perplexity-web-mcp-cli 0.14.0__tar.gz → 0.14.2__tar.gz

{perplexity_web_mcp_cli-0.14.0 → perplexity_web_mcp_cli-0.14.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: perplexity-web-mcp-cli
-Version: 0.14.0
+Version: 0.14.2
 Summary: CLI, MCP server, and Anthropic/OpenAI API-compatible interface for Perplexity AI.
 Keywords: perplexity,ai,mcp,anthropic,api,client
 Author: Jacob BD
@@ -46,7 +46,6 @@ Description-Content-Type: text/markdown
 [![Total downloads](https://static.pepy.tech/badge/perplexity-web-mcp-cli)](https://pepy.tech/projects/perplexity-web-mcp-cli)
 [![Python](https://img.shields.io/pypi/pyversions/perplexity-web-mcp-cli)](https://pypi.org/project/perplexity-web-mcp-cli/)
 [![License](https://img.shields.io/pypi/l/perplexity-web-mcp-cli)](https://github.com/jacob-bd/perplexity-web-mcp/blob/main/LICENSE)
-[![Buy Me a Coffee](https://img.shields.io/badge/Buy%20Me%20a%20Coffee-FFDD00?style=flat-square&logo=buy-me-a-coffee&logoColor=black)](https://buymeacoffee.com/jacobbd)
 
 <p align="center">
   <a href="https://youtu.be/9xyClDvmoZ0">
@@ -62,14 +61,14 @@ Description-Content-Type: text/markdown
 
 MCP server, CLI, and API-compatible interface for Perplexity AI's web interface.
 
-Use your Perplexity Pro/Max subscription to access premium models (Sonar 2, GPT-5.4, GPT-5.5, Claude Sonnet 5.0, Claude Opus 4.8, Gemini 3.1 Pro, GLM 5.2, Kimi K2.6, Nemotron 3 Ultra) from the terminal, through MCP tools, or as an API endpoint.
+Use your Perplexity Pro/Max subscription to access premium models (Sonar 2, GPT-5.6 Terra, GPT-5.6 Sol, Gemini 3.1 Pro, Claude Sonnet 5, Claude Opus 4.8, GLM 5.2, Kimi K2.6, Grok 4.5, and Nemotron 3 Ultra) from the terminal, through MCP tools, or as an API endpoint.
 
 ## Features
 
 - **CLI**: Query Perplexity models directly from the terminal (`pwm ask`, `pwm council`, `pwm research`, `pwm chat`)
 - **MCP Server**: MCP tools for AI agents with citations, rate limit checking, and multi-turn context
 - **API Server**: Drop-in Anthropic Messages API and OpenAI Chat Completions API
-- **9 Models**: Sonar 2, GPT-5.4, GPT-5.5, Claude Sonnet 5.0, Claude Opus 4.8, Gemini 3.1 Pro, GLM 5.2, Kimi K2.6, Nemotron 3 Ultra
+- **10 Models**: Sonar 2, GPT-5.6 Terra, GPT-5.6 Sol, Gemini 3.1 Pro, Claude Sonnet 5, Claude Opus 4.8, GLM 5.2, Kimi K2.6, Grok 4.5, and Nemotron 3 Ultra
 - **Thinking Mode**: Extended thinking support for all compatible models
 - **Deep Research**: Full support for Perplexity's Deep Research mode
 - **Multi-Turn Conversations**: State-preserved threaded conversations for both MCP and CLI REPL
@@ -174,7 +173,7 @@ pwm ask "What is quantum computing?"
 **Choose a specific model** with `-m` (see [Models](#models) for the full list):
 
 ```bash
-pwm ask "Compare React and Vue" -m gpt54
+pwm ask "Compare React and Vue" -m gpt56_terra
 ```
 
 ```bash
@@ -244,13 +243,13 @@ pwm research "NVIDIA competitive landscape" -s finance --json
 Query multiple models in parallel and get a synthesized consensus. Each model costs 1 Pro Search. Default synthesis uses Sonar 2 (also 1 Pro Search).
 
