pi-lore

内容来源：README.md（说明文档） · 原始地址 · 查看安装指南

原始内容

# Lore

Compiler-aware tools that help coding agents understand and modify Haskell projects while keeping large-codebase context pressure low.

Lore loads a project through the GHC API and exposes structured operations for project discovery, targeted source retrieval, symbol navigation, compilation diagnostics, typeclass instances, type inference, code evaluation, tests, and dead-code analysis. The tools let agents retrieve project facts and source slices instead of repeatedly filling context with whole files, broad text-search output, or raw build logs.

The [tool guide](https://github.com/catdarick/lore/blob/HEAD/docs/Tools.md) is the canonical tool reference. It lists every shared Lore tool, explains the context-pressure benefit, and links to per-tool input/output examples.

## Choose a frontend

| Frontend | Best fit | More details |
| --- | --- | --- |
| `pi-lore` | Pi users who want managed server setup, branch-aware definition memory, recovery summaries, settings, status, and usage statistics. | [`pi-lore/README.md`](https://github.com/catdarick/lore/blob/HEAD/pi-lore/README.md) |
| `lore-mcp` | MCP clients that can launch a local stdio server directly. | [`lore-mcp/README.md`](https://github.com/catdarick/lore/blob/HEAD/lore-mcp/README.md) |
| `lore-cli` | Shell, scripting, CI, or interactive terminal exploration. | [`lore-tools-cli/README.md`](https://github.com/catdarick/lore/blob/HEAD/lore-tools-cli/README.md) |

## Quick start for a target project

After choosing a frontend, add a small `lore.yaml` at the root of the Haskell project being inspected. This keeps Lore's output focused from the first run and avoids spending context on noisy test output or repeated definitions.

```yaml
session:
  project-root: .
  ghc-work-dir: .lore-work

  # Keep test-tool output focused. These defaults are useful for Hspec.
  default-test-args:
    - --format=failed-examples
    - --no-color

# Add public modules, plugin entry points, framework callbacks, or other
# externally-called code that should not be reported as dead.
dead-code:
  alive-modules:
    - MyLibrary.Public
  alive-symbols:
    - MyLibrary.runServer

# Add project vocabulary so symbol search matches local naming conventions.
symbol-search:
  synonym-groups:
    - [account, profile]
    - [author, writer]

mcp:
  # Raw lore-mcp users: recommended, because it avoids repeating unchanged
  # definitions in later getDefinitions responses. pi-lore users do not need
  # to set this here because pi-lore enables and manages it automatically.
  enable-definition-knowledge-cache: true

  tools:
    # Raw lore-mcp users only: enable this if the MCP client summarizes or
    # compacts chats, then instruct the agent to call it only after such a
    # summarization/reset. Otherwise restart lore-mcp after summarization.
    # pi-lore users should leave this disabled; pi-lore tracks resets itself.
    notifyKnowledgeReset: true

    # Disable tools that should not be exposed in this project.
    executeCode: false
```

Recommended first checks:

1. Verify the `lore-mcp` or `lore-cli` binary was built with the exact GHC version used by the target project.
2. Run `discoverProject` to confirm Lore sees the expected packages and components.
3. Run `reloadHomeModules` to load the project and surface focused GHC diagnostics.
4. Run a narrow `runTestSuite` call only after test defaults are configured to keep output concise.

For `pi-lore` users, definition-knowledge caching and reset tracking are automatic; no extra `lore.yaml` cache settings are needed. For raw `lore-mcp` users, enabling `enable-definition-knowledge-cache` is recommended. If the MCP client summarizes or compacts chats, also enable `notifyKnowledgeReset` and instruct the agent to call it only after that summarization/reset; otherwise restart `lore-mcp` after summarization.

Use the [`lore-mcp` configuration guide](https://github.com/catdarick/lore/blob/HEAD/lore-mcp/README.md#configuration) for configuration precedence, environment variables, and path behavior. Use the [tool guide](https://github.com/catdarick/lore/blob/HEAD/docs/Tools.md) for per-tool inputs, outputs, and tool-owned configuration semantics.

