Compare commits

..

1 Commits

Author SHA1 Message Date
Dai Ha 7954a4d399 CB-161: match a pid's ancestry against a pane, not just its shell/foreground pid
CI / contract (pull_request) Successful in 56s
CI / build (pull_request) Successful in 1m39s
paneOwnsPid only compared a pid directly against a pane's shell_pid and
foreground_processes pids, so a grandchild process a member spawns (a python3
or curl helper opening its own MCP connection) matched no pane. Connection
identity then fell through to loopback-trust and resolved that connection as
the primary -- a worker->primary privilege escalation.

terminalForPid now walks the caller's ancestry once (bounded at 32
generations, with a cycle guard) via an injectable ParentResolver seam, and
paneOwnsPid checks pane pids against that ancestor set instead of the raw
pid. Production wiring defaults to ProcessHandle; tests drive a fake
pid->parent map.
2026-08-31 14:04:05 +07:00
14 changed files with 281 additions and 542 deletions
+131
View File
@@ -0,0 +1,131 @@
# CB-161: fix the pane-ownership ancestry gap in PaneLocator
## The bug
`PaneLocator.paneOwnsPid` matched a pid against only two things: a pane's `shell_pid`, and
the pids in `foreground_processes`. A grandchild process a member spawns (for example a
`python3` or `curl` helper that opens its own connection to `127.0.0.1:8765`) matched no
pane. `ConnectionIdentity.resolve` then fell through to loopback-trust and resolved that
connection as the **primary** — a worker→primary privilege escalation.
## The fix
`fleetd/src/main/java/dev/ltms/fleet/herdr/PaneLocator.java`:
- `terminalForPid` now builds the caller's **ancestor pid set once** (the pid itself, then its
parents up the tree) before scanning any pane, and reuses that same set across every `HerdrClient`
searched (the CB-185 two-daemon path). It does not walk the tree once per pane.
- The walk is bounded at 32 generations (`MAX_ANCESTRY_DEPTH`) and also stops the moment it revisits
a pid it has already recorded (cycle guard), so a cycle or a pathologically deep tree cannot hang
identity resolution, which runs on every MCP call.
- `paneOwnsPid` (renamed `paneOwnsAnyOf`) now checks whether a pane's `shell_pid` or any
`foreground_processes` pid is a **member of that ancestor set**, instead of comparing the raw pid
directly. A pid that IS the shell pid or a foreground pid still matches (ancestry sets always
contain the pid itself) — no regression there.
- A vanished ancestor (`ParentResolver.parentOf` returning empty) ends the walk quietly, not as an
error — matches the brief.
### The seam: `ParentResolver`
New file `fleetd/src/main/java/dev/ltms/fleet/herdr/ParentResolver.java` — a one-method interface
(`OptionalLong parentOf(long pid)`) so the ancestry walk is testable without spawning real OS
processes. `ParentResolver.PROCESS_HANDLE` is the production implementation, backed by
`ProcessHandle.of(pid)` / `.parent()`. `PaneLocator` gained two new constructor overloads
(`PaneLocator(HerdrClient, ParentResolver)` and `PaneLocator(HerdrClient, HerdrClient,
ParentResolver)`) that accept an injected resolver; the existing constructors delegate to these with
`ParentResolver.PROCESS_HANDLE`, so every existing call site (`Fleetd.java`, `ConnectionIdentity`
tests, the contract test) is unchanged and compiles as-is.
## Tests
New file `fleetd/src/test/java/dev/ltms/fleet/herdr/FakeParentResolver.java` — a fake `ParentResolver`
backed by an explicit pid→parent `Map`.
New tests added to `fleetd/src/test/java/dev/ltms/fleet/herdr/PaneLocatorTest.java` (using a new
small `OnePaneHerdr` fake nested in that test class, purpose-built so shell_pid and a foreground pid
can be given distinct values — the shared `FakeHerdr` fixture happens to use the same pid for both,
which would not isolate the two cases):
1. `stillResolvesAPidThatIsExactlyThePaneShellPid` — regression, no ancestry walk needed.
2. `stillResolvesAPidThatIsExactlyAForegroundPid` — regression, via the foreground-processes list.
3. `resolvesAGrandchildPidTwoLevelsBelowTheShellPid` — **the bug.** grandchild → child → shell_pid.
4. `nullForAPidWhoseAncestryMatchesNoPane` — a pid whose ancestry never reaches any pane still
resolves to `null`, so the real primary is not demoted to a worker.
5. `ancestryWalkTerminatesOnACycleInsteadOfHanging` — a 2-cycle fake parent map (100→101→100)
terminates and returns `null` instead of hanging.
6. `ancestrySetIsComputedOnceAcrossBothClientsInTheTwoDaemonConstructor` — wraps the fake resolver
with a call counter, uses the two-arg (CB-185) constructor with one client reporting no panes,
and asserts exactly 3 `parentOf` calls (not 6) — the ancestor set is walked once and reused
across both clients, not recomputed per client.
All existing `PaneLocatorTest` tests (the CB-185 two-daemon fallback tests) still pass unchanged.
### Proof that tests 3 and 5 fail without the fix
I temporarily reverted `ancestorsOf` twice, ran the specific test, then restored the real fix
(`git diff` is clean of these — they were never committed):
**Test 3, reverted to "no ancestry walk at all" (`return Set.of(pid);`):**
```
mvn -Dtest=PaneLocatorTest#resolvesAGrandchildPidTwoLevelsBelowTheShellPid test
...
org.opentest4j.AssertionFailedError: expected: <term_x> but was: <null>
[ERROR] PaneLocatorTest.resolvesAGrandchildPidTwoLevelsBelowTheShellPid:100 expected: <term_x> but was: <null>
[ERROR] Tests run: 1, Failures: 1, Errors: 0, Skipped: 0
[INFO] BUILD FAILURE
```
**Test 5, reverted to a naive walk (no `MAX_ANCESTRY_DEPTH` bound, no "already seen" cycle guard —
just `while (current > 0) { ancestry.add(current); ... }`):**
Ran `mvn -Dtest=PaneLocatorTest#ancestryWalkTerminatesOnACycleInsteadOfHanging test` in the
background and polled it. It was still running (stuck inside `PaneLocatorTest`, no test-run summary
line ever printed) after 20 seconds and had to be force-killed — confirming the naive walk hangs on
the 100↔101 cycle fixture. The real fix's depth bound and cycle guard are what make this test pass
in ~0.25s instead.
Both reverts were made only in-memory in my worktree (I kept a copy of the correct file at
`/tmp/PaneLocator.java.fixed` beforehand) and restored via `cp` before running the full build below.
`git status` / `git diff` at commit time show only the intended fix, not either revert.
## Full build
```
cd fleetd && mvn clean install
```
Ran to completion, unpiped. Final lines:
```
[INFO] Tests run: 1043, Failures: 0, Errors: 0, Skipped: 0
[INFO] BUILD SUCCESS
[INFO] Total time: 35.061 s
```
(main was 1037 tests; this branch adds 6 new tests in `PaneLocatorTest` = 1043.)
## Scope discipline — what I did NOT touch
Per the brief: only `PaneLocator`, its tests, and the minimal `ParentResolver` seam + its fake.
`Authz`, `ConnectionIdentity`'s fallthrough, and the loopback-trust policy are unchanged — those
stay the deliberate primary-only decision they were.
