Gemini CLI gets into "failed replace" loops and eats tokens like crazy

[708-145]

More often than not, it does a few edits in the right direction and at some point repeatedly fails to replace the code.
When that happens it easily eats 10-20M tokens before I hit a consumption limit on the free tier so it's Google's money that's wasted.
I'm sure the Google team's AI assistant sees these cases in their logs as well but maybe it's worth mentioning here for awareness as well.

[708-145]

Just found the following. Seems the developers are already on it, just not working too well, yet.

Here's the corrected replace operation:
 │ x  replace {"old_string":"    ggml_cgraph * build_graph(const llm_graph_params & params) const;\n\nbool llama_model::load_tenso…  │
 │                                                                                                                                   │
 │    Failed to edit, 0 occurrences found for old_string in                                                                          │
 │    /mnt/c/Users/tobia/Desktop/tobifun/llama.cpp/src/llama-model.cpp. No edits made. The exact text in old_string was              │
 │    not found. Ensure you're not escaping content incorrectly and check whitespace, indentation, and context. Use                  │
 │    read_file tool to verify.  

│  Model Usage                  Reqs   Input Tokens  Output Tokens                                    │
│  ───────────────────────────────────────────────────────────────                                    │
│  gemini-2.5-flash               71     21,327,944         10,850  

[acoliver]

I'd love to see a test case of this. The edit tool has worked pretty well for me, but I haven't tried a lot of C. Could you supply a reference file or link and a prompt?

[yashasvigirdhar]

+1, I've seen it too many times when it just goes into the loop of failed to replace content. Then it also tries to read and rerwrite the entire file but fails there as well sometimes.

[acoliver]

Like I said I'd love to see a test case. I might be able to tweak the edit tool.

[Manamama]

Weeks ago I spent some hours on A/B/B1/B2/C testing. A summary of a summary (advice for Gemini ♊ Cloud AI about Gemini CLI) is here: https://github.com/Manamama/Puzzles_for_AIs/blob/main/docs/guides/gemini_editing_tools_summary.md - I use it to ground ♊ now and then in the cases as such, or to have ♊ switch the tools to avoid ➿.

[708-145]

Seems the current version found a way around:

I understand the replace tool is being problematic. I will use sed to replace the entire llama_model_load function in
  src/llama.cpp with the clean version.I understand the replace tool is not working as expected. I will use sed to replace the
  entire llama_model_load function in src/llama.cpp with a clean version that does not contain any of the problematic checksum
  verification logic. This should ensure all remnants are removed.
 
 │ ✔  Shell sed -i '/static int llama_model_load(const std::string & fname, std::vector<std::string> & splits, llama_model & model…  │
 
✦ The sed command executed successfully. This should have removed all the problematic checksum verification code from
  src/llama.cpp.

I'll provide a failing example in a few minutes.

[acoliver]

If you used qwen3 it likes sed alot.

[708-145]

Starting from commit 7a152e002725d93c1ebaedce9ad15099c0e8201f in branch https://github.com/708-145/llama.cpp/tree/dev
For debug I have a small local seq_Qwen3-0.6B.gguf prepared.

Gemini CLI prompt:

When called with ./llama-cli -m seq_Qwen3-0.6B.gguf -p "hi" -n 10 --no-mmap
I see checksum_key_id is always -1. 
in common.cpp :	LOG_INF("checksum key '%s', checksum key id %" PRId64 ".\n", checksum_key.c_str(), checksum_key_id);
Please fix the access to the model context: KV data and tensor data are needed for this check.
params.check_tensors is something else entirely and not to be used for this check.
compile with cmake --build build --config Release -j8 --target llama-cli

It got into a loop and eventually gave up. In the end I fixed it myself.

[acoliver]

I'll take a look

[acoliver]

This was 2.5 pro? It matters because the model makes diff tools choices.

[708-145]

It happened with pro and flash. I think the replace tool usage is or was broken.
When I prompt the exact file and line it's better

[708-145]

Just saw a lot of this again. I reads the file again and again.
I think what's missing is a re_read_file operation that removes all previous occurrences in the context.

Failed to edit, 0 occurrences found for old_string in                                                                          │
 │    /mnt/c/Users/tobia/Desktop/tobifun/llama.cpp/src/llama-quant.cpp. No edits made. The exact text in old_string was              │
 │    not found. Ensure you're not escaping content incorrectly and check whitespace, indentation, and context. Use                  │
 │    read_file tool to verify.

[708-145]

During the loops it computes the same thing over and over so at least the caching works. 👍

│  Model Usage                  Reqs   Input Tokens  Output Tokens                                                                             │
│  ───────────────────────────────────────────────────────────────                                                                             │
│  gemini-2.5-pro                 81     12,664,615         36,005                                                                             │
│  gemini-2.5-flash              115     36,239,022         49,641                                                                             │
│                                                                                                                                              │
│  Savings Highlight: 31,879,285 (65.2%) of input tokens were served from the cache, reducing costs.

[j0826]

This is still not fixed??? Wow.