fleet/fleetd
3
0
Fork 0
Code Issues 73 Pull Requests Actions Packages Projects Releases 1 Wiki Activity

fleetd #249: withhold agentSessionId when the cwd is not a provisioned worktree
CI / contract (push) Successful in 1m22s
Details
CI / build (push) Successful in 1m22s
Details

Browse Source 
A member spawned without a worktree inherits the lead's cwd, which holds
many old opencode session rows. sessionIdForDirectory picks the most
recently updated row for that directory, so a brand-new member - which
has not written its own row yet - resolves to somebody else's session.
Measured: a row three days old, from a different profile.

The damage was at the tool surface. fleet_list told the lead that
agentSessionId is the id to pass as resumeSessionId, so acting on it
would resume a stranger's conversation, with foreign context, and
nothing to distinguish that from a correct resume.

Fixed by refusing to answer rather than by making the heuristic smarter.
#234 already established the heuristic cannot be made reliable at that
layer, and its javadoc records why, so the SQL is untouched.

agentSessionId() now returns null for a non-provisioned cwd, and
spawn() refuses a resumeSessionId request for one outright, before
anything starts. isProvisionedWorktree moved to HerdrPeerLauncher so
both adapters share it. fleet_list and fleet_spawn descriptions no
longer describe the id as always safe to resume.

Verified rather than taken on trust:
- the refusal reaches the lead as a readable message, not a stack trace
  - FleetMcp.spawn already catches IllegalArgumentException and returns
  error(e.getMessage()).
- the message tells the lead to pass fleet_spawn{worktree:<slug>}, which
  is valid: worktree is typed string, 'true' or a ticket slug.
- 13 existing tests moved off a placeholder "/work/dir" onto a real
  provisioned-worktree fixture. They cover #175/#234 model-mismatch
  machinery and would otherwise have tripped the new gate incidentally.

Worker's mutation evidence, both reverted and diff-confirmed:
- gate at OpenCodeLauncher:812 -> if(false): RED at
  OpenCodeLauncherTest:448, expected <null> but was <ses_someone_elses>.
- resume refusal at OpenCodeLauncher:683 -> 'false &&': RED at
  OpenCodeLauncherTest:379, expected IllegalArgumentException.
Both 0 compile errors.

Closes #249. PR #253.
This commit is contained in:
Dai Ha
2026-09-03 12:55:29 +07:00
parent e897e5257b 2afa3652bb
commit c4deef08be
5 changed files with 222 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fleetd/src/main/java/dev/ltms/fleet/mcp/FleetMcp.java
+14 -5
View File
@@ -1253,7 +1253,11 @@ public final class FleetMcp {
+ "worktree:<ticket-slug> to provision an isolated git worktree. Pass resumeSessionId "
+ "to relaunch onto a prior conversation instead of starting cold — this requires an "
+ "explicit profile whose backend supports it (fleet_list shows agentSessionId for "
+ "resumable members), and is refused otherwise rather than silently starting fresh. "
+ "resumable members; it is absent for a member fleetd cannot reliably re-identify, "
+ "e.g. an opencode member spawned without a worktree), and is refused otherwise "
+ "rather than silently starting fresh. For an opencode profile, resumeSessionId "
+ "itself also requires worktree:true/<slug> on THIS spawn — without one fleetd can "
+ "never re-verify which conversation it actually resumed (fleetd #249). "
+ "sessionName gives the member a display name in its own UI when the backend supports "
+ "one. Returns the member's sessionId (use with fleet_send) and paneId (use with "
+ "fleet_stop).",
@@ -1264,7 +1268,7 @@ public final class FleetMcp {
"worktree", Map.of("type", "string", "description", "'true' or a ticket slug — requests an isolated git worktree"),
"ticket", stringProp("Ticket slug when worktree:true"),
"sessionName", stringProp("Logical display name for the member's own session, when its backend supports one"),
"resumeSessionId", stringProp("A prior member's agentSessionId (from fleet_list) to resume — requires an explicit profile that supports it")),
"resumeSessionId", stringProp("A prior member's agentSessionId (from fleet_list) to resume — requires an explicit profile that supports it, and (for opencode) a worktree on this spawn too")),
List.of()));
}
@@ -1285,9 +1289,14 @@ public final class FleetMcp {
+ "discover a peer lead without being told its address. 'members' are the "
+ "sessions delegated to — each with sessionId, paneId, role (architect/dev/"
+ "reviewer), profile (the backend it runs on), state, optional "
+ "worktree/branch/owner/agentSessionId (the id to pass as fleet_spawn's "
+ "resumeSessionId to relaunch onto that same conversation, when the backend "
+ "supports it), and live herdr status. An empty 'members' "
+ "worktree/branch/owner/agentSessionId, and live herdr status. agentSessionId, "
+ "when present, is the id to pass as fleet_spawn's resumeSessionId to relaunch "
+ "onto that same conversation. It is ABSENT — not a guess — for a member fleetd "
+ "cannot reliably re-identify: some backends (e.g. opencode) resolve it from the "
+ "member's working directory, which only uniquely identifies a member when it "
+ "was spawned into its own fleetd-provisioned worktree (worktree:true/<slug>); a "
+ "member spawned without one shares its directory with others and never reports "
+ "an id, however long it runs (fleetd #249). An empty 'members' "
+ "means no members are spawned; it says nothing about peers. When capacity "
+ "facts are configured, a 'capacity' row per profile also reports free: 0 for "
+ "a quarantined profile's credential (see fleet_profiles), whatever its "
fleetd/src/main/java/dev/ltms/fleet/member/ClaudeCodeLauncher.java
+3 -29
View File
@@ -505,7 +505,7 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
* peer that starts without the seed still starts; it just may hit the dialog fleetd #149
* describes.