 ```bash
-# Default: GPT-5.4, Claude Sonnet, Gemini Pro + Sonar 2 synthesis (4 Pro Searches)
+# Default: GPT-5.6 Terra, Claude Sonnet, Gemini Pro + Sonar 2 synthesis (4 Pro Searches)
 pwm council "What are best practices for microservices?"
 ```
 
 ```bash
 # Custom model selection
-pwm council "Compare Rust and Go" -m gpt54,claude_sonnet
+pwm council "Compare Rust and Go" -m gpt56_terra,claude_sonnet
 ```
 
 ```bash
@@ -285,7 +284,7 @@ Seamlessly launch external AI tools connected to the Perplexity API server. This
 
 ```bash
 pwm hack claude            # Launch Claude Code
-pwm hack claude -m gpt54   # Launch Claude Code with a specific model
+pwm hack claude -m gpt56_terra   # Launch Claude Code with a specific model
 ```
 
 ### MCP Setup
@@ -337,8 +336,10 @@ pwm --ai                   # Print comprehensive AI-optimized reference
 | `auto`          | Perplexity | No       | Auto-selects best model                          |
 | `sonar`         | Perplexity | No       | Sonar 2 (latest in-house; API id `experimental`) |
 | `deep_research` | Perplexity | No       | Monthly quota, in-depth reports                  |
-| `gpt54`         | OpenAI     | Toggle   | GPT-5.4                                          |
-| `claude_sonnet` | Anthropic  | Toggle   | Claude Sonnet 5.0                                |
+| `gpt56_terra`   | OpenAI     | Toggle   | GPT-5.6 Terra                                    |
+| `gpt56_sol`     | OpenAI     | Toggle   | GPT-5.6 Sol (Max tier required)                  |
+| `grok45`        | xAI        | Toggle   | Grok 4.5                                         |
+| `claude_sonnet` | Anthropic  | Toggle   | Claude Sonnet 5                                |
 | `claude_opus`   | Anthropic  | Toggle   | Claude Opus 4.8 (Max tier required)              |
 | `gemini_pro`    | Google     | Always   | Gemini 3.1 Pro                                   |
 | `nemotron`      | NVIDIA     | Always   | Nemotron 3 Ultra 550B                            |
@@ -424,9 +425,10 @@ claude mcp add perplexity pwm-mcp
 | `pplx_ask`                                        | Quick Q&A (auto-selects best model)         |
 | `pplx_deep_research`                              | In-depth reports with sources               |
 | `pplx_sonar`                                      | Perplexity Sonar 2 (1 Pro Search)           |
-| `pplx_gpt54` / `pplx_gpt54_thinking`              | GPT-5.4                                     |
-| `pplx_gpt55` / `pplx_gpt55_thinking`              | GPT-5.5 (Max tier)                          |
-| `pplx_claude_sonnet` / `pplx_claude_sonnet_think` | Claude Sonnet 5.0                           |
+| `pplx_gpt56_terra` / `pplx_gpt56_terra_thinking`              | GPT-5.6 Terra                                     |
+| `pplx_gpt56_sol` / `pplx_gpt56_sol_thinking`              | GPT-5.6 Sol (Max tier)                          |
+| `pplx_grok45` / `pplx_grok45_thinking`                    | Grok 4.5                                        |
+| `pplx_claude_sonnet` / `pplx_claude_sonnet_think` | Claude Sonnet 5                           |
 | `pplx_claude_opus` / `pplx_claude_opus_think`     | Claude Opus 4.8 (Max tier)                  |
 | `pplx_gemini_pro_think`                           | Gemini 3.1 Pro (thinking always on)         |
 | `pplx_nemotron_thinking`                          | Nemotron 3 Ultra (thinking always on)       |
@@ -474,7 +476,7 @@ pwm api
 ```bash
 export ANTHROPIC_BASE_URL=http://localhost:8080
 export ANTHROPIC_API_KEY=perplexity
-claude --model gpt-5.4
+claude --model gpt-5.6-terra
 ```
 
