mirror of
https://github.com/QwenLM/qwen-code.git
synced 2026-11-07 06:19:43 -10:00
main
5
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 jinyeandQwen-Coder
|
f3af3dd857 |
refactor(managed-agent): Split server into Maven modules (#15871)
* refactor(managed-agent): split server into Maven modules Separate contracts, orchestration, persistence, adapters, assembly, and verification while retaining the existing service delivery model. Move tests and build consumers together and document local evidence and pending CI acceptance. Co-authored-by: Qwen-Coder <[email protected]> * refactor(managed-agent): reduce configuration and SDK boilerplate Co-authored-by: Qwen-Coder <[email protected]> * refactor(managed-agent): narrow contracts and simplify test scaffolding Co-authored-by: Qwen-Coder <[email protected]> * fix(managed-agent): repair verification after module migration Resolve shared fixtures and worker paths from the Maven module root, update CI packaging assertions, and align existing test fixtures with the extracted contracts. Restore restricted Workspace constructors and remove redundant test visibility changes without altering inherited MySQL coverage. Co-authored-by: Qwen-Coder <[email protected]> * test(managed-agent): align remaining Harness and OSS fixtures Preserve SDK refusal messages and register the managed OSS adapter at fixture injection points. Keep the existing retry and client shutdown assertions intact. Co-authored-by: Qwen-Coder <[email protected]> * style(managed-agent): order Hosted cleanup Harness imports Co-authored-by: Qwen-Coder <[email protected]> * test(managed-agent): fix recovery smoke fixture and import order Supply the required transcript coverage in the recovery browser fixture and wait for history before testing a complete handoff. Preserve its existing safety and submission assertions. Order the cleanup coordinator test imports for Checkstyle. Co-authored-by: Qwen-Coder <[email protected]> --------- Co-authored-by: Qwen-Coder <[email protected]> |
||
|
Shaojin Wenand易良
|
e7a67ee4e3 |
feat(managed-agent): give the Hosted Harness its Broker through the environment (#15244)
* feat(managed-agent): give the Hosted Harness its Broker through the environment
The Hosted Harness reached the Runtime Broker only through command-line
options, so the shipped unit expanded the Broker token into argv, where
any local user can read it, and the token was stored again in
server.env for the control plane. A busy Harness port also moved the
Harness to another port that the control plane could not reach.
harness.env now carries QWEN_MANAGED_RUNTIME_BROKER_URL and
QWEN_MANAGED_RUNTIME_BROKER_TOKEN, which qwen serve already reads and
deletes from its environment, and the unit no longer passes Broker
arguments. Every secret gate lists the token too: core's child and
${VAR} denylist, the ACP scrub list, the channel worker environment and
the control plane's Runtime worker filter. The Hosted Harness profile
fails on a busy port instead of moving.
The Harness credentials file may carry the Broker token for the control
plane, refusing a second direct copy and binding nothing on a conflict;
both Broker lines follow the strict credential rules. Preflight adds a
broker-url row that compares the file's URL with the Broker's address
and warns when the file has none, and the broker row's fix names the
file only when startup read it. The bootstrap, templates, systemd unit
and guides keep the token in harness.env only and tell upgraded hosts
to replace QWEN_HOSTED_HARNESS_BROKER_ARGS.
* fix(managed-agent): check the Harness Broker URL against the actual listener
The broker-url preflight row matched loopback addresses by class, so a
Broker bound to 127.0.0.1 passed a file URL on 127.0.0.2. It failed any
URL that differed from the bind address, although a proxy or TLS front
end may forward it. And it ignored the scheme once a port was given, so
https to the plain-HTTP Broker passed.
The row now passes only a direct http URL: the Broker's port, and a host
that resolves to the bound address itself or, for a wildcard bind, to an
address of this machine. https to that same listener fails, since the
embedded Broker serves plain HTTP. Any other URL warns that a proxy may
forward it, which preflight cannot verify, so it no longer fails --check.
* fix(managed-agent): hold the Harness Broker URL to the Harness's own rules
- broker-url first applies the rules managed-runtime-broker-url.ts applies
at Harness startup (an http or https origin, plain http only on loopback),
reads the host of a registry authority such as an underscore name, and
names the direct loopback URL, or the TLS-terminating proxy for a
non-loopback bind, in its fix.
- A broker check fails a non-loopback bind without
runtime-broker.allow-non-loopback, which the Broker refuses at startup.