One thing spotted but out of scope, noted only: `terminalForPid` still issues one
`pane.process_info` herdr call **per pane** on every resolution (no caching) — the class javadoc
already flags this as "the obvious optimization once wired into `ClaudeCodeLauncher`." Not touched.
## What I could NOT check
- No IDE tooling (no `ide_diagnostics` / IntelliJ inspections) — `mvn clean install` is my only
verification, as the worker charter expects.
- Cannot restart or query the live `fleetd` daemon — this fix is unverified against a real herdr
process tree. The pre-existing `PaneLocatorContractTest` (tagged `contract`, run with
`-Pcontract`) exercises `PaneLocator` against a real herdr daemon and a real shell pid (not a
grandchild), but I did not run it — it requires a live herdr socket, which I don't have in this
worktree, and it wasn't part of the standard `mvn clean install` run above (it's a separate
Maven profile). I did not add a grandchild-pid variant of it, since the brief scoped tests to
unit tests against a fake resolver.
- I did not attempt to reproduce the original privilege-escalation scenario end-to-end (spawning a
real grandchild process against a live daemon) — the ticket says that half is already measured
and confirmed, and I have no way to drive a live daemon from this worktree.
-18
View File
@@ -28,7 +28,6 @@
<testcontainers.version>1.20.4</testcontainers.version>
<commons-compress.version>1.27.1</commons-compress.version>
<commons-lang3.version>3.18.0</commons-lang3.version>
<sqlite-jdbc.version>3.53.4.0</sqlite-jdbc.version>
</properties>
<!--
@@ -45,12 +44,6 @@
3.0-rc5; bumping Jackson 3 to the patched 3.2.x breaks the SDK (annotation mismatch).
Only the loopback /mcp endpoint parses this JSON, from trusted local Claude clients.
The 11.0.23 -> 11.0.25 bump did clear jetty CVE-2024-8184 (5.9) and CVE-2024-6763.
fleetd #206: org.xerial:sqlite-jdbc 3.53.4.0 (added for OpenCodeSessionDiscovery) — the
only known advisory against this artifact is CVE-2023-32697 (RCE via an attacker-controlled
JDBC URL), fixed in 3.41.2.2; 3.53.4.0 is well past that fix and OSV.dev reports no open
advisory against it. Checked via the OSV.dev API (no Mend.io/JetBrains IDE MCP mount
available from this worktree) on 2026-08-31.
-->
<!-- Force the latest patched Jetty 11.x across all Javalin-pulled Jetty modules (no version
@@ -130,17 +123,6 @@
<version>${amqp.version}</version>
</dependency>
<!-- fleetd #206: opencode moved its session store from a JSON tree to SQLite
(opencode.db). This is the JDBC driver OpenCodeSessionDiscovery uses to read it
read-only. Ships bundled native libraries (linux/mac/windows, several archs), so it
is a heavier jar than most deps here — see the pom's dependency-security note below
for the size/CVE tradeoff actually measured. -->
<dependency>
<groupId>org.xerial</groupId>
<artifactId>sqlite-jdbc</artifactId>
<version>${sqlite-jdbc.version}</version>
</dependency>
<!-- Logging -->
<dependency>
<groupId>org.slf4j</groupId>
@@ -177,14 +177,14 @@ public final class Fleetd {
claudeProfiles, cfg.effectiveDefaultProfile(), System::getenv,
cfg.spawnReadyTimeoutMs(), cfg.spawnReadyPollMs(),
() -> config.get().fleet(),
() -> config.get().memberCredentials(), null, config::get));
() -> config.get().memberCredentials()));
}
if (!opencodeProfiles.isEmpty()) {
adapters.add(new OpenCodeLauncher(router.memberAgents(), router.memberSpaces(),
opencodeProfiles, cfg.effectiveDefaultProfile(), System::getenv,
cfg.spawnReadyTimeoutMs(), cfg.spawnReadyPollMs(),
() -> config.get().fleet(),
() -> config.get().memberCredentials(), config::get));
() -> config.get().memberCredentials()));
}
AtomicReference<Function<String, Integer>> liveCountRef = new AtomicReference<>(_ -> 0);
// CB-578 stage B: one quarantine tracker for the whole daemon, shared between the launcher
@@ -94,26 +94,13 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
public ClaudeCodeLauncher(AgentControl agents, WorkspaceControl spaces, SubscriptionGuard guard,
Map<String, FleetConfig.Profile> profiles, String defaultProfile,
Function<String, String> env,
long spawnReadyTimeoutMs, long spawnReadyPollMs,
Supplier<FleetConfig.Fleet> fleet,
Supplier<FleetConfig.MemberCredentials> memberCredentials) {
this(agents, spaces, guard, profiles, defaultProfile, env, spawnReadyTimeoutMs, spawnReadyPollMs,
fleet, memberCredentials, null, null);
}
/** Production constructor, plus the live config for URI environment exclusions. */
public ClaudeCodeLauncher(AgentControl agents, WorkspaceControl spaces, SubscriptionGuard guard,
Map<String, FleetConfig.Profile> profiles, String defaultProfile,
Function<String, String> env,
long spawnReadyTimeoutMs, long spawnReadyPollMs,
Supplier<FleetConfig.Fleet> fleet,
Supplier<FleetConfig.MemberCredentials> memberCredentials,
Supplier<Set<String>> hostEnvNames,
Supplier<FleetConfig> config) {
long spawnReadyTimeoutMs, long spawnReadyPollMs,
Supplier<FleetConfig.Fleet> fleet,
Supplier<FleetConfig.MemberCredentials> memberCredentials) {
this(agents, spaces, guard, profiles, defaultProfile, env,
spawnReadyTimeoutMs,
System::currentTimeMillis, () -> sleepUninterruptibly(spawnReadyPollMs),
fleet, memberCredentials, hostEnvNames, config);
fleet, memberCredentials);
}
/**
@@ -166,22 +153,10 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
Function<String, String> env,
long spawnReadyTimeoutMs,
LongSupplier nowMillis, Runnable sleeper,
Supplier<FleetConfig.Fleet> fleet,
Supplier<FleetConfig.MemberCredentials> memberCredentials) {
this(agents, spaces, guard, profiles, defaultProfile, env, spawnReadyTimeoutMs, nowMillis, sleeper,
fleet, memberCredentials, null, null);
}
/** Full testability constructor, plus the live config for URI environment exclusions. */
public ClaudeCodeLauncher(AgentControl agents, WorkspaceControl spaces, SubscriptionGuard guard,
Map<String, FleetConfig.Profile> profiles, String defaultProfile,
Function<String, String> env, long spawnReadyTimeoutMs,
LongSupplier nowMillis, Runnable sleeper, Supplier<FleetConfig.Fleet> fleet,
Supplier<FleetConfig.MemberCredentials> memberCredentials,
Supplier<Set<String>> hostEnvNames,
Supplier<FleetConfig> config) {
Supplier<FleetConfig.Fleet> fleet,
Supplier<FleetConfig.MemberCredentials> memberCredentials) {
super(NAME_PREFIX, agents, spaces, profiles, defaultProfile, env,
spawnReadyTimeoutMs, nowMillis, sleeper, fleet, memberCredentials, hostEnvNames, config);
spawnReadyTimeoutMs, nowMillis, sleeper, fleet, memberCredentials);
this.guard = guard;
}
@@ -199,7 +174,7 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
Supplier<FleetConfig.MemberCredentials> memberCredentials,
Supplier<Set<String>> hostEnvNames) {
super(NAME_PREFIX, agents, spaces, profiles, defaultProfile, env,