*
* <p><b>Gated to a provisioned worktree</b> ({@link #isProvisionedWorktree}) — see that
* <p><b>Gated to a provisioned worktree</b> ({@link HerdrPeerLauncher#isProvisionedWorktree}) — see that
* method's javadoc for the incident that made this gate mandatory, not optional: this must
* never run against a real checkout or an un-configured fallback cwd, only the exact
* always-fresh-directory population fleetd #149 describes.
@@ -516,7 +516,7 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
* sibling-temp-file + {@code ATOMIC_MOVE}, never a truncate-in-place) so a crash mid-write or a
* concurrent reader never observes a half-written file, and through {@link #TRUST_JSON_LOCK} so
* two concurrent spawns' entries both survive instead of the second write silently discarding
* the first. Both exist because of a real incident: see {@link #isProvisionedWorktree}'s javadoc
* the first. Both exist because of a real incident: see {@link HerdrPeerLauncher#isProvisionedWorktree}'s javadoc
* and {@link #writeAtomically}'s javadoc.
*
* @param configDir the profile's {@code CLAUDE_CONFIG_DIR} ({@code cfg.configDir()}), or
@@ -588,7 +588,7 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
* <p><b>fleetd #149 incident.</b> The original implementation used
* {@code Files.writeString(target, content)} directly, which truncates {@code target} in place
* before writing the replacement bytes. Combined with an ungated {@code cwd} (see
* {@link #isProvisionedWorktree}'s javadoc), a mutation-testing run hit that truncation window
* {@link HerdrPeerLauncher#isProvisionedWorktree}'s javadoc), a mutation-testing run hit that truncation window
* against the operator's real {@code ~/.claude.json} and left it at 178 bytes. The gate closes
* <em>which file</em> this can ever target; this closes <em>how</em> the target is written, so
* that even a legitimate write against a real, live, concurrently-read {@code .claude.json}
@@ -638,32 +638,6 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
}
}
/**
* Whether {@code cwd} is a fleetd-provisioned git worktree — signalled the same way
* {@link #writeIdeOverlay} already gates on: a {@code .git} that is a <strong>regular
* file</strong> holding a {@code gitdir:} pointer, as opposed to a real checkout's {@code .git}
* <strong>directory</strong>. {@code null}/blank never qualifies.
*
* <p>Shared by every write that must land only in a worktree fleetd itself created for a
* member — never in a real checkout, an arbitrary configured directory, or (see the incident
* below) the daemon's own fallback cwd.
*
* <p><b>fleetd #149 incident.</b> {@link #seedTrustDialog} originally ran unconditionally on
* any non-blank {@code cwd}. Most of this launcher's OWN tests spawn a profile with no
* {@code cwd} configured, so the base class's {@code resolveCwd} falls through to the real
* {@code user.dir} — and with no {@code configDir} either (also the common case in this
* file's fixtures), the seed's target falls through the same way to the real
* {@code ~/.claude.json}. Running this repo's own test suite corrupted the operator's actual
* config file (it shrank from ~72 KB to a single seeded entry) the first time a mutation
* happened to make the write non-additive. Gating both cwd-targeted writes on "this is a
* worktree fleetd provisioned" — exactly the population fleetd #149 describes
* ({@code worktree: true} always lands in a brand-new directory) — makes that class of write
* impossible against a real checkout or an untouched fallback cwd, in production or in tests.