- The Harness credentials file requires the Broker URL and token together,
refuses a key containing whitespace such as `export KEY=value` (systemd
ignores it, a shell assigns it), and lets the URL bracket an IPv6 literal.
- qwen serve warns when --managed-runtime-broker-token is on the command
line; serve tests cover the environment fallback, its removal and flag
precedence.
- The upgrade runbook names the server.env token to delete, the server
README lists the Broker prefix among the stripped ones, and the bootstrap
test checks stdout for every generated secret.
* fix(managed-agent): read the Harness Broker URL the way the Harness parses it
The broker-url preflight row predicted the Harness with java.net.URI and a
hand-written loopback check, which disagree with Node's WHATWG URL parser in
both directions. It failed http://127.1:4182, http:/127.0.0.1:4182 and
http://127.0.0.1:4182/./, which the Harness takes as its loopback origin, and
passed http://::1:4182, http://127.0.0.08:4182 and port 99999, which the
Harness refuses before it starts.
- HarnessBrokerUrl follows the URL Standard for the part of an http(s) URL a
credentials-file value can reach: the scheme, the slashes after it, the
IPv4, IPv6 and ASCII-domain host parsers, the port and the path's dot
segments. An xn-- label needs IDNA, so the row warns that it cannot read
that URL instead of failing.
- managed-runtime-broker-url.fixtures.json records the Harness's verdict for
each spelling; the CLI test holds the Harness to it and HarnessBrokerUrlTest
holds preflight to it.
- The fix's direct URL is built from the address the bind resolves to, so a
::ffff:127.0.0.1 bind yields http://127.0.0.1:4182.
- Docs: the upgrade backs up server.env and harness.env and a rollback
restores them; the guide names the row's no-URL warning; the operator pages
allow [ and ] in the URL key; the server README states the URL/token pairing
rule.
* fix(managed-agent): keep the Broker pair across a daemon update restart
A daemon self-update replaces `qwen serve` through execve, which keeps
the PID, so systemd never hands the replacement the Broker pair from
harness.env again, and serve has already removed the pair from its own
environment. Hand the update restarter the resolved pair and add it to
the replacement's environment only, beside the server token.
Also from the same review round:
- preflight decides the plain-http, path and ends-in-a-number refusals
for a host with an xn-- label without IDNA, and leaves only an https
URL to such a host unclassified;
- the Java fixture loop allows only the two documented approximations, a
scheme or path read before the host and an xn-- host read without
IDNA; the shared fixture gains cases for both;
- the CLI fixture loop fails on a verdict it does not know instead of
passing on any throw;
- the broker-url row brackets an IPv6 Broker address;
- the credentials file refuses an empty Broker URL or token, which the
Harness takes as set and then refuses to start with;
- the deployment guide no longer says a wildcard Broker bind is reached
through a proxy.
* fix(managed-agent): warn on an IPv6 URL for an IPv4 wildcard Broker bind
The JDK binds 0.0.0.0 to IPv4 alone under -Djava.net.preferIPv4Stack=true,
so a Harness given http://[::1]:P cannot reach a Broker bound there, while
the broker-url row passed it as served. Preflight cannot see the server's
JVM options, which may come from an environment it does not share, so a
wildcard bind now serves only addresses of its own family, and an IPv6 URL
to the IPv4 wildcard warns and names the IPv4 loopback URL. An https URL
there still fails. The JDK binds :: only on a dual-stack socket, so the
IPv6 wildcard keeps serving IPv4 too.
* fix(managed-agent): warn on an https Broker URL across address families
When the Broker binds the IPv4 wildcard and the URL names an IPv6
address, the Broker's own socket takes it only from a dual-stack JVM.
Under -Djava.net.preferIPv4Stack=true another listener, such as a TLS
proxy, may hold that address and forward to the Broker, so preflight
cannot tell an https URL there apart from a working deployment. It now
warns for both schemes and fails https only for an address the Broker's
plain-HTTP socket takes in either mode.
