dehub/SPEC.md

# SPEC

This document describes the dehub protocol.

This document assumes that the reader is familiar with git, both conceptually
and in practical use of the git tool. All references to a git-specific concept
retain their meaning; dehub concepts build upon git concepts, but do not
override them.

## dehub Project

A dehub project is comprised of:

* A collection of files and directories.

* Meta actions related to those files, e.g. discussion, proposed changes, etc.

* Configuration defining which meta actions are allowed under which
  circumstances.

All of these components are housed in a git repository. A dehub project does not
require a central repository location (a "remote"), though it may use one if
desired.

## Commit Payload

All commits in a dehub project contain a payload. The payload is encoded into
the commit message as a YAML object. Here is the general structure of a commit
message containing a payload:

```
Human readable message head

---
# Three dashes indicate the start of the yaml body.

type: type of the payload # Always required
fingerprint: std-base-64 string # Always required
credentials:[...] # Not required but usually present

type_specific_field_a: valueA
type_specific_field_b: valueB
```

The message head is a human readable description of what is being committed, and
is terminated at the first newline. Everything after the message head must be
valid YAML which encodes the payload.

### Fingerprint

Each payload object contains a `fingerprint` field. The fingerprint is an opaque
byte string encoded using standard base-64. The algorithm used to generate the
fingerprint will depend on the payload type, and can be found in each type's
sub-section in this document.

### Credential

The `credentials` field is not requires, but in practice will be found on almost
every payload.  The field's value will be an array of credential objects. Only
one credential object is currently supported, `pgp_signature`:

```yaml
type: pgp_signature

# One of these fields is required. If account_id is present, it relates the
# signature to a pgp_public_key signifier defined for that account in the config
# (see the Signifier sub-section). Otherwise, the public key will be included in
# the credential itself as the value of pub_key_body.
account_id: some_user_id # Optional
pub_key_body: inlined ASCII-armored pgp public key

# the ID (pgp fingerprint) of the key used to generate the signature
pub_key_id: XXX

# a signature of the payload's unencoded fingerprint, encoded using standard
# base-64
body: std-base-64 signature
```

### Payload Types

#### Change Payload

A change payload encompasses a set of changes to the files in the project. To
construct the change payload one must reference the file tree of the commit
which houses the payload as well as the file tree of its parent commit;
specifically one must take the difference between them.

A change payload looks like this:

```yaml
type: change
fingerprint: std-base-64 string
credentials: [...]
description: |-
    The description will generally start with a single line, followed by a long-form body

    The description corresponds to the body of a commit message in a "normal"
    git repo. It gives a more-or-less long-form explanation of the changes being
    made to the project's files.    
```

##### Change Payload Fingerprint

The unencoded fingerprint of a change payload is calculated as follows:

* Concatenate the following:
  * A uvarint indicating the number of bytes in the description string.
  * The description string.
  * A uvarint indicating the number of files changed between this commit and
    its parent.
  * For each file changed, ordered lexographically-ascending based on its full
    relative path within the git repo:
    * A uvarint indicating the length of the full relative path of the file
      within the repo, as a string.
    * The full relative path of the file within the repo, as a string.
    * A little-endian uint32 representing the previous file mode of the file
      (or 0 if the file is not present in the parent commit's tree).
    * The 20-byte SHA1 hash of the contents of the previous version of the file
      (or 20 0 bytes if the file is not present in the parent commit's tree).
    * A little-endian uint32 representing the new file mode of the file (or 0
      if the file is not present in the current commit's tree).
    * The 20-byte SHA1 hash of the contents of the new version of the file (or
      20 0 bytes if the file is not present in the current commit's tree).
* Calculate the SHA-256 hash of the concatenation result.
* Prepend a 0 byte to the result of the SHA-256 hash.

This unencoded fingerprint is then standard base-64 encoded, and that is used as
the value of the fingerprint field.

#### Comment Payload

A comment payload encompasses no file changes, and is used only to contain a
comment made by a single user.