*/
private static boolean isProvisionedWorktree(String cwd) {
return cwd != null && !cwd.isBlank() && Files.isRegularFile(Path.of(cwd, ".git"));
}
/** {@code s}, or {@code null} when {@code s} is null/blank — the charter-presence test used above. */
private static String nonBlank(String s) {
return (s == null || s.isBlank()) ? null : s;
fleetd/src/main/java/dev/ltms/fleet/member/HerdrPeerLauncher.java
+36
View File
@@ -358,6 +358,42 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
return Files.isRegularFile(candidate) ? candidate : null;
}
/**
* Whether {@code cwd} is a fleetd-provisioned git worktree — signalled the same way
* {@code ClaudeCodeLauncher#writeIdeOverlay} already gates on: a {@code .git} that is a
* <strong>regular file</strong> holding a {@code gitdir:} pointer, as opposed to a real
* checkout's {@code .git} <strong>directory</strong>. {@code null}/blank never qualifies.
*
* <p>Shared by every write (and, since fleetd #249, every identity read) that must land only
* in a worktree fleetd itself created for a member — never in a real checkout, an arbitrary
* configured directory, or (see the incident below) the daemon's own fallback cwd. Package-
* private (not {@code protected}) on purpose: {@link ClaudeCodeLauncher} and
* {@link OpenCodeLauncher} both call it, and same-package visibility is enough — no subclass
* outside this package needs it.
*
* <p><b>fleetd #149 incident.</b> {@code ClaudeCodeLauncher#seedTrustDialog} originally ran
* unconditionally on any non-blank {@code cwd}. Most of that launcher's OWN tests spawn a
* profile with no {@code cwd} configured, so the base class's {@code resolveCwd} falls
* through to the real {@code user.dir} — and with no {@code configDir} either (also the
* common case in that file's fixtures), the seed's target falls through the same way to the
* real {@code ~/.claude.json}. Running this repo's own test suite corrupted the operator's
* actual config file (it shrank from ~72 KB to a single seeded entry) the first time a
* mutation happened to make the write non-additive. Gating both cwd-targeted writes on "this
* is a worktree fleetd provisioned" — exactly the population fleetd #149 describes
* ({@code worktree: true} always lands in a brand-new directory) — makes that class of write
* impossible against a real checkout or an untouched fallback cwd, in production or in tests.
*
* <p><b>fleetd #249.</b> The same reasoning extends to a READ: {@code
* OpenCodeSessionDiscovery#sessionIdForDirectory} keys on {@code directory}, a heuristic that
* is only reliable when the directory is unique to this member — i.e., exactly the population
* this gate identifies. {@link OpenCodeLauncher} uses it to withhold {@code agentSessionId()}
* (report absence rather than a guess) and to refuse a {@code resumeSessionId} spawn that
* cannot be resolved reliably going forward.
*/
static boolean isProvisionedWorktree(String cwd) {
return cwd != null && !cwd.isBlank() && Files.isRegularFile(Path.of(cwd, ".git"));
}
// --- profile surface -----------------------------------------------------------------------
/** The configured peer profile names (what {@code spawn(profile)} accepts). */
fleetd/src/main/java/dev/ltms/fleet/member/OpenCodeLauncher.java
+42 -1
View File
@@ -672,13 +672,29 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
/** Add lazy on-disk session discovery to the base handle. */
@Override
public PeerHandle spawn(SpawnRequest req) {
String cwd = effectiveCwd(req);
// fleetd #249: refuse rather than silently resume into unverifiable territory. opencode's
// `-s <id>` flag itself resumes precisely — the resolved id is what fails, not the resume —
// but resolvedSessionId() below can never confirm (or later re-report) this handle's own
// identity without a fleetd-provisioned worktree (isProvisionedWorktree(cwd)), because the
// directory is shared and sessionIdForDirectory's "most recently updated row" heuristic can
// pick a sibling's session. Refusing here, before anything spawns, beats letting the member
// start and only then discovering fleetd can never again verify who it actually is.
if (req.resumeSessionId() != null && !req.resumeSessionId().isBlank()
&& !isProvisionedWorktree(cwd)) {
throw new IllegalArgumentException("resumeSessionId requires a fleetd-provisioned "
+ "worktree for an opencode profile — without one, this member's cwd is shared "
+ "with other sessions, so fleetd can never reliably confirm (now or later) which "
+ "conversation it is actually running (fleetd #249). Pass fleet_spawn{worktree:"
+ "<ticket-slug>} to resume this member.");
}
PeerHandle inner = super.spawn(req);
// fleetd #175: the same profile config buildLaunch resolved for this spawn (requireProfile
// is deterministic on req.profileName(), so re-resolving here costs a map lookup, not a
// second decision) — SessionAwareHandle needs cfg.model() to know what THIS session should
// be running.