## Requirements

### Building Lore

- GHC 9.6 or newer within the package's supported bounds.
- Cabal.
- `hpack` when regenerating Cabal files from `package.yaml` or when a target project requires its generated Cabal file to be refreshed.
- Node.js 24 or newer only when developing or packaging `pi-lore`.

### Inspecting a project

The target project's build tool and compiler must be available where Lore runs. Lore supports Stack and Cabal projects; provider detection and path behavior are documented in the [`lore-mcp` configuration guide](https://github.com/catdarick/lore/blob/HEAD/lore-mcp/README.md#configuration).

## GHC compatibility

Lore links against the GHC API, so a `lore-mcp` or `lore-cli` binary must be built with the **same full GHC version** as the project it inspects.

For example, a binary built with GHC 9.6.5 must not be used for a project running GHC 9.6.7.

Check a project compiler with:

```bash
# Cabal project
cabal exec --write-ghc-environment-files=never -- ghc --numeric-version

# Stack project
stack exec -- ghc --numeric-version
```

Check a Lore server binary with:

```bash
lore-mcp --version-json
```

`pi-lore` validates downloaded and manually configured servers before starting them.

## Configuration overview

`lore-mcp` and `lore-cli` read optional project configuration from `lore.yaml`. A frontend such as `pi-lore` may provide environment overrides when it starts the server.

For `lore-mcp` and `lore-cli`, configuration precedence is:

```text
built-in defaults < lore.yaml < environment variables
```

| Configuration area | Canonical doc |
| --- | --- |
| Haskell session, project root, GHC work directory, dead-code roots, symbol-search synonyms, MCP tool enablement, feedback, and custom command tools. | [`lore-mcp` configuration](https://github.com/catdarick/lore/blob/HEAD/lore-mcp/README.md#configuration) |
| Pi-specific startup, managed binary selection, tool proxying, recovery, timeout, and state settings. | [`pi-lore` configuration](https://github.com/catdarick/lore/blob/HEAD/pi-lore/README.md#configuration) |
| Per-tool input, output, examples, and tool-owned configuration semantics. | [Tool guide](https://github.com/catdarick/lore/blob/HEAD/docs/Tools.md) |

## Build the repository

Clone the repository and build all Cabal packages:

```bash
git clone https://github.com/catdarick/lore.git
cd lore

cabal build all
```

Build only the end-user executables:

```bash
cabal build exe:lore-mcp exe:lore-cli
```

Find the resulting executable paths:

```bash
cabal list-bin exe:lore-mcp
cabal list-bin exe:lore-cli
```

Run the Haskell test suites:

```bash
cabal test all
```

Build an optimized server binary:

```bash
cabal build exe:lore-mcp --enable-optimization=2
```

The repository also contains Stack configuration for supported development workflows, but Cabal is the primary build path documented here.

## Repository layout

| Path | Purpose |
| --- | --- |
| [`lore/`](https://github.com/catdarick/lore/tree/HEAD/lore/) | Core GHC session, loading, analysis, interpreter, configuration, and project support. |
| [`lore-tools/`](https://github.com/catdarick/lore/tree/HEAD/lore-tools/) | Shared tool operations, structured results, and rendering used by frontends. |
| [`lore-mcp/`](https://github.com/catdarick/lore/tree/HEAD/lore-mcp/) | MCP protocol server, tool schemas, MCP configuration, and custom command tools. |
| [`lore-tools-cli/`](https://github.com/catdarick/lore/tree/HEAD/lore-tools-cli/) | Interactive and single-command terminal frontend exposed as `lore-cli`. |
| [`pi-lore/`](https://github.com/catdarick/lore/tree/HEAD/pi-lore/) | Pi package that manages `lore-mcp` and adds Pi-specific context and recovery behavior. |
| [`docs/Tools.md`](https://github.com/catdarick/lore/blob/HEAD/docs/Tools.md) | Canonical shared tool guide and links to per-tool pages. |
| [`scripts/`](https://github.com/catdarick/lore/tree/HEAD/scripts/) | Repository maintenance and release scripts. |
| [`.github/workflows/`](https://github.com/catdarick/lore/tree/HEAD/.github/workflows/) | Build, release, binary packaging, and npm publication workflows. |