A comment payload looks like this:

```yaml:
type: comment
fingerprint: std-base-64 string
credentials: [...]
comment: |-
    Hey all, how's it going?

    Just wanted to pop by and say howdy.
```

The message head of a comment payload will generally be a truncated form of the
comment itself.

##### Comment Payload Fingerprint

The unencoded fingerprint of a comment payload is calculated as follows:

* Concatenate the following:
  * A uvarint indicating the number of bytes in the comment string.
  * The comment string.
* Calculate the SHA-256 hash of the concatenation result.
* Prepend a 0 byte to the result of the SHA-256 hash.

This unencoded fingerprint is then standard base-64 encoded, and that is used as
the value of the fingerprint field.

#### Credential Payload

A credential payload contains only one or more credentials for an arbitrary
fingerprint. Credential payloads can be combined with other payloads of the same
fingerprint to create a new payload with many credentials.

A credential payload looks like this:

```yaml
type: credential
fingerprint: std-base-64 string
credentials: [...]

# This field is not required, but can be helpful in situations where the
# fingerprint was generated based on multiple change payloads
fingerprint_commits:
    - commit hash
    - commit hash
    - commit hash
```

## Project Configuration

The `.dehub` directory contains all meta information related to the dehub
project. All files within `.dehub` are tracked by the git repo like any other
files in the project.

### config.yml

The `.dehub/config.yml` file contains a yaml encoded configuration object:

```yaml
accounts: [...]
access_controls: [...]
```

Both fields are described in their own sub-section below.

#### Account

An account defines a specific user of the project. Every account has an ID; no
two accounts within a project may share the same ID.

An account looks like this:

```yaml
id: some_string
signifiers: [...]
```

##### Signifier

A signifier is used to signify that an account has taken some action. The most
common use-case is to prove that an account created a particular credential. An
account may have more than one signifier.

Currently there is only one signifier type, `pgp_public_key`:

```yaml
type: pgp_public_key

# Path to ASCII-armored pgp public key, relative to repo root.
path: .dehub/account.asc
```

or

```yaml
type: pgp_public_key
body: inlined ASCII-armored pgp public key
```

#### Access Control

An access control allows or denies a particular commit from becoming a part of
the project. Each action control has an action (allow or deny) and a set of
filters:

```yaml
action: allow # or deny
filters: [...]
```

When a verifying a commit against a project's access controls, each access
control's filters are applied to the commit in the order they appear in the
configuration. The first access control for which all filters match is found,
and its action is taken.

An access control with no filters matches all commits.

##### Filters

There are many kinds of access control filters. Any filter can be applied to a
commit, with no other input, and produce a boolean value. All filters have a
`type` field which indicates their type.

###### Signature Filter

A filter of type `signature` asserts that a commit's payload contains signature
credentials with certain properties. A signature filter must have one of these
fields, which define the set of users or accounts whose signatures are
applicable.

* `account_ids: [...]` - an array of account IDs, each having been defined in the
  accounts section of the configuration.

* `any_account: true` - matches any account defined in the accounts section of
  the configuration.

* `any: true` - matches any signature, whether or not its signifier has been
  defined in the configuration.

A `count` field may also be included. Its value may be a number or a string
indicating a percent (e.g. `"50%"`). If not included it will be assumed to be
`1`.

The count indicates how many accounts from the specified set must have a
signature included. If a percent is given then that will be multiplied against
the size of the set (rounded up) to determine the necessary number.

Here are some example signature filters, and explanations for each:

```yaml
# requires that 2 of the 3 specified accounts has a signature credential on
# the commit.
type: signature
account_ids:
    - amy
    - bill
    - colleen
count: 2
```

```yaml
# requires that every account defined in the configuration has a signature
# credential on the commit.
type: signature
any_account: true
count: 100%
```

```yaml
# requires at least one signature credential, not necessarily from an account.
type: signature
any: true
```

###### Branch Filter

A filter of type `branch` matches the commit based on which branch in the repo
it is being or has been committed to. Matching is performed on the short name
of the branch, using globstar pattern matching.

A branch filter can have one or multiple patterns defined. The filter will match
if at least one defined pattern matches the short form of the branch name.