FleetConfig.Profile cfg = requireProfile(req.profileName());
return new SessionAwareHandle(inner, discovery, effectiveCwd(req), cfg,
return new SessionAwareHandle(inner, discovery, cwd, cfg,
this::memberHerdrSocketConfigured, discoveryUnavailableWarned, exhaustionSink);
}
@@ -718,6 +734,17 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
* the directory right now."
*/
private final AtomicReference<String> resolvedSessionId = new AtomicReference<>();
/**
* fleetd #249: whether {@link #cwd} is a fleetd-provisioned git worktree
* ({@link HerdrPeerLauncher#isProvisionedWorktree}), computed once at spawn time since
* {@code cwd} never changes for this handle. When {@code false} the directory is shared
* with other sessions (the default no-worktree spawn inherits the lead's own cwd), so
* {@link OpenCodeSessionDiscovery#sessionIdForDirectory}'s "most recently updated row for
* this directory" heuristic can and does pick another session's row — see that class's
* javadoc. {@link #agentSessionId()} refuses to guess in that case: it reports absent
* rather than a possibly-foreign id.
*/
private final boolean worktreeProvisioned;
SessionAwareHandle(PeerHandle delegate, OpenCodeSessionDiscovery discovery, String cwd,
FleetConfig.Profile cfg,
@@ -731,6 +758,7 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
this.discoveryUnavailable = discoveryUnavailable;
this.discoveryUnavailableWarned = discoveryUnavailableWarned;
this.exhaustionSink = exhaustionSink;
this.worktreeProvisioned = isProvisionedWorktree(cwd);
}
@Override
@@ -760,6 +788,9 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
// built from (see OpenCodeLauncher#defaultDiscoveryRoot's javadoc for the full
// reasoning). Scanning fleetd's own $HOME under that config would only ever find "no
// row" and read as "resume unsupported" — declare it unavailable instead, once, loudly.
// Checked before the fleetd #249 worktree gate below: this OS-user mismatch makes
// discovery unusable regardless of whether cwd happens to be a provisioned worktree, so
// it earns the one-time WARN either way.
if (discoveryUnavailable.getAsBoolean()) {
if (discoveryUnavailableWarned.compareAndSet(false, true)) {
log.warn("opencode session discovery unavailable: memberHerdrSocket is "
@@ -771,6 +802,16 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
}
return null;
}
// fleetd #249: cwd is shared with other sessions unless fleetd itself provisioned this
// worktree, and sessionIdForDirectory's directory-keyed heuristic cannot tell this
// member's row apart from a sibling's in that case (measured: a three-day-old row from
// a different profile). Refuse to guess — absent is the honest answer, and it is what
// this codebase already returns elsewhere for absent evidence (fleetd #175's UNKNOWN).
// No WARN here: unlike discoveryUnavailable above, this is the ordinary, expected shape
// of the large majority of spawns (no worktree requested), not a configuration gap.
if (!worktreeProvisioned) {
return null;
}
// fleetd #234: once resolved, stay resolved. Re-deriving from `directory` on every call
// would let this handle's identity drift to a sibling session that later shares the
// same cwd and writes a newer row — see resolvedSessionId's javadoc.
fleetd/src/test/java/dev/ltms/fleet/member/OpenCodeLauncherTest.java
+127 -31
View File
@@ -323,11 +323,39 @@ class OpenCodeLauncherTest {
// --- CB-547: resume + post-hoc session discovery --------------------------------------------
/**
* Give {@code dir} the exact signature {@link HerdrPeerLauncher#isProvisionedWorktree} checks
* for: a {@code .git} REGULAR FILE, never a directory. Content is never parsed by that gate, so
* any {@code gitdir:} pointer is fine. Mirrors {@code ClaudeCodeLauncherTest}'s helper of the
* same shape (fleetd #249).
