1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
//! Replacing only the harness of a session that keeps its environment.
//!
//! A move whose destination target, attached resources, and resource
//! allocation are unchanged does not rebuild anything: the source is
//! checkpointed and sealed but its target is kept (the record stays `Closing`
//! with its target), and this module then takes the old harness out of that
//! target and puts the new one in. The container or worker root, the
//! workspace, untracked files, and any build caches survive.
use anyhow::{Context, Result, ensure};
use tokio_util::sync::CancellationToken;
use mj_core::state::{MaterializedSession, SessionState};
use mj_transcript::projection::materialized_session_from_canonical;
use super::{
Controller, RestoreIntoTarget, WorkerRootReset, backend_locator, native_continuity_preserved,
now, projection_rebuild_required, utility_handoff_while_cancellable, verify_resume_checkpoint,
};
use crate::controller::removable_profile_root;
use crate::targets::CommandExecutor;
impl Controller {
/// Replace the harness of a sealed session without rebuilding its target.
///
/// The session must be the one [`Controller::close_session_for_move`] left
/// behind for an in-place swap: `Closing`, with its verified checkpoint and
/// its target still on the record. On success the session is `Running` on
/// `profile_id` in the same target. On any failure the session is torn down
/// to `Stopped` with its verified checkpoint, because a half-written worker
/// root is never retried in place.
// The target gate below is an ordinary lock held across the restore on
// purpose; see the comment where it is taken.
#[allow(
clippy::await_holding_lock,
reason = "the target gate deliberately spans the whole in-place restore"
)]
pub(in crate::controller) async fn restore_session_in_place(
&mut self,
session_id: &str,
profile_id: &str,
executor: &(impl CommandExecutor + Sync),
) -> Result<MaterializedSession> {
let previous = self
.state
.sessions
.get(session_id)
.with_context(|| format!("unknown session {session_id}"))?
.clone();
ensure!(
previous.state == SessionState::Closing,
"session {session_id} is not sealed for an in-place harness replacement"
);
let locator = previous
.target
.as_ref()
.context("an in-place harness replacement has no target to replace it in")?;
let checkpoint = previous
.checkpoint
.as_ref()
.context("session has no checkpoint")?;
let verified_archive = verify_resume_checkpoint(session_id, checkpoint)?;
let profile = self
.config
.profiles
.get(profile_id)
.with_context(|| format!("unknown profile {profile_id:?}"))?
.clone();
ensure!(profile.enabled, "profile {profile_id:?} is disabled");
let target_template = self
.config
.targets
.get(&previous.target_template_id)
.with_context(|| format!("unknown target template {:?}", previous.target_template_id))?
.clone();
self.validate_muse_resume_destination(
&previous,
profile.kind,
&previous.target_template_id,
)?;
ensure!(
profile.kind != mj_core::config::HarnessKind::Muse
|| previous.additional_mounts.is_empty(),
"Muse Code ACP supports one workspace root; attached directories are unsupported"
);
// Resolving the worker binary is local and costs microseconds. A swap
// that could never install a worker fails before the old harness is
// removed from the target.
crate::controller::worker_binary::preflight_worker_binary(&target_template)?;
// The directory the *source* profile owns inside the target. It is
// read from the record's own profile, before the record names the
// destination one.
let backend = backend_locator(locator, &previous, &self.config)?;
let source_profile = self
.config
.profiles
.get(&previous.last_profile)
.with_context(|| {
format!(
"session profile {:?} is missing; it names the profile home to remove",
previous.last_profile
)
})?;
let previous_profile_root = removable_profile_root(&backend, session_id, source_profile);
let archive_manifest = &verified_archive.manifest;
let canonical_session = std::sync::Arc::clone(&verified_archive.canonical_session);
let native_continuity =
native_continuity_preserved(profile.kind, archive_manifest.session.harness_kind);
// A move always seals its source behind a barrier and never replays the
// interrupted prompt itself; queued work is admitted afterwards by
// `admit_move_queue`.
let discard_queued_prompts = true;
let context_bytes = crate::handoff::profile_handoff_bytes(Some(&profile));
let utility_config = (!native_continuity).then(|| self.config.clone());
let stored_frontier = crate::database::materialized_event_frontier(session_id)
.unwrap_or_else(|error| {
tracing::warn!(
session_id,
error = format!("{error:#}"),
"could not read the stored projection frontier; rebuilding it from the archive"
);
None
});
let rebuild_projection = projection_rebuild_required(
stored_frontier
.as_ref()
.map(|(ordinal, digest)| (*ordinal, digest.as_str())),
canonical_session.event_frontier,
&canonical_session.event_frontier_digest,
);
let projection_build = rebuild_projection.then(|| {
let canonical = std::sync::Arc::clone(&canonical_session);
let session_id = session_id.to_owned();
tokio::task::spawn_blocking(move || {
materialized_session_from_canonical(session_id, &canonical)
})
});
// One record transition, and the crash boundary of the whole swap:
// before it, recovery finishes the interrupted close; after it,
// recovery rolls the resume back to `Stopped`. The target stays on the
// record, because it is the environment being kept.
