Repository navigation
Conversation
Three onboarding/inventory tests timed out at Bun's default 5 s in the full local suite: `hack agent onboard` on an empty repo, and the Python-only and other-ecosystem inventory fixtures. None of these fixtures has a package.json, so findRepoRootForInit walked up past the temp dir. With TMPDIR=/tmp and a stray /tmp/package.json on the host, the repo root resolved to /tmp. The inventory then globbed all of /tmp and described the wrong repo. Bound each fixture with an empty `.git` (tests/helpers/repo-boundary.ts, matching the existing retained-* fixtures), assert the onboard prompt's inventory root is the fixture, and add a host-independent guard. The guard nests a bounded and an unbounded fixture under an ancestor package.json and checks where the walk-up stops. Timeouts and product code are unchanged. HACK-1137 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The prerelease installer bounded archive members by the sum of its allowlist sizes, so the 17-member next.4/next.5 layout filled the bound exactly. Replace it with a fixed 64-member hard ceiling (MAX_ARCHIVE_MEMBERS) and a distinct "Too many archive entries." refusal. Every other check is unchanged: each member must still be a distinct allowlisted, non-traversing, bounded regular file in a complete group, with checksums, modes, identity and signatures verified before selection. Add the manager shipped by 5.0.0-next.3 through next.5 (e2e2bcba) to MANAGER_PREDECESSORS with the fixture prerelease-manager-storage-v5.py, so those channels can adopt this installer through upgrade-manager. HACK-1124 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…atives Add WU04 negative-case coverage for the node job journal: - real SQLITE_FULL on terminal publication and cancellation leaves the prior receipt, generation and operation set untouched; the same requests apply exactly once after space returns - cancel and completion serialize under real SQLite write-lock contention; terminal receipts are final and an early cancel is never reported as an achieved cancellation - the same operation ID with a changed fixture, deadline, capability set, generation or action conflicts without effect, and a cancel operation ID cannot be re-pointed at another job - a supervisor crash after its claim, before or after spawn persistence, is quarantined without replay or signalling the recorded PID - terminal receipts, generation and retry identities survive both an in-process journal reopen and a node service restart Refs HACK-1117. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Register Docker-tier scenarios that start copies of the checked-in examples/patterns fixtures in the hosted docker-e2e shards: - pattern-p01-dependency-bootstrap (HACK-1145): cold and warm starts, a linked worktree, a bun.lock change and an installer killed mid-run; running consumers only ever read complete payloads. - pattern-p02-portable-toolchain (HACK-1146): hack run --profile from the primary checkout and a linked worktree with a .git pointer file; failing checks and exact exit statuses propagate. - pattern-p07-capability-overlays (HACK-1151): new laptop/runner overlay fixture; value-free env explain sources, container metadata, upstream reachability and no browser opener. Cards name their scenario in e2eScenario; static tests require it to be a registered, hinted Docker-tier scenario and check the P07 overlay split. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add `hack setup migrate` for moving one scope at a time from standalone Hack skill and rule files to the skills-only `hack` plugin (HACK-1132). - The default preview is read-only. It lists each standalone target as remove, blocked, preserve or absent, using the W11 ownership classifier. Customized, unknown and symlinked paths are preserved. - Replacement checks read client configuration per scope: Claude Code settings precedence plus an installed copy carrying the replacement skills; Codex `[plugins."hack@hack-dance"] enabled = true`. Cursor is not inspectable and needs an explicit --skip-plugin-check. - --apply is all or nothing. It refuses without writing when a target is blocked, commits a 0600 backup manifest with exact bytes and mode before removing anything, and rechecks bytes before each removal. - --rollback restores only into empty paths, never overwrites, refuses manifests whose paths, digests or bytes are not Hack-owned targets, and recovers an interrupted apply. - Hooks, MCP entries and AGENTS.md/CLAUDE.md sections are untouched. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Docker Hub shorthand such as alpine:3.21 was refused as registry_image, and only after native foreground startup had already booted the VM. - Add a pure shared normalizer (provider::image_reference): alpine -> docker.io/library/alpine:latest, user/img -> docker.io/user/img:latest, index.docker.io -> docker.io; tags, digests, registries and ports are kept. Uppercase repositories, uppercase/non-sha256 digests, bad ports, empty and over-long references refuse with closed value-free reasons. - Compose plan admission validates with the same normalizer and keeps the authored image text, so existing plan identities are unchanged. - ensure-image, resolve-image and fetch-image accept shorthand. - New pure `runtime admit-image --reference REF --json`; native foreground up admits every authored service image before hooks, VM or provider start. Refs HACK-1116. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add a library contract for long-lived helpers shared across branches on one loopback port (HACK-1142, WU29). Starters are serialized by the projects-registry lock protocol; a listener is adopted only after its identity endpoint names the expected helper and branch upstream; owned helpers are recorded as PID plus process birth time and spawned behind an exec gate that only opens once the claim is durable, so a lost starter never leaves an unclaimed child. Readiness is the identity answer, bounded by backoff and one deadline, with cancellation stopping the owned child. Release drops one holder and stops an owned helper's birth-verified tree only for the last holder; adopted, foreign and stale-PID processes are never signalled. Failures are a closed, value-free set. Portable process fixtures cover adoption, wrong upstream, foreign listeners, concurrent independent starters, stale/reused PIDs, child restart, cancellation during retry, placeholder liveness, lost parent before and after the gate, descendants, impostor identity, invalid claims and an unavailable census. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d cleanup Ordinary graph cleanup retired every bound environment slot but left each immutable intent in run/environment-leases, so admission, cleanup and restore scanned every historical generation of a repeatedly started graph. After `graph cleanup` commits, the same engine mutation lease now compacts the run's records. Each must match a receipt container that is absent and a guest slot that is absent, or nothing is written. Exact intent digests move into one value-free per-run tombstone bound to the VM incarnation, and only then are the active records unlinked. Prior boots fold into one sealed summary each; the current boot keeps per-slot reservations. A retired slot, or any slot of a sealed boot, is never recorded, delivered, retired as live or counted again. Pending, live, changed, misbound or foreign-incarnation records refuse the whole compaction before its first write; other runs' and standalone records are untouched. Exact residue of an interrupted compaction reads as retired and is completed by the next cleanup or by archive's resume-only pass; changed residue refuses. Recovery flows that carry an exact environment inventory keep using the fenced path without compaction. Refs HACK-1193 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…touching app data Add an opt-in, single-target quarantine for a regenerable cache directory inside an ordinary owned volume of a stopped, data-retaining graph (HACK-1198). - Compose volumes may declare `x-hack-disposable-cache` (at most eight distinct, non-nested, volume-relative directories). Shared dependency cache volumes and paths that are or contain a mounted subpath are refused. The declaration is recorded on the admitted volume resource; absent declarations keep plan and receipt bytes unchanged. - `graph inspect-cache-quarantine` measures the exact target under the provider lease and returns a selection digest binding receipt bytes, volume incarnation, root and target directory identities and measured bytes. - `graph quarantine-cache --expect-selection` renames the directory on the same filesystem into a Hack-owned `hack-quarantine` sibling of the volume's `_data` root. Nothing is deleted; reclaimed bytes are reported as zero. - A durable private record is committed before the guest effect and completed after the guest syncs the move; replay resumes only the identical selection. - Refuses active graphs, incomplete journals, undeclared or traversing paths, foreign labels, changed volume incarnation, any container consumer, symlinked or replaced targets, ambiguous destinations and torn records. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…can controls HACK-1127 (WU14) F9. - tests/env-secret-sentinel-scan.test.ts: CI synthetic-sentinel scan of a primary checkout and linked worktree after env add/materialize/exec, up (Compose generation), env apply, logs, list, explain and doctor plus a careless `git add -A` commit. Zero raw/hex/base64 hits required in output, child argv, tracked files, Git objects, HACK_HOME, HOME caches and runtime files; plaintext only in ignored, untracked, 0600 .hack/.env and compose.env.override.yml. - The modern compose.env.override.yml (decrypted values) is now written private (0600, atomic, symlink refusal). - encrypted_file first-use key bootstrap publishes the key exclusively (link(2) of a fsynced 0600 sibling) so concurrent first users adopt one key instead of replacing a key the store was already encrypted with. - Project key creation runs under a cooperative lock in the git common dir when no key exists, key files are published complete, and a read that overlaps a publication is retried briefly; primary and linked checkouts now converge. - Cross-context v1 controls: moves to another key are refused value-free; same-key moves are pinned as accepted and documented as residual risk. - docs/security/env-secrets*.md, docs/env.md, docs/architecture.md updated. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The hosted run showed the legacy hack run path printing its internal DNS/TLS notices on stdout ahead of the probe output. Match the probe's own git= and checkout= lines instead of the whole stream. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…and lost starts Read process births in UTC so clients with different TZ values agree on a live owned helper; previously a second client saw it as stale, re-adopted it and the last release leaked it. A stale owned claim whose PID still answers as the helper is now kept (claim-invalid) instead of adopted. Report deadline (not foreign-listener) when lock contention or the deadline cut the port probe short. Stop a helper left in phase starting by a lost starter once it still cannot become ready. Rethrow observer errors after closing the gate or stopping the started helper, and report state errors raised under the lock as claim-invalid/state-unavailable rather than lock-unavailable. Bound the claim read by bytes before reading it. Add a lock barrier to the concurrent-starters fixture and regression tests for every documented check; register the suite in the Linux process job. Refs: HACK-1142 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…y platform The hosted Linux job showed that a refused loopback fetch does not carry the macOS ECONNREFUSED code there, so every free port was misread as a foreign listener. A failed connection now never proves absence by its error code: the loopback bind probe decides (occupied is foreign-listener, free spawns). Add a listener that drops every connection to the foreign-listener fixture. Refs: HACK-1142 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…solution Route admit/inspect/quarantine through a Port trait so a synthetic provider runs the shipped orchestration and guest scripts against a fixture volume. Add resolve-cache-quarantine, which settles a pending record as quarantined or abandoned without moving bytes, so a restart or a persistently refusing move no longer blocks a volume. Require the same findmnt mount ID before mv so it cannot copy-then-unlink, report retained quarantine bytes on every result, bound records per volume, and add an ignored native bad-cache fixture. Document the commands in the runtime README and the storage-inventory undercount. Refs: HACK-1198 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Plan-time image admission now runs `runtime admit-image` before lifecycle hooks. The branch startup fixture's fake runtime treated every unknown call as a runtime effect and failed it, so admission refused before hooks ran. Answer admission like the runtime does, without touching the native-ran tripwire, so the test still proves hooks run before runtime up. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
argv is parsed once against the union of every command's options keyed by long name. Declaring `setup migrate --rollback <manifest>` as a value option retyped the existing boolean `config adopt --rollback`, so `hack config adopt --rollback --json` consumed `--json` as its value and failed with a usage error. That broke every legacy-compose adoption rollback in the docker E2E shards. Rename the migration flag to `--rollback-manifest <manifest>` and add a command-tree guard that each long option keeps one value type, plus a parse regression for `config adopt --rollback --json`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Detached startup ran one `docker compose up -d` under the 90 s HACK_COMPOSE_STARTUP_TIMEOUT_MS budget, so a first-time image build or large pull was killed before any container started. Split detached startup into an image phase (`up --no-start`, HACK_COMPOSE_IMAGE_TIMEOUT_MS, default 30 min) and a start phase (`up -d --no-build`, the existing startup budget). Each phase keeps its own owned process-group deadline, and E_STARTUP_TIMEOUT now names the phase, its budget and the variable that raises it. Refs HACK-1274 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Replace wall-clock guesses in three tests that failed intermittently on loaded hosts (HACK-1280): - The hanging-probe test now