*/
private static void markAsProvisionedWorktree(Path dir) throws IOException {
Files.writeString(dir.resolve(".git"), "gitdir: /tmp/not-a-real-gitdir");
}
/**
* A fresh subdirectory of {@code configRoot}, marked as a provisioned worktree (fleetd #249),
* for tests that predate this gate and stood in a bare {@code "/work/dir"} string as their
* member's cwd — a directory that never existed on disk and, post-#249, would never pass
* {@link HerdrPeerLauncher#isProvisionedWorktree} either. Those tests are about the model
* mismatch / late-resolve machinery (fleetd #175/#234/#209), not about the worktree gate
* itself, so they need a cwd the gate accepts without changing what each test demonstrates.
*/
private static String provisionedWorkDir(Path configRoot) throws IOException {
Path dir = Files.createDirectories(configRoot.resolve("work-dir"));
markAsProvisionedWorktree(dir);
return dir.toString();
}
@Test
void aResumeSpawnPassesTheSessionIdAsDashS(@TempDir Path root) {
void aResumeSpawnIntoAProvisionedWorktreePassesTheSessionIdAsDashS(@TempDir Path root,
@TempDir Path worktree)
throws Exception {
markAsProvisionedWorktree(worktree);
FakeHerdr herdr = new FakeHerdr();
service(herdr, root, opencodeCfg("google/gemini-2.5-pro", null, null))
.spawn(new SpawnRequest(null, null, null, null, "ses_41b79fc90ffeI9E8uZv6VprUn2"));
.spawn(new SpawnRequest(null, worktree.toString(), null, null,
"ses_41b79fc90ffeI9E8uZv6VprUn2"));
List<String> args = startArgs(herdr);
int s = args.indexOf("-s");
@@ -336,6 +364,27 @@ class OpenCodeLauncherTest {
"the resume target id follows -s");
}
/**
* fleetd #249 acceptance criterion 3: without a fleetd-provisioned worktree, the member's cwd
* is shared with other sessions, so fleetd can never reliably confirm (now or later via {@link
* OpenCodeSessionDiscovery}) which conversation it is actually running. Refuse the spawn itself
* rather than silently launching opencode's {@code -s <id>} into unverifiable territory.
*/
@Test
void aResumeSpawnWithoutAProvisionedWorktreeIsRefused(@TempDir Path root) {
FakeHerdr herdr = new FakeHerdr();
OpenCodeLauncher launcher = service(herdr, root,
opencodeCfg("google/gemini-2.5-pro", null, null));
IllegalArgumentException e = assertThrows(IllegalArgumentException.class, () ->
launcher.spawn(new SpawnRequest(null, null, null, null,
"ses_41b79fc90ffeI9E8uZv6VprUn2")));
assertTrue(e.getMessage().contains("worktree"), e.getMessage());
assertFalse(herdr.called("agent.start"),
"the refusal must happen before anything spawns — no pane, no process");
}
@Test
void aFreshSpawnCarriesNoSessionFlag(@TempDir Path root) {
FakeHerdr herdr = new FakeHerdr();
@@ -348,25 +397,59 @@ class OpenCodeLauncherTest {
@Test
void theHandleDiscoversTheSessionIdForTheWorkersCwdOnlyAfterItAppears(@TempDir Path root,
@TempDir Path discRoot)
@TempDir Path discRoot,
@TempDir Path worktree)
throws Exception {
markAsProvisionedWorktree(worktree);
FakeHerdr herdr = new FakeHerdr();
OpenCodeLauncher launcher = new OpenCodeLauncher(new AgentControl(herdr),
new WorkspaceControl(herdr), Map.of("gemini", opencodeCfg(null, null, null)),
"gemini", _ -> null, 0, System::currentTimeMillis, () -> { }, root, discRoot);
PeerHandle handle = launcher.spawn(new SpawnRequest(null, "/work/dir", null));
PeerHandle handle = launcher.spawn(new SpawnRequest(null, worktree.toString(), null));
// opencode writes the record only when the session is first persisted — the instant the
// pane is ready it does not exist, so agentSessionId() is null (never a spawn failure).
assertNull(handle.agentSessionId(), "no record yet → null, not a spawn-time block");
// Once the record appears (here: same cwd), lazy discovery resolves it — the handle's
// session id matches its own worktree, not another's.
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_resolved", "/work/dir", 1000L);
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_resolved", worktree.toString(), 1000L);
assertEquals("ses_resolved", handle.agentSessionId(),
"agentSessionId() re-scans and picks up a record that has since been written");
}
/**
* fleetd #249 acceptance criterion 1, exercised through the real caller path (the handle
* {@code fleet_list} actually reads), not {@link OpenCodeSessionDiscovery} directly. Without a
* fleetd-provisioned worktree the member's cwd is shared — the default no-worktree spawn
* inherits the lead's own long-lived cwd — so even once a matching row appears (here:
* simulating another profile's session that happens to share the directory) the handle must
* report absence rather than guess. Measured real-world case (2026-09-03): the row it would
* otherwise pick was three days old and belonged to a different profile.
