ )
+ third_post.update!(raw: "Also remember AMBER-FOX-42")
+ recent_reply = Fabricate(:post, topic: pm, user: bot_user, raw: "Noted")
+ PostCustomPrompt.create!(
+ post_id: recent_reply.id,
+ custom_prompt: [
+ %w[{"arguments":{}} recent tool_call read],
+ ["AMBER-FOX-42 stored", "recent", "tool", "read"],
+ ["Noted", bot_user.username],
+ ],
+ )
+ question = Fabricate(:post, topic: pm, user: user, raw: "Which codeword is stored?")
+ claude_2.update!(max_prompt_tokens: 16_000)
+ agent_record =
+ Fabricate(:ai_agent, max_turn_tokens: 4000, compression_threshold: 50, tools: [])
+ dropping_playground =
+ described_class.new(
+ DiscourseAi::Agents::Bot.as(
+ bot_user,
+ agent: agent_record.class_instance.new,
+ model: claude_2,
+ ),
+ )
+
+ DiscourseAi::Completions::Llm.with_prepared_responses(
+ ["", "AMBER-FOX-42", "Still AMBER-FOX-42"],
+ ) do |_, _, prompts, options|
+ dropping_playground.reply_to(question, stream_reply: false, auto_set_title: false)
+ followup = Fabricate(:post, topic: pm, user: user, raw: "Repeat it")
+ dropping_playground.reply_to(followup, stream_reply: false, auto_set_title: false)
+
+ expect(options.map { |option| option[:feature_name] }).to eq(
+ %w[context_compression bot bot],
+ )
+ [prompts[1], prompts.last].each do |prompt|
+ transcript = prompt.messages.to_s
+ expect(transcript).to include(
+ DiscourseAi::Completions::ContextPreparation::HISTORY_DROPPED_NOTICE,
+ )
+ expect(transcript.scan(third_post.raw).size).to eq(1)
+ expect(transcript.scan("AMBER-FOX-42 stored").size).to eq(1)
+ expect(transcript).not_to include(large_result)
+ end
+ expect(prompts.last.messages.to_s).to include(question.raw, followup.raw)
+ end
+ end
end
describe "#title_playground with a multipart model response" do
diff --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.rb
index b469196dacf..be5551e8221 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
@@ -1156,7 +1156,7 @@ RSpec.describe DiscourseAi::AiBot::StreamReplyCustomToolsSession do
end
end
- it "defaults work to the model context window" do
+ it "defaults work to DEFAULT_MAX_TURN_TOKENS" do
ai_agent.update!(max_turn_tokens: nil)
tool_call =
@@ -1181,7 +1181,7 @@ RSpec.describe DiscourseAi::AiBot::StreamReplyCustomToolsSession do
if (tracker = kwargs[:execution_context]&.token_usage_tracker)
tracker.add_effective(request: 90_000, response: 30_000)
kwargs[:execution_context].work_budget.debit(
- llm.max_prompt_tokens,
+ DiscourseAi::Agents::Bot::DEFAULT_MAX_TURN_TOKENS,
event_id: SecureRandom.uuid,
)
end