hyperhive/hive-c0re/src/flake_check.rs
2026-05-25 23:52:00 +02:00

431 lines
16 KiB
Rust

//! Pre-apply validation for agent `flake.lock` files (closes part of #317).
//!
//! Every `request_apply_commit` lands a `proposal/<id>` tag in the
//! agent's applied repo before the operator sees the approval. We
//! parse `flake.lock` from that tag's tree and reject the request if
//! two or more nodes share an identical `original` field — that
//! signals a missing `inputs.<X>.inputs.nixpkgs.follows = "nixpkgs"`
//! directive in `flake.nix` and would inflate meta's lock with
//! duplicates after deploy.
//!
//! Per mara's scope note on #317 (comment 4189): the check runs on
//! the agent repo, not meta, and catches *new* violations only.
//! Existing agents whose lock already has duplicates are out of
//! scope here and get a coordinated config-change pass via the
//! manager instead.
use std::collections::BTreeMap;
use std::fmt::Write as _;
use std::path::Path;
use anyhow::{Context, Result};
use serde_json::Value;
use tokio::process::Command;
use crate::lifecycle::git_command;
/// One group of `flake.lock` nodes that all share the same canonical
/// `original` reference. Surfaced in the rejection message so the
/// operator (and the manager that submitted the apply) can see
/// exactly which input pair needs a `follows` directive.
#[derive(Debug, Clone)]
pub struct DuplicateGroup {
/// One of the original `Value`s from the lock — used for
/// pretty-printing in the error message. Canonical equivalence
/// is enforced by the `BTreeMap` key in `duplicate_groups`, so
/// we don't need to keep the canonicalised form on the struct.
pub original: Value,
/// Names of the flake.lock nodes that share this `original`,
/// sorted for stable error output.
pub keys: Vec<String>,
}
/// Read `flake.lock` from `<tag>:flake.lock` in `repo`. Returns
/// `Ok(None)` when the file isn't tracked in that tag (no inputs ⇒
/// nothing to dedup); `Err` only on real git plumbing failures.
async fn read_lock_at_tag(repo: &Path, tag: &str) -> Result<Option<String>> {
let spec = format!("{tag}:flake.lock");
let out = git_command()
.current_dir(repo)
.args(["show", &spec])
.output()
.await
.with_context(|| format!("git show {spec} in {}", repo.display()))?;
if !out.status.success() {
let stderr = String::from_utf8_lossy(&out.stderr);
// git uses two different messages for "path not in tree"
// depending on whether the path also collides with an on-disk
// file. Both translate to "no flake.lock in this commit" —
// a legitimate, dedup-clean state for an agent with empty
// `inputs = { }`. Any other git failure (permission denied,
// ref-not-found, etc.) propagates as a hard error rather than
// being silently swallowed.
if stderr.contains("does not exist")
|| stderr.contains("exists on disk, but not in")
{
return Ok(None);
}
anyhow::bail!("git show {spec} failed: {}", stderr.trim());
}
Ok(Some(String::from_utf8_lossy(&out.stdout).into_owned()))
}
/// Recursively serialise `v` with object keys sorted, so two
/// JSON values that differ only in key insertion order produce the
/// same string. `serde_json::Value` preserves `IndexMap` order by
/// default, which is fine for parsing but breaks our group-by-key
/// idea — hence this hand-rolled canonicaliser.
fn canonical_json(v: &Value) -> String {
match v {
Value::Object(map) => {
let mut keys: Vec<&String> = map.keys().collect();
keys.sort();
let mut s = String::from("{");
for (i, k) in keys.iter().enumerate() {
if i > 0 {
s.push(',');
}
s.push_str(&serde_json::to_string(k).unwrap_or_default());
s.push(':');
s.push_str(&canonical_json(&map[*k]));
}
s.push('}');
s
}
Value::Array(arr) => {
let mut s = String::from("[");
for (i, x) in arr.iter().enumerate() {
if i > 0 {
s.push(',');
}
s.push_str(&canonical_json(x));
}
s.push(']');
s
}
_ => serde_json::to_string(v).unwrap_or_default(),
}
}
/// Parse `raw` (a `flake.lock` JSON document) and return every group
/// of nodes whose `original` field is identical. Nodes without an
/// `original` (the synthetic `root`, or anomalous entries) are
/// skipped. Groups with only one member are filtered out — only true
/// duplicates surface.
