I think the emerging best way is to do "agentic search" over files. If you think about it, Claude Code is quite good at navigating large codebases and finding the required context for a problem.
Further, instead of polluting the context of your main agent, you can run a subagent to do search and retrieve the important bits of information and report back to your main agent. This is what Claude Code does if you use the keyword "explore". It starts a subagent with Haiku which reads ten of thousands of tokens in seconds.
From my experience the only shortcoming of this approach right now is that it's slow, and sometimes haiku misses some details in what it reads. These will get better very soon (in one or two generations, we will likely see opus 4.5 level intelligence at haiku speeds/price). For now, if not missing a detail is important for your usecase, you can give the output from the first subagent to a second one and ask the second one to find important details the first one missed. I've found this additional step to catch most things the first search missed. You can try this for yourself with Claude Code: ask it to create a plan for your spec, and then pass the plan to a second Claude Code session and ask it to find gaps and missing files from the plan.
Gemini 3 Flash is very good at the search task (it benchmarks quite close to 3 Pro in coding tasks but is much faster). I believe Amp switch to Gemini Flash for their search agent because it is better.
I very much doubt this. I've been using gemini whenever I get hit by codex limits over the past week. 3 pro is very good - a bit behind codex but still very useful. I've tried 3 flash several times and each time what I got back was complete garbage. After the third or fourth attempt I stopped trying.
I actually defended Gemini CLI when someone said this a few days ago. Then Murphy's law hit me with bug after bug, all of which have been reported many times already going back over year or more. I keep having to totally clear out the ~/.gemini folder and then it works again for a while.
Specifically for coding agents, one issue is how to continue work when almost fill the context window.
Compaction always loses information, so I use an alternative approach that works extremely well, based on this almost silly idea — your original session file itself is the golden source of truth with all details, so why not directly leverage it?
So I built the aichat feature in my Claude-code-tools repo with exactly this sort of thought; the aichat rollover option puts you in a fresh session, with the original session path injected, and you use sub agents to recover any arbitrary detail at any time. Now I keep auto-compact turned off and don’t compact ever.
It’s a relatively simple idea; no elaborate “memory” artifacts, no discipline or system to follow, work until 95%+ context usage.
The tool (with the related plugins) makes it seamless: first type “>resume” in your session (this copies session id to clipboard), then quit and run
aichat resume <pasted session id>
And this launched a TUI offering a few ways to resume your work, one of which is “rollover”; this puts you in a new session with the original session jsonl path injected.
And in the new session say something like,
“There is a chat session log file path shown to you; Use subagents strategically to extract details of the task we were working on at the end of it”, or use the /recover-context slash command. If it doesn’t quite get all of it, prompt it again for specific details.
There’s also an aichat search command for rust/tantivy based fast full text search to search across sessions, with a TUI for humans and a CLI/JSON mode for agents/subagents. The latter ( and the corresponding skill and sub agent) can be used to recover arbitrary detailed context about past work.
Every time you send a request to a model you're already providing all of the context history along with it. To edit the context, just send a different context history. You can send whatever you want as history, it's entirely up to you and entirely arbitrary.
We only think in conversational turns because that's what we've expected a conversation to 'look like'. But that's just a very deeply ingrained convention.
Forget that there is such a thing as 'turns' in a LLM convo for now, imagine that it's all 'one-shot'.
So you ask A, it responds A1.
But when you and B, and expect B1 - which depends on A and A1 already being in the convo history - consider that you are actually sending that again anyhow.
Behind the scenes when you think you're sending just 'B' (next prompt) you're actually sending A + A1 + B aka including the history.
A and A1 are usually 'cached' but that's not the simplest way to do it, the caching is an optimization.
Without caching the model would just process all of A + A1 + B and B1 in return just the same.
And then A + A1 + B + B1 + C and expect C1 in return.
It just so happens it will cache the state of the convo at your previous turn, and so it's optimized but the key insight is that you can send whatever context you want at any time.
If after you send A + A1 + B + B1 + C and get C1, if you want to then send A + B + C + D and expect D1 ... (basically sending the prompts with no responses) - you can totally do that. It will have to re-process all of that aka no cached state, but it will definitely do it for you.
Heck you can send Z + A + X, or A + A1 + X + Y - or whatever you want.