 Alternatively, launch Claude Code seamlessly using the `hack` command, which automatically starts the API server and configures the environment for you:
@@ -503,15 +505,17 @@ export OPENAI_API_KEY=dummy
 codex -m sonar --local-provider lmstudio
 ```
 
-Our server's `MODEL_MAP` will seamlessly intercept `sonar` (or any other mapped names like `gemini-pro`, `nemotron`, `glm-5.2`, `claude-sonnet-5-0`) and correctly route it to Perplexity's API. You can also create an alias in your shell to make this easier: `alias codex-pplx="codex --local-provider lmstudio"`.
+Our server's `MODEL_MAP` will seamlessly intercept `sonar` (or any other mapped names like `gemini-pro`, `nemotron`, `glm-5.2`, `claude-sonnet-5`) and correctly route it to Perplexity's API. You can also create an alias in your shell to make this easier: `alias codex-pplx="codex --local-provider lmstudio"`.
 
 ### API Model Names
 
 | API Name                        | Perplexity Model   | Thinking |
 | ------------------------------- | ------------------ | -------- |
 | `perplexity-auto`               | Best (auto-select) | No       |
-| `gpt-5.4`                       | GPT-5.4            | Toggle   |
-| `claude-sonnet-5-0`             | Claude Sonnet 5.0  | Toggle   |
+| `gpt-5.6-terra`                 | GPT-5.6 Terra      | Toggle   |
+| `gpt-5.6-sol`                   | GPT-5.6 Sol        | Toggle   |
+| `grok-4.5`                      | Grok 4.5           | Toggle   |
+| `claude-sonnet-5`               | Claude Sonnet 5    | Toggle   |
 | `claude-opus-4-8`               | Claude Opus 4.8    | Toggle   |
 | `gemini-3.1-pro`                | Gemini 3.1 Pro     | Always   |
 | `glm-5.2` / `glm52`             | GLM 5.2            | Always   |
@@ -622,6 +626,12 @@ The skill follows Anthropic's Agent Skills open standard and works across any co
 
 Originally forked from [perplexity-webui-scraper](https://github.com/henrique-coder/perplexity-webui-scraper) by [henrique-coder](https://github.com/henrique-coder).
 
+## Support
+
+If Perplexity Web MCP saves you time or money, you can help cover the AI bills that go into building and testing it. Any support is hugely appreciated. 🙏
+
+<a href="https://buymeacoffee.com/jacobbd"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="42"></a>
+
 ## License
 
 MIT

{perplexity_web_mcp_cli-0.14.0 → perplexity_web_mcp_cli-0.14.2}/README.md

@@ -9,7 +9,6 @@
 [![Total downloads](https://static.pepy.tech/badge/perplexity-web-mcp-cli)](https://pepy.tech/projects/perplexity-web-mcp-cli)
 [![Python](https://img.shields.io/pypi/pyversions/perplexity-web-mcp-cli)](https://pypi.org/project/perplexity-web-mcp-cli/)
 [![License](https://img.shields.io/pypi/l/perplexity-web-mcp-cli)](https://github.com/jacob-bd/perplexity-web-mcp/blob/main/LICENSE)
-[![Buy Me a Coffee](https://img.shields.io/badge/Buy%20Me%20a%20Coffee-FFDD00?style=flat-square&logo=buy-me-a-coffee&logoColor=black)](https://buymeacoffee.com/jacobbd)
 
 <p align="center">
   <a href="https://youtu.be/9xyClDvmoZ0">
@@ -25,14 +24,14 @@
 
 MCP server, CLI, and API-compatible interface for Perplexity AI's web interface.
 
-Use your Perplexity Pro/Max subscription to access premium models (Sonar 2, GPT-5.4, GPT-5.5, Claude Sonnet 5.0, Claude Opus 4.8, Gemini 3.1 Pro, GLM 5.2, Kimi K2.6, Nemotron 3 Ultra) from the terminal, through MCP tools, or as an API endpoint.
+Use your Perplexity Pro/Max subscription to access premium models (Sonar 2, GPT-5.6 Terra, GPT-5.6 Sol, Gemini 3.1 Pro, Claude Sonnet 5, Claude Opus 4.8, GLM 5.2, Kimi K2.6, Grok 4.5, and Nemotron 3 Ultra) from the terminal, through MCP tools, or as an API endpoint.
 
 ## Features
 
 - **CLI**: Query Perplexity models directly from the terminal (`pwm ask`, `pwm council`, `pwm research`, `pwm chat`)
 - **MCP Server**: MCP tools for AI agents with citations, rate limit checking, and multi-turn context
 - **API Server**: Drop-in Anthropic Messages API and OpenAI Chat Completions API
-- **9 Models**: Sonar 2, GPT-5.4, GPT-5.5, Claude Sonnet 5.0, Claude Opus 4.8, Gemini 3.1 Pro, GLM 5.2, Kimi K2.6, Nemotron 3 Ultra
+- **10 Models**: Sonar 2, GPT-5.6 Terra, GPT-5.6 Sol, Gemini 3.1 Pro, Claude Sonnet 5, Claude Opus 4.8, GLM 5.2, Kimi K2.6, Grok 4.5, and Nemotron 3 Ultra
 - **Thinking Mode**: Extended thinking support for all compatible models
 - **Deep Research**: Full support for Perplexity's Deep Research mode
 - **Multi-Turn Conversations**: State-preserved threaded conversations for both MCP and CLI REPL
@@ -137,7 +136,7 @@ pwm ask "What is quantum computing?"
 **Choose a specific model** with `-m` (see [Models](#models) for the full list):
 