spawnReadyTimeoutMs, nowMillis, sleeper, fleet, memberCredentials, hostEnvNames, null);
spawnReadyTimeoutMs, nowMillis, sleeper, fleet, memberCredentials, hostEnvNames);
this.guard = guard;
}
@@ -170,8 +170,6 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
/** Guards {@link #warnNonZsh} to one WARN per launcher instance, not one per spawn. */
private final AtomicBoolean nonZshShellWarned = new AtomicBoolean();
/** Live config provides URI environment names that must never enter member panes. */
private final Supplier<FleetConfig> config;
/**
* @param namePrefix label prefix for this peer kind (drives naming and reap)
@@ -240,19 +238,8 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
long spawnReadyTimeoutMs,
LongSupplier nowMillis, Runnable sleeper,
Supplier<FleetConfig.Fleet> fleet,
Supplier<FleetConfig.MemberCredentials> memberCredentials,
Supplier<Set<String>> hostEnvNames) {
this(namePrefix, agents, spaces, profiles, defaultProfile, env, spawnReadyTimeoutMs, nowMillis,
sleeper, fleet, memberCredentials, hostEnvNames, null);
}
/** As above, plus the live full config for secret-bearing URI environment names. */
protected HerdrPeerLauncher(String namePrefix, AgentControl agents, WorkspaceControl spaces,
Map<String, FleetConfig.Profile> profiles, String defaultProfile,
Function<String, String> env, long spawnReadyTimeoutMs,
LongSupplier nowMillis, Runnable sleeper, Supplier<FleetConfig.Fleet> fleet,
Supplier<FleetConfig.MemberCredentials> memberCredentials,
Supplier<Set<String>> hostEnvNames, Supplier<FleetConfig> config) {
Supplier<Set<String>> hostEnvNames) {
this.fleet = fleet;
this.namePrefix = namePrefix;
this.agents = agents;
@@ -265,7 +252,6 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
this.sleeper = sleeper;
this.memberCredentials = memberCredentials;
this.hostEnvNames = hostEnvNames != null ? hostEnvNames : () -> System.getenv().keySet();
this.config = config;
}
// --- adapter seams -------------------------------------------------------------------------
@@ -1042,11 +1028,9 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
}
/** Put {@link #BLOCKED_CREDENTIAL_SENTINEL} over every blocked name in the pane-creation env map. */
private void overlayBlockedCredentials(Map<String, String> workerEnv,
FleetConfig.MemberCredentials creds) {
Set<String> blocked = new java.util.TreeSet<>(creds.blockedSet());
blocked.addAll(brokerUriEnvNames());
for (String name : blocked) {
private static void overlayBlockedCredentials(Map<String, String> workerEnv,
FleetConfig.MemberCredentials creds) {
for (String name : creds.blockedSet()) {
workerEnv.put(name, BLOCKED_CREDENTIAL_SENTINEL);
}
}
@@ -1113,21 +1097,15 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
* ssh-agent handle when explicitly allowed, and the exact keys of THIS launch's own env map.
*/
private Set<String> derivedAllowedNames(FleetConfig.MemberCredentials creds, Launch launch) {
Set<String> brokerUriEnvNames = brokerUriEnvNames();
Set<String> allowed = new java.util.TreeSet<>(
MemberEnvAllowList.derive(profiles.values(), creds.allowSet(), brokerUriEnvNames));
MemberEnvAllowList.derive(profiles.values(), creds.allowSet()));
if (creds.sshAuthSockAllowed()) {
allowed.add(SSH_AUTH_SOCK);
} // blocked by default: absent from the set ⇒ blanked by the scrub like any other name
allowed.addAll(launch.env().keySet());
allowed.removeAll(brokerUriEnvNames);
return allowed;
}
private Set<String> brokerUriEnvNames() {
return MemberEnvAllowList.brokerUriEnvNames(config == null ? null : config.get());
}
/**
* CB-633 follow-up: one INFO line per allow-list spawn WHOSE SCRUB ACTUALLY RUNS, so an operator
* can read a single log line and know the scrub ran and how much of the visible environment it
@@ -40,9 +40,8 @@ import java.util.TreeSet;
* operator's own explicit list. Before this, {@code policy: allow-list} silently ignored every name
* an operator wrote under {@code allow:} unless a profile happened to carry it too, which meant
* turning the policy on could blank credentials working members already depended on. {@code
* SSH_AUTH_SOCK} and configured broker URI environment names are exceptions: even when the operator
* lists them under {@code allow:}, they are excluded here. {@code SSH_AUTH_SOCK} is added back ONLY
* by the caller when {@code sshAuthSock: allow} is explicitly set
* SSH_AUTH_SOCK} is the one exception: even when the operator lists it under {@code allow:}, it is
* excluded here and added back ONLY by the caller when {@code sshAuthSock: allow} is explicitly set
* (see {@link #SSH_AUTH_SOCK}'s javadoc) — it is a live handle to the operator's own ssh-agent, not
* a value, so treating it like any other allow-listed name would hand a member every key the
* operator's agent holds the moment they typed the name under {@code allow:} for an unrelated
@@ -107,16 +106,6 @@ public final class MemberEnvAllowList {
* run-to-run.
*/
public static Set<String> derive(Collection<FleetConfig.Profile> profiles, Set<String> configuredAllow) {
return derive(profiles, configuredAllow, Set.of());
}
/**
* As {@link #derive(Collection, Set)}, while excluding names that fleetd knows carry credentials.
* A configured broker URI contains its AMQP password inline, so it must never reach a member,
* even when an operator put its variable name in {@code memberCredentials.allow:}.
*/
public static Set<String> derive(Collection<FleetConfig.Profile> profiles, Set<String> configuredAllow,
Set<String> excludedNames) {
Set<String> derived = new TreeSet<>(INFRASTRUCTURE_PASSTHROUGH);
if (profiles != null) {
for (FleetConfig.Profile p : profiles) {
@@ -135,37 +124,9 @@ public final class MemberEnvAllowList {
}
}
}
if (excludedNames != null) {
derived.removeAll(excludedNames);
}
return Set.copyOf(derived);
}
/**
* The host environment names whose values are AMQP URIs with inline passwords. Both broker
* connections belong to fleetd, never to a member pane. Blank and absent configuration changes
* nothing.
*
* <p><b>How strong this exclusion is depends on the policy, and the difference matters.</b>
* Under {@code policy: allow-list} it is enforced by the generated ZDOTDIR scrub, which runs
* AFTER the pane's shell has sourced the operator's chain — so a login shell that re-exports the
* name is still blanked. Under the deny-list policy there is no scrub: the name is only removed
* from the pre-shell env map, and a login shell that sources the operator's secret store
* re-exports it. That is the long-standing weakness of deny-list (a sourced file can undo it),
* not something this exclusion introduces, but it means deny-list deployments do NOT get this
* guarantee. The same caveat applies to the non-zsh path, which has no scrub at all — see
* {@code HerdrPeerLauncher#applyEnvironmentAllowListPolicy}.