A branch filter with only one pattern can be defined like this:

```yaml
type: branch
pattern: some_branch
```

A branch filter with multiple patterns can be defined like this:

```yaml
type: branch
patterns:
    - some_branch
    - branch*glob
    - amy/**
```

###### Files Changed Filter

A filter of type `files_changed` matches the commit based on which files were
changed between the tree of the commit's parent and the commit's tree. Matching
is performed on the paths of the changed files, relative to the repo root.

A files changed filter can have one or multiple patterns defined. The filter
will match if at least one defined pattern matches for every file changed.

(TODO this may change to be: `The filter will match if any of the changed files
matches at least one defined pattern.`)

A files changed filter with only one pattern can be defined like this:

```yaml
type: files_changed
pattern: .dehub/*
```

A files changed filter with multiple patterns can be defined like this:

```yaml
type: files_changed
patterns:
    - some/dir/*
    - foo_files_*
    - **.jpg
```

###### Payload Type Filter

A filter of type `payload_type` matches a commit based on the type of its
payload. A payload type filter can have one or more types defined. The filter
will match if the commit's payload type matches at least one of the defined
types.

A payload type filter with only one matching type can be defined like this:

```yaml
type: payload_type
payload_type: comment
```

A payload type filter with multiple matching types can be defined like this:

```yaml
type: payload_type
payload_types:
    - comment
    - change
```

###### Commit Attributes Filter

A filter of type `commit_attributes` matches a commit based on certain
attributes it has. A commit attributes filter may have one or more fields
defined, each corresponding to a different attribute the commit may have. If
more than one field is defined then all corresponding attributes on the commit
must match for the filter to match.

Currently the only possible attribute is `non_fast_forward: true`, which matches
a commit which is not an ancestor of the HEAD of the branch it's being pushed
onto. This attribute only makes sense in the context of a pre-receive git hook.

A commit attributes filter looks like this:

```yaml
type: commit_attributes
non_fast_forward: true
```

###### Not Filter

A filter of type `not` matches a commit using the negation of a sub-filter,
defined within the not filter. If the sub-filter returns true for the commit,
then the not filter returns false, and vice-versa.

A not filter looks like this:

```
type: not
filter:
    # a branch filter is used as the sub-filter in this example
    type: branch
    pattern: main
```

##### Default Access Controls

These access controls will be implicitly appended to the list defined in the
configuration:

```yaml
# Any account may add any commit to any non-main branch, provided there is at
# least one signature credential. This includes non-fast-forwards.
- action: allow
  filters:
  - type: not
    filter:
      type: branch
      pattern: main
  - type: signature
    any_account: true
    count: 1

# Non-fast-forwards are denied in all other cases. In effect, one cannot
# force-push onto the main branch.
- action: deny
  filters:
  - type: commit_attributes
    non_fast_forward: true

# Any account may add any change commit to the main branch, provided there is
# at least one signature credential.
- action: allow
  filters:
  - type: branch
    pattern: main
  - type: payload_type
    payload_type: change
  - type: signature
    any_account: true
    count: 1

# All other actions are denied.
- action: deny
```

These default access controls provide a useful baseline of requirements that all
projects will (hopefully) find useful in their infancy.

## Commit Verification

The dehub protocol is designed such that every commit is "verifiable". A
verifiable commit has the following properties:

* Its fingerprint is correctly formed.
* All of its credentials are correctly formed.
    * If they are signatures, they are valid signatures of the commit's
      unencoded fingerprint.
* The project's access controls allow the commit.

The project's configuration is referenced frequently when verifying a commit,
such as when determining which access controls to apply and discovering
signifiers of accounts. In all cases the configuration as defined in the
commit's _parent_ is used when verifying that commit. The exception is the prime
commit, which uses its own configuration.

### Prime Commit

The prime commit is the trusted seed of the project. When a user clones and
verifies a dehub project they must, implicitly or explicitly, trust the contents
of the prime commit. All other commits must be ancestors of the prime commit.

Manually specifying a prime commit is not currently spec'd, but it will be.

By default the prime commit is the root commit of the `main` branch.