Compare commits

...

2 Commits

Author SHA1 Message Date
Dai Ha be6c45ff78 CB-617 review: warn instead of refuse on conflicting autoCompactWindow
CI / shell-tests (pull_request) Failing after 9s
CI / contract (pull_request) Successful in 49s
CI / build (pull_request) Failing after 2m9s
rejectConflictingAutoCompactWindows threw and stopped fleetd from starting when a
Claude Code profile's autoCompactWindow flag and CLAUDE_CODE_AUTO_COMPACT_WINDOW
env var disagreed. Under launchd that is a restart loop, and the config that
would fix it (fleetd.yaml) is gitignored, so the cause is invisible on the host
where it bites (measured live: 4 profiles on this host trip it, including the
lead's own profile and the one every worker spawns on).

Renamed to warnConflictingAutoCompactWindows: it now logs a WARN naming each
offending profile with BOTH values (autoCompactWindow=... and
env.CLAUDE_CODE_AUTO_COMPACT_WINDOW=...) instead of throwing, so the daemon
starts and an operator can fix the config without reading the source. Equal
values still load silently.

Also reworded ClaudeCodeArguments' javadoc, which stated as fact that the env
var takes precedence over the flag. That was never measured, and this host's
own fleetd.yaml comment asserts the opposite — the javadoc no longer picks a
side.
2026-09-22 10:13:14 +07:00
Dai Ha e99cb70a8b CB-617: pass auto-compact window to leads 2026-09-22 10:12:56 +07:00
7 changed files with 227 additions and 35 deletions
+5 -5
View File
@@ -209,13 +209,13 @@ herdrSocket: ~/.config/herdr/herdr.sock
# e.g. `env DISPLAY=:10.0 idea {dir}`. Best-effort: a failure is logged, never
# fails the spawn. Omit to open the member's module by hand. There is no close
# half yet — an opened module stays open until the operator closes it.
# autoCompactWindow → opt-in, default off. A bounded token window that forces a spawned member to
# compact its context instead of running on the backend's own default and dying
# mid-turn (losing its fleet_reply — the whole point of the turn — with it).
# autoCompactWindow → opt-in, default off. A bounded token window that forces a launched Claude Code
# lead or member to compact its context instead of running on the backend's own
# default. A member that runs out of context can die mid-turn and lose its fleet_reply.
# Validated at config load to [100000, 1000000] — the band Claude Code's own
# --autocompact flag accepts.
# CROSS-BACKEND SEMANTICS DIFFER: on claude-code this is a launch-time
# `--autocompact <tokens>` flag — the member compacts AT this window. opencode
# `--autocompact <tokens>` flag — the Claude Code session compacts AT this window. opencode
# has no equivalent flag (it only forces `compaction.auto: true`, unconditionally,
# already), so this is instead applied as the model's `limit.context` in the
# generated opencode.json — the member compacts WITHIN this window, not exactly
@@ -415,7 +415,7 @@ profiles:
# ideMcpUrl: http://127.0.0.1:29170/index-mcp/streamable-http # opt-in (CB-634): IDE code intelligence, pinned to the worktree
# ideProjectDir: fleetd # CB-634: module dir the IDE opens + the overlay pins (this repo's pom is in fleetd/)
# ideOpenCommand: env DISPLAY=:10.0 idea {dir} # CB-634 auto-open: opens {dir} in the IDE at spawn; omit to open by hand
# autoCompactWindow: 250000 # opt-in: bound member context; claude-code compacts AT this, opencode within it (model limit.context)
# autoCompactWindow: 250000 # opt-in: bound Claude Code lead/member context; claude-code compacts AT this, opencode within it (model limit.context)
gx11: # a second backend, so `placement: weighted` has a choice
baseUrl: http://gx01.gw:8000 # self-hosted; ccs handles the model + token
placement: tab
@@ -462,8 +462,8 @@ public record FleetConfig(
* profile that does not opt in. Read live off the current config, so it is
* HOT: a change takes effect on the next exhaustion classification / spawn,
* no restart needed.
