feat: relationship model (requires/repos/predicates) + git sync

A derived, mostly-computed model of how programs, deployments, and repos relate,
plus the git-sync surfaces that motivated it. See docs/relationships.md.

Core:
- `requires: [{kind, ref, version?, bind?}]` on programs + deployments — one
  precondition relation; `system_dependencies` is its `{kind: system}` alias.
  kind fixes meaning + check (system=installed, deployment=exists).
- relations.py: derives repos (git toplevel / monorepo), fan-in, and the
  predicates functional?/fresh?/deployed? — nothing stored.
- env is generated FROM a `{kind: deployment, bind}` requirement (target URL →
  consumer env), never scraped back into one; explicit defaults.env still wins.
- git.py: working-copy status/pull, repo toplevel + remote url.

Surfaces:
- `castle graph` + GET /graph — the relationship diagnostic.
- GET /repos, /repos/{key}/git|sync — repo-scoped sync (a repo is the sync unit;
  a monorepo backs several programs). GET /programs/{name}/git|sync + repo context.
- Dashboard: Graph screen, program-page git status + repo-aware Sync, and a
  monorepo banner on Programs.

Governing principle: predicates are derived; encode only the non-derivable, as a
node or edge property. Pull-only sync — converge stays a separate step.
This commit is contained in:
2026-07-05 10:55:42 -07:00
parent 3c566540aa
commit add356dcf2
24 changed files with 1669 additions and 4 deletions

View File

@@ -244,7 +244,9 @@ def apply(
return "restart"
return "unchanged"
result = ApplyResult(registry=NodeRegistry(node=_node_config(config), deployed=desired))
result = ApplyResult(
registry=NodeRegistry(node=_node_config(config), deployed=desired)
)
if plan:
# No writes: for systemd, predict the new unit bytes by rendering to a string
@@ -469,6 +471,36 @@ def _public_url(
return None
def _target_url(config: CastleConfig, target_name: str) -> str | None:
"""The base URL another deployment is reachable at — how a ``{kind: deployment,
bind: VAR}`` requirement projects its target into the consumer's env."""
dep = config.deployments.get(target_name)
if dep is None:
return None
expose = getattr(dep, "expose", None)
http = getattr(expose, "http", None) if expose else None
tport = http.internal.port if http else None
return _public_url(config, target_name, getattr(dep, "http_exposed", False), tport)
def _requires_env(config: CastleConfig, name: str, config_key: str) -> dict[str, str]:
"""Env generated FROM a deployment's ``requires`` — a ``{kind: deployment,
bind: VAR}`` requirement sets ``VAR`` to the target's URL. Env is derived from
the dependency, never scraped back into one (see docs/relationships.md)."""
dep = config.deployments[name]
prog = config.programs.get(config_key)
reqs = list(getattr(dep, "requires", []) or [])
if prog:
reqs += list(prog.requires)
out: dict[str, str] = {}
for r in reqs:
if r.kind == "deployment" and r.bind:
url = _target_url(config, r.ref)
if url:
out[r.bind] = url
return out
def _supabase_app_schemas(config: CastleConfig) -> str:
"""The ``${supabase_app_schemas}`` placeholder: each registered supabase app's
own schema, comma-prefixed and joined (or '' when there are none).
@@ -621,8 +653,14 @@ def _build_deployed(
# names to castle's computed values. Secret-bearing vars split out to a
# mode-0600 file (never in the unit or argv).
raw_env = dict(dep.defaults.env) if (dep.defaults and dep.defaults.env) else {}
# Env generated from `requires` ({kind: deployment, bind: VAR} → target URL).
# An explicit defaults.env value always wins — a hand-set var is never clobbered.
for var, url in _requires_env(config, name, config_key).items():
raw_env.setdefault(var, url)
public_url = _public_url(config, name, expose, port)
ctx = _env_context(name, config_key, port, public_url, _supabase_app_schemas(config))
ctx = _env_context(
name, config_key, port, public_url, _supabase_app_schemas(config)
)
# ${tls_*}: paths to castle-materialized cert files for a TLS-material TCP
# service. The deployment maps them into its own config (mount ${tls_dir} for a
# container, or reference ${tls_cert}/${tls_key} directly for a native service).
@@ -647,7 +685,12 @@ def _build_deployed(
_ensure_python_tool(config, dep.program, messages)
run_cmd = _build_run_cmd(
name, run, env, messages, source_dir, secret_env_file=secret_env_file,
name,
run,
env,
messages,
source_dir,
secret_env_file=secret_env_file,
placeholders=ctx,
)
stop_cmd = _build_stop_cmd(name, run, source_dir)

