Skip to content

refactor(config,docs): inline max_tokens default and note task-budgets beta #32

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
hakula139 merged 2 commits into from
Apr 24, 2026
Merged

refactor(config,docs): inline max_tokens default and note task-budgets beta #32

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3b469d9
refactor(config): inline DEFAULT_MAX_TOKENS as 16_000
hakula139 Apr 24, 2026
b1ca79a
docs(research): note task-budgets-2026-03-13 beta in anthropic-api
hakula139 Apr 24, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 8 additions & 13 deletions crates/oxide-code/src/config.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,7 +18,6 @@ use crate::util::env;

const DEFAULT_MODEL: &str = "claude-opus-4-7";
const DEFAULT_BASE_URL: &str = "https://api.anthropic.com";
const DEFAULT_MAX_TOKENS: u32 = 16384;

#[derive(Debug, Clone)]
pub enum Auth {
Expand Down Expand Up @@ -243,15 +242,13 @@ impl Config {
}
}

/// Per-effort `max_tokens` default. Matches claude-code 2.1.119's
/// observed values: 64 K for the top two tiers (xhigh / max), 32 K
/// for high, the legacy 16 384 for everything else. Users override
/// via `ANTHROPIC_MAX_TOKENS` / `[client].max_tokens`.
/// Per-effort `max_tokens` default; overridden by
/// `ANTHROPIC_MAX_TOKENS` / `[client].max_tokens`.
fn default_max_tokens(effort: Option<Effort>) -> u32 {
match effort {
Some(Effort::Xhigh | Effort::Max) => 64_000,
Some(Effort::High) => 32_000,
_ => DEFAULT_MAX_TOKENS,
_ => 16_000,
}
}

Expand Down Expand Up @@ -416,10 +413,8 @@ mod tests {

#[tokio::test]
async fn load_defaults_apply_when_no_config_and_no_env() {
// Default model (Opus 4.7) supports `xhigh`, so both `effort`
// and `max_tokens` derive from that ceiling — matches the
// claude-code 2.1.119 packet capture. Prompt cache defaults
// to 1h (opt-out via `OX_PROMPT_CACHE_TTL=5m`).
// Opus 4.7 supports `xhigh`, so both `effort` and `max_tokens`
// derive from that ceiling. Prompt cache defaults to 1h.
let dir = tempfile::tempdir().unwrap();
let config = temp_env::async_with_vars(env_vars(vec![xdg(&dir)]), Config::load())
.await
Expand Down Expand Up @@ -713,9 +708,9 @@ mod tests {
assert_eq!(default_max_tokens(Some(Effort::Max)), 64_000);
assert_eq!(default_max_tokens(Some(Effort::Xhigh)), 64_000);
assert_eq!(default_max_tokens(Some(Effort::High)), 32_000);
assert_eq!(default_max_tokens(Some(Effort::Medium)), DEFAULT_MAX_TOKENS);
assert_eq!(default_max_tokens(Some(Effort::Low)), DEFAULT_MAX_TOKENS);
assert_eq!(default_max_tokens(None), DEFAULT_MAX_TOKENS);
assert_eq!(default_max_tokens(Some(Effort::Medium)), 16_000);
assert_eq!(default_max_tokens(Some(Effort::Low)), 16_000);
assert_eq!(default_max_tokens(None), 16_000);
}

// ── Config::load / prompt_cache_ttl ──
Expand Down
2 changes: 1 addition & 1 deletion docs/guide/configuration.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,7 +61,7 @@ Tier guide (from the [Opus 4.7 migration guide](https://platform.claude.com/docs

#### `max_tokens` — response ceiling

When unset, oxide-code derives `max_tokens` from the resolved `effort` to match the claude-code reference: 64 000 for `xhigh` / `max`, 32 000 for `high`, 16 384 otherwise. Setting `max_tokens` explicitly (via TOML or `ANTHROPIC_MAX_TOKENS`) overrides the derivation.
When unset, oxide-code derives `max_tokens` from the resolved `effort`: 64 000 for `xhigh` / `max`, 32 000 for `high`, 16 000 otherwise. Setting `max_tokens` explicitly (via TOML or `ANTHROPIC_MAX_TOKENS`) overrides the derivation.

#### `prompt_cache_ttl` — cache duration

Expand Down
21 changes: 11 additions & 10 deletions docs/research/anthropic-api.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,15 +49,16 @@ anthropic-beta: claude-code-20250219,oauth-2025-04-20

Additional useful betas:

| Header | Purpose |
| --------------------------------- | -------------------------------------------------- |
| `interleaved-thinking-2025-05-14` | Extended thinking support |
| `context-1m-2025-08-07` | 1M context window |
| `context-management-2025-06-27` | Context management |
| `prompt-caching-scope-2026-01-05` | Prompt caching |
| `effort-2025-11-24` | Effort control |
| `structured-outputs-2025-12-15` | JSON-schema-constrained responses (one-shot calls) |
| `advanced-tool-use-2025-11-20` | Tool search (first-party only) |
| Header | Purpose |
| --------------------------------- | ------------------------------------------------------------ |
| `interleaved-thinking-2025-05-14` | Extended thinking support |
| `context-1m-2025-08-07` | 1M context window |
| `context-management-2025-06-27` | Context management |
| `prompt-caching-scope-2026-01-05` | Prompt caching |
| `effort-2025-11-24` | Effort control |
| `structured-outputs-2025-12-15` | JSON-schema-constrained responses (one-shot calls) |
| `advanced-tool-use-2025-11-20` | Tool search (first-party only) |
| `task-budgets-2026-03-13` | Advisory token budget across an agentic loop (Opus 4.7 only) |

#### Per-model beta sets

Expand Down Expand Up @@ -232,7 +233,7 @@ GA as of Opus 4.6. Controls the intelligence-vs-latency tier of agentic turns vi
- **The `effort-2025-11-24` beta header is necessary but not sufficient.** oxide-code used to send the header without the body field; the header became a no-op and the model ran at an undefined default.
- **Per-model ceiling.** `max` is Opus-only; Sonnet 4.6 400s on it. `xhigh` is Opus 4.7-only. The `Capabilities::effort_max` / `effort_xhigh` flags encode this; `Capabilities::clamp_effort` clamps a user pick down to the highest supported level at or below it.
- **Per-model default.** claude-code 2.1.119 sends `xhigh` on Opus 4.7, `high` on Opus 4.6 and Sonnet 4.6, omits the field entirely on earlier models. oxide-code mirrors this via `Capabilities::default_effort`.
- **`max_tokens` should scale with effort.** claude-code uses 64 K on Opus 4.7 at `xhigh`, 32 K on Sonnet 4.6 at `high`. oxide-code's `default_max_tokens(effort)` applies the same scaling when the user hasn't set `ANTHROPIC_MAX_TOKENS` explicitly.
- **`max_tokens` should scale with effort.** claude-code uses 64 K on Opus 4.7 at `xhigh`, 32 K on Sonnet 4.6 at `high`. oxide-code's `default_max_tokens(effort)` matches the upper tiers and uses 16 K otherwise when the user hasn't set `ANTHROPIC_MAX_TOKENS` explicitly.

### `context_management.edits`

Expand Down