* @param autoCompactWindow opt-in per-profile token window that forces a spawned member to
* auto-compact its context at (Claude Code) or within (opencode) a bound the
* @param autoCompactWindow opt-in per-profile token window that forces a launched Claude Code
* session to auto-compact its context at, or an opencode session within, a bound the
* operator chooses, instead of the backend's own default. {@code null} (the
* default) leaves today's behaviour exactly — opencode already forces
* {@code compaction.auto: true} unconditionally (CB-523) but has no absolute
@@ -1875,6 +1875,7 @@ public record FleetConfig(
rejectDuplicateMemberSlots(yaml);
rejectNegativeMaxLoad(yaml);
rejectAutoCompactWindowOutOfRange(yaml);
warnConflictingAutoCompactWindows(yaml);
rejectMalformedProfilePatterns(yaml);
rejectUnknownKind(yaml);
rejectUnknownAuthMode(yaml);
@@ -2225,6 +2226,71 @@ public record FleetConfig(
}
}
/**
* Warn (never refuse to start) about a Claude Code profile whose auto-compaction flag and
* environment setting disagree.
*
* <p>Renamed from {@code rejectConflictingAutoCompactWindows} (fleetd #601 review, measured
* 2026-09-22): that method threw {@link IllegalStateException}, so {@link #load(Path)} refused
* to start on a config carrying this conflict. On this host, four profiles trip it, including
* the lead's own profile and the one every worker spawns on — so the throw is not a rare edge
* case. Under launchd, a throw inside {@code load()} is a restart loop, not an error an operator
* reads once, and the config that would fix it ({@code fleetd/fleetd.yaml}) is gitignored, so
* the cause is invisible on the host where it bites. A WARN gives the operator the same
* information — which profiles, and now both values, so they can fix it without reading the
* source — without ever taking the fleet down.
*
* <p>Which of the two inputs Claude Code actually follows when they disagree is intentionally
* <em>not</em> asserted here. {@code ClaudeCodeArguments}'s javadoc used to state the
* environment variable always wins; nobody had measured that, and this host's own
* {@code fleetd.yaml} asserts the opposite in a comment. This method only detects and reports
* the disagreement — see {@link dev.ltms.fleet.launch.ClaudeCodeArguments}.
*
* <p>Equal values never warn: either input then produces the same session window, so there is
* nothing to reconcile.
*/
static void warnConflictingAutoCompactWindows(String yaml) {
Map<?, ?> raw;
try {
raw = YAML.readValue(yaml, Map.class);
} catch (IOException | IllegalArgumentException e) {
return;
}
if (raw == null || !(raw.get("profiles") instanceof Map<?, ?> profiles)) {
return;
}
List<String> names = new ArrayList<>();
List<String> detail = new ArrayList<>();
for (Map.Entry<?, ?> entry : profiles.entrySet()) {
if (!(entry.getValue() instanceof Map<?, ?> profile)
|| !(profile.get("autoCompactWindow") instanceof Number window)
|| !(profile.get("env") instanceof Map<?, ?> env)
|| !env.containsKey("CLAUDE_CODE_AUTO_COMPACT_WINDOW")) {
continue;
}
Object kind = profile.get("kind");
boolean claudeCode = kind == null || String.valueOf(kind).isBlank()
|| Profile.KIND_CLAUDE_CODE.equalsIgnoreCase(String.valueOf(kind));
Object envValue = env.get("CLAUDE_CODE_AUTO_COMPACT_WINDOW");
if (claudeCode && !String.valueOf(window).equals(String.valueOf(envValue))) {
String name = String.valueOf(entry.getKey());
names.add(name);
detail.add(name + " (autoCompactWindow=" + window
+ ", env.CLAUDE_CODE_AUTO_COMPACT_WINDOW=" + envValue + ")");
}
}
if (names.isEmpty()) {
return;
}
names.sort(String::compareTo);
detail.sort(String::compareTo);
log.warn("Claude Code profile(s) {} set disagreeing autoCompactWindow and env."
