Brave provider and named credentials

This release adds Brave as a first-class web provider and standardizes provider credentials so integrations can support multiple API keys. It also improves search mode guidance, research prompt display, and dependency compatibility.

💥 Breaking changes

Named provider credentials

Provider API credentials now live under the credentials map. Existing configuration files that use provider-specific credential keys such as apiKey or apiToken are rewritten to credentials.api when pi reads the config file.

Before:

{
  "providers": {
    "exa": {
      "apiKey": "EXA_API_KEY"
    },
    "cloudflare": {
      "apiToken": "CLOUDFLARE_API_TOKEN",
      "accountId": "CLOUDFLARE_ACCOUNT_ID"
    }
  }
}

After:

{
  "providers": {
    "exa": {
      "credentials": {
        "api": "EXA_API_KEY"
      }
    },
    "cloudflare": {
      "credentials": {
        "api": "CLOUDFLARE_API_TOKEN"
      },
      "accountId": "CLOUDFLARE_ACCOUNT_ID"
    }
  }
}

The automatic migration applies this change:

• apiKey → credentials.api
• apiToken → credentials.api

Non-credential fields such as baseUrl, accountId, options, and settings stay at their current locations.

This enables providers that require multiple credentials while keeping single-key providers consistent.

By @mavam and @codex in #18.

🚀 Features

Brave web provider

pi-web-providers can now route web_search, web_answer, and web_research through Brave.

Configure Brave with separate credentials for Search and Answers:

{
  "tools": {
    "search": "brave",
    "answer": "brave",
    "research": "brave"
  },
  "providers": {
    "brave": {
      "credentials": {
        "search": "BRAVE_SEARCH_API_KEY",
        "answers": "BRAVE_ANSWERS_API_KEY"
      }
    }
  }
}

web_search uses Brave Web Search by default to return source titles, URLs, snippets, and metadata. It also supports query-time modes for specialized Brave endpoints:

{
  "mode": "llm_context"
}

The llm_context mode uses Brave LLM Context for query-based retrieval and places extracted source chunks directly in search snippets so agents can read them from normal web_search output. The images mode maps Brave Image Search results to their source pages while preserving image URLs and dimensions as metadata. The places mode maps Brave Place Search results to source-like entries with place metadata such as address, categories, rating, and provider URLs.

web_answer and web_research use Brave Answers. Answers stream grounded responses with citations and render a Sources: section when Brave returns citation data. Research enables Brave's research mode, normalizes the request to enable_research=true, stream=true, and enable_citations=false, and returns the generated report through pi's normal web_research artifact flow.

web_contents remains separate: use a URL-fetch-capable provider for page extraction, while Brave LLM Context remains available as a search/retrieval mode for query-grounded context.

By @mavam and @codex in #18.

🔧 Changes

Clearer search mode guidance

Search tools now describe provider-specific modes more clearly so agents can make better use of each provider's native API options.

For example, local business queries can select places-style search where available, current-event queries can select news modes, and direct lookups can avoid heavier exploratory modes. This improves result relevance without requiring users to know the provider-specific option names.

By @mavam in #18.

🐞 Bug fixes

Dependency compatibility updates

Package dependencies are up to date with their latest compatible releases.

This update bumps the provider SDKs and development toolchain used by pi-web-providers:

Package	From	To
@anthropic-ai/claude-agent-sdk	^0.2.119	^0.2.123
@tavily/core	^0.7.2	^0.7.3
openai	^6.34.0	^6.35.0
@mariozechner/pi-ai	^0.70.2	^0.70.6
@mariozechner/pi-coding-agent	^0.70.2	^0.70.6
@mariozechner/pi-tui	^0.70.2	^0.70.6

By @mavam in #18.

Full expanded web research prompts

Expanded web_research calls now provide a useful Markdown detail view without letting long prompts dominate the transcript.

Collapsed tool calls stay compact:

web_research What is pi coding agent? Provide a concise overview...
Started web research via Brave (ctrl+o to expand)

Expanding the dispatch result now shows the full submitted brief and report path in a structured Markdown layout:

Started web research via Brave.

## Research brief

What is pi coding agent? Provide a concise overview of its purpose, main features, modes of operation, customization options, and where to find documentation.

## Report path

`.pi/artifacts/research/...md`

This makes the collapsed state scannable while preserving the exact research brief for review when expanded.

By @mavam and @codex in #18.