---------
Co-authored-by: 易良 <[email protected]>
|
||
|
jinyeandQwen-Coder
|
28415ddff3 |
refactor(managed-agent): Consolidate server and broker sources (#14277)
* refactor(managed-agent): Consolidate server and broker sources Move the combined Java application beside sdk-java and remove the standalone Runtime Broker artifact. Update build, development, packaging and CI references while preserving Java package identities, runtime protocols and schema resources. Co-authored-by: Qwen-Coder <[email protected]> * docs(managed-agent): Preserve historical evidence and verification guidance Restore commit-pinned H3 probe inputs and align the Java verification environment contract with the consolidated server layout. Co-authored-by: Qwen-Coder <[email protected]> * fix(ci): Run hosted Broker tests under Bash Keep Maven property arguments intact on Windows and stop before Broker tests when the SDK install fails. Exercise both successful and failed SDK installation in the workflow regression tests. Co-authored-by: Qwen-Coder <[email protected]> * fix(ci): Quote the hosted Broker shell value Follow the repository yamllint single-quote rule without changing the parsed workflow configuration. Co-authored-by: Qwen-Coder <[email protected]> --------- Co-authored-by: Qwen-Coder <[email protected]> |
||
|
|
8679b7ca75 |
fix(managed-agent): e2e runner and image robustness from the #12692 R2 review (#13344)
* fix(managed-agent): e2e runner and image robustness from the #12692 R2 review Six R2 review follow-ups on #12692, re-verified against the current script (post-#13263) before patching: - waitUntil surfaces the last predicate error in its timeout message instead of swallowing it, and every iteration is bounded by the remaining deadline so a hung predicate can no longer outlive timeoutMs several-fold. - crashProcess refuses to kill a child that already died BY SIGNAL (exitCode stays null but signalCode is set), instead of throwing a raw ESRCH that hides the descriptive pre-exit diagnostic. - An interrupt delivered after the last poll now exits 1 with the signal name; success JSON can no longer appear for a run the operator explicitly stopped. - runMysql and mysqladmin pass --no-defaults first, so a developer's ~/.my.cnf client options cannot auth-fail the scratch empty-password server (both mysqld invocations already did). - The packaged-jar lookup is version-independent in both the runner (exactly-one jar assertion over target/) and the Dockerfile COPY (wildcard that fails on zero or multiple matches). - The image listens beyond loopback by default: ENV QWEN_MANAGED_AGENT_SERVER_ADDRESS=0.0.0.0 and QWEN_MANAGED_AGENT_RUNTIME_BROKER_HOST=0.0.0.0, so a documented docker run -p no longer refuses traffic; the application.yml loopback defaults stay host-appropriate. * fix(scripts-e2e): unbreak the failover lanes from the #13344 review Three findings on the R2-fix follow-up, each verified against the CI failure on hosted-harness-mysql before patching: - The packaged-jar lookup assumed exactly one qwen-managed-agent-server-*.jar in target/, but the pom attaches classified repackage executions (workspace-bundle, operator-recovery) besides the unclassified server jar, so the cardinality check could never pass and every test:e2e:managed-*-failover step died at module scope. The script now reads the project <version> from the pom — the one following the module's own <artifactId>, not the <parent> block's — and names the unclassified jar exactly. - The Dockerfile wildcard COPY fails the image build outright: multiple matches into a file destination is a Docker error, and no workflow builds this image to catch it. The build stage now selects the unclassified jar by classifier exclusion, fails unless exactly one exists, and publishes it under the fixed name the runtime stage COPYs. - waitUntil's stall timer was discarded live: a fast success idled the runner up to the largest timeout after the success JSON. The timer is unref'd. Script suite 14/14 green (new pins: pom-version jar naming, the classifier-exclusion line, stallTimer.unref()); Prettier and ESLint clean. The pom regex and the Dockerfile selection loop were both executed locally against a three-jar target before committing. * fix(scripts-e2e): harden the image guard and record the publish consequence Second review round on the jar/bind changes, each verified before patching: - The glob default arm trusted sh's non-nullglob behaviour: with zero matches the pattern itself reaches the counting arm, passes count=1 vacuously, and the build dies at cp with a glob in the message. The loop now breaks on a non-existent match candidate and the guard names the expectation loudly (expected exactly one unclassified server jar, found N). - The wide bind was necessary (-p DNATs to the bridge IP) but its consequence went unrecorded while README.md:6-8 still says the server listens on the loopback deliberately: the API has no HTTP authentication, so tenancy is whatever