The dependency direction is approximately:

```text
lore
  ↓
lore-tools
  ├─→ lore-mcp
  └─→ lore-tools-cli

lore-mcp ← managed by → pi-lore
```

## Development workflow

After changing Haskell code:

```bash
cabal build all
cabal test all
```

After changing the Pi package:

```bash
cd pi-lore
npm test
npm run validate:package
```

When modifying `package.yaml`, regenerate the corresponding Cabal file with the repository's supported `hpack` version before committing both source-of-truth and generated manifest changes.

Useful repository tasks are also available through `Taskfile.yml`, including formatting, building, release version updates, and MCP Inspector startup.

## Security

Lore can compile project code, evaluate expressions, run test suites, and execute configured shell-command tools. These capabilities are appropriate only for projects and configuration that the developer trusts.

For direct MCP integrations, review which tools are enabled before exposing the server to an agent. `runTestSuite`, custom tools, and evaluation features can be disabled in `lore.yaml`.

## License

BSD-3-Clause.

Lore

Compiler-aware tools that help coding agents understand and modify Haskell projects while keeping large-codebase context pressure low.

Lore loads a project through the GHC API and exposes structured operations for project discovery, targeted source retrieval, symbol navigation, compilation diagnostics, typeclass instances, type inference, code evaluation, tests, and dead-code analysis. The tools let agents retrieve project facts and source slices instead of repeatedly filling context with whole files, broad text-search output, or raw build logs.

The tool guide is the canonical tool reference. It lists every shared Lore tool, explains the context-pressure benefit, and links to per-tool input/output examples.

Choose a frontend

Frontend	Best fit	More details
`pi-lore`	Pi users who want managed server setup, branch-aware definition memory, recovery summaries, settings, status, and usage statistics.	`pi-lore/README.md`
`lore-mcp`	MCP clients that can launch a local stdio server directly.	`lore-mcp/README.md`
`lore-cli`	Shell, scripting, CI, or interactive terminal exploration.	`lore-tools-cli/README.md`

Quick start for a target project

After choosing a frontend, add a small lore.yaml at the root of the Haskell project being inspected. This keeps Lore's output focused from the first run and avoids spending context on noisy test output or repeated definitions.

session:
  project-root: .
  ghc-work-dir: .lore-work

  # Keep test-tool output focused. These defaults are useful for Hspec.
  default-test-args:
    - --format=failed-examples
    - --no-color

# Add public modules, plugin entry points, framework callbacks, or other
# externally-called code that should not be reported as dead.
dead-code:
  alive-modules:
    - MyLibrary.Public
  alive-symbols:
    - MyLibrary.runServer

# Add project vocabulary so symbol search matches local naming conventions.
symbol-search:
  synonym-groups:
    - [account, profile]
    - [author, writer]

mcp:
  # Raw lore-mcp users: recommended, because it avoids repeating unchanged
  # definitions in later getDefinitions responses. pi-lore users do not need
  # to set this here because pi-lore enables and manages it automatically.
  enable-definition-knowledge-cache: true

  tools:
    # Raw lore-mcp users only: enable this if the MCP client summarizes or
    # compacts chats, then instruct the agent to call it only after such a
    # summarization/reset. Otherwise restart lore-mcp after summarization.
    # pi-lore users should leave this disabled; pi-lore tracks resets itself.
    notifyKnowledgeReset: true

    # Disable tools that should not be exposed in this project.
    executeCode: false

Recommended first checks:

Verify the lore-mcp or lore-cli binary was built with the exact GHC version used by the target project.
Run discoverProject to confirm Lore sees the expected packages and components.
Run reloadHomeModules to load the project and surface focused GHC diagnostics.
Run a narrow runTestSuite call only after test defaults are configured to keep output concise.

For pi-lore users, definition-knowledge caching and reset tracking are automatic; no extra lore.yaml cache settings are needed. For raw lore-mcp users, enabling enable-definition-knowledge-cache is recommended. If the MCP client summarizes or compacts chats, also enable notifyKnowledgeReset and instruct the agent to call it only after that summarization/reset; otherwise restart lore-mcp after summarization.