///
/// Pure function, no I/O — covered by the unit tests below.
pub fn duplicate_groups(raw: &str) -> Result<Vec<DuplicateGroup>> {
let json: Value = serde_json::from_str(raw).context("parse flake.lock")?;
let Some(nodes) = json.get("nodes").and_then(|v| v.as_object()) else {
return Ok(Vec::new());
};
let mut groups: BTreeMap<String, DuplicateGroup> = BTreeMap::new();
for (name, node) in nodes {
let Some(original) = node.get("original") else {
continue;
};
let key = canonical_json(original);
let entry = groups.entry(key).or_insert_with(|| DuplicateGroup {
original: original.clone(),
keys: Vec::new(),
});
entry.keys.push(name.clone());
}
let mut dups: Vec<DuplicateGroup> = groups
.into_values()
.filter(|g| g.keys.len() > 1)
.collect();
for g in &mut dups {
g.keys.sort();
}
Ok(dups)
}
/// Re-derive the agent's `flake.lock` from its `flake.nix` (in a
/// throw-away worktree at the proposal tag) and reject the apply when
/// the result differs from what's committed — that means the manager
/// edited `flake.nix` but didn't commit the regenerated lock, so the
/// shipped state lies about what nix will actually fetch.
///
/// Plain `nix flake lock` (no `--update-input` flags) only fills in
/// MISSING entries; it never refreshes existing ones. So a lock that
/// matches its `flake.nix` round-trips to a no-op, and any diff is a
/// real "stale lock" signal.
///
/// Materialises the proposal tag into a temp worktree under
/// `std::env::temp_dir()` to avoid touching `applied/<n>/main` while
/// the check runs. Cleanup is unconditional via the inner-fn pattern.
///
/// Returns `Ok(())` when in sync (or there's no `flake.nix` at all);
/// `Err` with a human-readable message on stale lock or nix tooling
/// failure.
pub async fn check_lock_in_sync(repo: &Path, tag: &str, approval_id: i64) -> Result<()> {
let suffix = std::time::SystemTime::now()
.duration_since(std::time::UNIX_EPOCH)
.map(|d| d.as_nanos())
.unwrap_or(0);
let tmp_dir = std::env::temp_dir().join(format!("hive-flake-check-{approval_id}-{suffix}"));
// Detached worktree at the proposal tag — gives us a clean, mutable
// copy of the agent's tree without disturbing whatever's currently
// checked out on `applied/<n>/main`.
let out = git_command()
.current_dir(repo)
.args([
"worktree",
"add",
"--detach",
&tmp_dir.to_string_lossy(),
tag,
])
.output()
.await
.with_context(|| format!("git worktree add {} {tag}", tmp_dir.display()))?;
if !out.status.success() {
anyhow::bail!(
"git worktree add failed: {}",
String::from_utf8_lossy(&out.stderr).trim()
);
}
let result = lock_in_sync_inner(&tmp_dir).await;
// Best-effort cleanup. `git worktree remove --force` handles the
// common case; `remove_dir_all` mops up if git decided the worktree
// is half-gone (or if the inner work bailed before nix touched the
// tree). Failures here are logged, not propagated — the check's
// result is what matters.
if let Err(e) = remove_worktree(repo, &tmp_dir).await {
tracing::warn!(
worktree = %tmp_dir.display(),
error = %format!("{e:#}"),
"flake_check: temp worktree cleanup failed"
);
}
result
}
async fn lock_in_sync_inner(worktree: &Path) -> Result<()> {
// No `flake.nix` means there's nothing for nix to lock — skip the
// check cleanly (the dedup check will likewise no-op).
