@@ -1,9 +1,16 @@
11 # TODO
22
3+## Write blog posts
4+
5+- Codex being big dumb and lazy without strong guidance in system prompt/instructions
6+- Keeping up codebase health when using agents to develop an applications. Avoid regressions, bad tests, lint etc.
7+
38 ## Help agent edits error less:
49
510 - 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.
611
12+---
13+
714 ## `/model` thinking-effort toggle
815
916 The idea specifies: *"/model allows the user to pick a model as well as thinking
@@ -14,15 +21,15 @@ so thinking effort is never forwarded to the SDK. The `settings` table (via
1421 `getSetting`/`setSetting` in `src/session/db.ts`) is already available for
1522 persistence alongside `preferred_model`.
1623
17-**Fix:**
18-- Add `thinkingEffort?: "low" | "medium" | "high"` to `runTurn()` options and pass
19- it as `providerOptions` — e.g. for Anthropic:
20- `{ anthropic: { thinking: { type: "enabled", budgetTokens: N } } }`.
21-- Expose it in `/model` (e.g. `/model --effort high`) and persist to `settings`
22- with key `"thinking_effort"`.
23-- Read it back in `agent.ts` (alongside `getPreferredModel()`) and thread it into
24- every `runTurn()` call including subagents.
25-- Gate display in `/model` listing to models that support it (Claude 3.5+ Sonnet /
26- Opus, o-series OpenAI models). Add a capability flag in `providers.ts`.
24+Research model capabilities and if we can fetch them automatically from providers (from the model list maybe?)
25+For the models that support reasoning, show the options when listing models with `/models`
2726
2827 ---
28+
29+## LSP Diagnostics
30+
31+We should have a closed loop feedback for LSP diagnostics on edits/reads.
32+
33+This could potentially be a very big slowdown, waiting for updated diagnostics every edit. This also has bias downsides, stale diagnostics confuse the LLM, and can cause negative distractions.
34+
35+We need to do research first, but maybe we can leverage the hooks feature to achieve a similar result without the performance penalties? Needs brainstorming