The two items aimed at pointing GUI at a real veloxd instead of mockd.
rpc/event_hub — per-subscription fan-out shared by both transports.
subscribe() registers a connection with no interest; set_filter()
(session.subscribe, replaces not adds) turns on event kinds and an
optional per-task id filter; publish() delivers a pre-built
notification to every matching subscriber. session.subscribe on both
UdsServer and WsServer now does the real thing — registers/updates a
subscription, tears it down in close_conn.
sched/scheduler — the on_engine_state hook now actually publishes:
- transition() is the one place a task's row changes state; it reads
the store's own prior row for previousState (authoritative
regardless of engine/scheduler timing), writes the error columns,
and — when a hub is supplied — publishes event.task.state with
{taskId, state, previousState, summary, error}. Wired into every
transition: scheduler-driven (admission -> probing, resume ->
connecting, pause) and engine-reported (on_engine_state).
- progress_snapshot(): one row per task the engine is tracking
(EnginePort::progress(), a new interface method backed by
DownloadHandle::progress()), plus a store side-effect
(Tasks::update_progress) so download.list/get stay current between
state transitions. Returns rows; does NOT publish itself — batching
into one array message is the caller's job, per the schema's
x-maxRateHz: 4 and AGENT-DAEMON.md item 5 ("one message per task per
tick burns a core"). main.cpp's 250 ms timerfd is that caller: one
event.task.progress per tick, only when there's something to say.
dispatcher::on_download_add now publishes event.task.added (schema:
"summary is always present so a client can insert the row without a
follow-up download.get").
store/categories, store/queues — the two D3 handlers GUI's panels
call. category.list projects the six seeded built-ins; queue.list
derives taskIds from tasks.queue_id/queue_position (Queue's own schema
note: a queue's stored row never carries membership, download.update
/ queue.reorder do).
Verified live end to end against tools/testserver: a subscribed client
sees event.task.added on add, then the full event.task.state sequence
(queued -> probing -> connecting -> downloading -> assembling ->
verifying -> complete) with correct previousState at every step, and
real category.list / queue.list results.
Tests: event_hub (filter-by-kind, filter-by-task-id, replace-not-add,
unsubscribe), store_categories_queues, plus new sched_scheduler cases
for event.task.state publishing and progress_snapshot's store
side-effect. 38 daemon/cli tests green; sched_scheduler / event_hub /
ws_server / uds_roundtrip TSan-clean.
deferrals.md: D5 mostly closed (event.task.removed and the
still-unpublished events wait on their owning D3 handlers); D3 down to
the remaining download.* verbs, rules/settings/limiter/schedule,
queue mutation, category mutation, grabber, media.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
CORE stage 8 merged, so vdm::Engine is linkable. This closes D4a and
narrows D4b: `velox add <url>` now actually downloads.
- sched/engine_port_core.hpp — the real EnginePort: forwards to a live
vdm::Engine, keeps the DownloadHandle per task for pause/resume/
cancel/provide_auth/decide/refresh_url, drives set_task_order /
set_max_active_segments / set_host_segment_cap via
engine.segment_budget(). CORE confirmed the admission model: DAEMON
decides when to start(); the engine's own download_task calls
register_task/set_want internally — DAEMON never touches per-task
budget calls. EnginePort gains release(TaskId) so the port drops a
handle when the task goes terminal.
- rpc/event_loop — EventLoop::post(fn): thread-safe, runs fn on the
loop thread next iteration. The marshaller for engine-thread
callbacks.
- main.cpp — constructs vdm::Engine + EnginePortCore + Scheduler
(post_to_loop = loop.post). At startup: reconcile_after_restart()
(ADR 0013 §5), reload_config(), tick(). A 1 s timerfd on the loop
re-runs tick() (schedule windows, missed nudges); download.add nudges
via dispatcher.set_on_mutation.
End-to-end verified against tools/testserver: `velox add
http://127.0.0.1:.../file/512K` -> task queued -> scheduler admits ->
engine downloads 524288 bytes -> complete, file on disk. First
byte-path all the way through the project.
safepath-adversarial.md: re-verified per its own note — CORE landed
O_NOFOLLOW on the target open (core/src/io/sparse_file.cpp), so the
leaf-symlink TOCTOU is now closed; residual is down to one
intermediate-dir gap (documented post-M1 chase).
