xonecas.io
64875
Mc

mini-coder / commit f5e7538

chore: todo updates

Sean2026-03-01 f5e7538d9bbc970b205f79b2bb1c0fd1c631fc42 parent 346d9cf

1 file changed

TODO.md+15−53
@@ -1,68 +1,30 @@
11 # TODO
22
3-## Help agent edits error less:
4-
5-- Agents sometimes include `|` at the end of the hash, this is because how we display the hashed to them, let's check if the anchor last char is `|` and strip it if it instead of letting it error.
6-
7-## Implement WebContent and WebSearch using exa.ai api
8-
9-Why not just keep using them via their MCP? Simple answer: their tools descriptions/schema is bloated with params that don't concern the llm. They are also misleading in their naming conventions, futher worsening the issue. Let's use their api (let's auto discover the api key like we do for other providers "EXA_API_KEY") and if Exa is enabled, the tools are given to the LLM.
10-
11-https://exa.ai/docs/reference/search-quickstart
12-https://exa.ai/docs/reference/contents-quickstart
13-**Ask the user for text versions if you can't find these pages**
14-
15-## Agent get's stuck on thinking
16-
17-- When using Google models, there is a rare occurence where the LLM seems to be stuck on thinking. There is no output for several minutes, and the only solution is to use CTRL+c to interrupt.
18-- This is not a infinite reasoning loop, there are requests being made, the app is "stuck", or "hanguing".
19-- Setting a thinkingConfig is also not a fix, 8000 is the default.
3+## Handle reaching max tool calls better:
204
21-## Fucking shit ctrl+c still doesn't work how the fuck is this possible:
5+- Force the LLM to respond with a status update if tool calls are exhausted.
6+- On the last toolcall before max append a `<system-message>` block telling the llm to give a status update and list
7+what is left todo.
228
23-Agent got stuck on thinking after calling the shell tool, using ctrl+c exitted the app... This is un-acceptable, see the git history, ctrl+c has been broken and there have been several claims to fixing it. The truth is that it doesn't work at all. I'm fucking done with this bullshit, either it works this time, or we are moving interrupt to a different key. This is fucking stupid...
9+## DB error after thinking for a long time?
2410
2511 ```
26-◆ I have updated `src/tools/exa.ts` to address both issues:
27-
28-1. Context Window Vulnerabilities: Added a constraint to the URLs array length in the `ExaContentSchema` by chaining `.max(3)` so the agent won't fetch more than 3 URLs at a time. Also, changed `text: true` to `text: { maxCharacters: 10000 }` to avoid returning overly large payloads.
29-2. Swallowed Error Details: Extracted the error body (`await response.text()`) and included it in the thrown `Error` for both `webSearchTool` and `webContentTool`, providing detailed visibility into API failures.
30-
31-All code checks and tests passed successfully.
32-gemini-3.1-pro mm7kimxq ~/src/mini-coder ⎇ main ctx 46.4k/1000.0k 5% ↑1033.7k ↓8.6k
33-▶ ok make a commit
34- $ $ git add src/tools/exa.ts src/agent/tools.ts && git commit -m "feat: i…
35- ✔ 0
36- │ [main a30a472] feat: implement native Exa web search and content tools
37- │ 2 files changed, 102 insertions(+), 2 deletions(-)
38- │ create mode 100644 src/tools/exa.ts
39- $ $ git push
40- ✔ 0
41- │ To github.com:sacenox/mini-coder.git
42- │ 9348b39..a30a472 main -> main
43-error: script "dev" exited with code 130
12+ ← read src/llm-api/turn.ts:1+100
13+ · src/llm-api/turn.ts lines 1–100 of 189 (truncated)
14+✖ FOREIGN KEY constraint failed
4415 ```
4516
46-## Subagents don't have access to MCP tools?
47-
48-Investigate if subagents are able to see the mcp tools. The user reported that after adding the exa mcp tools, that the subagents claimed to not be able to search online, but the main agent could.
17+Please investigate and report back with evidence of the root cause.
4918
50-## Investigate missing output?
51-
52-```
53-Summary of the design:
54-1. Skills (`~/.agents/skills/*/SKILL.md` + `.agents/skills/*/SKILL.md`): `@skill-name` in prompt → loads skill content into context. Tab-autocomplete from both global and local skill dirs.
55-2. Custom Agents (`~/.agents/agents/*.md` + `.agents/agents/*.md`): `@agent-name` in prompt → runs that agent as a subagent with its custom system prompt/model. YAML frontmatter with `description`, `model`.
56-^C3. Tab completion in input.ts should show `@skill-name` and `@agent-name` suggestions alongside file completions.
19+## Help agent edits error less:
5720
58-[BIG WAIT HERE, nothing happening for 10m+, user used CTRL+c to return to prompt]
21+- Agents sometimes include `|` at the end of the hash, this is because how we display the hashed to them, let's check if the anchor last char is `|` and strip it if it instead of letting it error.
5922
60-claude-sonnet-4-6 mm1v6tyb ~/src/mini-coder ⎇ main ctx 55.9k/200.0k 28% ↑182.1k ↓949
61-▶ What happened?
62-◆ I'm mid-way through exploring the codebase to implement your `@` references feature for skills and custom agents. Here's what I've gathered so far:
63-```
23+## Agent get's stuck on thinking
6424
65-LLM said it was exploring but no output or tool calls were seen. Session continued normally after the "What happened?" message. (mm1v6tyb session id)
25+- When using Google models, there is a rare occurence where the LLM seems to be stuck on thinking. There is no output for several minutes, and the only solution is to use CTRL+c to interrupt.
26+- This is not a infinite reasoning loop, there are requests being made, the app is "stuck", or "hanguing".
27+- Setting a thinkingConfig is also not a fix, 8000 is the default.
6628
6729 ## `/model` thinking-effort toggle