*/
public static Set<String> brokerUriEnvNames(FleetConfig config) {
if (config == null) {
return Set.of();
}
Set<String> names = new TreeSet<>();
addUriEnvIfPresent(names, config.broker());
addUriEnvIfPresent(names, config.coordinator());
return Set.copyOf(names);
}
/**
* Whether {@code name} survives the scrub when {@code allowedNames} is the derived set: an exact
* match, or an infrastructure-prefixed name ({@code LC_*}). Prefix rules live ONLY here and in
@@ -185,16 +146,4 @@ public final class MemberEnvAllowList {
into.add(name);
}
}
private static void addUriEnvIfPresent(Set<String> into, FleetConfig.Broker broker) {
if (broker != null && broker.hasUriEnv()) {
into.add(broker.uriEnv());
}
}
private static void addUriEnvIfPresent(Set<String> into, FleetConfig.Coordinator coordinator) {
if (coordinator != null && coordinator.hasUriEnv()) {
into.add(coordinator.uriEnv());
}
}
}
@@ -116,23 +116,12 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
public OpenCodeLauncher(AgentControl agents, WorkspaceControl spaces,
Map<String, FleetConfig.Profile> profiles, String defaultProfile,
Function<String, String> env,
long spawnReadyTimeoutMs, long spawnReadyPollMs,
Supplier<FleetConfig.Fleet> fleet,
Supplier<FleetConfig.MemberCredentials> memberCredentials) {
this(agents, spaces, profiles, defaultProfile, env, spawnReadyTimeoutMs, spawnReadyPollMs,
fleet, memberCredentials, null);
}
/** Production constructor, plus the live config for URI environment exclusions. */
public OpenCodeLauncher(AgentControl agents, WorkspaceControl spaces,
Map<String, FleetConfig.Profile> profiles, String defaultProfile,
Function<String, String> env, long spawnReadyTimeoutMs, long spawnReadyPollMs,
Supplier<FleetConfig.Fleet> fleet,
Supplier<FleetConfig.MemberCredentials> memberCredentials,
Supplier<FleetConfig> config) {
long spawnReadyTimeoutMs, long spawnReadyPollMs,
Supplier<FleetConfig.Fleet> fleet,
Supplier<FleetConfig.MemberCredentials> memberCredentials) {
this(agents, spaces, profiles, defaultProfile, env, spawnReadyTimeoutMs,
System::currentTimeMillis, () -> sleepUninterruptibly(spawnReadyPollMs),
defaultConfigRoot(), defaultDiscoveryRoot(), fleet, memberCredentials, config);
defaultConfigRoot(), defaultDiscoveryRoot(), fleet, memberCredentials);
}
/**
@@ -193,20 +182,8 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
Path configRoot, Path discoveryRoot,
Supplier<FleetConfig.Fleet> fleet,
Supplier<FleetConfig.MemberCredentials> memberCredentials) {
this(agents, spaces, profiles, defaultProfile, env, spawnReadyTimeoutMs, nowMillis, sleeper,
configRoot, discoveryRoot, fleet, memberCredentials, null);
}
/** Full testability constructor, plus the live config for URI environment exclusions. */
public OpenCodeLauncher(AgentControl agents, WorkspaceControl spaces,
Map<String, FleetConfig.Profile> profiles, String defaultProfile,
Function<String, String> env, long spawnReadyTimeoutMs,
LongSupplier nowMillis, Runnable sleeper, Path configRoot, Path discoveryRoot,
Supplier<FleetConfig.Fleet> fleet,
Supplier<FleetConfig.MemberCredentials> memberCredentials,
Supplier<FleetConfig> config) {
super(NAME_PREFIX, agents, spaces, profiles, defaultProfile, env,
spawnReadyTimeoutMs, nowMillis, sleeper, fleet, memberCredentials, null, config);
spawnReadyTimeoutMs, nowMillis, sleeper, fleet, memberCredentials);
this.configRoot = configRoot;
this.discovery = new OpenCodeSessionDiscovery(discoveryRoot);
}
@@ -1,87 +1,58 @@
package dev.ltms.fleet.member;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import org.sqlite.SQLiteConfig;
import com.fasterxml.jackson.databind.JsonNode;
import com.fasterxml.jackson.databind.ObjectMapper;
import java.io.IOException;
import java.nio.file.Files;
import java.nio.file.Path;
import java.sql.Connection;
import java.sql.PreparedStatement;
import java.sql.ResultSet;
import java.sql.SQLException;
import java.util.concurrent.atomic.AtomicBoolean;
import java.util.stream.Stream;
/**
* Resolves the opencode session id for a fleetd worker from opencode's on-disk storage — the
* only place this adapter touches opencode's private layout, and deliberately the <em>only</em>
* class that does.
*
* <p><strong>Why this is isolated behind one seam.</strong> The layout is version-coupled and not
* a stable contract: opencode persists its session state in a SQLite database at
* {@code <storageRoot>/opencode.db} (a {@code session} table, one row per session, keyed by id and
* carrying a {@code directory} column). That schema can move between opencode releases exactly
* like the JSON-file layout it replaced did (opencode migrated off a one-JSON-file-per-session
* tree under {@code <storageRoot>/storage/session/<projectID>/ses_*.json} in January 2026 — that
* tree is now a frozen migration artefact nothing writes, which is why this class no longer reads
* it). opencode also ships a headless HTTP server that may supersede both of these entirely.
* Everything this adapter knows about that private storage — its shape and column names — lives
* here, so a layout change, or a switch to the HTTP server, changes exactly one class and nothing
* in {@link OpenCodeLauncher}.
* <p><strong>Why this is isolated behind one seam.</strong> The layout is version-coupled and not a
* stable contract: opencode writes one JSON file per session under
* {@code <storageRoot>/session/<projectID>/<ses_*.json>}, and each record carries a
* {@code "version"} field (e.g. {@code "1.1.31"}), so the exact directory shape, file naming, and
* field names can move between opencode releases. opencode also ships a headless HTTP server that
* may supersede file scanning entirely. Everything this adapter knows about that private storage —
* its shape, naming, and field names — lives here, so a layout change, or a switch to the HTTP
* server, changes exactly one class and nothing in {@link OpenCodeLauncher}.
*
* <p>The determinism that makes this useful is structural, not a guess: every fleetd worker runs
* in its own unique git worktree, so the row's {@code directory} (its project root) equals the
* in its own unique git worktree, so the record's {@code directory} (its project root) equals the
* worker's cwd identifies <em>its</em> session unambiguously. We match on {@code directory} rather
* than diffing {@code opencode session list} before/after — that races under concurrent spawns, and
* the CLI listing does not even show the directory.
*
* <p>All reads are best-effort and never throw: a missing or unreadable database, a query that
* fails, or a directory with no row yet all yield {@code null}, and the caller (the session
* handle) treats that as "identity not resolved yet" and retries later. The database is opened
* read-only and never written to: opencode itself may be running and writing it concurrently (WAL
* mode), and this class must never disturb that.
* <p>All reads are best-effort and never throw: a missing or unreadable storage root, a record that
* fails to parse, or a directory with no record yet all yield {@code null}, and the caller (the
* session handle) treats that as "identity not resolved yet" and retries later.
*/
final class OpenCodeSessionDiscovery {
private static final Logger log = LoggerFactory.getLogger(OpenCodeSessionDiscovery.class);
private final Path storageRoot; // e.g. ~/.local/share/opencode (injectable for tests)
private final Path databasePath;
private final AtomicBoolean warnedMissingDatabase = new AtomicBoolean(false);
private final ObjectMapper json;
OpenCodeSessionDiscovery(Path storageRoot) {
this.storageRoot = storageRoot;
this.databasePath = storageRoot.resolve("opencode.db");
this.json = new ObjectMapper();
}
/**
* A connection to {@link #databasePath} opened with SQLite's {@code SQLITE_OPEN_READONLY}
* flag: it never creates the file, never writes, and never touches WAL or journal mode.
