pi-coralbricks-provider 1.0.8 → 1.0.10

package/README.md

@@ -23,7 +23,7 @@ _A [pi](https://github.com/earendil-works/pi-coding-agent) provider extension fo
 - **Per-family thinking levels** — zai-style `thinking` control for GLM (including a *real* off switch), `reasoning_effort` for Kimi K3 and GPT-OSS
 - **Accurate cost tracking** — pricing mirrors Coral's published rates, and cached reads are **$0 on every model**, so pi's computed cost matches Coral's own `usage.cost` to the token
 - **Self-healing model sync** — stale-while-revalidate from the authenticated `/v1/models` (or the unauthenticated [public catalog](https://www.coralbricks.ai/api/public/models) before auth), hot-swapped at session start
-- **synbad-validated** — [synbad](https://github.com/synthetic-lab/synbad) tool-calling and reasoning-parsing evals pass 15/15 on GLM 5.2, GLM 5.3, and Kimi K3 in both unary and streaming modes
+- **synbad-validated** — [synbad](https://github.com/synthetic-lab/synbad) tool-calling and reasoning-parsing evals pass 15/15 on GLM 5.3 and Kimi K3 in both unary and streaming modes
 
 ## Installation
 
@@ -80,7 +80,6 @@ pi
 
 | Model | Context | Vision | Reasoning | Input $/M | Cache Read $/M | Output $/M |
 |-------|---------|--------|-----------|-----------|-----------------|------------|
-| GLM 5.2 FP4 | 1.0M | ❌ | ✅ | $1.12 | — | $4.40 |
 | GLM 5.3 Flash | 1.0M | ✅ | ✅ | $0.15 | — | $0.50 |
 | GLM 5.3 FP4 | 1.0M | ❌ | ✅ | $1.12 | — | $4.40 |
 | GPT-OSS 120B | 131K | ❌ | ✅ | $0.12 | — | $0.60 |
@@ -130,12 +129,12 @@ Verified against the live gateway:
 
 | Model | Format | off | low | medium | high | max |
 |-------|--------|-----|-----|--------|------|-----|
-| GLM 5.2 FP4 | `thinking: {type}` + `reasoning_effort` | ✅ | — | ✅ | ✅ | ✅ |
 | GLM 5.3 FP4 | `thinking: {type}` + `reasoning_effort` | ✅ | ✅ | — | ✅ | ✅ |
+| GLM 5.3 Flash | `thinking: {type}` + `reasoning_effort` | ✅ | ✅ | — | ✅ | ✅ |
 | Kimi K3 | `reasoning_effort` | — | ✅ | — | ✅ | ✅ |
 | GPT-OSS 120B | `reasoning_effort` | — | ✅ | ✅ | ✅ | — |
 
-- **GLM** accepts zai-style `thinking: {type: "disabled"}` on Coral — pi's *off* level genuinely disables thinking (the upstream Z.ai API does not, so this differs from the canonical Z.ai map).
+- **GLM** accepts zai-style `thinking: {type: "disabled"}` on Coral — pi's *off* level turns thinking off, though a short preamble of a few dozen tokens can still appear (the upstream Z.ai API has no off at all, so this differs from the canonical Z.ai map).
 - **Kimi K3** always thinks: `reasoning_effort: "none"` is accepted but doesn't disable reasoning, so *off* is intentionally not offered. Assistant replays include `reasoning_content`.
 - **GPT-OSS** exposes the standard low/medium/high reasoning efforts; reasoning arrives in `reasoning_content`.
 - Coral streams a duplicate `reasoning` field alongside `reasoning_content`; pi dedupes these automatically.
@@ -146,7 +145,7 @@ Coral's gateway follows the OpenAI Chat Completions API:
 
 - **`supportsStore: false`** / **`supportsDeveloperRole: false`** — all models; Coral serves open models on the classic roles.
 - **`maxTokensField: "max_tokens"`** — all models.
-- **`thinkingFormat: "zai"`** — GLM 5.2/5.3: `thinking: {type: "enabled"|"disabled"}` toggles reasoning, `reasoning_effort` picks the depth.
+- **`thinkingFormat: "zai"`** — GLM 5.3: `thinking: {type: "enabled"|"disabled"}` toggles reasoning, `reasoning_effort` picks the depth.
 - **`thinkingFormat: "openai"`** — Kimi K3 and GPT-OSS 120B: `reasoning_effort` drives thinking depth.
 - **`supportsStrictMode: false`** — Kimi K3 (no strict JSON-schema tool definitions).
 - **`requiresReasoningContentOnAssistantMessages: true`** — Kimi K3.
@@ -160,16 +159,19 @@ Merge order: `[live|cache|embedded] → patch.json → custom-models.json`
 