So in that sense - what you are really sending (if you're using the simplest form API), is sending 'a bunch of content' and 'expecting a response'. That's it. Everything is actually 'one shot' (prefill => response) and that's it. It feels conversational but structural and operational convention.
So the very simple answer to your question is: send whatever context you want. That's it.
If you know you will be pruning or otherwise reusing the context across multiple threads, the best place for context that will be retained is at the beginning due to prompt caching - it will reduce the cost and improve the speed.
If not, inserting new context any place other than at the end will cause cache misses and therefore slow down the response and increase cost.
Models also have some bias for tokens at start and end of the context window, so potentially there is a reason to put important instructions in one of those places.
I open 4 chat windows with Gemini 3.0 Pro. I paste in all file contents to each window. I ask them "which files would an AI need to do $TASK effectively?"
Each of the 4 responses will disagree, despite some overlap. I take the union of the 4 responses as the canonical set of files that an implementer would need to see.
This reduces the risk of missing key files, while increasing the risk of including marginally important files. An easy trade-off.
Then I paste the subset of files into GPT 5.2 Pro, and give it $TASK.
You could replace the upstream process with N codex sessions instead of N gemini chat windows. It doesn't matter.
This process can be automated with structured json outputs, but I haven't bothered yet.
It uses much inference compute. But it's better than missing key inputs and wasting time with hallucinated output.
That sounds cumbersome and even more wasteful than my own method of simply dumping a fixed selection of project code in Gemini for each set of requests. Is there any benefit to pruning?
What works best for me using Claude Code is to let the CC engineer its own context. You need to provide it with tools that it can use to engineer its context. CC comes with a lot of tools already (grep, sed, curl, etc), but for specific domain you may want to add more, e.g., access to a database, a cms, a parser for a bespoke language, etc.
With these i'll mostly just give it questions: what are some approaches to implement x, what are the pros and cons, what libraries are available to handle x? What data would you need to create x screen, or y report? And then let it google it, or run queries on your data.
I'll have it create markdown documents or skills to persist the insights it comes back with that will be useful in the future.
LLMs are pretty good at plan/do/check/act: create a plan (maybe to run a query to see what tables you have in your database), run the query, understand the output, and then determine the next step.
Your main goal should be to enable the PDCA loop of the LLM through tools you provide.
Tool calling + recursion seems to be the answer. Two tools are for manipulating the logical call stack - call/return. The trick is to not permit use of any meaningful tools at the root of recursion, but to always make their descriptions available. For instance, the root can't QueryWidgets or ExecuteShell, but any descendant of it can.
These constraints result in token-hungry activity being confined to child scopes that are fully isolated from their parents. The only way to communicate between stack frames is by way of the arguments to call() and return(). Theoretically, recursive dispatch gives us exponential scaling of effective context size as we descend into the call graph. It also helps to isolate bad trips and potentially learn from them.
We ran into this while building GTWY.ai. What worked for us wasn’t trying to keep a single model “continuously informed”, but breaking work into smaller steps with explicit context passed between them. Long-lived context drifted fast. Short-lived, well-scoped context stayed predictable.
Check out "cursor-mirror", this extended Anthropic Skill I've developed as a part of MOOLLM, which will tell you all about how cursor assembles its context:
See yourself think. Introspection tools for Cursor IDE — 47 read-only commands to inspect conversations, tool calls, context assembly, and agent reasoning from Cursor's internal SQLite databases.
By Don Hopkins, Leela AI — Part of MOOLLM
The Problem
LLM agents are black boxes. You prompt, they respond, you have no idea what happened inside. Context assembly? Opaque. Tool selection? Hidden. Reasoning? Buried in thinking blocks you can't access.
Cursor stores everything in SQLite. This tool opens those databases.
The Science
"You can't think about thinking without thinking about thinking about something." — Seymour Papert, Mindstorms: Children, Computers, and Powerful Ideas (Basic Books, 1980), p. 137
Papert's insight: metacognition requires concrete artifacts. Abstract introspection is empty. You need something to inspect.
This connects to three traditions:
Constructionism (Papert, 1980) — Learning happens through building inspectable artifacts. The Logo turtle wasn't about drawing; it was about making geometry visible so children could debug their mental models. cursor-mirror makes agent behavior visible so you can debug your mental model of how Cursor works.