+ "CLAUDE_CODE_AUTO_COMPACT_WINDOW — the daemon starts anyway. Fix by "
+ "removing one key or setting equal values on each: {}. Which input Claude "
+ "Code actually follows when they disagree is not verified here.",
names, String.join(", ", detail));
}
/**
* Reject a profile whose {@code errorPattern} (fleetd #201 Unit 5) or {@code exhaustedPattern}
* (CB-578 stage A) is not a valid Java regex, naming the profile, the key, and the parser's own
@@ -0,0 +1,35 @@
package dev.ltms.fleet.launch;
import dev.ltms.fleet.config.FleetConfig;
import java.util.ArrayList;
import java.util.List;
/** Arguments shared by every fleetd path that starts Claude Code. */
public final class ClaudeCodeArguments {
private ClaudeCodeArguments() {
}
/**
* Append the configured Claude Code auto-compaction window when the profile opts in.
*
* <p>This flag and the environment variable {@code CLAUDE_CODE_AUTO_COMPACT_WINDOW} can
* disagree. Which one Claude Code actually follows when they do is NOT verified here — this
* javadoc used to claim the environment variable always wins, but nobody had measured that, and
* this host's own {@code fleetd.yaml} asserts the opposite in a comment. So this javadoc no
* longer picks a side. {@link FleetConfig#load(java.nio.file.Path)} only WARNS when a Claude
* Code profile sets both to different values (see {@code
* FleetConfig.warnConflictingAutoCompactWindows}) — it does not stop the daemon from starting,
* and a launched session may end up honouring either window.
*/
public static List<String> withAutoCompactWindow(List<String> argv, FleetConfig.Profile profile) {
if (profile.autoCompactWindow() == null) {
return argv;
}
List<String> withAutoCompact = new ArrayList<>(argv);
withAutoCompact.add("--autocompact");
withAutoCompact.add(String.valueOf(profile.autoCompactWindow()));
return withAutoCompact;
}
}
@@ -8,6 +8,7 @@ import dev.ltms.fleet.herdr.PendingCloseMarker;
import dev.ltms.fleet.herdr.Tab;
import dev.ltms.fleet.herdr.Workspace;
import dev.ltms.fleet.herdr.WorkspaceControl;
import dev.ltms.fleet.launch.ClaudeCodeArguments;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
@@ -358,7 +359,8 @@ public final class LeadLauncher {
}
/**
* The lead's argv: the profile's own command, the model pin, and the bridge MCP mount.
* The lead's argv: the profile's own command, the model and auto-compaction pins, and the bridge
* MCP mount.
*
* <p>No {@code --append-system-prompt}. That flag carries the worker reply charter, and a lead
* is not a worker — it reads its orchestration rules from the project's {@code CLAUDE.md} like
@@ -380,7 +382,7 @@ public final class LeadLauncher {
argv.add("--model");
argv.add(profile.model());
}
return argv;
return profile.isOpenCode() ? argv : ClaudeCodeArguments.withAutoCompactWindow(argv, profile);
}
/**
@@ -8,6 +8,7 @@ import dev.ltms.fleet.guard.SubscriptionGuard;
import dev.ltms.fleet.herdr.Agent;
import dev.ltms.fleet.herdr.AgentControl;
import dev.ltms.fleet.herdr.WorkspaceControl;
import dev.ltms.fleet.launch.ClaudeCodeArguments;
import dev.ltms.fleet.peer.Capability;
import dev.ltms.fleet.peer.PeerLauncher;
import org.slf4j.Logger;
@@ -298,7 +299,7 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
// has neither MCP nor a charter — session flags must be added into a list we own.