if !worktree.join("flake.nix").exists() {
return Ok(());
}
let committed = tokio::fs::read_to_string(worktree.join("flake.lock"))
.await
.ok();
// `--extra-experimental-features` mirrors `meta::nix` for hosts
// that haven't already enabled flakes in `nix.conf`. Plain
// `nix flake lock` (no `--update-input`) fills missing entries but
// never refreshes existing ones — exactly the semantics we want.
let out = Command::new("nix")
.current_dir(worktree)
.args([
"--extra-experimental-features",
"nix-command flakes",
"flake",
"lock",
])
.output()
.await
.with_context(|| format!("nix flake lock in {}", worktree.display()))?;
if !out.status.success() {
anyhow::bail!(
"nix flake lock failed: {}",
String::from_utf8_lossy(&out.stderr).trim()
);
}
let regenerated = tokio::fs::read_to_string(worktree.join("flake.lock"))
.await
.ok();
// An agent that declares inputs in flake.nix but ships no
// flake.lock at all hits this branch (committed = None,
// regenerated = Some(...)). That's a deliberate reject: every
// agent with inputs MUST commit its lock, otherwise meta's
// dedup pass has nothing to introspect and the broken state
// leaks downstream. Treated identically to a stale lock.
if committed.as_deref() != regenerated.as_deref() {
anyhow::bail!(
"flake.lock is out of sync with flake.nix — `nix flake lock` produces a different lock. \
Run `nix flake lock` in your agent config, commit the result, and re-submit request_apply_commit."
);
}
Ok(())
}
async fn remove_worktree(repo: &Path, worktree: &Path) -> Result<()> {
let out = git_command()
.current_dir(repo)
.args([
"worktree",
"remove",
"--force",
&worktree.to_string_lossy(),
])
.output()
.await
.with_context(|| format!("git worktree remove {}", worktree.display()))?;
if !out.status.success() {
// `git worktree remove` already errored — still try the raw
// rmdir so we don't leak the dir on disk. Surface the original
// git stderr for context.
let _ = tokio::fs::remove_dir_all(worktree).await;
anyhow::bail!(
"git worktree remove failed: {}",
String::from_utf8_lossy(&out.stderr).trim()
);
}
// git removed the worktree's metadata but the dir itself may
// linger on stripped-down git versions — best-effort clean.
let _ = tokio::fs::remove_dir_all(worktree).await;
Ok(())
}
/// Run the dedup check against the agent's freshly-applied tree.
///
/// `Ok(())` means either the commit doesn't carry a `flake.lock` (no
/// inputs declared) or every node has a unique `original`. `Err`
/// carries a multi-line message listing every duplicate group with
/// the offending node names, suitable for surfacing on the failed
/// approval row.
pub async fn check_no_duplicate_inputs(repo: &Path, tag: &str) -> Result<()> {
let Some(raw) = read_lock_at_tag(repo, tag).await? else {
return Ok(());
};
let dups = duplicate_groups(&raw)?;
if dups.is_empty() {
return Ok(());
}
let mut msg = String::from(
"flake.lock has duplicate flake inputs — add a `follows` directive in flake.nix to collapse them:\n",
);
for g in &dups {
let original = serde_json::to_string(&g.original).unwrap_or_else(|_| "?".into());
let _ = writeln!(msg, " - {original} → nodes [{}]", g.keys.join(", "));
}
anyhow::bail!("{}", msg.trim_end());
}
#[cfg(test)]
mod tests {
use super::*;
const CLEAN_LOCK: &str = r#"{
"nodes": {
"nixpkgs": {
"locked": {"rev": "aaa"},
"original": {"owner": "NixOS", "repo": "nixpkgs", "ref": "nixos-25.11", "type": "github"}