Society of Mind (Minsky, 1986) — Intelligence emerges from interacting agents. Minsky's "K-lines" are activation patterns that recall mental states. cursor-mirror lets you see these patterns: which tools activated, what context was assembled, how the agent reasoned.
Schema Mechanism (Drescher, 1991) — Made-Up Minds describes how agents learn causal models through Context → Action → Result schemas. cursor-mirror provides the data for schema refinement: what context was assembled, what action was taken, what result occurred.
Here is the design and exploration and hacking session in which I iteratively designed and developed it, using MOOLLM's Constructionist "PLAY-LEARN-LIFT" methodology:
Currently only supports Cursor running on Mac, but I'd be happy to accept PRs for Linux and Windows support. Look at the cursor-chat-relection.md document to see how I had Cursor analyze its own directories, files, and sqlite databases and JSON schemas. Also looking for help developing mirrors and MOOLMM kernel drivers for other orchestrators like Claud Code, etc.
There is no such thing as continuous context. There is only context that you start and stop, which is the same as typing those words in the prompt. To make anything carry over to a second thread, it must be included in the second thread's context.
Rules are just context, too, and all elaborate AI control systems boil down to these contexts and tool calls.
In other words, you can rig it up anyway you like. Only the context in the actual thread (or "continuation," as it used to be called) is sent to the model, which has no memory or context outside that prompt.
> Furthermore, all of the major LLM APIs reward you for re-sending the same context with only appended data in the form of lower token costs (caching).
There's a little more flexibility than that. You can strip of some trailing context before appending some new context. This allows you to keep the 'long-term context' minimal, while still making good use of the cache.
i dont understand why these questions are so common? is it not obvious how one should use these capabilities? i compose my context in md file and send it through API. i wrote a simple lms.exe to send context and append response to the same file. why doesn't everyone else do that? i never believed in agents that compose their own context like Cursor. and i always pass the lowest reasoning value parameter to the API I can. why doesn't anyone else do this? you become dependent on a tool, you're already dependent some of you on fancy IDEs and agents. we're already dependent on top 3 vendors and openai is the only one that no one complains about from API key configuration side. you're gonna become dependent not only on LLMs but on the tooling as well? no thanks. anyone with a different opinion regarding this down to exact work flow, you are walking down the wrong path. you have to become efficient at converting electricity to text. admit it, some are just better than others while some will never get it at all. you know you won't because you know people in your life that never change their opinions about something, or always get into car accidents because they're a bad driver. you cant change these people and you might be one of them.
I think the emerging best way is to do "agentic search" over files. If you think about it, Claude Code is quite good at navigating large codebases and finding the required context for a problem.
Further, instead of polluting the context of your main agent, you can run a subagent to do search and retrieve the important bits of information and report back to your main agent. This is what Claude Code does if you use the keyword "explore". It starts a subagent with Haiku which reads ten of thousands of tokens in seconds.
From my experience the only shortcoming of this approach right now is that it's slow, and sometimes haiku misses some details in what it reads. These will get better very soon (in one or two generations, we will likely see opus 4.5 level intelligence at haiku speeds/price). For now, if not missing a detail is important for your usecase, you can give the output from the first subagent to a second one and ask the second one to find important details the first one missed. I've found this additional step to catch most things the first search missed. You can try this for yourself with Claude Code: ask it to create a plan for your spec, and then pass the plan to a second Claude Code session and ask it to find gaps and missing files from the plan.
Gemini 3 Flash is very good at the search task (it benchmarks quite close to 3 Pro in coding tasks but is much faster). I believe Amp switch to Gemini Flash for their search agent because it is better.
I very much doubt this. I've been using gemini whenever I get hit by codex limits over the past week. 3 pro is very good - a bit behind codex but still very useful. I've tried 3 flash several times and each time what I got back was complete garbage. After the third or fourth attempt I stopped trying.
Gemini the model is good, Gemini the framework around the model is a steaming pile of crap.
Yup. Gemini CLI needs a lot of work.
I actually defended Gemini CLI when someone said this a few days ago. Then Murphy's law hit me with bug after bug, all of which have been reported many times already going back over year or more. I keep having to totally clear out the ~/.gemini folder and then it works again for a while.
You can also create per-project agents with specific "expertise" in different parts of the code.
