xonecas.io
64875
Mc

mini-coder / commit 38c8609

chore: TODO updates

xonecas2026-03-11 38c86093a91bdd44ca1f60c7d70d65d0a07c942b parent 35738e1

1 file changed

TODO.md+62−1
@@ -1,5 +1,56 @@
11 # TODO
22
3+## Worktree issues with subagents
4+
5+Subagents don't see the unstaged changes of the repo, see bellow:
6+
7+Agent implemented a change in the system prompt, then the user requested two reviews, one without arguments that should have defaulted to unstaged changes, but the reviewer looked at the reasoning commit.
8+
9+Then when prompted to review unstaged changes directly in another /review call, it clearly doesn't see any changes other than the node_modules folder.
10+
11+Output bellow:
12+
13+```
14+Recommended practice for GPT models (incl. codex) in our stack
15+
16+· ✅ Keep GPT/codex on Responses endpoint.
17+· ✅ Use one system-guidance channel per request:
18+ · either top-level `instructions` (preferred for Responses),
19+ · or developer/system message in `input/system`.
20+· ❌ Don’t send the same long prompt in both `instructions` and `system` simultaneously.
21+
22+So the change we made (`system` only when `!useInstructions`) aligns with recommended practice and should reduce prompt-shape confusion without affecting history.
23+gpt-5.3-codex ✦ medium 🤔 mmm8yarz @orc ~/src/mini-coder ⎇ feat/reasoning-display-control ctx 167.4k/400.0k 42% ↑ 6056.0k ↓ 13.1k
24+▶ /review
25+· review [~/.agents/commands/review.md]
26+
27+⠹ subagentNo actionable issues found in the recent changes.
28+
29+I checked correctness against the feature goal (reasoning display toggle + wiring), looked for regressions/duplication/dead patterns, and didn’t find any bugs or clear performance problems.
30+◆ Review summary: no actionable issues found.
31+
32+· Changes appear correct against the intended goal.
33+· No regressions identified.
34+· No obvious duplication/dead code patterns introduced.
35+· No clear performance concerns observed.
36+
37+So from this review pass, the update looks good to keep.
38+gpt-5.3-codex ✦ medium 🤔 mmm8yarz @orc ~/src/mini-coder ⎇ feat/reasoning-display-control ctx 167.7k/400.0k 42% ↑ 6223.7k ↓ 13.2k
39+▶ /review unstaged changes
40+· review [~/.agents/commands/review.md]
41+
42+· Repository hygiene issue: There is an untracked `node_modules` directory (`?? node_modules`). This should not be part of repo changes; remove it from the working tree (or ensure it’s ignored) before committing/reviewing further.
43+◆ Review found one actionable repo-hygiene issue:
44+
45+· There is an untracked `node_modules` directory (`?? node_modules`).
46+
47+This should not be included in repo changes. Remove it from the working tree (or ensure it’s ignored) before continuing with commit/review.
48+gpt-5.3-codex ✦ medium 🤔 mmm8yarz @orc ~/src/mini-coder ⎇ feat/reasoning-display-control ctx 167.8k/400.0k 42% ↑ 6391.5k ↓ 13.2k
49+▶
50+```
51+
52+---
53+
354 ## Bug with chat completions models (kimi 2.5 for example)
455
556 See bellow, there is the thinking spiner before each line, this doesn't happen with other models.
@@ -61,7 +112,15 @@ kimi-k2.5 ✦ high mmm7tm8p ~/src/mini-coder ⎇ main ctx 37.9k/262.1k 14%
61112
62113 ---
63114
64-## Gpt 5.4 starts printing tool calls as the assistant message:
115+## Gpt 5.4, Gpt 5.3 codex both starts printing tool calls as the assistant message:
116+
117+A fix was made:
118+
119+· Commit: `35738e1`
120+· Message: `fix(llm-api): avoid duplicate system prompt on responses models`
121+· Scope: `src/llm-api/turn.ts` only
122+
123+Currently testing with gpt models to see if the fix is effective.
65124
66125 ```
67126 ← read src/llm-api/turn.test.ts:108+50
@@ -97,6 +156,8 @@ error: script "dev" exited with code 130
97156
98157 I was not able to reproduce the same behaviour with opencode using the same provider. Which points to an issue with our code.
99158
159+Let's investigate.
160+
100161 ---
101162
102163 ## Gpt codex context window maxinng out