I noticed a current-date/context-binding issue in Kagi Assistant.
In a GLM-4.7 reasoning trace, the model recognized a system/runtime timestamp of 2026-05-06, but then stated:
“The system context provides a timestamp: 2026-05-06. This is a date in the future relative to the current date.”
If 2026-05-06 is the current runtime date, it should not be treated as future. This kind of reasoning trace is also commonly hidden in collapsed thinking sections where people may not see it.
I also saw Research Assistant generate a stale-year query during a current-context task:
Kagi bug bounty program security vulnerability disclosure 2025
I attached a screenshot highlighting the stale-year Research query.
Steps to reproduce:
- Open Kagi Assistant.
- Use GLM-4.7 reasoning or Research Assistant.
- Ask a current/time-sensitive query.
- Inspect the reasoning/research trace and generated searches.
- Check whether the model consistently uses the actual current date/year.
This matters because current-date binding affects prices, subscriptions, software versions, laws, model availability, recent events, and “latest/current/still” queries.
Full GLM-4.7 trace:
@agent::decide-agree? #AI~printing-press
GLM-4.7 (reasoning)
Thinking
The user has invoked a specific agent command: @agent::decide-agree?. This indicates they want me to act as a decision-making or evaluative agent regarding a proposition.
The tag #AI~printing-press likely specifies the topic or domain of the proposition. It suggests a hypothetical scenario or an analogy between AI and the printing press.
The user has not provided a specific statement to agree or disagree with, just the agent invocation and context.
The system context provides a timestamp: 2026-05-06. This is a date in the future relative to the current date.
Interpretation of the Request:
The user is likely asking for an evaluation or decision on a topic related to "AI and the printing press" (perhaps comparing their impact, regulation, or dissemination of information) within the context of the year 2026.
However, there is no explicit question or proposition to "decide-agree" upon.
Usually, in such a framework, if I am asked to "decide-agree", I need a proposition.
Possibility A: The user forgot to paste the proposition.
Possibility B: The proposition is implicit in the tag (e.g., "Is AI the new printing press?" or "Should AI be regulated like the printing press?").
Possibility C: The user is testing the agent syntax to see how it responds to missing input.
Given the specific tag #AI~printing-press, a common debate is whether AI will have a democratizing effect like the printing press or a disruptive one, or perhaps regarding information integrity. Since I must respond in the same language as the query (which is English, though mostly system syntax), I will assume English.
Hypothesis regarding the Proposition:
The most implicit proposition associated with "AI printing press" analogies is usually something like: "Artificial Intelligence will have a democratizing impact on information dissemination similar to the printing press," or conversely, "AI poses a risk of mass misinformation unlike the printing press."
However, without the proposition, I cannot strictly "decide-agree".
Strategy: //removed empty new line below
Acknowledge the agent invocation.
Note the context (2026).
Politely ask for the specific proposition or statement that requires a decision/agreement.
Alternatively, provide a general analysis of the AI vs. Printing Press analogy in 2026 to be helpful, but keep it brief as the instruction says "Be concise". Actually, since the instruction is "decide-agree", I should probably treat the prompt as incomplete or ask for clarification.
Wait, looking at the prompt structure again: @agent::decide-agree? #AI~printing-press. This looks like a command. Perhaps the "proposition" is the tag itself? No, that's unlikely.
Let's look at the system instructions. "Be concise in your replies." "Always respond in the same language..." "USE PRIMARY INTERFACE LANGUAGE..."
If I simply return an error or "Please provide input", I might be failing the "persona" or the specific agent task.
Let's assume the user wants me to simulate the decision process for a hypothetical proposition related to that tag. But that's guessing.
Let's ask for the proposition. It is the safest and most logical action for a "decide-agree" agent.
Refined Strategy:
I will signal that I am ready to act as the decide-agree agent regarding the #AI~printing-press topic but I need the specific statement to evaluate.
Drafting the response:
"To provide a decision or agreement evaluation on the topic of #AI~printing-press, I require the specific proposition or statement you would like me to assess. Please provide the text of the proposition."
Self-Correction:
The prompt is @agent::decide-agree?. This is a question syntax. Maybe it's asking "Do I agree that... [missing text]?".
_**Let's look at the provided timestamp: 2026. This suggests a futuristic context.**_
Maybe the user made a mistake and this is a test case. I will treat it as a request for the missing input.
Wait, let's look closer at the prompt format.
It starts with @agent::. This is a command to the system.
Then decide-agree?.
Then #AI~printing-press.
This looks like a query to a database of opinions or a specific logic model where the topic is the argument.
If I must force a "decision", based on common current discourse (extrapolated to 2026), the consensus is usually that AI is a transformative technology comparable to the printing press, often leading to "Agree" on the significance/disruption scale.
But "decide-agree" is binary-ish.
Let's try to be helpful. I'll ask for the proposition.
Alternative interpretation:
Is #AI~printing-press the proposition? "AI [is like/a] printing-press".
If so, the answer is usually "Agree" in terms of impact scale, but "Disagree" in terms of mechanics.
Without the text, I can't decide.
I will ask for the proposition clearly.
Constraint Check:
Response language: English.
Format: Standard markdown.
Concise.
Final Plan:
State that I am the decision agent.
State that the proposition is missing.
Ask the user to provide it.
I am ready to act as the decide-agree agent for the topic #AIprinting-press.
However, you have not provided a specific proposition or statement for me to evaluate. Please provide the text of the proposition you would like me to decide on.
I expected Kagi Assistant to consistently bind to the current runtime date across final answers, reasoning traces, Research Assistant traces, generated search queries, and source selection.
If the runtime/system context says the current date is 2026-05-06, the model should treat 2026-05-06 as current, not future.
Research Assistant should avoid stale-year search anchors, such as 2025, unless the user explicitly asks about that year or the search context requires it.
Desired behavior:
- The current runtime date should be treated as authoritative.
- Reasoning traces should not contradict the injected runtime date.
- Research Assistant should generate current-year/current-date searches for current-context tasks.
- If the model is uncertain about the current date, it should say so rather than silently anchoring to the wrong temporal frame.
- Kagi may want to add eval tests for current-date binding, temporal anchoring, and consistency between runtime date, reasoning traces, generated searches, and final answers.