},
"nixpkgs-unstable": {
"locked": {"rev": "bbb"},
"original": {"owner": "NixOS", "repo": "nixpkgs", "ref": "nixpkgs-unstable", "type": "github"}
},
"root": {"inputs": {"nixpkgs": "nixpkgs"}}
},
"root": "root",
"version": 7
}"#;
const DUPLICATE_LOCK: &str = r#"{
"nodes": {
"nixpkgs": {
"locked": {"rev": "aaa"},
"original": {"owner": "NixOS", "repo": "nixpkgs", "ref": "nixos-25.11", "type": "github"}
},
"nixpkgs_2": {
"locked": {"rev": "ccc"},
"original": {"owner": "NixOS", "repo": "nixpkgs", "ref": "nixos-25.11", "type": "github"}
},
"nixpkgs_3": {
"locked": {"rev": "ddd"},
"original": {"owner": "NixOS", "repo": "nixpkgs", "ref": "nixos-25.11", "type": "github"}
},
"treefmt-nix": {
"locked": {"rev": "eee"},
"original": {"owner": "numtide", "repo": "treefmt-nix", "type": "github"}
},
"treefmt-nix_2": {
"locked": {"rev": "fff"},
"original": {"type": "github", "owner": "numtide", "repo": "treefmt-nix"}
},
"root": {"inputs": {"nixpkgs": "nixpkgs"}}
},
"root": "root",
"version": 7
}"#;
#[test]
fn clean_lock_has_no_duplicates() {
let dups = duplicate_groups(CLEAN_LOCK).expect("parse");
assert!(dups.is_empty(), "expected no dups, got {dups:#?}");
}
#[test]
fn duplicate_lock_reports_groups() {
let dups = duplicate_groups(DUPLICATE_LOCK).expect("parse");
assert_eq!(dups.len(), 2, "expected nixpkgs + treefmt-nix groups");
// Order is BTreeMap-stable: sorted by canonical_json key. The
// numtide treefmt-nix key sorts before the NixOS nixpkgs one
// because `numtide` < `NixOS` lexicographically (case-sensitive,
// capitals come first... wait — capital N is 0x4e, lowercase n
// is 0x6e, so capitals come first). So nixpkgs group sorts
// first. Verify by content instead of position to avoid coupling
// to that subtlety.
let nixpkgs_group = dups
.iter()
.find(|g| g.original.get("ref").is_some())
.expect("nixpkgs group present");
assert_eq!(
nixpkgs_group.keys,
vec!["nixpkgs", "nixpkgs_2", "nixpkgs_3"]
);
let treefmt_group = dups
.iter()
.find(|g| g.original.get("ref").is_none())
.expect("treefmt-nix group present");
assert_eq!(treefmt_group.keys, vec!["treefmt-nix", "treefmt-nix_2"]);
}
#[test]
fn key_order_in_original_does_not_matter() {
// The treefmt-nix and treefmt-nix_2 entries above use different
// key orderings for `original` ({owner,repo,type} vs
// {type,owner,repo}); duplicate_groups should still merge them.
let dups = duplicate_groups(DUPLICATE_LOCK).expect("parse");
let treefmt = dups
.iter()
.find(|g| g.keys.iter().any(|k| k == "treefmt-nix"))
.expect("treefmt-nix group");
assert!(treefmt.keys.contains(&"treefmt-nix_2".to_owned()));
}
#[test]
fn nodes_without_original_are_ignored() {
// The synthetic `root` node has no `original` and must not be
// grouped against anything.
let dups = duplicate_groups(CLEAN_LOCK).expect("parse");
assert!(dups.iter().all(|g| !g.keys.iter().any(|k| k == "root")));
}
#[test]
fn missing_nodes_object_is_not_an_error() {
// A lock file that's syntactically JSON but lacks `nodes` (e.g.
// a partial test fixture) should fail open — no dups reported.
let raw = r#"{"root": "root", "version": 7}"#;
let dups = duplicate_groups(raw).expect("parse");
assert!(dups.is_empty());
}
#[test]
fn invalid_json_is_a_hard_error() {
let err = duplicate_groups("not-json-at-all").unwrap_err();
assert!(format!("{err:#}").contains("parse flake.lock"));
}
}