{
let record = self.state.sessions.get_mut(session_id).unwrap();
record.harness_kind = profile.kind;
record.last_profile = profile_id.to_string();
record.native_session_id =
native_continuity.then(|| archive_manifest.session.native_session_id.clone());
record.state = SessionState::Provisioning;
record.updated_at = now();
record.last_error = None;
}
crate::database::save_session(&self.state.sessions[session_id])?;
let result = async {
// A cross-harness swap has no provisioning to overlap with, so the
// handoff is compacted here, before the target is touched. It
// watches the executor for cancellation itself.
let utility_handoff = match utility_config.as_ref() {
Some(config) => Some(
utility_handoff_while_cancellable(
session_id,
config,
&canonical_session,
context_bytes,
executor,
CancellationToken::new(),
)
.await
.context("prepare the cross-harness handoff")?,
),
None => None,
};
// The reset that empties the target of the old harness runs inside
// `restore_into_target`, immediately before the new worker binary
// is installed. Hold the target lock across the whole call so no
// background recovery can act on the worker while it has neither
// harness installed.
let target_mutex = crate::recovery_gate::worker_target_mutex(session_id);
// Holding this ordinary lock across the restore is the point: it is
// the same gate a destroy holds, and the swap is exactly the window
// where a background worker recovery would act on a worker root
// that has no harness in it. A move runs on its own runtime through
// `block_on`, and every other holder takes the lock in synchronous
// code, so nothing here can park the lock on an unscheduled task.
let _target_guard = target_mutex.lock().map_err(|_| {
anyhow::anyhow!("worker target ownership lock poisoned for {session_id}")
})?;
self.restore_into_target(
session_id,
RestoreIntoTarget {
profile: &profile,
archive: &verified_archive,
restored_archive: &verified_archive.archive_path,
resumed_project_directory: previous.project_directory.clone(),
resumed_container_workspace: previous.container_workspace.clone(),
// The repositories are already in the workspace, untouched
// by the swap; only the harness state is restored.
restore_repositories: false,
primary_repository_root_from_conversion: false,
native_continuity,
discard_queued_prompts,
replay_queue: false,
utility_handoff,
projection_build,
resume_notices: Vec::new(),
// EC2-only, and in-place eligibility requires unchanged
// attached resources, so they are already on the instance.
install_attached_resources: false,
worker_root_reset: WorkerRootReset::InPlace {
previous_profile_root,
},
retire_after_ready: None,
},
executor,
)
.await
}
.await;
match result {
Ok(materialized) => Ok(materialized),
Err(error) => {
// Put back whatever this swap could have written to the durable
// projection. Both branches restore archived content, so they
// are correct whether or not the write had happened.
if rebuild_projection {
match materialized_session_from_canonical(session_id, &canonical_session) {
Ok(previous_projection) => {
if let Err(restore_error) =
crate::database::save_materialized_session(&previous_projection)
{
tracing::error!(
session_id,
error = format!("{restore_error:#}"),
"could not restore the durable projection after an in-place move failed"
);
}
}
Err(restore_error) => tracing::error!(
session_id,
error = format!("{restore_error:#}"),
"could not rebuild the durable projection after an in-place move failed"
),
}
} else if let Err(restore_error) =
crate::database::replace_materialized_queued_prompts(
session_id,
&mj_transcript::projection::materialized_queued_prompts_from_canonical(
&canonical_session.queued_prompts,
),
)
{
tracing::error!(
session_id,
error = format!("{restore_error:#}"),
"could not restore queued prompts after an in-place move failed"
);
}
// Never retry in place: the rollback tears the target down and
// leaves the session `Stopped` with its verified checkpoint,
// which is what the fresh path resumes from. The managed
// checkout is still present and still owned by the session, so
// the rollback retires it exactly as a recreated one.
Err(self.rollback_failed_resume(session_id, &previous, true, error, executor)?)
}
}
}
}