36 daemon/cli tests green; scheduler + uds_roundtrip TSan-clean.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
The Scheduler that D4 was waiting on. Built against CORE's engine
HEADERS (now in main); the real EnginePort and the veloxd wiring wait
for lane/core's stage-8 bodies to reach main (deferrals.md D4a/D4b) —
core/src/task/ is still .gitkeep there, so linking vdm::Engine now
would be an unresolved symbol.
- sched/engine_port — the abstract seam: start/pause/resume/cancel/
provide_auth/decide/refresh_url + the ADR 0011 admission config
(set_task_order / set_max_active_segments / set_host_segment_cap).
Keeps the Scheduler testable without a live engine and the daemon
unbound from the concrete vdm::Engine.
- sched/fake_engine_port — a recording impl for tests.
- sched/scheduler:
* owns the wire-UUID <-> vdm::TaskId map.
* tick(): snapshot queues (schedule window evaluated with an
injectable clock) + non-terminal tasks -> governor.evaluate ->
apply. to_start builds a vdm::task::DownloadSpec from the row and
calls EnginePort::start; to_resume -> resume(); to_pause ->
pause() + writes the pause_reason; priority_order -> set_task_order
over the mapped engine ids. `new` tasks are parked (startMode
manual) and skipped.
* on_engine_state(wire_id, state, err): projects an engine
transition onto the store row (state, pause_reason='auto' when an
error rides a paused transition per ADR 0013 §2, flattened error
columns) so the next tick sees ground truth. This is also the hook
event.task.state will fire from (D5).
* reconcile_after_restart(): CORE-owned states -> queued, paused
keeps its reason (ADR 0013 §5).
* reload_config(): reads connection.maxConcurrentDownloads /
maxActiveSegments + a daemon-local host-cap map, pushes caps to
the engine, updates the governor.
* Deps: injectable local-now clock and a post_to_loop marshaller
(engine callbacks arrive on engine threads; default runs inline
for tests).
Test veloxd.sched_scheduler (ASan+UBSan and TSan clean): admission +
ordering, a slot freeing on completion, queue-stop -> pause
(queue_stopped) then queue-restart -> resume (not a fresh start),
engine auto-pause -> pause_reason 'auto' + never auto-resumed,
reconcile_after_restart, reload_config caps push. 35 daemon/cli tests
green.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
Stage 8 of the CORE build order: the bodies behind the DownloadSpec /
callback API reviewed in core/docs/engine-api-m1.md. Wires probe -> segment
workers -> WriteBuffer -> SparseFile -> .veloxpart.meta -> retry/backoff ->
SegmentBudget -> RateLimiter -> callbacks into one event-driven machine.
- Engine (src/engine.cpp): owns HttpClient, Prober, SegmentBudget,
RateLimiter and one timer jthread (min-heap of scheduled fns). start()
builds a task and returns a DownloadHandle; ~Impl quiesces every task
before joining the timer so no callback fires during teardown.
- DownloadTaskState (src/task/download_task.cpp): one `mu` task lock; a
shared_mutex over the worker map for the curl write path; callbacks
collected under `mu` and fired after release via a separate deferred
queue; weak_from_this() in every async hop. State machine over the
CORE-owned EngineState subset, auto-pause on 401/407 and on a 200 where
206 was expected, validated resume via If-Range.
- digest (src/task/digest.cpp): OpenSSL EVP hash_file() for the optional
post-download checksum; links OpenSSL::Crypto PRIVATE.
- Segmenter::release_segment(): hand a paused segment back to the pool
unassigned so resume's assign_slot() picks it up instead of splitting a
still-"assigned" range and orphaning its front half.
- DownloadHandle now names the real control block (vdm::task::
DownloadTaskState, defined only in the engine TU) via a namespace-scope
fwd decl and a public-but-effectively-engine-only ctor, replacing the
nested State/friend pair. Every public signature is unchanged; DAEMON
(vdm-79) confirmed sched/ names only the public API.
Fixes found while building the end-to-end suite (tests/task/engine_test.cpp,
9 cases against tools/testserver, green under ASan/UBSan and TSan):
- a dropped connection lost its unflushed WriteBuffer tail while advance()
had already counted those bytes as done -> a retry resumed past an
unwritten hole. Flush on the failure path.