* opencode may be running and writing this database concurrently, and this class must never
* disturb it.
*
* <p>Package-private so a test can hold the connection and prove it refuses a write. That is
* the only way to pin this property: making the file unwritable does <em>not</em> work,
* because SQLite silently downgrades a read-write open of an unwritable file to read-only, so
* such a test passes whether or not the flag is set.
*/
Connection openReadOnly() throws SQLException {
SQLiteConfig config = new SQLiteConfig();
config.setReadOnly(true);
return config.createConnection("jdbc:sqlite:" + databasePath);
}
/**
* The opencode session id whose row references {@code directory} (the worker's cwd), or
* {@code null} when no row matches yet. When several rows share the directory — e.g. repeated
* spawns into the same worktree — the row with the highest {@code time_updated} wins: it is
* The opencode session id whose record references {@code directory} (the worker's cwd), or
* {@code null} when no record matches yet. When several records share the directory — e.g.
* repeated spawns into the same worktree — the <em>most recently modified</em> one wins: it is
* the session the pane most likely corresponds to.
*
* <p>Never throws: a missing {@code opencode.db}, a locked/unreadable database, a query
* failure, or a directory that has not been persisted yet all resolve to {@code null} rather
* than failing a spawn. A fleetd worker's session row is written lazily (when the session is
* first persisted), so {@code null} here is the normal answer right after the pane is ready,
* and the caller retries later.
* <p>Never throws: a missing {@code storageRoot}, an unreadable/malformed record, or a
* directory that has not been persisted yet all resolve to {@code null} rather than failing a
* spawn. A fleetd worker's session record is written lazily (when the session is first
* persisted), so {@code null} here is the normal answer right after the pane is ready, and the
* caller retries later.
*
* @param directory the worker's cwd, as resolved for this spawn
* @return the matching session id, or {@code null} if none is known yet
@@ -90,31 +61,62 @@ final class OpenCodeSessionDiscovery {
if (directory == null || directory.isBlank()) {
return null;
}
if (!Files.isRegularFile(databasePath)) {
if (warnedMissingDatabase.compareAndSet(false, true)) {
log.warn("opencode session database not found at {} — opencode's on-disk layout "
+ "may have moved again; session discovery will keep returning null",
databasePath);
}
Path sessionRoot = storageRoot.resolve("session");
if (!Files.isDirectory(sessionRoot)) {
return null;
}
String sql = "SELECT id FROM session WHERE directory = ? ORDER BY time_updated DESC LIMIT 1";
try (Connection connection = openReadOnly();
PreparedStatement statement = connection.prepareStatement(sql)) {
statement.setString(1, directory);
try (ResultSet rows = statement.executeQuery()) {
if (rows.next()) {
return rows.getString("id");
String best = null;
long bestMtime = Long.MIN_VALUE;
try (Stream<Path> projectDirs = Files.list(sessionRoot)) {
for (Path projectDir : projectDirs.filter(Files::isDirectory).toList()) {
try (Stream<Path> records = Files.list(projectDir)) {
for (Path record : records.toList()) {
String id = matchId(record, directory);
if (id == null) {
continue;
}
long mtime = lastModifiedEpochMillis(record);
if (mtime > bestMtime) {
bestMtime = mtime;
best = id;
}
}
} catch (IOException ignored) {
// one project dir unreadable — skip it; another may still match
}
}
} catch (SQLException e) {
// Locked, corrupt, or otherwise unreadable — never fatal to a spawn. Not the
// "database moved" signal (the file exists), so this stays below WARN.
log.debug("opencode session database unreadable at {}: {}", databasePath, e.toString());
} catch (IOException ignored) {
// storage root vanished or became unreadable — "no session known yet"
return null;
}
log.debug("no opencode session row for directory (root={}, directory={})",
storageRoot, directory);
return null;
return best;
}
/**
* The record's session id when it references {@code directory}, else {@code null}. A record
* that is not JSON, lacks {@code id}/{@code directory}, or points at a different directory is
* simply not our session; a malformed one is skipped, never fatal.
*/
private String matchId(Path record, String directory) {
try {
JsonNode node = json.readTree(record.toFile());
JsonNode id = node == null ? null : node.get("id");
JsonNode dir = node == null ? null : node.get("directory");
if (id == null || dir == null || !directory.equals(dir.asText())) {
return null;
}
return id.asText();
} catch (IOException e) {
return null;
}
}
/** The record's last-modified epoch ms, or {@code Long.MIN_VALUE} if unreadable (never wins). */
private static long lastModifiedEpochMillis(Path record) {
try {
return Files.getLastModifiedTime(record).toMillis();
} catch (IOException e) {
return Long.MIN_VALUE;
}
}
}
@@ -366,40 +366,21 @@ public final class MessageService {
}
/**
* Route a worker's explicit {@code fleet_reply}: resolve an open send, complete an async ticket
* still parked waiting on this exact turn's answer, or — only once neither applies — queue it in
* the inbox. Unlike the bare {@link Rendezvous#resolve}, a no-waiter result is <em>not</em> a
* failure — the reply is held for later drain.
* Route a worker's explicit {@code fleet_reply}: resolve an open send, or queue it in the
* inbox if no send is currently open. Unlike the bare {@link Rendezvous#resolve}, a no-waiter
* result is <em>not</em> a failure — the reply is held for later drain.
*
* <p><strong>Do NOT use this for mid-turn questions.</strong> {@code fleet_ask} /
* {@link Rendezvous#resolveQuestion} must keep today's {@code NO_WAITER} behaviour — questions
* are interactive and must never be queued.
*
* @return always {@code true} — the reply resolved a live send, completed a parked ticket, or
* was queued
* @return always {@code true} — the reply either resolved a live send or was queued
*/
public boolean reply(String session, String content) {
if (rendezvous.resolve(session, content)) {
count(FleetMetrics.REPLIES, "path", "rendezvous");
return true; // a live send took it — unchanged fast path
}
// #137: no live rendezvous waiter, but this may be the worker's real fleet_reply resuming a
// turn that {@link #answer} already gave up waiting on. answer()'s own bounded wait (the
// primary's fleet_send{turnId} call, capped well under a minute) can time out and close its
// waiter long before the worker — now actually resuming real work — finishes and replies. That
// reply used to have nowhere to land but the session inbox, leaving the async ticket's future
// unresolved forever: fleet_poll{ticket} stayed PENDING until fleet_stop's abandon() forced it
// FAILED with a misleading "session released before it replied" reason, even though the reply
// had, in fact, arrived. Completing the matching ticket directly here means fleet_poll{ticket}
// sees the real reply instead.
Task orphan = askAnsweredAsyncTask(session);
if (orphan != null && orphan.future.complete(new Reply(Outcome.REPLIED, content))) {
if (orphan.turnId != null) {
asyncTasksByTurn.remove(orphan.turnId, orphan);
}
count(FleetMetrics.REPLIES, "path", "async-recovered");
return true; // the ticket itself took it — no inbox stranding at all
}
inbox.publish(session, UUID.randomUUID().toString(), content);
// CB-640: record the stranding itself (not just the reply text) so fleet health can see a
// worker whose replies keep missing their waiter, not only the queue depth this leaves behind.
@@ -413,24 +394,6 @@ public final class MessageService {
return true; // held, not lost
}
/**
* The still-open async task on {@code target} whose {@code fleet_ask} was already answered — its
* {@link Task#turnId} is stamped but its {@link Task#question} was cleared by {@link #answer} —
* yet whose future is not resolved yet (#137). {@code null} if no such task exists, including the
* common case where {@code target}'s worker never used {@code fleet_ask} at all (a task that was
* never asked has {@code turnId == null}, so it can never match here and only ever completes
* through the ordinary rendezvous fast path in {@link #reply}).