List<String> argv = mutableArgv(argvWithFleet(cfg, spec));
String agentSessionId = applySessionIdentity(argv, spec.sessionName(), spec.resumeSessionId());
return new Launch(workerEnv, argvWithAutoCompact(argvWithModel(argv, cfg), cfg), agentSessionId);
return new Launch(workerEnv, ClaudeCodeArguments.withAutoCompactWindow(argvWithModel(argv, cfg), cfg), agentSessionId);
}
/**
@@ -908,30 +909,6 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
return withModel;
}
/**
* Pin a bounded auto-compaction window on the command line via {@code --autocompact <tokens>},
* opt-in per profile (CB-634's sibling ticket: a member that runs out of context dies mid-turn
* and its {@code fleet_reply} — the whole point of the turn — is lost with it; opencode already
* forces {@code compaction.auto: true} unconditionally, CB-523, but Claude Code has no equivalent
* and runs at the backend's own default window).
*
* <p>Mirrors {@link #argvWithModel}: appended after it, so it survives the {@code ccs <profile>}
* wrapper the same way {@code --model} does, and outranks env/settings and the operator's own
* {@code argv}. Verified: {@code claude 2.1.241 --help} lists {@code --autocompact <auto|tokens>}
* (either the literal {@code auto}, or an integer 100k–1M) — {@link FleetConfig#load} rejects a
* configured value outside that band before this ever runs, so the flag Claude Code receives here
* is always in range.
*/
private static List<String> argvWithAutoCompact(List<String> argv, FleetConfig.Profile cfg) {
if (cfg.autoCompactWindow() == null) {
return argv;
}
List<String> withAutoCompact = mutableArgv(argv);
withAutoCompact.add("--autocompact");
withAutoCompact.add(String.valueOf(cfg.autoCompactWindow()));
return withAutoCompact;
}
// --- Agent-returning convenience spawns (used by callers/tests that want the herdr Agent) ---
/** Spawn a worker for the default profile in the resolved default cwd. */
@@ -1,8 +1,11 @@
package dev.ltms.fleet.config;
import ch.qos.logback.classic.Level;
import ch.qos.logback.classic.spi.ILoggingEvent;
import dev.ltms.fleet.auth.MemberRegistry;
import dev.ltms.fleet.msg.LeadMailbox;
import dev.ltms.fleet.peer.MemberRole;
import dev.ltms.fleet.testing.CapturedLog;
import org.junit.jupiter.api.Test;
import org.junit.jupiter.api.io.TempDir;
@@ -93,6 +96,72 @@ class FleetConfigTest {
"unset means off — today's behaviour, unchanged");
}
/**
* fleetd #601 review (measured 2026-09-22): this guard used to throw {@link
* IllegalStateException} and refuse to start. On a host with the conflict configured, that
* turned into a launchd restart loop with no readable cause, since {@code fleetd.yaml} is
* gitignored. It must now WARN and let the daemon start, and the warning must carry both values
* so an operator can fix the config without reading the source. Pinning the log line itself (via
* {@link CapturedLog}) rather than a snippet of production source text — the latter is the
* anti-pattern this repo avoids; the former is the actual observable behaviour a reader (or an
* alert on the log) depends on.
*/
@Test
void aClaudeProfileWithConflictingAutoCompactFlagAndEnvironmentWindowLoadsAndWarnsWithBothValues(
@TempDir Path dir) throws Exception {
Path f = dir.resolve("conflicting-auto-compact-window.yaml");
Files.writeString(f, """
profiles:
claude-profile:
autoCompactWindow: 250000
env:
CLAUDE_CODE_AUTO_COMPACT_WINDOW: "300000"
""");
FleetConfig cfg;
List<String> warnings;
try (CapturedLog log = CapturedLog.at(FleetConfig.class, Level.WARN)) {
cfg = FleetConfig.load(f);
warnings = log.events().stream().map(ILoggingEvent::getFormattedMessage).toList();
}
assertEquals(250_000, cfg.profiles().get("claude-profile").autoCompactWindow(),
"the disagreement is reported, not corrected — the flag value still loads as-is");
assertEquals(1, warnings.size(), "exactly one warning for the one conflicting profile: " + warnings);
String warning = warnings.get(0);
assertTrue(warning.contains("claude-profile"), "names the offending profile: " + warning);
assertTrue(warning.contains("autoCompactWindow=250000"), "names the flag value: " + warning);
assertTrue(warning.contains("CLAUDE_CODE_AUTO_COMPACT_WINDOW=300000"), "names the env value: " + warning);
}
/**
* The negative probe paired with the test above (per fleetd #601 review): a warning that fires
* on every load and a warning that never fires read the same from a single test, so both must be
* checked. No conflict here — the flag and the env value agree — so no warning should be logged.