X-Qwen-Tenant-Id says, and docker run -p on a shared host publishes that surface. README's container section and the Dockerfile ENV comment now say this plainly; the pre-existing private-ingress gate stays with the wider lifecycle deferrals (#13302 family). - Two pins watch the bind and the guard in the script suite (both go red against the old text); the Reviewer Test Plan again matches the current head. Suite 15/15 green; Prettier applied. * fix(scripts-e2e): close the second round's eight follow-ups Each verified against the runner before patching: - The synthetic 'predicate stalled' rejection no longer overwrites the real predicate error; undici's cause (ECONNREFUSED) is what the diagnostic surfaces instead of 'fetch failed'. - MySQL client invocations run with an isolated empty HOME: MySQL honours ~/.mylogin.cnf even with option files disabled, so a developer's mysql_config_editor credential would otherwise defeat the flag at exactly the scratch empty-password servers it was for. spawnSync probes carry their own timeouts, as a sync-blocking call prevents the race deadline from firing inside an iteration. - One 'has this child exited' spelling (childExited) replaces the four drifted phrasings that produced the original ESRCH. - The AST extractor exists once; the port-allocation test calls the shared helper instead of maintaining a second copy. - The two surfaces that pick the server jar are named similarly in test intent (pom-derived exact name in the runner, classifier exclusion with a loud guard in the Dockerfile) so their ownership no longer contradicts, and the pom regex is executed against the real pom in the suite rather than only grepped. Suite 16/16 green; ESLint clean; Prettier applied. * fix(scripts-e2e): opt the image bind in explicitly and close round three Third review round, each finding reproduced or probed before patching: - The published image no longer binds the managed-agent control plane on 0.0.0.0 by default (R1-1): the module has no HTTP authentication, so the ENV pair published a tenant API anyone routing to the port could drive. The image keeps the application.yml loopback defaults; publishing is now an explicit docker run -e opt-in documented in the README's container section, including the no--p bridge exposure. - .dockerignore excludes target/ in both spellings, so a developer's stale Maven output cannot ride the build context into the stage the jar cardinality guard inspects. - waitUntil: the stall sentinel is compared by identity and recorded only when no real predicate error was seen; an answered (falsy) iteration clears a stale earlier error; and the remaining budget is handed to the predicate so synchronous probes bound themselves. - runMysql surfaces the spawn-level reason (a timeout no longer reads as a bare 'MySQL command failed: '), and the isolated MySQL client HOME hangs off the runner's one reclaimed scratch root. - Tests: extraction lists name childExited so the early-exit branch is reachable; the pom regex is lifted from the runner source and pinned by shape instead of a copied literal (a version bump stays green); the double transpile is gone. Suite 23/23 green (7 existing + 16 added by this PR); build, typecheck and lint clean; 14 mutation probes each red their witness. * fix(scripts-e2e): floor the last MySQL probe and close round four (#13344) Fourth review round, each finding reproduced or probed before patching: - R4-1 (Critical): the container-publishing recipe named QWEN_MANAGED_AGENT_RUNTIME_BROKER_HOST=0.0.0.0 without ALLOW_NON_LOOPBACK=true, which the broker requires for a non-loopback bind — an operator enabling the broker got a crash-looping container. Dropped the broker line from the README recipe and the Dockerfile comment and pointed at the Broker deployment section; publishing the API never needs the broker's HTTP face. A pairing pin now reds any documented broker wildcard bind that omits the flag. - R1-8: waitUntil handed the final poll a remaining budget below one client round trip, so the deadline message named a wedged client when the lease was simply unexpired. runMysql floors its spawn timeout at 2s, re-admitting up to ~2s of overshoot per wedged poll against the ~10s the budget threading removed. - R3-1: the mysqlClientHome pin was a whole-file count, blind to a fifth client invocation without the override. Assert per call site over the runner's AST: every spawnSync mysql/mysqladmin/mysqld passes --no-defaults, and the mysql/mysqladmin clients use the isolated HOME (the mysqld launches deliberately keep the real one). - R3-3: the loopback-by-default guarantee is pinned at application.yml's own default, not only at the Dockerfile declining to override it. - R3-2: the waitUntil extraction list lived in six pasted copies; one WAIT_UNTIL_DEPS inventory with one loadWaitUntil() loader now serves all six, and the loader takes injected new Function parameters (used by the crashProcess and runMysql cases). - R4-2: the