 ## Inference-Quality Testing
 
-Validated with [synbad](https://github.com/synthetic-lab/synbad) — Synthetic's tool-calling and reasoning-parsing eval suite for LLM inference providers (`--count 1`, `--reasoning-effort high`, 15 evals per run):
+Validated with [synbad](https://github.com/synthetic-lab/synbad) — Synthetic's tool-calling and reasoning-parsing eval suite for LLM inference providers (`--count 1`, `--reasoning-effort high`; the GLM 5.3/Kimi runs used the 15-eval suite, the GPT-OSS and GLM 5.3 Flash runs on 2026-09-15 the current 13-eval suite):
 
 | Model | Unary | Stream | Notes |
 |-------|-------|--------|-------|
-| GLM 5.2 FP4 | 15/15 ✅ | 15/15 ✅ | |
 | GLM 5.3 FP4 | 15/15 ✅ | 15/15 ✅ | |
+| GLM 5.3 Flash | 13/13 ✅ | 12/13 ⚠️ | re-run 2026-09-15; see below |
 | Kimi K3 | 15/15 ✅ | 15/15 ✅ | |
-| GPT-OSS 120B | 14/15 ⚠️ | pending re-eval | gateway tool-call index bug is fixed upstream |
+| GPT-OSS 120B | 11/13 ⚠️ | 12/13 ⚠️ | re-run 2026-09-15; see below |
+| GLM 5.2 FP4 (retired) | 15/15 ✅ | 15/15 ✅ | evaluated before Coral retired the model |
 
-The remaining gpt-oss quirk is model-side: it answers "Paris and London" with a single batched tool call even when `parallel_tool_calls: true` — not a gateway bug.
+The gpt-oss misses are model-side, not transport: `tools/parallel-tool` fails in both modes because it answers "Paris and London" with one call after another instead of two parallel calls even when `parallel_tool_calls: true`; `reasoning/reasoning-claude-tool-call` (a prompt that asks the model to "put the tool call inside your thinking") occasionally gets exactly that in unary mode — the model narrates the call in its reasoning and answers in text — while streaming passed it every time.
+
+GLM 5.3 Flash's streaming miss is `reasoning/reasoning-parsing`: on the trivial prompt it uses, Flash occasionally answers without a think block (about 1 in 5 streams and 1 in 8 unary responses in our sampling); the answer itself is unaffected, and a second full unary pass was 13/13.
 
 Reproduce:
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-coralbricks-provider",
-  "version": "1.0.8",
+  "version": "1.0.10",
   "description": "CoralBricks provider extension for pi - Access GLM 5.3, GLM 5.3 Flash, Kimi K3, and GPT-OSS 120B through the Coral Inference API",
   "author": "monotykamary",
   "homepage": "https://github.com/monotykamary/pi-coralbricks-provider#readme",

package/scripts/update-models.js

@@ -478,27 +478,6 @@ function updateDeprecatedModels(modelsJsonPath, newModels) {
   }
 }
 
-function withDeprecatedForReadme(models) {
-  const deprecatedPath = path.join(path.dirname(MODELS_JSON_PATH), 'deprecated-models.json');
-  let deprecated = {};
-  try {
-    const parsed = JSON.parse(fs.readFileSync(deprecatedPath, 'utf8'));
-    if (parsed && typeof parsed === 'object' && !Array.isArray(parsed)) deprecated = parsed;
-  } catch { /* no graveyard yet */ }
-  const now = Date.now();
-  const seen = new Set(models.map((m) => m.id));
-  const extras = [];
-  for (const entry of Object.values(deprecated)) {
-    if (!entry || !entry.id || seen.has(entry.id)) continue;
-    const removedAt = Date.parse(entry.deprecatedAt || '');
-    if (Number.isNaN(removedAt) || now - removedAt > DEPRECATED_MODEL_TTL_MS) continue;
-    const m = { ...entry };
-    delete m.deprecatedAt;
-    extras.push(m);
-  }
-  return extras.length > 0 ? [...models, ...extras] : models;
-}
-
 async function main() {
   try {
     // Regenerate the derived README table from local source data without an API
@@ -507,7 +486,7 @@ async function main() {
       const baseModels = loadJson(MODELS_JSON_PATH);
       const patchData = loadJson(PATCH_JSON_PATH);
       const customModels = loadJson(CUSTOM_MODELS_JSON_PATH);
-      const readmeBase = withDeprecatedForReadme(Array.isArray(baseModels) ? baseModels : []);
+      const readmeBase = Array.isArray(baseModels) ? baseModels : [];
       const readmeModels = buildModels(readmeBase, Array.isArray(customModels) ? customModels : [], patchData);
       readmeModels.sort((a, b) => a.name.localeCompare(b.name));
       updateReadme(readmeModels);
@@ -541,10 +520,11 @@ async function main() {
     updateDeprecatedModels(MODELS_JSON_PATH, models);
     saveJson(MODELS_JSON_PATH, models);
 
-    // Build full model list for README: (base + grace-period deprecated) → patch → custom.
+    // Build the README model list: base → patch → custom. Grace-period models stay
+    // servable at runtime (index.ts appends them) but are not listed as available.
     const patchData = loadJson(PATCH_JSON_PATH);
     const customModels = loadJson(CUSTOM_MODELS_JSON_PATH);
-    const readmeBase = withDeprecatedForReadme(models);
+    const readmeBase = models;
     const readmeModels = buildModels(readmeBase, Array.isArray(customModels) ? customModels : [], patchData);
     readmeModels.sort((a, b) => a.name.localeCompare(b.name));