 ```bash
-pwm ask "Compare React and Vue" -m gpt54
+pwm ask "Compare React and Vue" -m gpt56_terra
 ```
 
 ```bash
@@ -207,13 +206,13 @@ pwm research "NVIDIA competitive landscape" -s finance --json
 Query multiple models in parallel and get a synthesized consensus. Each model costs 1 Pro Search. Default synthesis uses Sonar 2 (also 1 Pro Search).
 
 ```bash
-# Default: GPT-5.4, Claude Sonnet, Gemini Pro + Sonar 2 synthesis (4 Pro Searches)
+# Default: GPT-5.6 Terra, Claude Sonnet, Gemini Pro + Sonar 2 synthesis (4 Pro Searches)
 pwm council "What are best practices for microservices?"
 ```
 
 ```bash
 # Custom model selection
-pwm council "Compare Rust and Go" -m gpt54,claude_sonnet
+pwm council "Compare Rust and Go" -m gpt56_terra,claude_sonnet
 ```
 
 ```bash
@@ -248,7 +247,7 @@ Seamlessly launch external AI tools connected to the Perplexity API server. This
 
 ```bash
 pwm hack claude            # Launch Claude Code
-pwm hack claude -m gpt54   # Launch Claude Code with a specific model
+pwm hack claude -m gpt56_terra   # Launch Claude Code with a specific model
 ```
 
 ### MCP Setup
@@ -300,8 +299,10 @@ pwm --ai                   # Print comprehensive AI-optimized reference
 | `auto`          | Perplexity | No       | Auto-selects best model                          |
 | `sonar`         | Perplexity | No       | Sonar 2 (latest in-house; API id `experimental`) |
 | `deep_research` | Perplexity | No       | Monthly quota, in-depth reports                  |
-| `gpt54`         | OpenAI     | Toggle   | GPT-5.4                                          |
-| `claude_sonnet` | Anthropic  | Toggle   | Claude Sonnet 5.0                                |
+| `gpt56_terra`   | OpenAI     | Toggle   | GPT-5.6 Terra                                    |
+| `gpt56_sol`     | OpenAI     | Toggle   | GPT-5.6 Sol (Max tier required)                  |
+| `grok45`        | xAI        | Toggle   | Grok 4.5                                         |
+| `claude_sonnet` | Anthropic  | Toggle   | Claude Sonnet 5                                |
 | `claude_opus`   | Anthropic  | Toggle   | Claude Opus 4.8 (Max tier required)              |
 | `gemini_pro`    | Google     | Always   | Gemini 3.1 Pro                                   |
 | `nemotron`      | NVIDIA     | Always   | Nemotron 3 Ultra 550B                            |
@@ -387,9 +388,10 @@ claude mcp add perplexity pwm-mcp
 | `pplx_ask`                                        | Quick Q&A (auto-selects best model)         |
 | `pplx_deep_research`                              | In-depth reports with sources               |
 | `pplx_sonar`                                      | Perplexity Sonar 2 (1 Pro Search)           |
-| `pplx_gpt54` / `pplx_gpt54_thinking`              | GPT-5.4                                     |
-| `pplx_gpt55` / `pplx_gpt55_thinking`              | GPT-5.5 (Max tier)                          |
-| `pplx_claude_sonnet` / `pplx_claude_sonnet_think` | Claude Sonnet 5.0                           |
+| `pplx_gpt56_terra` / `pplx_gpt56_terra_thinking`              | GPT-5.6 Terra                                     |
+| `pplx_gpt56_sol` / `pplx_gpt56_sol_thinking`              | GPT-5.6 Sol (Max tier)                          |
+| `pplx_grok45` / `pplx_grok45_thinking`                    | Grok 4.5                                        |
+| `pplx_claude_sonnet` / `pplx_claude_sonnet_think` | Claude Sonnet 5                           |
 | `pplx_claude_opus` / `pplx_claude_opus_think`     | Claude Opus 4.8 (Max tier)                  |
 | `pplx_gemini_pro_think`                           | Gemini 3.1 Pro (thinking always on)         |
 | `pplx_nemotron_thinking`                          | Nemotron 3 Ultra (thinking always on)       |
@@ -437,7 +439,7 @@ pwm api
 ```bash
 export ANTHROPIC_BASE_URL=http://localhost:8080
 export ANTHROPIC_API_KEY=perplexity
-claude --model gpt-5.4
+claude --model gpt-5.6-terra
 ```
 
 Alternatively, launch Claude Code seamlessly using the `hack` command, which automatically starts the API server and configures the environment for you:
@@ -466,15 +468,17 @@ export OPENAI_API_KEY=dummy
 codex -m sonar --local-provider lmstudio
 ```
 