*/
private Task askAnsweredAsyncTask(String target) {
for (Task task : tasks.values()) {
if (target.equals(task.target) && task.question == null && task.turnId != null
&& !task.future.isDone()) {
return task;
}
}
return null;
}
/** Record a counter sample when a registry is wired; a no-op in unit tests. */
private void count(String name, String... labels) {
if (metrics != null) {
@@ -472,43 +435,19 @@ public final class MessageService {
* <p>Resolving the waiter as a failure — rather than letting it time out — also means the
* outcome is counted, so a torn-down delegation stops being invisible to {@code /metrics}.
*
* <p><strong>#137 defence in depth.</strong> {@link #reply} already hands a worker's real
* {@code fleet_reply} straight to the async ticket it belongs to whenever one is still parked
* waiting for it (see {@link #askAnsweredAsyncTask}), so by the time a session is released its
* tasks are normally already resolved — this loop's {@code complete} calls are then harmless
* no-ops (a {@link CompletableFuture} can only resolve once). But should some other path someday
* strand a reply in the inbox without completing its ticket, checking
* {@link #hasStrandedReply(String)} here — before ever writing a failure — means a torn-down
* session whose worker in fact replied is still reported {@code REPLIED} with that reply's own
* text, never the misleading "the worker session was released before it replied" (which also
* means the snapshot/worktree recovery hint that follows it never prints once a reply exists).
*
* @return true if a live waiter or an async task was failed (never true for one recovered as a
* reply — see the note above)
* @return true if a live waiter was failed
*/
public boolean abandon(String target, String reason) {
boolean hadStrandedReply = hasStrandedReply(target);
// CB-640: the session is gone — nothing will ever accept or deliver into it now.
strandedReplies.remove(target);
queuedDeliveries.remove(target);
CompletableFuture<Rendezvous.Resolution> waiter = rendezvous.currentWaiter(target);
boolean failed = waiter != null && !waiter.isDone() && rendezvous.resolveFailure(waiter, reason);
boolean asyncFailed = false;
Reply recovered = null; // lazily drained at most once, only if a task actually needs it
for (Task task : tasks.values()) {
if (!target.equals(task.target) || task.question != null || task.future.isDone()) {
continue;
}
if (hadStrandedReply && recovered == null) {
recovered = recoverStrandedReply(target);
}
Reply outcome = recovered != null ? recovered : new Reply(Outcome.WORKER_FAILED, reason);
if (task.future.complete(outcome)) {
if (outcome.outcome() == Outcome.WORKER_FAILED) {
asyncFailed = true;
} else if (task.turnId != null) {
asyncTasksByTurn.remove(task.turnId, task);
}
if (target.equals(task.target) && task.question == null
&& task.future.complete(new Reply(Outcome.WORKER_FAILED, reason))) {
asyncFailed = true;
}
}
if (failed) {
@@ -517,21 +456,6 @@ public final class MessageService {
return failed || asyncFailed;
}
/**
* Drain {@code target}'s inbox and hand its content back as a {@link Outcome#REPLIED} result
* (#137 defence in depth for {@link #abandon}) — {@code null} if it turned out empty (the
* stranding fact raced away, e.g. a lead's own {@code fleet_poll} on the raw session already
* drained it first). When more than one message is queued, only the newest is the worker's actual
* final answer ({@link #drainReplies} returns them oldest-first).
*/
private Reply recoverStrandedReply(String target) {
var messages = drainReplies(target);
if (messages.isEmpty()) {
return null;
}
return new Reply(Outcome.REPLIED, messages.get(messages.size() - 1).content());
}
/**
* Acknowledge a specific reply by {@code msgId} for {@code target}. Removes it from the inbox
* so that a subsequent drain or peek no longer returns it.
@@ -161,34 +161,6 @@ class HerdrPeerLauncherAllowListWiringTest {
+ "it under allow: — sshAuthSock is unset here, so it defaults to block");
}
@Test
void brokerUriEnvStaysBlockedWhenListedInMemberCredentialsAllow() {
FakeHerdr herdr = new FakeHerdr();
WiringLauncher launcher = new WiringLauncher(herdr,
allowListWithAllow(List.of("BROKER_CONNECTION_URI")), "/bin/zsh", null,
() -> config("BROKER_CONNECTION_URI"));
launcher.spawn(new SpawnRequest("test", null, null, null, null, MemberRole.DEV));
Path dir = Path.of(launcher.env.get("ZDOTDIR"));
assertFalse(readAll(dir.resolve(EnvAllowListScrub.SCRUB_FILE)).contains("'BROKER_CONNECTION_URI'"),
"broker.uriEnv must not reach a member even when listed in memberCredentials.allow:");
}
@Test
void brokerUriEnvIsDeniedUnderTheDenyListPolicyEvenWhenAllowed() {
FakeHerdr herdr = new FakeHerdr();
WiringLauncher launcher = new WiringLauncher(herdr,
() -> new FleetConfig.MemberCredentials(null, List.of("BROKER_CONNECTION_URI"), List.of(), null),
"/bin/bash", null,
() -> config("BROKER_CONNECTION_URI"));
launcher.spawn(new SpawnRequest("test", null, null, null, null, MemberRole.DEV));
assertEquals("blocked-by-fleetd-cb596-see-gitea-issue-82", launcher.env.get("BROKER_CONNECTION_URI"),
"the deny-list overlay must deny broker.uriEnv even when allow: names it");
}
/**
* CB-633 follow-up criterion 3: on every allow-list spawn the daemon logs one INFO line, shaped
* "member credentials: allowed N of M", with real counts — not constants. Real path: the count
@@ -288,24 +260,15 @@ class HerdrPeerLauncherAllowListWiringTest {
/** Plus an injectable {@code hostEnvNames} source, for the "allowed N of M" log line test. */
WiringLauncher(FakeHerdr herdr, Supplier<FleetConfig.MemberCredentials> creds, String shell,
Supplier<Set<String>> hostEnvNames) {
this(herdr, creds, shell, hostEnvNames, null);
}
WiringLauncher(FakeHerdr herdr, Supplier<FleetConfig.MemberCredentials> creds, String shell,
Supplier<Set<String>> hostEnvNames, Supplier<FleetConfig> config) {
Supplier<Set<String>> hostEnvNames) {
super("test", new AgentControl(herdr), new WorkspaceControl(herdr),
Map.of("test", profile()), "test",
name -> "SHELL".equals(name) ? shell : null,
0, () -> 0L, () -> { }, null, creds, hostEnvNames, config);
0, () -> 0L, () -> { }, null, creds, hostEnvNames);
}
@Override
protected Launch buildLaunch(FleetConfig.Profile cfg, LaunchSpec spec) {
Map<String, String> launchEnv = baseEnv(cfg);
launchEnv.putAll(env);
env.clear();
env.putAll(launchEnv);
return new Launch(env, List.of("test"));
}
@@ -315,12 +278,6 @@ class HerdrPeerLauncherAllowListWiringTest {
}
}
private static FleetConfig config(String brokerUriEnv) {
return new FleetConfig(null, null, null, Map.of(), null, null, null, null, null,
new FleetConfig.Broker(null, brokerUriEnv, null), null, null, null, null, null,
null, null, null, null, null).withDefaults();
}
/** The generated directory is a temp directory; make sure the test does not leave a pile. */
@Test
void theGeneratedDirectoryIsRemovedWhenThePaneIsStopped() {
@@ -121,34 +121,6 @@ class MemberEnvAllowListTest {
assertTrue(derived.contains("OTHER_NAME"), "other allow: names are unaffected");
}
@Test
void configuredBrokerAndCoordinatorUriEnvNamesAreExcludedEvenWhenAllowed() {