- when the byte counters hit total while other workers were still live,
teardown dropped their buffered tails. Now: cancel them and let each
worker's own seg_finished drain it (the `assembling` state), last one
starts verification -- no cross-thread buffer access.
- seg_head() let a 401 with credentials present abort before libcurl's
resend; now it proceeds once and acts on the final status.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HPPSGhiArbvQgwC2DNiURS
run_all() takes an optional substring list; vtest_main forwards argv and a
comma-separated $VT_ONLY. No filter => run everything, as before. Makes
iterating on one slow end-to-end case (the engine suite) practical
without a framework swap.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HPPSGhiArbvQgwC2DNiURS
libcurl with CURLAUTH_ANY answers a 401/407 by resending the request with
an Authorization header. Two spots in net/ cut that short:
- http_client's header callback delivered the response head exactly once
and latched `head_delivered`, so after an auth challenge the caller only
ever saw the 401 — never the 2xx of the authenticated resend. Reset the
latch when a fresh status line follows a delivered 401/407 (redirects
never reach that path — their head is suppressed).
- the prober's head callbacks return DataAction::abort to skip the body,
which also aborts the transfer mid-handshake. Return `proceed` for a
401/407 when credentials were supplied, so curl's resend can run; the
real status lands on the next header block.
Also give ProbeRequest an `auth` field (default scheme == none) and pass
it through base_request(), so a re-probe after a 401 can present the
credentials the user just entered. No behaviour change when no auth is
configured.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HPPSGhiArbvQgwC2DNiURS
VeloxDispatcher now takes a store::Db& and three handlers are real:
- download.list -> store::Tasks::list (filter / sort / paging in SQL) ->
to_summary per row. No more empty-table stub.
- download.add -> resolve saveDir (spec, else saveTo.defaultDir; ~ expanded)
and the leaf (spec.filename, else the URL's last segment percent-decoded,
else download.bin) -> fs::resolve_target against canonicalize_root'd
saveTo.allowedRoots. Any path-destination failure is -32011 with the
*original* saveDir in data.path. On success a TaskRow is inserted in
state `queued` (or `new` for startMode "manual") and {taskId, state}
returned. The scheduler that would then admit it is D4.
- download.get -> store::Tasks::get; a real -32010 + data.taskId for an
unknown id, else a TaskDetail (segmentDetail empty until the engine
segments the task, which the schema permits).
util/time.hpp: now_iso() factored out of ws_server.cpp.
main.cpp constructs the dispatcher with the opened db. The three
integration tests build an in-memory migrated db for it; velox.client
now drives the full slice through the CLI — add outside roots -> -32011
with data.path, add into an allowed root -> a task that download.list
shows and download.get details, unknown id -> -32010. Verified with the
real binaries: velox add persists, velox ls shows it, it survives a
daemon restart, /etc is refused.
ASan+UBSan and TSan clean; 34 daemon/cli tests green. deferrals.md:
D2 down to just download.probe; D3 down to categories/queues/rules/
settings/limiter/schedule.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
DAEMON's safepath-adversarial.md accepts a TOCTOU residual between its
canonicalise-and-check and the download starting, on the stated grounds
that CORE's O_NOFOLLOW open of the final file closes it. That flag was
never actually set: SparseFile::open used O_WRONLY|O_CREAT|O_CLOEXEC, so
a symlink swapped in as the final path component after DAEMON's check
would be followed and redirect our pwrites outside the allowed roots.
Add O_NOFOLLOW. A symlinked leaf now fails the open with ELOOP, which
errno_to_error already maps to Error::path_rejected. Regular files and
the O_CREAT of a fresh part file are unaffected; resume (existing regular
part file) is unaffected. Test that a symlinked destination is rejected
rather than silently followed, and that the link target is never touched.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HPPSGhiArbvQgwC2DNiURS
The residual section claimed the leaf/component TOCTOU is "closed in
practice by CORE's O_NOFOLLOW open of the final file". Verified: it is
not — core/src/io/sparse_file.cpp:77 opens O_WRONLY|O_CREAT|O_CLOEXEC,
no O_NOFOLLOW, no O_EXCL. Requested the flags from CORE via PKG/QA.
Doc now states the residual is currently OPEN, names the file:line and
flags checked and the date, says what actually limits exposure today
(0700 parent dirs), and flags this as the boundary where a reader
stops checking. Step 5 reworded the same way. Re-verify the flags when
the CORE change lands.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
The read/write surface the dispatcher handlers need, so download.add
persists and download.list / download.get project real rows when the
engine lands.
