pi-fireworks-provider 1.0.2 → 1.1.1

package/README.md

@@ -4,7 +4,7 @@
 
 **31+ models through [Fireworks AI](https://fireworks.ai/)**
 
-_Kimi, MiniMax, GLM, DeepSeek, GPT-OSS — unified OpenAI-compatible API for [pi](https://github.com/earendil-works/pi-coding-agent)._
+_Kimi, MiniMax, GLM, DeepSeek, GPT-OSS — via Fireworks AI's Anthropic Messages and OpenAI-compatible endpoints for [pi](https://github.com/earendil-works/pi-coding-agent)._
 
 [![pi extension](https://img.shields.io/badge/pi-extension-blueviolet)](https://github.com/earendil-works/pi-coding-agent)
 [![license](https://img.shields.io/badge/license-MIT-blue)](./LICENSE)
@@ -16,7 +16,10 @@ _Kimi, MiniMax, GLM, DeepSeek, GPT-OSS — unified OpenAI-compatible API for [pi
 ## Features
 
 - **35+ AI Models** including Kimi K2.5, MiniMax M2.5, GLM 4.5/4.7/5, DeepSeek V3.1/V3.2, DeepSeek V4 Flash, and GPT-OSS
-- **Unified API** via Fireworks AI's OpenAI-compatible completions endpoint
+- **Dual API support** via Fireworks AI's Anthropic Messages and OpenAI-compatible completions endpoints (per-model routing, matching pi core's Fireworks provider)
+- **Service tiers** — toggle Fireworks `priority` vs `standard` per request on supported models (with priority pricing reflected in cost tracking), via a keybinding, `/fireworks-tier`, and a footer status area
+- **Preserved thinking** — toggle Fireworks' `reasoning_history: "preserved"` so prior assistant reasoning is retained across turns (better multi-turn recall; uses more tokens), via the `/fireworks-settings` panel, with a model-select notification. Matches neuralwatt/makora's settings-only UX, adapted to Fireworks' single global `reasoning_history` knob
+- **Settings panel** — `/fireworks-settings` (TUI) to configure preserved thinking, service tier, and display preferences; persisted to `~/.pi/agent/extensions/fireworks.json`
 - **Cost Tracking** with per-model pricing for budget management
 - **Reasoning Models** support for advanced reasoning capabilities
 - **Vision Support** for image-capable models
@@ -105,6 +108,69 @@ pi
 | Qwen3 VL 30B A3B Thinking | Text + Image | 262K | 0 | Free | Free |
 *Costs are per million tokens. Prices subject to change - check [fireworks.ai](https://fireworks.ai) for current pricing.*
 
+## Service Tiers
+
+Fireworks exposes a `service_tier` request field (`standard` | `priority`) on its chat-completions endpoint. The **priority** tier trades higher per-token pricing for higher throughput / lower latency. This is orthogonal to the `-fast`/`-turbo` router model IDs (which are separate models) — service tiers apply to the base models below.
+
+| Model | Priority Uncached Input | Priority Cached Input | Priority Output |
+| --- | --- | --- | --- |
+| GLM 5.2 | $1.75/M | $0.175/M | $5.5/M |
+| Kimi K2.7 Code | $1.43/M | $0.29/M | $6/M |
+| Minimax M3 | $0.45/M | $0.09/M | $1.8/M |
+| DeepSeek V4 Pro | $2.61/M | $0.218/M | $5.22/M |
+| Kimi K2.6 | $1.5/M | $0.22/M | $6/M |
+| MiniMax M2.7 | $0.45/M | $0.09/M | $1.8/M |
+| GLM 5.1 | $2.1/M | $0.39/M | $6.6/M |
+| GPT OSS 120B | $0.18/M | $0.018/M | $0.72/M |
+| DeepSeek V4 Flash | $0.21/M | $0.045/M | $0.42/M |
+
+*Priority pricing is roughly 1.2–1.5× the standard rate. `cacheWrite` is not tiered.*
+
+**Switching tiers:**
+
+- **Keybinding:** `ctrl+shift+l` (default) toggles `standard` ↔ `priority` for the active supported model. No-op with an info notice for unsupported models.
+- **Command:** `/fireworks-tier standard|priority|toggle`.
+- **Status area:** a dim `tier: standard` / `tier: ⚡priority` line is shown in the footer for supported models while a Fireworks model is active.
+
+The selection is persisted per session (survives `/reload` and resume). When `priority` is active, `service_tier: "priority"` is injected into every request and finalized cost is recomputed against the priority rates above.
+
+**Configuration** — `~/.pi/agent/extensions/fireworks.json` (created with defaults on first load):
+
+```json
+{
+  "serviceTier": {
+    "default": "standard",
+    "keybinding": "ctrl+shift+l",
+    "display": "statusbar"
+  },
+  "preserveThinking": {
+    "default": false
+  }
+}
+```
+
+- `serviceTier.default` — tier used until you toggle (`standard` | `priority`).
+- `serviceTier.keybinding` — any [pi key format](https://github.com/earendil-works/pi-coding-agent/blob/main/docs/keybindings.md) (e.g. `ctrl+shift+l`, `ctrl+shift+k`). Requires `/reload` after changing. On macOS browser terminals (localterm), avoid `alt`/`ctrl+alt` (Option produces special chars) and `ctrl+shift+t/w/n/c/v` (browser/localterm tab + copy/paste shortcuts).
+- `serviceTier.display` — `statusbar` (footer status area) or `off` (hide the tier indicator).
+- `preserveThinking.default` — whether `reasoning_history: "preserved"` is injected (`true` | `false`, default `false`). Also settable via `/fireworks-settings`.
+
+> **Note:** The OpenAI completions endpoint accepts `service_tier` directly (per Fireworks' API). The Anthropic Messages endpoint passes the top-level field through as an extra. If a supported Anthropic-routed model rejects it, file an issue so we can gate injection by API.
+
+## Preserved Thinking
+
+Fireworks exposes a top-level `reasoning_history` request parameter. The only accepted value is `"preserved"`; omitting it (the default) means prior assistant reasoning is **stripped** from the model's context each turn. Setting `reasoning_history: "preserved"` makes Fireworks render prior assistant reasoning into the model's context, improving multi-turn recall at the cost of extra tokens. See [the Fireworks reasoning guide](https://docs.fireworks.ai/guides/reasoning#preserved-thinking).
+
+This works on **both** transports — the OpenAI completions endpoint (assistant `reasoning_content` field) and the Anthropic Messages endpoint (assistant `thinking` content blocks, for which Fireworks returns a `signature` so pi-ai replays them). pi-ai already replays the reasoning field/block on prior assistant turns; this extension's only job is injecting the top-level `reasoning_history: "preserved"` flag that makes Fireworks honor it.
+
+Unlike neuralwatt/makora (which use per-model vLLM `chat_template_kwargs` flags like `preserve_thinking`/`clear_thinking`), Fireworks' knob is a single global parameter that applies to every reasoning model, so we expose it as one on/off toggle rather than a per-model submenu.
+
+**Toggle it:**
+
+- **`/fireworks-settings`** (TUI) → *Preserved thinking* → `on` / `off`. Takes effect immediately and persists as the default for future sessions.
+- A dim **model-select notification** tells you the current state when you switch to a Fireworks reasoning model (`Preserved thinking ON for …` / `… OFF for …`).
+
+Preserved thinking is **off by default** to match pi core and Fireworks' default (stripped). There is intentionally no `/fireworks-preserve` command or keybinding — it's settings-panel-only, mirroring neuralwatt/makora.
+
 ## Usage
 