FleetConfig config = config("BROKER_CONNECTION_URI", "COORDINATOR_CONNECTION_URI");
Set<String> excluded = MemberEnvAllowList.brokerUriEnvNames(config);
Set<String> derived = MemberEnvAllowList.derive(List.of(),
Set.of("BROKER_CONNECTION_URI", "COORDINATOR_CONNECTION_URI", "OTHER_NAME"), excluded);
assertFalse(derived.contains("BROKER_CONNECTION_URI"),
"broker.uriEnv is secret-bearing and must not ride in on allow:");
assertFalse(derived.contains("COORDINATOR_CONNECTION_URI"),
"coordinator.uriEnv has the same inline-password shape");
assertTrue(derived.contains("OTHER_NAME"), "unrelated allow: entries are unaffected");
}
@Test
void absentOrBlankBrokerUriEnvAddsNoExclusions() {
assertTrue(MemberEnvAllowList.brokerUriEnvNames(config(null, null)).isEmpty());
assertTrue(MemberEnvAllowList.brokerUriEnvNames(config(" ", "")).isEmpty());
}
private static FleetConfig config(String brokerUriEnv, String coordinatorUriEnv) {
return new FleetConfig(null, null, null, Map.of(), null, null, null, null, null,
new FleetConfig.Broker(null, brokerUriEnv, null), null, null, null, null, null,
null, null, null, null,
new FleetConfig.Coordinator(null, coordinatorUriEnv, null, null)).withDefaults();
}
/** {@code LC_*} categories are infrastructure by prefix; everything else needs an exact match. */
@Test
void keepsMatchesExactlyPlusTheLocalePrefixRule() {
@@ -334,7 +334,8 @@ class OpenCodeLauncherTest {
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, "p1", "ses_a.json",
"ses_resolved", "/work/dir", 1000L);
assertEquals("ses_resolved", handle.agentSessionId(),
"agentSessionId() re-scans and picks up a record that has since been written");
}
@@ -5,89 +5,68 @@ import org.junit.jupiter.api.io.TempDir;
import java.nio.file.Files;
import java.nio.file.Path;
import java.sql.Connection;
import java.sql.DriverManager;
import java.sql.PreparedStatement;
import java.sql.SQLException;
import java.sql.Statement;
import java.nio.file.attribute.FileTime;
import static org.junit.jupiter.api.Assertions.*;
/**
* {@link OpenCodeSessionDiscovery} matches an opencode session row by the worker's cwd (its
* {@code directory}) against opencode's {@code opencode.db} SQLite database. These tests build a
* SYNTHETIC database themselves, in a JUnit temp directory — never the operator's real
* {@code ~/.local/share/opencode/opencode.db}, which a live opencode process may be writing.
* {@link OpenCodeSessionDiscovery} matches an opencode session record by the worker's cwd (its
* {@code directory}) against opencode's on-disk storage. These tests populate a TEMP storage root
* themselves — never the operator's real {@code ~/.local/share/opencode}.
*/
class OpenCodeSessionDiscoveryTest {
/**
* Create {@code <root>/opencode.db} with a minimal {@code session} table (just the columns
* {@link OpenCodeSessionDiscovery} reads: {@code id}, {@code directory}, {@code time_updated})
* and insert one row. Static so {@link OpenCodeLauncherTest} can reuse it.
* Write a session record {@code {"id":..., "directory":...}} under
* {@code <root>/session/<projectID>/<fileName>} and stamp it with a known last-modified time,
* so "most recently modified wins" is deterministic. Static so the launcher test can reuse it.
*/
static void writeRecord(Path root, String id, String directory, long timeUpdated) throws Exception {
Path db = root.resolve("opencode.db");
try (Connection connection = DriverManager.getConnection("jdbc:sqlite:" + db)) {
try (Statement statement = connection.createStatement()) {
statement.execute("CREATE TABLE IF NOT EXISTS session ("
+ "id TEXT PRIMARY KEY, directory TEXT, time_updated INTEGER)");
}
// Bound parameters, not string interpolation: the class under test uses a
// PreparedStatement, and a hand-escaped INSERT here is a pattern someone copies out.
try (PreparedStatement insert = connection.prepareStatement(
"INSERT INTO session (id, directory, time_updated) VALUES (?, ?, ?)")) {
insert.setString(1, id);
insert.setString(2, directory);
insert.setLong(3, timeUpdated);
insert.executeUpdate();
}
}
static void writeRecord(Path root, String projectId, String fileName, String id,
String directory, long lastModifiedEpochMillis) throws Exception {
Path dir = root.resolve("session").resolve(projectId);
Files.createDirectories(dir);
Path file = dir.resolve(fileName);
Files.writeString(file, "{\"id\":\"" + id + "\",\"directory\":\"" + directory
+ "\",\"projectID\":\"" + projectId + "\",\"version\":\"1.1.31\"}");
Files.setLastModifiedTime(file, FileTime.fromMillis(lastModifiedEpochMillis));
}
@Test
void findsTheRowWhoseDirectoryEqualsTheCwd(@TempDir Path root) throws Exception {
writeRecord(root, "ses_aaa", "/w/a", 1000L);
writeRecord(root, "ses_bbb", "/w/b", 2000L);
void findsTheRecordWhoseDirectoryEqualsTheCwd(@TempDir Path root) throws Exception {
writeRecord(root, "p1", "ses_a.json", "ses_aaa", "/w/a", 1000L);
writeRecord(root, "p2", "ses_b.json", "ses_bbb", "/w/b", 2000L);
assertEquals("ses_bbb", new OpenCodeSessionDiscovery(root).sessionIdForDirectory("/w/b"),
"the row whose directory equals the cwd is the one found");
"the record whose directory equals the cwd is the one found");
assertEquals("ses_aaa", new OpenCodeSessionDiscovery(root).sessionIdForDirectory("/w/a"));
}
@Test
void aNonMatchingDirectoryYieldsNullRatherThanAMismatch(@TempDir Path root) throws Exception {
writeRecord(root, "ses_aaa", "/w/a", 1000L);
writeRecord(root, "p1", "ses_a.json", "ses_aaa", "/w/a", 1000L);
assertNull(new OpenCodeSessionDiscovery(root).sessionIdForDirectory("/w/other"),
"no row for this cwd yet → null, not a wrong session");
"no record for this cwd yet → null, not a wrong session");
}
@Test
void prefersTheMostRecentlyUpdatedRowWhenSeveralMatch(@TempDir Path root) throws Exception {
writeRecord(root, "ses_old", "/w/a", 1000L);
writeRecord(root, "ses_new", "/w/a", 5000L);
void prefersTheMostRecentlyModifiedRecordWhenSeveralMatch(@TempDir Path root) throws Exception {
writeRecord(root, "p1", "old.json", "ses_old", "/w/a", 1000L);
writeRecord(root, "p2", "new.json", "ses_new", "/w/a", 5000L);
assertEquals("ses_new", new OpenCodeSessionDiscovery(root).sessionIdForDirectory("/w/a"),
"the row with the highest time_updated for the cwd wins");
"the freshest record for the cwd wins");
}
@Test
void aMissingDatabaseYieldsNullWithoutThrowing(@TempDir Path root) {
// No opencode.db at all under the root.
void aMissingOrEmptyStorageRootYieldsNullWithoutThrowing(@TempDir Path root) throws Exception {
// Missing: no session dir at all under the root.
assertNull(new OpenCodeSessionDiscovery(root).sessionIdForDirectory("/w/a"));
}
@Test
void anEmptyDatabaseYieldsNullWithoutThrowing(@TempDir Path root) throws Exception {
Path db = root.resolve("opencode.db");
try (Connection connection = DriverManager.getConnection("jdbc:sqlite:" + db);
Statement statement = connection.createStatement()) {
statement.execute("CREATE TABLE session (id TEXT PRIMARY KEY, directory TEXT, "
+ "time_updated INTEGER)");
}
assertNull(new OpenCodeSessionDiscovery(root).sessionIdForDirectory("/w/a"));
// Present but empty: a session dir with nothing in it produces no match, not a throw.