- store/tasks — TaskRow (1:1 with the schema), insert / get / remove /
set_state / count, and list(filter, sort, offset, limit) that does
all the WHERE / ORDER BY / LIMIT in SQL (M1 DoD: a 1000-row list
never materialised client-side). Filter covers states / category /
queue / case-insensitive filename+url substring / date range; sort
is a whitelisted column + direction with NULLs last, default
newest-first; the enum spellings in a state IN (...) come from
proto::to_string, never from user text. to_summary() projects a row
onto proto::TaskSummary including the flattened error block when the
task failed / retry_wait / auto-paused.
- store/settings — key -> JSON-text with a built-in default table
mirroring Settings.schema.json / ADR 0012; get_raw / set_raw /
overrides plus typed get_int / get_string / get_string_array for the
governor config and saveTo.allowedRoots. Full settings.get/set wire
projection lands with those handlers.
- veloxd_store now links velox::proto + nlohmann_json for the
projection.
Test veloxd.store_tasks: insert/get round trip, PK duplicate rejected,
the error-block projection, list total+paging+sort+every filter,
set_state pause_reason clear-on-unpause, remove, and settings default
vs override. ASan+UBSan and TSan clean; 34 daemon/cli tests green.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
veloxd is the one process that turns an untrusted string into a
filesystem destination, and via capture.offer that string can come
from a web page. CLAUDE.md §4 and the M1 DoD both name this.
daemon/docs/safepath-adversarial.md is the spec, written before the
code the way EXT did for shouldCapture: 21 rows — .. traversal
(A1/A2), absolute-outside-roots (A3), prefix-match confusion (A4),
symlink-out (A7), TOCTOU on a created tail (A8), NUL/control bytes in
the leaf that CORE's fuzzer hit through Content-Disposition (A9/A10),
degenerate and overlong leaves (A11/A13), overlong dir component
(A14), symlinked root (A16), destination-is-a-file (A17), and the
legitimate cases that must still pass — non-ASCII (A18), redundant "."
(A19), trailing space/dot trimming (A20).
fs/safepath.cpp:
- sanitize_leaf: strip <0x20 and 0x7F, trim ws, strip trailing dots,
reject ""/"."/".."/contains-'/', cap 255 UTF-8 bytes on a codepoint
boundary. Mirrors core/src/net/content_disposition.cpp.
- canonicalize_root: expand ~ and realpath each allowedRoots entry
once, so a symlinked root resolves to its target.
- resolve_target: reject relative saveDir and any ".." component
lexically; if the dir exists, realpath + component-wise containment
(a symlink that escapes is caught, one that stays inside passes); if
a tail is missing, realpath+check the deepest existing ancestor then
create the tail via an openat/mkdirat O_NOFOLLOW walk and re-derive
the final path from the fd. Every failure is -32011 with data.path =
the *original* saveDir (never the resolved path). Residual TOCTOU on
a pre-existing intermediate dir is documented and closed by CORE's
O_NOFOLLOW open of the file.
veloxd_fs static lib; veloxd_rpc links it for the download.add wiring
next. Test veloxd.safepath is the adversarial table, on a real temp
tree. ASan+UBSan and TSan clean; 33 daemon/cli tests green.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
The scheduling brain, built against CORE's headers (vdm/engine.hpp,
vdm/segment/budget.hpp) — signatures only; the Engine bodies land in
CORE stage 8 and the Scheduler that wires governor <-> store <-> engine
<-> timer comes after that (daemon/docs/deferrals.md D4).
- sched/governor — a pure decision function. In: a snapshot of every
task's coarse RunState and every queue's state (schedule windows
pre-resolved). Out: {to_start, to_resume, to_pause, pause_reasons,
priority_order}. Enforces, all in TASK units per ADR 0011 §1:
connection.maxConcurrentDownloads; the min(that, maxActiveSegments)
clamp (§2); Queue.maxConcurrent; the per-host task cap (§4); and a
stopped queue / closed window runs nothing. Never touches a task
paused for `user` or CORE's `auto` (ADR 0013 §3) — only Schedule /
QueueStopped / AdmissionReconcile are auto-resumable. Deterministic:
main-list before queued-in-queue, then queue order, then FIFO, then
task_id.