 After loading the extension, use the `/model` command in pi to select your preferred model:
@@ -145,6 +211,18 @@ Add to your pi configuration for automatic loading:
 }
 ```
 
+## Development
+
+```bash
+pnpm install          # install dev tooling (vitest, knip, typescript)
+pnpm test             # run the test suite (vitest)
+pnpm run test:watch   # watch mode
+pnpm run lint:dead    # dead-code / unused-export scan (knip)
+pnpm run check        # typecheck (tsc) + tests + knip, all green or exit non-zero
+```
+
+Tests live in `tests/` and stub the `@earendil-works/pi-coding-agent` / `@earendil-works/pi-tui` peer dependencies (see `tests/__mocks__/`) so they run without the real pi packages installed. A per-run temp dir is used for `~/.pi/agent` (via `PI_CODING_AGENT_DIR` in `tests/vitest.setup.ts`) so config/cache reads and writes never touch your real environment.
+
 ## License
 
 MIT

package/custom-models.json

@@ -147,5 +147,72 @@
     ],
     "contextWindow": 262000,
     "maxTokens": 0
+  },
+  {
+    "id": "accounts/fireworks/models/qwen3p7-plus",
+    "name": "Qwen 3.7 Plus",
+    "reasoning": true,
+    "cost": {
+      "input": 0.4,
+      "output": 1.6,
+      "cacheRead": 0.08,
+      "cacheWrite": 0
+    },
+    "input": [
+      "text",
+      "image"
+    ],
+    "contextWindow": 262144,
+    "maxTokens": 65536
+  },
+  {
+    "id": "accounts/fireworks/routers/glm-5p2-fast",
+    "name": "GLM 5.2 Fast",
+    "reasoning": true,
+    "cost": {
+      "input": 2.1,
+      "output": 6.6,
+      "cacheRead": 0.21,
+      "cacheWrite": 0
+    },
+    "input": [
+      "text"
+    ],
+    "contextWindow": 1048575,
+    "maxTokens": 131072
+  },
+  {
+    "id": "accounts/fireworks/routers/kimi-k2p6-fast",
+    "name": "Kimi K2.6 Fast",
+    "reasoning": true,
+    "cost": {
+      "input": 2,
+      "output": 8,
+      "cacheRead": 0.3,
+      "cacheWrite": 0
+    },
+    "input": [
+      "text",
+      "image"
+    ],
+    "contextWindow": 262000,
+    "maxTokens": 262000
+  },
+  {
+    "id": "accounts/fireworks/routers/kimi-k2p7-code-fast",
+    "name": "Kimi K2.7 Code Fast",
+    "reasoning": true,
+    "cost": {
+      "input": 1.9,
+      "output": 8,
+      "cacheRead": 0.38,
+      "cacheWrite": 0
+    },
+    "input": [
+      "text",
+      "image"
+    ],
+    "contextWindow": 262000,
+    "maxTokens": 262000
   }
 ]