Basically they're just few kilobytes of text that's given as extra context to "explore" agents when looking at specific parts of the code.
Specifically for coding agents, one issue is how to continue work when almost fill the context window.
Compaction always loses information, so I use an alternative approach that works extremely well, based on this almost silly idea — your original session file itself is the golden source of truth with all details, so why not directly leverage it?
So I built the aichat feature in my Claude-code-tools repo with exactly this sort of thought; the aichat rollover option puts you in a fresh session, with the original session path injected, and you use sub agents to recover any arbitrary detail at any time. Now I keep auto-compact turned off and don’t compact ever.
https://github.com/pchalasani/claude-code-tools?tab=readme-o...
It’s a relatively simple idea; no elaborate “memory” artifacts, no discipline or system to follow, work until 95%+ context usage.
The tool (with the related plugins) makes it seamless: first type “>resume” in your session (this copies session id to clipboard), then quit and run
And this launched a TUI offering a few ways to resume your work, one of which is “rollover”; this puts you in a new session with the original session jsonl path injected. And in the new session say something like,“There is a chat session log file path shown to you; Use subagents strategically to extract details of the task we were working on at the end of it”, or use the /recover-context slash command. If it doesn’t quite get all of it, prompt it again for specific details.
There’s also an aichat search command for rust/tantivy based fast full text search to search across sessions, with a TUI for humans and a CLI/JSON mode for agents/subagents. The latter ( and the corresponding skill and sub agent) can be used to recover arbitrary detailed context about past work.
Every time you send a request to a model you're already providing all of the context history along with it. To edit the context, just send a different context history. You can send whatever you want as history, it's entirely up to you and entirely arbitrary.
We only think in conversational turns because that's what we've expected a conversation to 'look like'. But that's just a very deeply ingrained convention.
Forget that there is such a thing as 'turns' in a LLM convo for now, imagine that it's all 'one-shot'.
So you ask A, it responds A1.
But when you and B, and expect B1 - which depends on A and A1 already being in the convo history - consider that you are actually sending that again anyhow.
Behind the scenes when you think you're sending just 'B' (next prompt) you're actually sending A + A1 + B aka including the history.
A and A1 are usually 'cached' but that's not the simplest way to do it, the caching is an optimization.
Without caching the model would just process all of A + A1 + B and B1 in return just the same.
And then A + A1 + B + B1 + C and expect C1 in return.
It just so happens it will cache the state of the convo at your previous turn, and so it's optimized but the key insight is that you can send whatever context you want at any time.
If after you send A + A1 + B + B1 + C and get C1, if you want to then send A + B + C + D and expect D1 ... (basically sending the prompts with no responses) - you can totally do that. It will have to re-process all of that aka no cached state, but it will definitely do it for you.
Heck you can send Z + A + X, or A + A1 + X + Y - or whatever you want.
So in that sense - what you are really sending (if you're using the simplest form API), is sending 'a bunch of content' and 'expecting a response'. That's it. Everything is actually 'one shot' (prefill => response) and that's it. It feels conversational but structural and operational convention.
So the very simple answer to your question is: send whatever context you want. That's it.
Bigger context makes responses slower.
Context is limited.
You do not want the cloud provider running a context compaction if you can control it a lot better.
There are even tips on when to ask the question like "send first the content then ask the question" vs. "ask the question then send the content"
This is how I view it as well.
And... and...
This results in a _very_ deep implication, which big companies may not be eager to let you see:
they are context processors
Take it for what it is.
What you are trying to say is they are plagiarists and training on the input?
We know that already I don’t know why have to be quiet or hint at it, in fact they have been quite explicit about it.
Or is there some other context to your statement? Anyway that’s my “take that for what you will”.
If you know you will be pruning or otherwise reusing the context across multiple threads, the best place for context that will be retained is at the beginning due to prompt caching - it will reduce the cost and improve the speed.
If not, inserting new context any place other than at the end will cause cache misses and therefore slow down the response and increase cost.
Models also have some bias for tokens at start and end of the context window, so potentially there is a reason to put important instructions in one of those places.
I wonder how far you can take that. Basically can you jam a bunch of garbage in the middle and still get useful results
I open 4 chat windows with Gemini 3.0 Pro. I paste in all file contents to each window. I ask them "which files would an AI need to do $TASK effectively?"