jar-selection guard is exercised as shell — the RUN stage's own text run under a probed sh against zero/one/three/two-unclassified fixture target/ populations, asserting exit status and message. Suite 24/24 green (7 existing + 17 added by this PR); build, typecheck and lint clean; seven mutation probes red their witnesses, with the guarded broker variant staying green as the control. * fix(scripts-e2e): gate the jar guard on cp and sharpen the MySQL scan (#13344) Fifth review round, each finding probed before patching: - R4-2 (Critical): the executable jar-guard case gated on sh alone, but its fixture ends in cp — on a Git-Bash-only Windows PATH sh.exe resolves without the coreutils, so the merge-queue lane would red at the status assertions, and where sh is absent the in-body return recorded a pass for a guard that never ran. The case now probes the capability it consumes (sh -c 'command -v cp') and skipIf-gates, so a host without either reports skipped, matching the repo's own hasSha256sum probe spelling. - R3-1: the per-call-site scan matched spawnSync only, missing the long-lived start(mysqld, ...) launch (3 of the 4 MySQL process launches); a collapsed population still passed the lower-bound assertion; and toContain pinned the flag's presence, not its position, though MySQL honors --no-defaults only as the first option. The scan now covers spawnSync/spawn/start, asserts the exact four-call inventory, and pins --no-defaults as each call's first array element. - R1-8: the floor pin exercised only the clamp side; the default-timeout call now also pins that a budget above the floor reaches spawnSync unchanged (10_000), so a Math.max-to-constant regression reds the case. R1-12 stays open for a maintainer: it asks for a PR-body edit (the stale Image bind bullet and the 15/15 count), which this loop cannot perform — it has no GitHub write credentials. Suite 24/24 green; build, typecheck, lint and prettier clean; six mutation probes red their witnesses (clamp-to-constant; --no-defaults dropped from the start() launch; a fifth unprotected client; the flag moved off index 0; cardinality guard disabled; a cp-less PATH reports the jar-guard case skipped). * fix(scripts-e2e): document a bootable publish recipe and close round five (#13344) Co-authored-by: Qwen-Coder <[email protected]> * fix(scripts-e2e): boot the documented container recipe and pin the round-two gaps (#13344) Co-authored-by: Qwen-Coder <[email protected]> * fix(scripts-e2e): pass the recipe's datasource password valueless and pin the round-three gaps (#13344) Co-authored-by: Qwen-Coder <[email protected]> * fix(scripts-e2e): name both valueless credentials in the publish recipe's export instruction (#13344) * fix(scripts-e2e): cover the third pom classifier in the image jar guard (#13344) The base-conflict merge from main added the workspace-migration repackage classifier to the pom while the Dockerfile's exclusion list stayed at two arms, so the cardinality guard counted two unclassified jars and the documented image build failed at the RUN stage with no CI lane building the image to catch it. Add the missing arm, extend the shell-fixture populations with the real four-artifact target/ layout, and derive the exclusion set from the pom's own <classifier> elements in the suite so a fourth classifier reddens the Test job instead of an operator's docker build. The three prose copies of the artifact count are retired rather than renumbered, so they cannot drift again. Also close the two remaining round-one test-hardening findings: the port-allocation test now loads through the shared load() helper instead of a second local new Function spelling, and the image ENV pins anchor on the instruction (case-insensitive, leading whitespace tolerated) rather than the conventional uppercase column-0 spelling. Co-authored-by: Qwen-Coder <[email protected]> * fix(scripts-e2e): defer the freeze wake record and name stalls in waitUntil errors (#13344) - The --freeze arm printed its fencedFormerWriter record eagerly inside the try, so a SIGINT right after it left a success-shaped document on stdout for a run that exits 1, and the record flipped to before the main result. Defer it like resultJson and print it after the main result, restoring the base ordering. - When the deadline's own poll stalled, waitUntil reported a stale earlier error as the final state. The message now reads "predicate stalled; last error before the stall: ..." so both facts survive. - README: the documented MySQL grants did not boot the server (Flyway's performance_schema.user_variables_by_thread probe is denied with ERROR 1142 and startup stops at 'foreign_key_checks'); add the SELECT grant to the one-liner and mention it in the container section. * fix(scripts-e2e): floor the waitUntil stall bound so a final in-flight poll can answer (#13344) Co-authored-by: Qwen-Coder <[email protected]> * fix(scripts-e2e): print the README's full grant set in the dev recipe hint (#13344) Co-authored-by: Qwen-Coder <[email protected]> * fix(scripts-e2e): re-key the in-flight poll test off the clock for re-entry margin (#13344) Co-authored-by: Qwen-Coder <[email protected]> --------- Co-authored-by: qwen-code-dev-bot <[email protected]> Co-authored-by: Qwen-Coder <[email protected]> Co-authored-by: qwen-code[bot] <qwen-code[bot]@users.noreply.github.com> |