-Our server's `MODEL_MAP` will seamlessly intercept `sonar` (or any other mapped names like `gemini-pro`, `nemotron`, `glm-5.2`, `claude-sonnet-5-0`) and correctly route it to Perplexity's API. You can also create an alias in your shell to make this easier: `alias codex-pplx="codex --local-provider lmstudio"`.
+Our server's `MODEL_MAP` will seamlessly intercept `sonar` (or any other mapped names like `gemini-pro`, `nemotron`, `glm-5.2`, `claude-sonnet-5`) and correctly route it to Perplexity's API. You can also create an alias in your shell to make this easier: `alias codex-pplx="codex --local-provider lmstudio"`.
 
 ### API Model Names
 
 | API Name                        | Perplexity Model   | Thinking |
 | ------------------------------- | ------------------ | -------- |
 | `perplexity-auto`               | Best (auto-select) | No       |
-| `gpt-5.4`                       | GPT-5.4            | Toggle   |
-| `claude-sonnet-5-0`             | Claude Sonnet 5.0  | Toggle   |
+| `gpt-5.6-terra`                 | GPT-5.6 Terra      | Toggle   |
+| `gpt-5.6-sol`                   | GPT-5.6 Sol        | Toggle   |
+| `grok-4.5`                      | Grok 4.5           | Toggle   |
+| `claude-sonnet-5`               | Claude Sonnet 5    | Toggle   |
 | `claude-opus-4-8`               | Claude Opus 4.8    | Toggle   |
 | `gemini-3.1-pro`                | Gemini 3.1 Pro     | Always   |
 | `glm-5.2` / `glm52`             | GLM 5.2            | Always   |
@@ -585,6 +589,12 @@ The skill follows Anthropic's Agent Skills open standard and works across any co
 
 Originally forked from [perplexity-webui-scraper](https://github.com/henrique-coder/perplexity-webui-scraper) by [henrique-coder](https://github.com/henrique-coder).
 
+## Support
+
+If Perplexity Web MCP saves you time or money, you can help cover the AI bills that go into building and testing it. Any support is hugely appreciated. 🙏
+
+<a href="https://buymeacoffee.com/jacobbd"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="42"></a>
+
 ## License
 
 MIT

{perplexity_web_mcp_cli-0.14.0 → perplexity_web_mcp_cli-0.14.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "perplexity-web-mcp-cli"
-version = "0.14.0"
+version = "0.14.2"
 description = "CLI, MCP server, and Anthropic/OpenAI API-compatible interface for Perplexity AI."
 authors = [{ name = "Jacob BD" }]
 license = "MIT"

{perplexity_web_mcp_cli-0.14.0 → perplexity_web_mcp_cli-0.14.2}/src/perplexity_web_mcp/api/server.py

@@ -20,10 +20,10 @@ Claude Code Integration:
     export ANTHROPIC_API_KEY=""
 
   Then run Claude Code with any supported model:
-    claude --model claude-sonnet-5-0      # Use Claude Sonnet 5.0 via Perplexity
-    claude --model gpt-5.4                # Use GPT-5.4 via Perplexity
+    claude --model claude-sonnet-5-0      # Use Claude Sonnet 5 via Perplexity
+    claude --model gpt-5.6-terra          # Use GPT-5.6 Terra via Perplexity
     claude --model perplexity-auto        # Use Perplexity's auto model selection
-    claude --model claude-3-5-sonnet      # Legacy name, maps to Claude Sonnet 5.0
+    claude --model claude-3-5-sonnet      # Legacy name, maps to Claude Sonnet 5
 """
 
 from __future__ import annotations
@@ -50,10 +50,7 @@ from perplexity_web_mcp import ConversationConfig, Models, Perplexity
 