*/
@Test
void theHandleNeverReportsAnIdForANonProvisionedCwdEvenAfterARowAppears(@TempDir Path root,
@TempDir Path discRoot)
throws Exception {
FakeHerdr herdr = new FakeHerdr();
OpenCodeLauncher launcher = new OpenCodeLauncher(new AgentControl(herdr),
new WorkspaceControl(herdr), Map.of("gemini", opencodeCfg(null, null, null)),
"gemini", _ -> null, 0, System::currentTimeMillis, () -> { }, root, discRoot);
// No markAsProvisionedWorktree — this cwd has no .git file, the shared-cwd shape a
// no-worktree spawn (or a real checkout) actually has.
String sharedCwd = root.resolve("shared-cwd").toString();
PeerHandle handle = launcher.spawn(new SpawnRequest(null, sharedCwd, null));
assertNull(handle.agentSessionId(), "no record yet → null, same as the provisioned case");
// A row for this exact directory now appears — e.g. a sibling member, or a stale session
// from days earlier, sharing the same unprovisioned cwd.
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_someone_elses", sharedCwd, 1000L);
assertNull(handle.agentSessionId(),
"a non-provisioned cwd must NEVER report an id, even once a row for it exists — "
+ "the row could belong to any other session sharing this directory");
}
@Test
void foreignWorkerMatchesOpencodePrefixButNotClaude() {
String nonce = "abc123";
@@ -920,6 +1003,7 @@ class OpenCodeLauncherTest {
@Test
void theRealSessionManagerLateResolvePathCatchesAModelMismatch(@TempDir Path configRoot,
@TempDir Path discRoot) throws Exception {
String workDir = provisionedWorkDir(configRoot);
FakeHerdr herdr = new FakeHerdr();
// xf's real shape (fleetd #175): weight:80, model "opencode/nemotron-3-ultra-free", no
// credentialId — the profile that actually escaped the fleet's accounting.
@@ -929,7 +1013,7 @@ class OpenCodeLauncherTest {
OpenCodeLauncher launcher = serviceWithSink(herdr, configRoot, discRoot, cfg, sink);
SessionManager sessions = new SessionManager(launcher);
MemberSession acquired = sessions.acquire(cfg.profile(), "/work/dir", null, null);
MemberSession acquired = sessions.acquire(cfg.profile(), workDir, null, null);
// Real late-resolve path, driven BEFORE opencode has written its session row — same shape
// as production the instant a pane goes ready.
@@ -940,7 +1024,7 @@ class OpenCodeLauncherTest {
// opencode writes its row late, running gpt-5.6-sol (a PAID credential) instead of the
// withdrawn free model the profile actually asked for — the exact fleetd #175 scenario.
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
"{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
// Drive the SAME real late-resolve path again: sessions.get() -> resolveAgentSessionId ->
@@ -959,12 +1043,13 @@ class OpenCodeLauncherTest {
@Test
void aProviderPrefixedModelMatchingBothIdAndProviderIsNotAMismatch(@TempDir Path configRoot,
@TempDir Path discRoot) throws Exception {
String workDir = provisionedWorkDir(configRoot);
List<String> exhausted = new ArrayList<>();
ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
.spawn(new SpawnRequest(null, "/work/dir", null));
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
.spawn(new SpawnRequest(null, workDir, null));
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
"{\"id\":\"gpt-5.6-terra\",\"providerID\":\"openai\"}");
assertEquals("ses_x", handle.agentSessionId());
@@ -975,12 +1060,13 @@ class OpenCodeLauncherTest {
@Test
void aGxProviderPrefixedModelMatchingBothIdAndProviderIsNotAMismatch(@TempDir Path configRoot,
@TempDir Path discRoot) throws Exception {
String workDir = provisionedWorkDir(configRoot);
List<String> exhausted = new ArrayList<>();
ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
FleetConfig.Profile cfg = opencodeCfg("gx/deepseek-v4-flash", null, null);
PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
.spawn(new SpawnRequest(null, "/work/dir", null));
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
.spawn(new SpawnRequest(null, workDir, null));
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
"{\"id\":\"deepseek-v4-flash\",\"providerID\":\"gx\"}");
assertEquals("ses_x", handle.agentSessionId());
@@ -997,12 +1083,13 @@ class OpenCodeLauncherTest {
@Test
void aMissingProviderIdInTheEvidenceIsUnknownNotAMismatchWhenTheIdMatches(
@TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
String workDir = provisionedWorkDir(configRoot);
List<String> exhausted = new ArrayList<>();
ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
.spawn(new SpawnRequest(null, "/work/dir", null));
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
.spawn(new SpawnRequest(null, workDir, null));
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
"{\"id\":\"gpt-5.6-terra\"}");
assertEquals("ses_x", handle.agentSessionId());
@@ -1018,12 +1105,13 @@ class OpenCodeLauncherTest {
@Test
void aMissingProviderIdInTheEvidenceStillCatchesARealIdMismatch(
@TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
String workDir = provisionedWorkDir(configRoot);
List<String> exhausted = new ArrayList<>();
ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
.spawn(new SpawnRequest(null, "/work/dir", null));
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
.spawn(new SpawnRequest(null, workDir, null));
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
"{\"id\":\"gpt-5.6-sol\"}");
assertEquals("ses_x", handle.agentSessionId());
@@ -1042,12 +1130,13 @@ class OpenCodeLauncherTest {
@Test
void aBareModelWithNoProviderPrefixMatchesOnIdAloneAndIsNotAMismatch(@TempDir Path configRoot,
@TempDir Path discRoot) throws Exception {
String workDir = provisionedWorkDir(configRoot);
List<String> exhausted = new ArrayList<>();
ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
FleetConfig.Profile cfg = opencodeCfg("deepseek-v4-flash", null, null);
PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
.spawn(new SpawnRequest(null, "/work/dir", null));
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
.spawn(new SpawnRequest(null, workDir, null));
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
"{\"id\":\"deepseek-v4-flash\",\"providerID\":\"gx\"}");
assertEquals("ses_x", handle.agentSessionId());
@@ -1064,6 +1153,7 @@ class OpenCodeLauncherTest {
@Test
void aRealIdMismatchLogsAnErrorNamingBothModelsAndQuarantinesThroughTheSink(
@TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
String workDir = provisionedWorkDir(configRoot);
List<String> exhausted = new ArrayList<>();
ExhaustionSink sink = (target, reason, profile) -> exhausted.add(target + "|" + reason);
FleetConfig.Profile cfg = opencodeCfg("opencode/nemotron-3-ultra-free", null, null);
@@ -1075,8 +1165,8 @@ class OpenCodeLauncherTest {
PeerHandle handle;
try {
handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
.spawn(new SpawnRequest(null, "/work/dir", null));
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
.spawn(new SpawnRequest(null, workDir, null));
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
"{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
assertEquals("ses_x", handle.agentSessionId());
} finally {
@@ -1111,11 +1201,12 @@ class OpenCodeLauncherTest {
@Test
void unknownOrUnparseableModelEvidenceNeverQuarantines(@TempDir Path configRoot,
@TempDir Path discRoot) throws Exception {
String workDir = provisionedWorkDir(configRoot);
List<String> exhausted = new ArrayList<>();
ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
.spawn(new SpawnRequest(null, "/work/dir", null));
.spawn(new SpawnRequest(null, workDir, null));
// No row yet at all.
assertNull(handle.agentSessionId());
@@ -1137,12 +1228,13 @@ class OpenCodeLauncherTest {
@Test
void aProfileWithNoConfiguredModelIsNeverCheckedForAMismatch(@TempDir Path configRoot,
@TempDir Path discRoot) throws Exception {
String workDir = provisionedWorkDir(configRoot);
List<String> exhausted = new ArrayList<>();
ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
FleetConfig.Profile cfg = opencodeCfg(null, null, null);
PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
.spawn(new SpawnRequest(null, "/work/dir", null));
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
.spawn(new SpawnRequest(null, workDir, null));
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
"{\"id\":\"anything-at-all\",\"providerID\":\"anyone\"}");
assertEquals("ses_x", handle.agentSessionId());
@@ -1166,21 +1258,22 @@ class OpenCodeLauncherTest {
@Test
void modelCheckReadsTheResolvedSessionsOwnRowNotWhateverIsNewestInTheSharedDirectory(
@TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
String workDir = provisionedWorkDir(configRoot);
List<String> exhausted = new ArrayList<>();
ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
.spawn(new SpawnRequest(null, "/work/dir", null));
.spawn(new SpawnRequest(null, workDir, null));
// Our own session's row, correctly matching the profile's requested model.