*/
@Test
void aClaudeProfileWithEqualAutoCompactFlagAndEnvironmentWindowLoadsWithNoWarning(@TempDir Path dir)
throws Exception {
Path f = dir.resolve("equal-auto-compact-window.yaml");
Files.writeString(f, """
profiles:
claude-profile:
autoCompactWindow: 250000
env:
CLAUDE_CODE_AUTO_COMPACT_WINDOW: "250000"
""");
FleetConfig cfg;
List<ILoggingEvent> events;
try (CapturedLog log = CapturedLog.at(FleetConfig.class, Level.WARN)) {
cfg = FleetConfig.load(f);
events = log.events();
}
assertEquals(250_000, cfg.profiles().get("claude-profile").autoCompactWindow());
assertTrue(events.isEmpty(), "equal values must not warn: " + events);
}
// ── fleetd #201 Unit 5: errorPattern ────────────────────────────────────────────────────────
@Test
@@ -32,13 +32,26 @@ class LeadLauncherTest {
Map.of("CLAUDE_CODE_AUTO_COMPACT_WINDOW", "300000"), null, null, true, null);
}
private static FleetConfig.Profile profileWithAutoCompactWindow(String kind) {
return new FleetConfig.Profile(
"opus", null, "claude-opus-5", null, "FLEETD_WORKER_TOKEN",
List.of("ccs", "ltms"), "tab", "fleet", null,
"http://127.0.0.1:8765/mcp", null, null,
null, null, kind, Map.of(), null, null, true, null, null,
null, null, null, 250_000);
}
private static FleetConfig configWith(FleetConfig.Leader lead) {
return configWith(lead, opusProfile());
}
private static FleetConfig configWith(FleetConfig.Leader lead, FleetConfig.Profile profile) {
Map<String, FleetConfig.Leader> leaders = new LinkedHashMap<>();
leaders.put("opus", lead);
FleetConfig.Fleet fleet =
new FleetConfig.Fleet(leaders, Map.of(), Map.of(), Map.of(), null);
return new FleetConfig(
null, null, Map.of("opus", opusProfile()), null, null, null, null, null,
null, null, Map.of("opus", profile), null, null, null, null, null,
null, null, fleet, null, "fixed", null).withDefaults();
}
@@ -330,6 +343,36 @@ class LeadLauncherTest {
"--model is appended last so it outranks the ccs wrapper (CB-533)");
}
@Test
void aClaudeLeadPassesItsConfiguredAutoCompactWindowToClaudeCode() {
FakeHerdr herdr = new FakeHerdr();
FleetConfig.Profile profile = profileWithAutoCompactWindow("claude-code");
launcher(herdr, configWith(lead("opus", "lead: opus", 1), profile)).ensureLeads();
List<String> args = startedArgs(herdr);
assertEquals("250000", args.get(args.indexOf("--autocompact") + 1));
}
@Test
void aClaudeLeadWithNoAutoCompactWindowGetsNoAutoCompactFlag() {
FakeHerdr herdr = new FakeHerdr();
launcher(herdr, configWith(lead("opus", "lead: opus", 1))).ensureLeads();
assertFalse(startedArgs(herdr).contains("--autocompact"));
}
@Test
void aNonClaudeLeadDoesNotGetAnAutoCompactFlag() {
FakeHerdr herdr = new FakeHerdr();
FleetConfig.Profile profile = profileWithAutoCompactWindow("opencode");
launcher(herdr, configWith(lead("opus", "lead: opus", 1), profile)).ensureLeads();
assertFalse(startedArgs(herdr).contains("--autocompact"));
}
/** A lead runs on the operator's subscription. Nothing may move it off. */
@Test
void theLeadEnvCarriesNoAnthropicBinding() {