 # Tool calling disabled for now - models don't reliably follow format instructions
 # from perplexity_web_mcp.api.tool_calling import (...)
-from perplexity_web_mcp.api.session_manager import (
-    ConversationManager,
-    distill_system_prompt,
-)
+from perplexity_web_mcp.api.session_manager import ConversationManager
 from perplexity_web_mcp.enums import CitationMode
 from perplexity_web_mcp.models import Model
 from perplexity_web_mcp.token_store import load_token
@@ -117,7 +114,7 @@ MODEL_MAP: dict[str, tuple[Model, Model | None]] = {
     "deep-research": (Models.DEEP_RESEARCH, None),
     # ==========================================================================
     # Anthropic Claude Models (via Perplexity)
-    # Claude Sonnet 5.0 - supports thinking toggle
+    # Claude Sonnet 5 - supports thinking toggle
     # Claude Opus 4.8 - supports thinking (requires Max subscription)
     # ==========================================================================
     # Current model names
@@ -157,16 +154,17 @@ MODEL_MAP: dict[str, tuple[Model, Model | None]] = {
     "sonnet": (Models.CLAUDE_50_SONNET, Models.CLAUDE_50_SONNET_THINKING),
     "opus": (Models.CLAUDE_48_OPUS, Models.CLAUDE_48_OPUS_THINKING),
     # ==========================================================================
-    # OpenAI GPT Models (via Perplexity) - support thinking toggle
+    # OpenAI GPT and xAI Grok Models (via Perplexity) - support thinking toggle
     # ==========================================================================
-    "gpt-5.4": (Models.GPT_54, Models.GPT_54_THINKING),
-    "gpt-5-4": (Models.GPT_54, Models.GPT_54_THINKING),
-    "gpt-54": (Models.GPT_54, Models.GPT_54_THINKING),
-    "gpt54": (Models.GPT_54, Models.GPT_54_THINKING),
-    "gpt-5.5": (Models.GPT_55, Models.GPT_55_THINKING),
-    "gpt-5-5": (Models.GPT_55, Models.GPT_55_THINKING),
-    "gpt-55": (Models.GPT_55, Models.GPT_55_THINKING),
-    "gpt55": (Models.GPT_55, Models.GPT_55_THINKING),
+    "gpt-5.6-terra": (Models.GPT_56_TERRA, Models.GPT_56_TERRA_THINKING),
+    "gpt-5-6-terra": (Models.GPT_56_TERRA, Models.GPT_56_TERRA_THINKING),
+    "gpt56_terra": (Models.GPT_56_TERRA, Models.GPT_56_TERRA_THINKING),
+    "gpt-5.6-sol": (Models.GPT_56_SOL, Models.GPT_56_SOL_THINKING),
+    "gpt-5-6-sol": (Models.GPT_56_SOL, Models.GPT_56_SOL_THINKING),
+    "gpt56_sol": (Models.GPT_56_SOL, Models.GPT_56_SOL_THINKING),
+    "grok-4.5": (Models.GROK_45, Models.GROK_45_THINKING),
+    "grok-4-5": (Models.GROK_45, Models.GROK_45_THINKING),
+    "grok45": (Models.GROK_45, Models.GROK_45_THINKING),
     # ==========================================================================
     # Google Gemini Models (via Perplexity)
     # Gemini 3.1 Pro: thinking ALWAYS enabled (no toggle in UI)
@@ -205,20 +203,22 @@ AVAILABLE_MODELS = [
     {"id": "perplexity-auto", "description": "Best - Automatically selects optimal model"},
     {"id": "perplexity-sonar", "description": "Sonar 2 - Perplexity's latest in-house model"},
     {"id": "perplexity-research", "description": "Deep Research - In-depth reports with sources"},
+    # OpenAI
+    {"id": "gpt-5.6-terra", "description": "GPT-5.6 Terra - OpenAI's versatile model, thinking toggle available"},
+    {"id": "gpt-5.6-sol", "description": "GPT-5.6 Sol - OpenAI's most powerful model, Max tier required"},
     # Google Gemini
     {"id": "gemini-3.1-pro", "description": "Gemini 3.1 Pro - Advanced, thinking always on"},
-    # OpenAI
-    {"id": "gpt-5.4", "description": "GPT-5.4 - OpenAI's versatile model, thinking toggle available"},
-    {"id": "gpt-5.5", "description": "GPT-5.5 - OpenAI's latest, Max tier required, thinking toggle available"},
     # Anthropic Claude
-    {"id": "claude-sonnet-5-0", "description": "Claude Sonnet 5.0 - Fast, thinking toggle available"},
+    {"id": "claude-sonnet-5", "description": "Claude Sonnet 5 - Fast, thinking toggle available"},
     {"id": "claude-opus-4-8", "description": "Claude Opus 4.8 - Advanced reasoning, Max tier required"},
     # Z.ai
     {"id": "glm-5.2", "description": "GLM 5.2 - Z.ai advanced model, thinking always on"},
-    # NVIDIA
-    {"id": "nemotron-3-ultra", "description": "Nemotron 3 Ultra - NVIDIA 120B, thinking always on"},
     # Moonshot
     {"id": "kimi-k2.6", "description": "Kimi K2.6 - Advanced, thinking toggle available"},
+    # xAI
+    {"id": "grok-4.5", "description": "Grok 4.5 - xAI's most advanced model, thinking toggle available"},
+    # NVIDIA
+    {"id": "nemotron-3-ultra", "description": "Nemotron 3 Ultra - NVIDIA 550B, thinking always on"},
 ]
 