159
core/src/castle_core/git.py Normal file
View File

@@ -0,0 +1,159 @@
"""Git working-copy status and sync for programs whose source is a git repo.
Programs that declare a ``repo:`` URL are cloned once (``castle program clone``);
this module lets a running castle *see how far behind* a working copy is and pull
later updates. It is intentionally pull-only — it touches files on disk and never
builds, applies, or restarts anything. Making the running artifact reflect the new
code (rebuild a frontend, restart a service) stays an explicit, separate step via
``castle apply`` / ``castle restart``.
Plain ``git`` via subprocess (matching ``castle program clone``); no GitPython.
"""
from __future__ import annotations
import subprocess
from dataclasses import dataclass
from pathlib import Path
# Bound network calls (fetch/pull) so an unreachable remote can't hang a request.
_FETCH_TIMEOUT = 20.0
_PULL_TIMEOUT = 60.0
@dataclass
class GitStatus:
"""A program working copy's git state. ``ahead``/``behind`` are relative to the
upstream tracking branch and reflect the *last fetch* (``git_status(fetch=True)``
refreshes them). ``None`` counts mean "no upstream to compare against"."""
is_repo: bool
branch: str | None = None
upstream: str | None = None
dirty: bool = False
ahead: int | None = None
behind: int | None = None
detached: bool = False
error: str | None = None
def _git(
source: Path, *args: str, timeout: float | None = None
) -> subprocess.CompletedProcess[str]:
"""Run ``git -C <source> <args>`` capturing text output."""
return subprocess.run(
["git", "-C", str(source), *args],
capture_output=True,
text=True,
timeout=timeout,
)
def is_git_repo(source: Path | None) -> bool:
"""True when ``source`` is inside a git working tree."""
if not source or not Path(source).is_dir():
return False
try:
r = _git(
Path(source), "rev-parse", "--is-inside-work-tree", timeout=_FETCH_TIMEOUT
)
except (OSError, subprocess.SubprocessError):
return False
return r.returncode == 0 and r.stdout.strip() == "true"
def toplevel(source: Path | None) -> str | None:
"""The absolute path of the git working copy ``source`` lives in, or None.
The natural identity of a *repo*: several programs whose sources share a
toplevel are the same working copy (a monorepo). Adopted single-program repos
are their own toplevel — the N=1 case."""
if not source or not Path(source).is_dir():
return None
try:
r = _git(Path(source), "rev-parse", "--show-toplevel", timeout=_FETCH_TIMEOUT)
except (OSError, subprocess.SubprocessError):
return None
return r.stdout.strip() or None if r.returncode == 0 else None
def remote_url(source: Path | None) -> str | None:
"""The ``origin`` remote URL of the working copy, or None (no remote)."""
if not is_git_repo(source):
return None
r = _git(Path(source), "remote", "get-url", "origin") # type: ignore[arg-type]
return r.stdout.strip() or None if r.returncode == 0 else None
def git_status(source: Path | None, fetch: bool = True) -> GitStatus:
"""The working copy's branch/dirty/ahead/behind state.
``fetch=True`` runs ``git fetch`` first (bounded, tolerant of an offline remote)
so ``behind`` reflects the real remote; on fetch failure the counts fall back to
the last-known values and ``error`` carries the reason. Never raises — a
non-repo returns ``GitStatus(is_repo=False)`` so callers can just hide the UI.
"""
if not is_git_repo(source):
return GitStatus(is_repo=False)
src = Path(source) # type: ignore[arg-type]
st = GitStatus(is_repo=True)
# Branch (or detached HEAD).
branch = _git(src, "rev-parse", "--abbrev-ref", "HEAD").stdout.strip()
if branch == "HEAD":
st.detached = True
else:
st.branch = branch
# Dirty working tree (staged, unstaged, or untracked).
st.dirty = bool(_git(src, "status", "--porcelain").stdout.strip())
# Best-effort refresh from the remote; failure is non-fatal (offline, no remote).
if fetch:
try:
fr = _git(src, "fetch", "--quiet", timeout=_FETCH_TIMEOUT)
if fr.returncode != 0:
st.error = (fr.stderr or fr.stdout).strip() or "git fetch failed"
except subprocess.TimeoutExpired:
st.error = "git fetch timed out"
except (OSError, subprocess.SubprocessError) as e:
st.error = str(e)
# Upstream tracking branch, then the ahead/behind split against it.
up = _git(src, "rev-parse", "--abbrev-ref", "--symbolic-full-name", "@{u}")
if up.returncode == 0 and up.stdout.strip():
st.upstream = up.stdout.strip()
counts = _git(src, "rev-list", "--left-right", "--count", "@{u}...HEAD")
if counts.returncode == 0:
parts = counts.stdout.split()
if len(parts) == 2:
st.behind, st.ahead = int(parts[0]), int(parts[1])
return st
def head(source: Path | None) -> str | None:
"""The working copy's current commit sha, or None if unavailable. Lets a caller
tell whether a ``pull`` actually advanced the tree (before != after)."""
if not is_git_repo(source):
return None
r = _git(Path(source), "rev-parse", "HEAD") # type: ignore[arg-type]
return r.stdout.strip() or None if r.returncode == 0 else None
def pull(source: Path | None) -> tuple[bool, str]:
"""Fast-forward the working copy to its upstream (``git pull --ff-only``).
``--ff-only`` is deliberate: it refuses to merge, so a dirty or diverged tree
fails cleanly with git's own message instead of creating a merge commit. Returns
``(ok, combined_output)``.
"""
if not is_git_repo(source):
return False, "not a git repository"
try:
r = _git(Path(source), "pull", "--ff-only", timeout=_PULL_TIMEOUT) # type: ignore[arg-type]
except subprocess.TimeoutExpired:
return False, "git pull timed out"
except (OSError, subprocess.SubprocessError) as e:
return False, str(e)
out = (r.stdout + r.stderr).strip()
return r.returncode == 0, out