Each of the 4 responses will disagree, despite some overlap. I take the union of the 4 responses as the canonical set of files that an implementer would need to see.
This reduces the risk of missing key files, while increasing the risk of including marginally important files. An easy trade-off.
Then I paste the subset of files into GPT 5.2 Pro, and give it $TASK.
You could replace the upstream process with N codex sessions instead of N gemini chat windows. It doesn't matter.
This process can be automated with structured json outputs, but I haven't bothered yet.
It uses much inference compute. But it's better than missing key inputs and wasting time with hallucinated output.
That sounds cumbersome and even more wasteful than my own method of simply dumping a fixed selection of project code in Gemini for each set of requests. Is there any benefit to pruning?
> Is there any benefit to pruning?
1- Better quality output due to pruning noise, while reducing the chances of missing key context.
2- Saving time/effort by not using my brain to decide which files to include.
3- ChatGPT 5.2 Pro only allows 60k tokens, so I have no choice sometimes.
It comes with costs as you identified. It's a trade-off that I am willing to pay.
What works best for me using Claude Code is to let the CC engineer its own context. You need to provide it with tools that it can use to engineer its context. CC comes with a lot of tools already (grep, sed, curl, etc), but for specific domain you may want to add more, e.g., access to a database, a cms, a parser for a bespoke language, etc.
With these i'll mostly just give it questions: what are some approaches to implement x, what are the pros and cons, what libraries are available to handle x? What data would you need to create x screen, or y report? And then let it google it, or run queries on your data.
I'll have it create markdown documents or skills to persist the insights it comes back with that will be useful in the future.
LLMs are pretty good at plan/do/check/act: create a plan (maybe to run a query to see what tables you have in your database), run the query, understand the output, and then determine the next step.
Your main goal should be to enable the PDCA loop of the LLM through tools you provide.
Tool calling + recursion seems to be the answer. Two tools are for manipulating the logical call stack - call/return. The trick is to not permit use of any meaningful tools at the root of recursion, but to always make their descriptions available. For instance, the root can't QueryWidgets or ExecuteShell, but any descendant of it can.
These constraints result in token-hungry activity being confined to child scopes that are fully isolated from their parents. The only way to communicate between stack frames is by way of the arguments to call() and return(). Theoretically, recursive dispatch gives us exponential scaling of effective context size as we descend into the call graph. It also helps to isolate bad trips and potentially learn from them.
This blog explains context folding ie tool calling and recursion - https://www.primeintellect.ai/blog/rlm
We ran into this while building GTWY.ai. What worked for us wasn’t trying to keep a single model “continuously informed”, but breaking work into smaller steps with explicit context passed between them. Long-lived context drifted fast. Short-lived, well-scoped context stayed predictable.
I've been building https://www.usesatori.sh/ to give persistent context to agents
Would be happy to onboard you personally.
I'm not OP, but send me an email. My address is in my HN profile. You and I are building the same thing, and I would love to have a chat.
What would you cover not-continuous?
Best methods I’ve observed -progressive loading (claude skills) & symbolic search (serena mcp)
> what according to you is the best way to provide context to a model.
Are you talking about manually or in an automated fashion?
Automated fashion would be what I'm curious on.
Check out "cursor-mirror", this extended Anthropic Skill I've developed as a part of MOOLLM, which will tell you all about how cursor assembles its context:
cursor-mirror skill: https://github.com/SimHacker/moollm/tree/main/skills/cursor-...
cursor-mirror
See yourself think. Introspection tools for Cursor IDE — 47 read-only commands to inspect conversations, tool calls, context assembly, and agent reasoning from Cursor's internal SQLite databases.
By Don Hopkins, Leela AI — Part of MOOLLM
The Problem
LLM agents are black boxes. You prompt, they respond, you have no idea what happened inside. Context assembly? Opaque. Tool selection? Hidden. Reasoning? Buried in thinking blocks you can't access.
Cursor stores everything in SQLite. This tool opens those databases.
The Science
"You can't think about thinking without thinking about thinking about something." — Seymour Papert, Mindstorms: Children, Computers, and Powerful Ideas (Basic Books, 1980), p. 137
Papert's insight: metacognition requires concrete artifacts. Abstract introspection is empty. You need something to inspect.