proves the kill only after the fake Docker publishes its pid through an atomic marker handshake. A probe timeout that fires before the handshake is inconclusive and escalates through named rungs (500 ms, 2 s, 8 s); every rung still requires the timeout refusal, and the reap check polls to a named deadline instead of a fixed 50 ms sleep. - The native compose ownership file sets a named 60 s default test budget: its scenarios launch many cold fake-Docker processes, and an expired test kept writing into the next test's shared root. - The initial forwarding trace test uses named bounds for each forwarded command (30 s) and for the test (150 s) instead of 3 s and 5 s. No product timeout or assertion changes. Refs HACK-1280 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every environment-bound graph the CLI creates is relay-owned, so ordinary graph cleanup never ran compaction for it. Foreground redelivery and fresh-owner restore now compact the run's retired history under the engine lease. Compaction runs after every prior slot is re-retired and after the acknowledged stopped receipt is superseded on disk, before the new generation allocates. Enrolled relay cleanup still does not compact, because the acknowledged-publisher check re-verifies its inventory. A compaction refusal after a committed cleanup is now Settled::Retained instead of an error. The records stay active and the next cleanup, restore or archive retries. graph::cleanup no longer reports a cleanup that succeeded on disk as failed. New tests cover the retire guard (live_intent), archived-copy refusal, residue with a pending record, the 16384 retired-slot cap, and 260 generations of a 14-service graph across 9 boots. A source-order wiring guard covers the trigger points. The README now describes the trigger points, refusal semantics, current-boot growth and tombstone retention. Refs: HACK-1193 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The "stop that cannot be confirmed" test froze its starter with
Subprocess.kill("SIGSTOP"). On macOS, the pinned Bun 1.4.2 delivers
SIGCONT for that call, so the starter kept running, reaped the helper
after TERM, and release reported `stopped`. CI uses the pinned Bun, so
the test failed there; local runs passed only on an unpinned Bun 1.4.3.
The scenario also showed a product defect. The census counted a zombie
(an exited, unreaped process) as live. A stop that left the helper a
zombie reported `stop-unconfirmed`. A `ready` claim whose helper crashed
under a suspended starter looked live, so other clients could not
restart the helper until that starter resumed. A zombie runs no code
and holds no listener, and its PID cannot be reused until it is reaped.
The census now reads `ps` state and treats zombies as exited.
The unconfirmed-stop test now injects the census and signal boundary
through a narrow `processes` seam on release, because no real process
can portably survive SIGKILL. A separate real-process test freezes the
starter with process.kill and asserts that the helper is an unreaped
zombie and that release reports `stopped`.
Refs HACK-1142
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Clients of a helper port coordinate only through the claim and lock under `<stateDir>/<name>`. The previous JSDoc called stateDir a "caller-owned private directory", which suggested one directory per caller. A client with its own directory finds no claim. It adopts the running owned helper, and the owner's last release then stops a helper that client still holds. Its starts are also not serialized with the owner's. The spec JSDoc, the module contract and the reference doc now state that every client of one port must pass the same stateDir and name, derived from the helper and port rather than the branch. Refs HACK-1142 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Contributor
Author
|
Integration check done: all 12 PRs were squashed into next in order. next b431f0e has the same tree as train/b18 (9001df2a). Closing; this PR was never meant to be merged. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Integration check only — do not merge this PR. It exists so CI runs on the combined result of the next merge train before each PR is squashed into
nextindividually.train/b18=next099877a plus these PR heads, merged in order:#358 f8cbd52, #343 e2dab3a, #339 11436cd, #338 536b348, #332 ebbdb9c, #333 e43075c, #336 a6c15bd, #345 254a7bf, #347 d5d96ed, #348 907726e, #349 0983d7a, #354 7b17177.
Each PR passed CI on its own head and its independent verifier. #341 conflicts with #336 in a generated plugin file and will follow separately after regeneration.
When this is green, the PRs are squashed in that order with
--match-head-commit, andnext's final tree is checked against this branch's tree (9001df2a). Then this PR is closed.🤖 Generated with Claude Code
Need help on this PR? Tag
@codesmith-botwith what you need.