Path emptyRoot = root.resolve("empty");
Files.createDirectories(emptyRoot.resolve("session"));
assertNull(new OpenCodeSessionDiscovery(emptyRoot).sessionIdForDirectory("/w/a"));
}
@Test
@@ -97,41 +76,15 @@ class OpenCodeSessionDiscoveryTest {
assertNull(discovery.sessionIdForDirectory(" "));
}
/**
* The one line standing between fleetd and writing the operator's live {@code opencode.db} —
* 841MB, with a running opencode writing it — is {@code config.setReadOnly(true)} in
* {@link OpenCodeSessionDiscovery#openReadOnly()}. Delete it and every other test in this class
* still passes, so this is the test that guards it.
*
* <p>It asks the connection to write, and requires a refusal. The obvious alternative — make
* the database file unwritable and check the read still works — proves nothing: SQLite silently
* downgrades a read-write open of an unwritable file to read-only, so that test passes either
* way. It was tried and watched pass with the flag removed.
*/
@Test
void theDatabaseIsOpenedReadOnly(@TempDir Path root) throws Exception {
writeRecord(root, "ses_aaa", "/w/a", 1000L);
void aMalformedRecordIsSkippedRatherThanFatal(@TempDir Path root) throws Exception {
// A record that fails to parse must not abort the scan of its siblings.
Path dir = root.resolve("session").resolve("p1");
Files.createDirectories(dir);
Files.writeString(dir.resolve("broken.json"), "{not valid json");
writeRecord(root, "p1", "good.json", "ses_good", "/w/a", 1000L);
try (Connection connection = new OpenCodeSessionDiscovery(root).openReadOnly();
Statement statement = connection.createStatement()) {
SQLException refused = assertThrows(SQLException.class,
() -> statement.executeUpdate("INSERT INTO session (id, directory, time_updated) "
+ "VALUES ('ses_zzz', '/w/z', 1)"),
"the connection must REFUSE a write — opencode is writing this database live");
assertTrue(refused.getMessage().toLowerCase().contains("readonly")
|| refused.getMessage().toLowerCase().contains("read-only"),
"the refusal must be about read-only, not some other error: " + refused.getMessage());
}
}
@Test
void aCorruptDatabaseFileYieldsNullWithoutThrowing(@TempDir Path root) throws Exception {
// A file at opencode.db that is not a SQLite database at all — the open/query must fail
// safe, never fatal to a spawn.
Path db = root.resolve("opencode.db");
Files.writeString(db, "this is not a sqlite database");
assertNull(new OpenCodeSessionDiscovery(root).sessionIdForDirectory("/w/a"),
"an unreadable database resolves to null, not an exception");
assertEquals("ses_good", new OpenCodeSessionDiscovery(root).sessionIdForDirectory("/w/a"),
"an unreadable record is skipped; a later valid one still matches");
}
}
@@ -710,68 +710,6 @@ class MessageServiceTest {
assertEquals(MessageService.Outcome.REPLIED, answer.get(5, TimeUnit.SECONDS).outcome());
}
// --- #137: a fleet_ask round-trip must not orphan the ticket's own reply -------------------
//
// The primary's fleet_send{turnId} answer call is itself bounded (a real MCP call, capped well
// under a minute) — far shorter than a resumed turn can genuinely take to finish real work. These
// drive the exact real delegation path (async send -> worker asks -> primary answers -> primary's
// own wait gives up -> worker's real fleet_reply arrives afterwards) rather than calling a reply
// sink directly, since the bug is specifically about which sink the resumed turn's reply reaches.
@Test
void aReplyAfterAnswerTimesOutStillCompletesTheAsyncTicket() throws Exception {
String ticket = messages.sendAsync(T, "task that asks");
awaitWaiting();
injectDelivery();
CompletableFuture<MessageService.AskResult> ask =
CompletableFuture.supplyAsync(() -> messages.ask(T, "which config?", 5000));
MessageService.TaskView asking = awaitTicketPhase(ticket, MessageService.Phase.ASKING);
// The primary answers, but its own bounded wait for the worker's resumed turn is short and
// expires before the worker (still genuinely working) gets back to it.
MessageService.Reply answerReply = messages.answer(asking.turnId(), "config.yaml", 150);
assertEquals("config.yaml", ask.get(5, TimeUnit.SECONDS).answer());
assertEquals(MessageService.Outcome.TIMED_OUT_WORKING, answerReply.outcome(),
"the primary's own bounded wait gives up before the worker finishes resuming");
// The worker keeps working past that window and only now calls fleet_reply.
assertTrue(messages.reply(T, "PR opened: https://example/pulls/42"));
MessageService.TaskView done = awaitTicketPhase(ticket, MessageService.Phase.DONE);
assertEquals("PR opened: https://example/pulls/42", done.reply(),
"fleet_poll{ticket} must return the worker's real reply, not stay pending forever");
assertEquals("reply", done.replySource());
assertFalse(messages.hasStrandedReply(T),
"the reply completed its own ticket directly and never touched the inbox");
}
@Test
void fleetStopAfterAnOrphanedReplyDoesNotFailTheTicket() throws Exception {
String ticket = messages.sendAsync(T, "task that asks");
awaitWaiting();
injectDelivery();
CompletableFuture<MessageService.AskResult> ask =
CompletableFuture.supplyAsync(() -> messages.ask(T, "which config?", 5000));
MessageService.TaskView asking = awaitTicketPhase(ticket, MessageService.Phase.ASKING);
MessageService.Reply answerReply = messages.answer(asking.turnId(), "config.yaml", 150);
assertEquals("config.yaml", ask.get(5, TimeUnit.SECONDS).answer());
assertEquals(MessageService.Outcome.TIMED_OUT_WORKING, answerReply.outcome());
assertTrue(messages.reply(T, "PR opened: https://example/pulls/42"));
// fleet_stop tears the worker's session down right after the reply landed — this must never
// report the misleading "the worker session was released before it replied": a reply is
// exactly what happened.
assertFalse(messages.abandon(T, "the worker session was released before it replied"),
"a reply already arrived, so nothing here is a genuine failure");
MessageService.TaskView view = awaitTicketPhase(ticket, MessageService.Phase.DONE);
assertEquals("PR opened: https://example/pulls/42", view.reply());
}
@Test
void unansweredAsyncQuestionReturnsTheTicketToPendingAndReleasesItsTarget() throws Exception {
String ticket = messages.sendAsync(T, "task that asks");