||
|
Shaojin WenandQwen-Coder
|
aca4d03c7e |
feat(managed-agent): Add dev:managed-agent one-shot dev launcher (#13218)
* feat(managed-agent): Add dev:managed-agent one-shot dev launcher Developing against the Managed dual-path WebShell previously meant starting an ordinary daemon, a private Hosted Harness, and the Spring control plane by hand, keeping their ports, tokens, and capability digest consistent by copy-paste. The new launcher mirrors daemon-dev: it serves the daemon and the Hosted Harness from TypeScript source with rotating tokens and digest, writes the Spring wiring to the git-ignored (0600) .qwen/managed-agent-dev/spring.env, waits for /actuator/health on the Java side (--skip-java-wait bypasses), then opens the WebShell Vite dev server with the Managed panel selected. Ports auto-increment when busy, and any child exiting aborts the group so the stack cannot half-start. Co-authored-by: Qwen-Coder <[email protected]> * fix(managed-agent): Harden dev:managed-agent against review findings Round-1 review found 4 Critical and 14 Suggestion-level issues in the launcher; this commit addresses all 18: - Windows teardown now uses taskkill /T so the dev.js grandchild that owns the sockets cannot be orphaned (R1-1). - Equal --daemon-port/--harness-port is rejected, the daemon probe excludes an explicit harness port, and both health gates assert responder identity through /capabilities (the harness must present this run's capability digest), so a busy-port bump can no longer be reported as a healthy mis-wired stack (R1-2). - The two serve children boot in serialized stages: dev.js destructively re-stages the shared browser-use runtime on every boot, and the parallel spawns raced on it (R1-3). - The token-bearing open path travels via QWEN_WEB_SHELL_OPEN_PATH, consumed as server.open in the web-shell Vite config, instead of a --open argv that npm echoed and cmd.exe would split on '&' (R1-4, R1-17). - spring.env moved to the OS temp directory (mode 0600) so the harness bearer token no longer sits inside the served workspace; a PowerShell sibling is written for Windows, a warning fires when a previous run's file (and a possibly stale Spring) exists, and the Spring recipe now carries the datasource exports and one-time sibling installs in both README and the startup banner (R1-5, R1-6, R1-11, R1-12, R1-14). - --tenant is validated against Spring's own pattern, waitForHttpOk carries the last HTTP status/error into its timeout, a Java timeout no longer tears down the healthy local stack, signal handlers live inside the entry guard, and the isMain guard is symlink-safe (R1-8, R1-9, R1-10, R1-16, R1-18). - Exported builders (kill plan, harness wiring, serve stages, web launch, URL) with mutation-witness tests pin the contracts the first round showed unpinned (R1-7, R1-13, R1-15). Co-authored-by: Qwen-Coder <[email protected]> Co-authored-by: Qwen-Coder <[email protected]> * fix(managed-agent): Probe the web port on Vite's bind family Vite binds 'localhost', which resolves to ::1 only on IPv6-first hosts, but the web-port probe listened on 127.0.0.1 and called a held port free — observed in a two-launcher dry run where the second launcher's probe passed and Vite silently bumped past the probed value (the printed URL then names the wrong port; the TTY URL also used the IPv4 spelling that cannot connect to an ::1-only bind). The probe now takes a host parameter and the web probe uses 'localhost', matching Vite's own bind attempt by construction; the printed URL uses the localhost spelling. A unit witness binds a server on 'localhost' and requires the probe to walk past it. Co-authored-by: Qwen-Coder <[email protected]> Co-authored-by: Qwen-Coder <[email protected]> * fix(managed-agent): Close round-2 review findings on dev:managed-agent The second review round (2 Critical, 8 Suggestions) found the round-1 remedies themselves creating new edges; this commit addresses all ten: - The env-file write is now fail-closed: the directory is keyed by uid and a hash of the checkout root (multi-user and multi-worktree runs no longer collide or cross-truncate), created 0700, verified owned, each target lstat-checked and refused when it is not our regular file (a planted symlink is bounced, its victim untouched), and 0600 is applied by chmod so it holds on the overwrite path too. POSIX-only enforcement: NTFS owns the boundary on Windows, where stat mode masks and chmod read-only would have broken the launcher outright (R1-11). - Teardown now covers SIGHUP plus a synchronous exit hook, so a closed terminal tab or dropped SSH session no