fix(agent-interface): align codex reasoning efforts

.changeset/codex-xhigh-reasoning.md

@@ -0,0 +1,5 @@
+---
+"@tangle-network/agent-interface": patch
+---
+
+Align Codex reasoning support with the live model catalog in harness capability helpers.

packages/agent-interface/src/agent-profile.ts

@@ -116,7 +116,8 @@ export interface AgentProfileResources {
  *   - `minimal`  — thinking ON, the lowest budget (distinct from `none`)
  *   - `low` / `medium` / `high` / `xhigh`
  *   - `ultracode` — maximum (claude-code's "ultracode" run mode; codex's `max` reconciles here).
- * A backend without a matching native tier clamps to its nearest (e.g. codex maps `xhigh`/`ultracode` → `high`).
+ * A backend without a matching native tier clamps to its nearest (e.g. codex maps `ultracode` → `xhigh`
+ * on models that support it).
  */
 export type ReasoningEffort =
   | "none"

packages/agent-interface/src/harness-capabilities.test.ts

@@ -116,7 +116,7 @@ describe("reasoning effort support", () => {
     // no-thinking runners
     expect(harnessReasoningEfforts("cli-base")).toEqual(["none"]);
     // clamp-based: `none` dropped (inert ≡ auto); capped at the adapter's real ceiling
-    expect(harnessReasoningEfforts("codex")).toEqual(["minimal", "low", "medium", "high"]);
+    expect(harnessReasoningEfforts("codex")).toEqual(["low", "medium", "high", "xhigh"]);
     expect(harnessReasoningEfforts("pi")).toEqual(["minimal", "low", "medium", "high", "xhigh"]);
     expect(harnessReasoningEfforts("openclaw")).toEqual(["minimal", "low", "medium", "high", "xhigh"]);
     // claude: real `--effort` ladder low…max (ultracode stands in for max); no none/minimal
@@ -144,7 +144,11 @@ describe("reasoning effort support", () => {
       "low",
       "medium",
     ]);
-    // model ceiling above the harness ceiling can't widen it
+    expect(reasoningEffortsFor("codex", { maxEffort: "high" })).toEqual([
+      "low",
+      "medium",
+      "high",
+    ]);
     expect(reasoningEffortsFor("codex", { maxEffort: "ultracode" })).toEqual(
       harnessReasoningEfforts("codex"),
     );

packages/agent-interface/src/harness-capabilities.ts

@@ -12,9 +12,10 @@ import { canonicalizeHarness, type HarnessType } from "./harness.js";
  * cli-bridge backends, the sandbox UI pickers, and the router all read one truth instead of each
  * hand-rolling a divergent copy.
  *
- * Grounded in cli-bridge's real backend clamps (codex caps at `high`, kimi is binary on/off, claude
- * carries the full range, cli-base has no agent) — NOT a guessed matrix. The per-MODEL reasoning
- * capability (does this specific model reason at all) is dynamic catalog data the caller supplies.
+ * Grounded in cli-bridge's real backend clamps (codex is `low..xhigh` for current models, kimi is
+ * binary on/off, claude carries the full range, cli-base has no agent) — NOT a guessed matrix. The
+ * per-MODEL reasoning capability (does this specific model reason at all) is dynamic catalog data the
+ * caller supplies.
  */
 
 /** low → high. `none` = thinking off; `ultracode` = max (claude-code mode). */
@@ -149,8 +150,8 @@ export function snapHarnessToModel(
 /**
  * The explicit reasoning-effort set a harness's runtime accepts when it ISN'T a plain `none…ceiling`
  * slice — grounded in the cli-bridge adapters (NOT the canonical ladder):
- *   - codex: `model_reasoning_effort` accepts `minimal|low|medium|high` (xhigh/ultracode clamp down);
- *     it has no real `none` (the flag is just omitted), so `none` is dropped — `auto` covers "default".
+ *   - codex: current models advertise `low|medium|high|xhigh`; `none` is omitted (use `auto`) and
+ *     legacy `minimal` requests clamp up to `low`.
  *   - claude-code: `--effort` accepts `low|medium|high|xhigh|max`. `ultracode` is the ladder's stand-in
  *     for claude's `max` (clamped at runtime); `minimal`→`low` and `none`/`auto`→no flag, so both are
  *     dropped as redundant.
@@ -161,7 +162,7 @@ export function snapHarnessToModel(
 const harnessReasoningEffortsOverride: Partial<
   Record<HarnessType, readonly ReasoningEffort[]>
 > = {
-  codex: ["minimal", "low", "medium", "high"],
+  codex: ["low", "medium", "high", "xhigh"],
   "claude-code": ["low", "medium", "high", "xhigh", "ultracode"],
   pi: ["minimal", "low", "medium", "high", "xhigh"],
   openclaw: ["minimal", "low", "medium", "high", "xhigh"],