View File

@@ -283,6 +283,27 @@ class Capability(BaseModel):
meta: dict[str, str] = Field(default_factory=dict)
class Requirement(BaseModel):
"""A precondition — something that must be true for a program/deployment to be
*functional*. The ``kind`` fixes both the meaning and how it's checked (there is
no separate purpose tag):
- ``system`` — a host package/binary must be installed (``ref`` = package).
- ``deployment`` — another deployment must exist/run (``ref`` = its name).
``version`` is reserved for a future constraint (unused now). ``bind`` (for a
``deployment`` requirement) names the env var castle projects the target's URL
into — env is derived *from* the requirement, never scraped back into it.
See docs/relationships.md. ``system_dependencies`` is the ``kind: system`` case.
"""
kind: Literal["system", "deployment"]
ref: str
version: str | None = None
bind: str | None = None
# ---------------------
# Defaults
# ---------------------
@@ -363,6 +384,10 @@ class ProgramSpec(BaseModel):
# Per-program dev verb overrides (declared verbs override the stack default).
commands: CommandsSpec | None = None
# `requires` is the general precondition relation (see docs/relationships.md).
# `system_dependencies` is kept as the `{kind: system}` alias/back-compat; both
# are merged when evaluating what a program requires.
requires: list[Requirement] = Field(default_factory=list)
system_dependencies: list[str] = Field(default_factory=list)
install_extras: list[str] = Field(default_factory=list)
version: str | None = None
@@ -406,6 +431,9 @@ class DeploymentBase(BaseModel):
)
description: str | None = None
defaults: DefaultsSpec | None = None
# Runtime preconditions (e.g. another deployment that must exist). See
# docs/relationships.md; merged with the program's `requires` when evaluated.
requires: list[Requirement] = Field(default_factory=list)
# Declared on/off state. `castle apply` converges reality to this: enabled
# deployments are activated (service started, tool installed, route served),
# disabled ones are deactivated but kept in the catalog. This is *desired