This connects to three traditions:
Constructionism (Papert, 1980) — Learning happens through building inspectable artifacts. The Logo turtle wasn't about drawing; it was about making geometry visible so children could debug their mental models. cursor-mirror makes agent behavior visible so you can debug your mental model of how Cursor works.
Society of Mind (Minsky, 1986) — Intelligence emerges from interacting agents. Minsky's "K-lines" are activation patterns that recall mental states. cursor-mirror lets you see these patterns: which tools activated, what context was assembled, how the agent reasoned.
Schema Mechanism (Drescher, 1991) — Made-Up Minds describes how agents learn causal models through Context → Action → Result schemas. cursor-mirror provides the data for schema refinement: what context was assembled, what action was taken, what result occurred.
What You Can Inspect:
Conversation Structure
Context Assembly
Tool Execution
Server Configuration
MCP Servers
Image Archaeology
Python Sister Script CLI Tool: cursor_mirror.py
cursor_mirror.py: https://github.com/SimHacker/moollm/blob/main/skills/cursor-...
Here is the design and exploration and hacking session in which I iteratively designed and developed it, using MOOLLM's Constructionist "PLAY-LEARN-LIFT" methodology:
cursor-chat-reflection.md: https://github.com/SimHacker/moollm/blob/main/examples/adven...
Look at the "Scene 19 — Context Assembly Deep Dive" section and messageRequestContext schema, and "Scene 23 — Orchestration Deep Dive" section!
PR-CURSOR-MIRROR-GENESIS.md: https://github.com/SimHacker/moollm/blob/main/designs/PR-CUR...
play-learn-lift skill: https://github.com/SimHacker/moollm/tree/main/skills/play-le...
MOOLLM Anthropic compatible extended meta skill skill: https://github.com/SimHacker/moollm/tree/main/skills/skill
Specifically you can check out ORCHESTRATION.yml and other "YAML Jazz" metadata in the directory:
ORCHESTRATION.yml: https://github.com/SimHacker/moollm/blob/main/skills/cursor-...
Currently only supports Cursor running on Mac, but I'd be happy to accept PRs for Linux and Windows support. Look at the cursor-chat-relection.md document to see how I had Cursor analyze its own directories, files, and sqlite databases and JSON schemas. Also looking for help developing mirrors and MOOLMM kernel drivers for other orchestrators like Claud Code, etc.
DATA-SCHEMAS.yml: https://github.com/SimHacker/moollm/blob/main/skills/cursor-...
It's called continuous learning. You can't do it with an LLM service but you can if in training mode with bigger hardware.
Cursor and AI coding doesn't do it. It uses agentic subtasks.
There is no such thing as continuous context. There is only context that you start and stop, which is the same as typing those words in the prompt. To make anything carry over to a second thread, it must be included in the second thread's context.
Rules are just context, too, and all elaborate AI control systems boil down to these contexts and tool calls.
In other words, you can rig it up anyway you like. Only the context in the actual thread (or "continuation," as it used to be called) is sent to the model, which has no memory or context outside that prompt.
Furthermore, all of the major LLM APIs reward you for re-sending the same context with only appended data in the form of lower token costs (caching).
There may be a day when we retroactively edit context, but the system in it's current state is not very supportive of that.
> Furthermore, all of the major LLM APIs reward you for re-sending the same context with only appended data in the form of lower token costs (caching).
There's a little more flexibility than that. You can strip of some trailing context before appending some new context. This allows you to keep the 'long-term context' minimal, while still making good use of the cache.
i dont understand why these questions are so common? is it not obvious how one should use these capabilities? i compose my context in md file and send it through API. i wrote a simple lms.exe to send context and append response to the same file. why doesn't everyone else do that? i never believed in agents that compose their own context like Cursor. and i always pass the lowest reasoning value parameter to the API I can. why doesn't anyone else do this? you become dependent on a tool, you're already dependent some of you on fancy IDEs and agents. we're already dependent on top 3 vendors and openai is the only one that no one complains about from API key configuration side. you're gonna become dependent not only on LLMs but on the tooling as well? no thanks. anyone with a different opinion regarding this down to exact work flow, you are walking down the wrong path. you have to become efficient at converting electricity to text. admit it, some are just better than others while some will never get it at all. you know you won't because you know people in your life that never change their opinions about something, or always get into car accidents because they're a bad driver. you cant change these people and you might be one of them.