longer orphans the credentialed serves that detached:true would strand (R2-1); registrations stay inside the entry guard and are pinned via TEARDOWN_SIGNALS. - Explicitly pinned ports are probed and rejected when busy, because qwen serve bumps on EADDRINUSE regardless, which previously meant a 60-second misleading timeout (R2-2); Vite gets --strictPort and a pinned --host so it fails closed instead of drifting past the probed port (R2-3). - waitForHttpOk captures err.cause.code (ECONNREFUSED now visible through undici's generic 'fetch failed') and clears stale accumulators so a fresher error supersedes an older response (R1-9 residual). - The Java gate applies its siblings' identity logic through springHealthExpectBody instead of certifying any 2xx (R2-6). - The PowerShell recipe line is uncommented and self-enabling (Set-ExecutionPolicy -Scope Process), emitted by an exported buildSpringRecipeLines that also carries the database/user creation step matching the reference e2e, mirrored into the README (R1-14, recipe gap). - Effectful decisions moved into exported, injectable helpers (executeKillPlan, runServeStages, ensurePortFree) with interleaving and argv witnesses, and nine loose oracles tightened (rotation, ::1 family fixture, daemon argv, launch env/argv, sharper retry message, the vite env-consumer case in web-shell, the Authorization header, the harness port in both base URLs, ambient-credential collision) (R2-4, R2-5). Dry-run witnesses: staged healthy with digest pairing, recipe lines and the env-file layout (0700 dir, 0600 file, no ps1 on POSIX) verified, pinned-busy fails in milliseconds with exit 1, and a live SIGHUP drains daemon, harness and Vite with no orphaned listeners. Co-authored-by: Qwen-Coder <[email protected]> Co-authored-by: Qwen-Coder <[email protected]> * fix(managed-agent): Skip the web-shell spawn once teardown has begun Self-audit after the round-2 fixes found two edges: a child dying during the Java wait now lands in the non-fatal continue branch, which would print the Managed URL and spawn Vite into a group teardown was already reaping (the exit hook caught it, but the output misleads); and killChild's re-entrancy comment overclaimed a no-op second pass — the group/taskkill plans do not set child.killed, so a second pass is a redundant signal to a dying tree, which is safe but not a no-op. Guard the spawn on shuttingDown and correct the comment. No behavior change on the happy paths; the script suite stays at 60/60. Co-authored-by: Qwen-Coder <[email protected]> Co-authored-by: Qwen-Coder <[email protected]> * fix(managed-agent): Close round-3 review findings on dev:managed-agent The third review round (4 Critical, 18 Suggestions), including three defects the round-2 fixes introduced, now addressed together with the review's batching recommendation: - shutdown() counted only signal-dead children (the >>continue<< predicate read 'signalCode === null'), so Ctrl-C/SIGHUP launched process.exit immediately, orphaned every live server, and a sole signal-killed child exited 0. The probe now reads isLiveChild, which the suite pins in both directions (R2-1). - The credential writer's checks resolved through symlinks: a planted link at the directory walked the lstat refusal, and a dangling link at a target skipped it entirely. The writer now lstats everything and refuses any symlink up front (R1-11), with planted-link victim- untouched cases. - Teardown now hears SIGQUIT (Ctrl-\) and escalates to SIGKILL after the 5 s grace instead of just exiting (R2-1/R3-1); the exit-hook comment narrows to what an in-process hook can promise instead of claiming coverage for unhandled signals and hard aborts. - Pinned-port checks moved from launcher start to each stage's own spawn step, so a port taken between the daemon's 60 s boot and the harness's spawn can no longer be silently bumped past (R2-2); the web port resolves at spawn time through an exported launch seam, so a squatter in the Java window cannot turn --strictPort into a full teardown (R2-3, R3-2). - The Java gate's teardown path is silent: no 'stays up' continuation line while children are reaping (R3-4), extracted into a runJavaGate seam with an injected health wait so the path is test-reachable. - healthFailureAction left only a literal-'continue' call site; the dead policy table, the unreachable abort branch and its describe are removed per the review's KISS option (R3-5). - The MySQL recipe grants the 'qwen' user at both 'localhost' and '127.0.0.1' (official images enable skip-name-resolve) with a container-topology note, and a pairing test ties the granted hosts/password to the exported SPRING_DATASOURCE_* literals (R2-7, R3-10). - Test-oracle batch (R3-3, R3-6..R3-9, R3-14, R3-15): default-runner witness for the kill plan, detached flag via an exported spawnDevProcess, first-run mkdir 0700 gate, stalled-responder rejection, both uid arms, stages[1].args pin, : prefix pins for the ps1 BASE_URL/digest lines, the printed ps1 path tied to the writer's filename, '503' replaced by the port-safe 'last response' suffix assertion with an in-range budget and a missing-finally close, the ephemeral-port-safe family fixture (property, not exact port), usable-on-Windows branch-shape cases decoupled from POSIX mode assertions, and parsed-actionable-line checks for the PowerShell recipe. ensurePortFree's unused probeHost parameter is gone (R3-13), and the env-transport comments now record the browser-hop argv the token still rides (R3-12). 