 
@@ -528,6 +528,7 @@ perplexity_semaphore: asyncio.Semaphore
 # Track last request time for rate limiting
 last_request_time: float = 0.0
 MIN_REQUEST_INTERVAL: float = 5.0  # Minimum seconds between Perplexity requests (Perplexity rate limits)
+API_CONTEXT_MAX_CHARS: int = int(os.getenv("PWM_API_CONTEXT_MAX_CHARS", "60000"))
 
 
 # =============================================================================
@@ -633,29 +634,85 @@ def messages_to_query(messages: list[MessageParam]) -> str:
     return "\n\n".join(parts)
 
 
+def bound_context_text(text: str | None, max_chars: int = API_CONTEXT_MAX_CHARS) -> str:
+    """Preserve prompt context within a bounded character budget."""
+    if not text:
+        return ""
+
+    stripped = text.strip()
+    if len(stripped) <= max_chars:
+        return stripped
+    if max_chars <= 0:
+        return ""
+
+    marker = f"\n\n[... {len(stripped)} characters of API context truncated to fit budget ...]\n\n"
+    available = max_chars - len(marker)
+    if available <= 0:
+        return stripped[:max_chars]
+
+    head_chars = available // 2
+    tail_chars = available - head_chars
+    return f"{stripped[:head_chars].rstrip()}{marker}{stripped[-tail_chars:].lstrip()}"
+
+
+def build_query_with_system_context(query: str, system_text: str | None) -> str:
+    """Prefix a user/conversation query with preserved system/workspace context."""
+    context = bound_context_text(system_text)
+    if not context:
+        return query
+
+    return f"[System instructions and workspace context]\n{context}\n\n[Conversation]\n{query}"
+
+
+def openai_messages_to_search_query(messages: list[OpenAIChatMessage]) -> str:
+    """Return the latest user text for Perplexity search initialization."""
+    for msg in reversed(messages):
+        if msg.role == "user":
+            text = msg.get_text().strip()
+            if text:
+                return text
+
+    return openai_messages_to_query(messages)[:500]
+
+
 def openai_messages_to_query(messages: list[OpenAIChatMessage]) -> str:
     """Convert OpenAI chat messages to Perplexity query.
 
-    Note: System messages are intentionally NOT included to avoid URL length
-    limits with Perplexity.
+    System and developer messages are preserved because agent runtimes use them
+    for workspace bootstrap context. URL length is handled separately by passing
+    a short init_query to Perplexity's search initialization endpoint.
     """
-    # Filter out system messages and get user/assistant messages only
-    conversation_msgs = [m for m in messages if m.role in ("user", "assistant")]
+    supported_msgs = [m for m in messages if m.role in ("system", "developer", "user", "assistant", "tool")]
 
     # For single user message, just return it directly
-    user_msgs = [m for m in conversation_msgs if m.role == "user"]
-    if len(user_msgs) == 1 and len(conversation_msgs) == 1:
+    user_msgs = [m for m in supported_msgs if m.role == "user"]
+    if len(user_msgs) == 1 and len(supported_msgs) == 1:
         return user_msgs[0].get_text()
 
     # Multi-turn: format as conversation
-    parts = []
-    for msg in conversation_msgs:
+    context_parts = []
+    conversation_parts = []
+    role_labels = {
+        "system": "System",
+        "developer": "Developer",
+        "user": "User",
+        "assistant": "Assistant",
+        "tool": "Tool",
+    }
+
+    for msg in supported_msgs:
         text = msg.get_text()
-        if msg.role == "user":
-            parts.append(f"User: {text}")
-        elif msg.role == "assistant":
-            parts.append(f"Assistant: {text}")
+        if not text:
+            continue
+        label = role_labels[msg.role]
+        formatted = f"{label}: {text}"
+        if msg.role in ("system", "developer"):
+            context_parts.append(formatted)
+        else:
+            conversation_parts.append(formatted)
 
+    context = bound_context_text("\n\n".join(context_parts))
+    parts = [part for part in (context, "\n\n".join(conversation_parts)) if part]
     return "\n\n".join(parts)
 
 
@@ -813,7 +870,8 @@ async def create_message(request: Request, body: MessagesRequest):
 