View File

@@ -0,0 +1,204 @@
"""The relationship model — derived, never stored. See docs/relationships.md.
Entities: **program**, **deployment**, **repo** (a repo is a git working copy;
programs sharing a toplevel form a monorepo). One encoded relation, **`requires`**
(a precondition, typed by ``kind``: ``system`` = must be installed, ``deployment``
= must exist). Everything else — repos, env wiring, fan-in, and the predicates
``functional?`` / ``fresh?`` / ``deployed?`` — is computed here on demand.
Governing rule: *predicates are derived; we encode only the non-derivable.* So this
module reads the encoded ``requires`` (plus ``system_dependencies`` as its
``kind: system`` alias) and derives the rest. It does **not** scrape env for
dependencies — env is generated *from* requirements, not the reverse.
"""
from __future__ import annotations
import shutil
from collections import Counter
from dataclasses import dataclass, field
from pathlib import Path
from castle_core import git
from castle_core.config import CastleConfig
from castle_core.manifest import Requirement
@dataclass
class Repo:
key: str # url-safe slug (basename of the working copy)
path: str # git toplevel
url: str | None
ref: str | None
programs: list[str]
deployments: list[str]
behind: int | None = None # commits behind upstream (None = unknown/no upstream)
dirty: bool = False
fresh: bool | None = None # derived: at latest and clean (None = not evaluated)
@property
def multi(self) -> bool:
"""A monorepo — more than one program shares this working copy."""
return len(self.programs) > 1
@dataclass
class Edge:
src: str # deployment name
dst: str # target: a package (system) or another deployment
kind: str # "system" | "deployment"
bind: str | None = None # env var to project the target URL into (deployment)
@dataclass
class Node:
name: str # deployment name
program: str | None
kind: str # service|job|tool|static|reference
repo: str | None
depended_on_by: int # distinct deployments that require this one (fan-in)
unmet: list[str] = field(default_factory=list) # unsatisfied requirements
functional: bool = True # derived: all requirements satisfied
fresh: bool | None = None # derived: its repo is at latest + clean
deployed: bool | None = None # derived: active in the registry (None = unknown)
@dataclass
class Model:
repos: list[Repo] = field(default_factory=list)
nodes: list[Node] = field(default_factory=list)
edges: list[Edge] = field(default_factory=list)
def _program_of(name: str, dep: object) -> str:
return getattr(dep, "program", None) or name
def _slug(name: str, used: set[str]) -> str:
base = name or "repo"
key, n = base, 2
while key in used:
key, n = f"{base}-{n}", n + 1
used.add(key)
return key
def derive_repos(config: CastleConfig) -> dict[str, Repo]:
"""Group programs by the git working copy their source lives in."""
by_top: dict[str, list[str]] = {}
for pname, prog in config.programs.items():
top = git.toplevel(prog.source) if prog.source else None
if top:
by_top.setdefault(top, []).append(pname)
used: set[str] = set()
repos: dict[str, Repo] = {}
for top, progs in sorted(by_top.items()):
progs = sorted(progs)
url = next(
(config.programs[p].repo for p in progs if config.programs[p].repo), None
) or git.remote_url(Path(top))
ref = (
next(
(config.programs[p].ref for p in progs if config.programs[p].ref), None
)
or git.git_status(Path(top), fetch=False).branch
)
deps = sorted(
d for d, dep in config.deployments.items() if _program_of(d, dep) in progs
)
repos[_slug(Path(top).name, used)] = Repo("", top, url, ref, progs, deps)
for key, repo in repos.items():
repo.key = key
return repos
def requirements_of(config: CastleConfig, dep_name: str) -> list[Requirement]:
"""The full requirement set for a deployment: its own ``requires`` plus its
program's ``requires`` and ``system_dependencies`` (the ``kind: system`` alias),
de-duplicated by (kind, ref)."""
dep = config.deployments[dep_name]
prog = config.programs.get(_program_of(dep_name, dep))
reqs: list[Requirement] = list(getattr(dep, "requires", []) or [])
if prog:
reqs += list(prog.requires)
reqs += [
Requirement(kind="system", ref=pkg) for pkg in prog.system_dependencies
]
seen: set[tuple[str, str]] = set()
out: list[Requirement] = []
for r in reqs:
if (r.kind, r.ref) not in seen:
seen.add((r.kind, r.ref))
out.append(r)
return out
def _check(config: CastleConfig, req: Requirement) -> bool:
"""Is a single requirement satisfied? (The check is fixed by its kind.)"""
if req.kind == "system":
return shutil.which(req.ref) is not None
if req.kind == "deployment":
return req.ref in config.deployments
return True
def build_model(
config: CastleConfig,
check: bool = True,
active: set[str] | None = None,
freshness: bool = False,
) -> Model:
"""Compute the relationship model.
- ``check`` (default): evaluate ``functional?`` (unmet requirements) via a live
``which`` / registry probe. ``check=False`` → pure structural model.
- ``active``: names of currently-active deployments → the ``deployed?``
predicate (left ``None`` when the caller has no runtime view).
- ``freshness``: also evaluate ``fresh?`` per repo (a ``git status``, no fetch —
last-known — so it stays a local, network-free probe over many repos)."""
from castle_core.manifest import kind_for
repos = derive_repos(config)
if freshness:
for repo in repos.values():
st = git.git_status(Path(repo.path), fetch=False)
repo.behind = st.behind
repo.dirty = st.dirty
repo.fresh = (st.behind == 0 or st.behind is None) and not st.dirty
repo_of = {p: key for key, r in repos.items() for p in r.programs}
fresh_of = {key: r.fresh for key, r in repos.items()}
edges: list[Edge] = []
for name in config.deployments:
for r in requirements_of(config, name):
edges.append(Edge(name, r.ref, r.kind, r.bind))
fan_in = Counter(e.dst for e in edges if e.kind == "deployment")
nodes: list[Node] = []
for name, dep in config.deployments.items():
unmet = (
[
f"{r.kind}:{r.ref}"
for r in requirements_of(config, name)
if not _check(config, r)
]
if check
else []
)
repo_key = repo_of.get(_program_of(name, dep))
nodes.append(
Node(
name=name,
program=_program_of(name, dep),
kind=kind_for(dep),
repo=repo_key,
depended_on_by=fan_in.get(name, 0),
unmet=unmet,
functional=not unmet,
fresh=fresh_of.get(repo_key) if (freshness and repo_key) else None,
deployed=(name in active) if active is not None else None,
)
)
return Model(repos=list(repos.values()), nodes=nodes, edges=edges)