78/78 in the script suite; live dry run healthy (dual-user recipe, argv clean), and a real SIGQUIT drains daemon, harness and Vite with no orphaned ports. Co-authored-by: Qwen-Coder <[email protected]> Co-authored-by: Qwen-Coder <[email protected]> * fix(managed-agent): Wire the Session Store every Managed Turn needs Real-stack verification on the PR thread found F1: since #12754 the Hosted Harness answers 400 invalid_managed_session_store to any POST /session without a managedSessionStore descriptor, and Spring only sends one when its HTTP Session Store is enabled — so the launcher's recipe brought up a perfectly healthy stack in which every Managed Turn failed with hosted_harness_rejected. The stub-Spring dry-run and env-name checks in earlier rounds never drove a real Turn through the wiring. Adopts the reviewed candidate patch (wenshao assets-pr13218 @ ef748da, measured 3/3 Turns against a freshly created DB) and adds the remaining notes: - spring.env/ps1 now also carry QWEN_MANAGED_AGENT_SESSION_STORE_* (enabled, base URL = --java-url since the Harness calls Spring back, and a dev workspace id), so the recipe by itself produces a usable Session Store (F1). - SERVER_PORT is emitted from --java-url for plain-http loopback URLs, so a non-default --java-url no longer leaves Spring on 8080 while the launcher waits elsewhere (N2). - Thrown messages no longer carry the '[managed-agent-dev]' prefix that main()'s catch prepends once (N3). - '--java-url' is trimmed before validation, so surrounding whitespace can't pass the checks and then fail every health attempt (N5). - The README states the hosted_harness_rejected diagnostic mapping (Harness 400 = Session Store not loaded; 401 = stale Spring) (N1). - The first-run 0700/0600 writer case is re-gated for POSIX semantics (R4-1); content/branch-shape cases keep running everywhere. The supervisor data point from N4 (SIGKILL of the launcher) stays a deliberate follow-up on the launcher family, unchanged. Co-authored-by: Qwen-Coder <[email protected]> Co-authored-by: Qwen-Coder <[email protected]> * fix(managed-agent): Gate the POSIX-arm writer case on the host, not the flag The round-5 review found the split's new case reproducing R4-1's exact failure on the Windows lane one level up: 'writes the env content and no ps1 sibling on the POSIX arm' injected isWinPlatform: false unguarded, so on a real Windows runner the production 0700 gate reads the synthesized 777 directory mode and writeSpringEnvFiles throws before either assertion — the mechanism and message R4-1 described, on the case I believed the split had made portable. It now gates on the host like its six siblings. Also adopts the round-3 verification's wording note: the two sibling installs are labeled once per clone, and re-run after pulling changes to qwencode or runtime-broker, since a snapshot install before a main pull leaves spring-boot:run compiling against stale artifacts (the banner builder and the README recipe say the same thing either way). Co-authored-by: Qwen-Coder <[email protected]> Co-authored-by: Qwen-Coder <[email protected]> * fix(managed-agent): Close the actionable slice of the round-4 deferred list Five items from the round-4 review body's deferred Suggestions were covered here because each is a small consistency hole in the seam this PR just built: - main() re-checks pinned ports at t0 before writeSpringEnvFiles runs, so a busy --daemon-port/--harness-port no longer rotates credentials and prints the banner before failing at stage spawn the same run (R4-14); the per-stage recheck stays for the boot window it was added for. - launchWebShell re-consults the teardown flag after resolveWebPort, so a launcher dying mid-probe no longer prints a Managed URL for a group the exit hook is about to reap (R4-10), with a witness that flips the flag mid-resolve. - findWebPort is now held against a stand-in bound to this host's localhost resolution at DEFAULT_WEB_PORT, so the family walk is operational rather than vacuous (R4-17). - installTeardownHandlers' captured handlers are asserted to be functions, not only named functions (R4-19). - The vi.mock child-process fixture spawns fake children with pid=undefined, so teardown code can never signal a real process through the registry it is pushed into (R4-20). Co-authored-by: Qwen-Coder <[email protected]> Co-authored-by: Qwen-Coder <[email protected]> --------- Co-authored-by: Qwen-Coder <[email protected]> |