Use the lore-mcp configuration guide for configuration precedence, environment variables, and path behavior. Use the tool guide for per-tool inputs, outputs, and tool-owned configuration semantics.

Requirements

Building Lore

GHC 9.6 or newer within the package's supported bounds.
Cabal.
hpack when regenerating Cabal files from package.yaml or when a target project requires its generated Cabal file to be refreshed.
Node.js 24 or newer only when developing or packaging pi-lore.

Inspecting a project

The target project's build tool and compiler must be available where Lore runs. Lore supports Stack and Cabal projects; provider detection and path behavior are documented in the lore-mcp configuration guide.

GHC compatibility

Lore links against the GHC API, so a lore-mcp or lore-cli binary must be built with the same full GHC version as the project it inspects.

For example, a binary built with GHC 9.6.5 must not be used for a project running GHC 9.6.7.

Check a project compiler with:

# Cabal project
cabal exec --write-ghc-environment-files=never -- ghc --numeric-version

# Stack project
stack exec -- ghc --numeric-version

Check a Lore server binary with:

lore-mcp --version-json

pi-lore validates downloaded and manually configured servers before starting them.

Configuration overview

lore-mcp and lore-cli read optional project configuration from lore.yaml. A frontend such as pi-lore may provide environment overrides when it starts the server.

For lore-mcp and lore-cli, configuration precedence is:

built-in defaults < lore.yaml < environment variables

Configuration area	Canonical doc
Haskell session, project root, GHC work directory, dead-code roots, symbol-search synonyms, MCP tool enablement, feedback, and custom command tools.	`lore-mcp` configuration
Pi-specific startup, managed binary selection, tool proxying, recovery, timeout, and state settings.	`pi-lore` configuration
Per-tool input, output, examples, and tool-owned configuration semantics.	Tool guide

Build the repository

Clone the repository and build all Cabal packages:

git clone https://github.com/catdarick/lore.git
cd lore

cabal build all

Build only the end-user executables:

cabal build exe:lore-mcp exe:lore-cli

Find the resulting executable paths:

cabal list-bin exe:lore-mcp
cabal list-bin exe:lore-cli

Run the Haskell test suites:

cabal test all

Build an optimized server binary:

cabal build exe:lore-mcp --enable-optimization=2

The repository also contains Stack configuration for supported development workflows, but Cabal is the primary build path documented here.

Repository layout

Path	Purpose
`lore/`	Core GHC session, loading, analysis, interpreter, configuration, and project support.
`lore-tools/`	Shared tool operations, structured results, and rendering used by frontends.
`lore-mcp/`	MCP protocol server, tool schemas, MCP configuration, and custom command tools.
`lore-tools-cli/`	Interactive and single-command terminal frontend exposed as `lore-cli`.
`pi-lore/`	Pi package that manages `lore-mcp` and adds Pi-specific context and recovery behavior.
`docs/Tools.md`	Canonical shared tool guide and links to per-tool pages.
`scripts/`	Repository maintenance and release scripts.
`.github/workflows/`	Build, release, binary packaging, and npm publication workflows.

The dependency direction is approximately:

lore
  ↓
lore-tools
  ├─→ lore-mcp
  └─→ lore-tools-cli

lore-mcp ← managed by → pi-lore

Development workflow

After changing Haskell code:

cabal build all
cabal test all

After changing the Pi package:

cd pi-lore
npm test
npm run validate:package

When modifying package.yaml, regenerate the corresponding Cabal file with the repository's supported hpack version before committing both source-of-truth and generated manifest changes.

Useful repository tasks are also available through Taskfile.yml, including formatting, building, release version updates, and MCP Inspector startup.

Security

Lore can compile project code, evaluate expressions, run test suites, and execute configured shell-command tools. These capabilities are appropriate only for projects and configuration that the developer trusts.

For direct MCP integrations, review which tools are enabled before exposing the server to an agent. runTestSuite, custom tools, and evaluation features can be disabled in lore.yaml.

License

BSD-3-Clause.