     # Convert messages to query (tool calling not supported via web UI)
     query = messages_to_query(body.messages)
-    input_tokens = estimate_tokens(query)
+    full_query = build_query_with_system_context(query, system_text)
+    input_tokens = estimate_tokens(full_query)
 
     # Generate response ID
     response_id = f"msg_{uuid.uuid4().hex[:24]}"
@@ -822,7 +880,7 @@ async def create_message(request: Request, body: MessagesRequest):
 
     if body.stream:
         return StreamingResponse(
-            stream_response(response_id, body.model, model, query, input_tokens, system_text),
+            stream_response(response_id, body.model, model, full_query, input_tokens, init_query=query),
             media_type="text/event-stream",
             headers={
                 "Cache-Control": "no-cache",
@@ -846,18 +904,12 @@ async def create_message(request: Request, body: MessagesRequest):
                 await asyncio.sleep(wait_needed)
             last_request_time = time_module.time()
 
-            # Prepend distilled system prompt to query (single ask)
-            full_query = query
-            if system_text:
-                distilled = distill_system_prompt(system_text)
-                full_query = f"[Instructions: {distilled}]\n\n{query}"
-
             fresh_client = Perplexity(session_token=config.session_token)
             conversation = fresh_client.create_conversation(
                 ConversationConfig(model=model, citation_mode=CitationMode.DEFAULT)
             )
             # Single ask with prepended system context
-            await asyncio.to_thread(conversation.ask, full_query)
+            await asyncio.to_thread(conversation.ask, full_query, init_query=query)
             fresh_client.close()
         answer = conversation.answer or ""
 
@@ -890,7 +942,7 @@ async def stream_response(
     model: Model,
     query: str,
     input_tokens: int,
-    system_text: str | None = None,
+    init_query: str | None = None,
 ) -> AsyncGenerator[str, None]:
     """Stream Anthropic-format SSE response.
 
@@ -960,13 +1012,6 @@ async def stream_response(
         last = ""
         max_retries = 3
 
-        # Prepend distilled system prompt to query (single ask instead of priming)
-        full_query = query
-        if system_text:
-            distilled = distill_system_prompt(system_text)
-            full_query = f"[Instructions: {distilled}]\n\n{query}"
-            logging.debug(f"Query with system context ({len(distilled)} chars)")
-
         for attempt in range(max_retries):
             try:
                 # Create fresh client for each request to avoid curl_cffi Session issues
@@ -976,7 +1021,7 @@ async def stream_response(
                 )
 
                 # Single ask with prepended system context
-                for resp in conversation.ask(full_query, stream=True):
+                for resp in conversation.ask(query, stream=True, init_query=init_query):
                     current = resp.answer or ""
                     if len(current) > len(last):
                         delta = current[len(last) :]
@@ -1112,6 +1157,7 @@ async def create_chat_completion(request: Request, body: OpenAIChatRequest):
 
     # Convert messages to query
     query = openai_messages_to_query(body.messages)
+    init_query = openai_messages_to_search_query(body.messages)
     input_tokens = estimate_tokens(query)
 
     # Generate response ID
@@ -1122,7 +1168,7 @@ async def create_chat_completion(request: Request, body: OpenAIChatRequest):
 
     if body.stream:
         return StreamingResponse(
-            stream_openai_response(response_id, body.model, model, query, created),
+            stream_openai_response(response_id, body.model, model, query, created, init_query),
             media_type="text/event-stream",
             headers={
                 "Cache-Control": "no-cache",
@@ -1141,7 +1187,7 @@ async def create_chat_completion(request: Request, body: OpenAIChatRequest):
         )
 
         # Run in thread to not block
-        await asyncio.to_thread(conversation.ask, query)
+        await asyncio.to_thread(conversation.ask, query, init_query=init_query)
         answer = conversation.answer or ""
 
         # Append citations if available
@@ -1178,6 +1224,7 @@ async def stream_openai_response(
     model: Model,
     query: str,
     created: int,
+    init_query: str | None = None,
 ) -> AsyncGenerator[str, None]:
     """Stream OpenAI-format SSE response.
 
@@ -1217,7 +1264,7 @@ async def stream_openai_response(
             conversation = client.create_conversation(
                 ConversationConfig(model=model, citation_mode=CitationMode.DEFAULT)
             )
-            for resp in conversation.ask(query, stream=True):
+            for resp in conversation.ask(query, stream=True, init_query=init_query):
                 current = resp.answer or ""
                 if len(current) > len(last):
                     delta = current[len(last) :]