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_ours", "/work/dir", 1000L,
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_ours", workDir, 1000L,
"{\"id\":\"gpt-5.6-terra\",\"providerID\":\"openai\"}");
assertEquals("ses_ours", handle.agentSessionId(), "resolves to our own session");
assertTrue(exhausted.isEmpty(), "matching model → no mismatch on first resolve: " + exhausted);
// A sibling member, spawned later into the SAME shared directory (no worktree, fleetd
// #234's default), writes a newer row running a totally different model.
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_sibling", "/work/dir", 9000L,
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_sibling", workDir, 9000L,
"{\"id\":\"deepseek-v4-flash\",\"providerID\":\"gx\"}");
assertEquals("ses_ours", handle.agentSessionId(),
@@ -1212,6 +1305,7 @@ class OpenCodeLauncherTest {
@Test
void aSpawnTimeModelMismatchActuallyQuarantinesTheCredentialThroughTheRealAcquirePath(
@TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
String workDir = provisionedWorkDir(configRoot);
FleetConfig.Profile cfg = opencodeCfgWithCredential(
"terra", "opencode/nemotron-3-ultra-free", "openai-shared");
Map<String, FleetConfig.Profile> profiles = Map.of(cfg.profile(), cfg);
@@ -1232,14 +1326,14 @@ class OpenCodeLauncherTest {
// The mismatching row exists BEFORE the spawn — reproducing fleetd #234's exact timing:
// opencode's session table already carries evidence by the moment acquire() first asks.
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
"{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
assertFalse(quarantine.isQuarantined("openai-shared"), "nothing quarantined before the spawn");
// The real production entrypoint: acquire() builds the MemberSession by calling
// handle.agentSessionId() BEFORE registry.put() runs.
MemberSession acquired = sessions.acquire(cfg.profile(), "/work/dir", null, null);
MemberSession acquired = sessions.acquire(cfg.profile(), workDir, null, null);
assertEquals("ses_x", acquired.agentSessionId(), "the id itself still resolves correctly");
assertTrue(quarantine.isQuarantined("openai-shared"),
@@ -1258,6 +1352,7 @@ class OpenCodeLauncherTest {
@Test
void aRosterOnlySinkSilentlyDropsTheSpawnTimeQuarantine(@TempDir Path configRoot,
@TempDir Path discRoot) throws Exception {
String workDir = provisionedWorkDir(configRoot);
FleetConfig.Profile cfg = opencodeCfgWithCredential(
"terra", "opencode/nemotron-3-ultra-free", "openai-shared");
BackendQuarantine quarantine = new BackendQuarantine(() -> 0L, TimeUnit.SECONDS.toNanos(1800));
@@ -1270,10 +1365,10 @@ class OpenCodeLauncherTest {
OpenCodeLauncher launcher = serviceWithSink(herdr, configRoot, discRoot, cfg, rosterOnlySink);
SessionManager sessions = new SessionManager(launcher);
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
"{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
MemberSession acquired = sessions.acquire(cfg.profile(), "/work/dir", null, null);
MemberSession acquired = sessions.acquire(cfg.profile(), workDir, null, null);
assertEquals("ses_x", acquired.agentSessionId(), "the id itself still resolves correctly");
assertFalse(quarantine.isQuarantined("openai-shared"),
@@ -1301,6 +1396,7 @@ class OpenCodeLauncherTest {
@Test
void theSpawnTimeQuarantineSurvivesTheFleetdStyleForwardingHop(@TempDir Path configRoot,
@TempDir Path discRoot) throws Exception {
String workDir = provisionedWorkDir(configRoot);
FleetConfig.Profile cfg = opencodeCfgWithCredential(
"terra", "opencode/nemotron-3-ultra-free", "openai-shared");
Map<String, FleetConfig.Profile> profiles = Map.of(cfg.profile(), cfg);
@@ -1332,12 +1428,12 @@ class OpenCodeLauncherTest {
};
exhaustionSinkRef.set(realSink);
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
"{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
assertFalse(quarantine.isQuarantined("openai-shared"), "nothing quarantined before the spawn");
MemberSession acquired = sessions.acquire(cfg.profile(), "/work/dir", null, null);
MemberSession acquired = sessions.acquire(cfg.profile(), workDir, null, null);
assertEquals("ses_x", acquired.agentSessionId(), "the id itself still resolves correctly");
assertTrue(quarantine.isQuarantined("openai-shared"),