- sched/schedule_window — window_open(Schedule, local tm): disabled =>
always open; `once` => date + time match; `periodic` => weekday in
daysOfWeek (empty = every day) + time in [start, stop); null start =>
midnight, null stop => end of day, stop < start => overnight window.
Pure; re-evaluated every tick, no cached instants.
- veloxd_sched static lib; veloxd links it (nothing calls it yet).
Tests (ASan+UBSan and TSan clean): veloxd.sched_window (10 window
cases incl. overnight, once, null bounds), veloxd.sched_governor
(global/clamp/per-queue/per-host caps, stop vs window pause reasons,
resume-only-governor-reasons, auth-pause untouched, admission
reconcile, determinism under shuffled input). 32 daemon/cli tests
green; full tree green.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
CORE resolved all four (lane/core@3da4cd6): vdm::TaskId is hashable and
DAEMON never constructs one; DAEMON mkdir -p's save_path's parent
(engine -> Error::path_rejected if missing); sha512 added as the 4th
Checksum::Algo so no -32602 at the RPC edge; on_state(cancelled) then
on_finished(Err{Error::canceled}) -- note the one-L spelling in the
error taxonomy. Also notes rate/token_bucket + Engine::rate_limiter()
for the limiter.set wiring.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
rate/token_bucket.hpp — a lazily-refilled TokenBucket (starts full: burst
then throttle, IDM behaviour; rate 0 == unlimited; burst caps idle
accumulation) and RateLimiter, the global -> per-queue -> per-task
hierarchy (docs/04 §6). acquire(task, n) peeks every applicable level and
commits on all-or-none so a blocked attempt never leaks tokens at a level
that had them; held under one mutex so a concurrent detach can't dangle
the bucket it's using. vdm/ids.hpp gains QueueId.
Tests: burst/refill/cap/unlimited for the bucket; tightest-level-binds,
no-partial-consumption, detach-safety, and an 8-thread aggregate-rate
check for the hierarchy. Green under ASan/UBSan and TSan.
Engine-API review (DAEMON signed off, no sched/ or dispatch rewrite):
- Engine::rate_limiter() accessor added (limiter.set -> set_global_limit).
- Checksum::Algo gains sha512 to match the wire Checksum set.
- DownloadSpec: DAEMON creates save_path's parent dir before start();
missing dir -> Error::path_rejected (made explicit).
- cancel(): documented to fire on_state(_, cancelled, nullopt) then
on_finished(Err{canceled}), in that order; download.cancel ==
cancel(false), download.remove == cancel(true).
- engine-api-m1.md: the five open questions resolved with DAEMON's
answers (probe_hint optional, single cancel flag, {restart,
keep_partial, abort} is the whole set, per-task 4 Hz is fine,
refresh_url restarts all segments after a validating re-probe).
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HPPSGhiArbvQgwC2DNiURS
Review of core/docs/engine-api-m1.md (lane/core@d6cf1fe) in
daemon/docs/engine-api-review.md. Sign off: nothing forces a sched/ or
RPC-dispatch rewrite; the split matches ADR 0011 and ADR 0013.
Answers: (1) keep probe_hint optional — DAEMON has a ProbeResult only
on the File Info path; (2) one cancel(discard_partial), download.remove
is cancel(true) + DAEMON-side row/file cleanup; (3) {restart,
keep_partial, abort} suffices if the engine owns the mechanical 416
re-probe/re-split; (4) per-task 4 Hz progress is fine — DAEMON
re-batches across tasks for event.task.progress anyway; (5) refresh_url
restarts all segments on the new URL (the signed-URL case), mirror
rotation is spec.mirrors not refresh_url.
Four things to confirm, none blocking: vdm::TaskId copy/hash semantics
and that DAEMON never constructs one; who mkdir -p's save_path's
parent; sha512 (in the wire Checksum, not the engine enum) rejected at
the RPC edge; on_finished(Err{cancelled}) code + ordering vs
on_state(_, cancelled, _).
Integration timing: wire after sched/ lands. sched/ builds against
these signatures in parallel with CORE stage 8.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
The extension's fallback transport (docs/05 §4). veloxd now also listens
on 127.0.0.1, first free port in 52000-52016, and writes it to
<runtime>/ws.port (0600).
- rpc/ws_frame — RFC 6455 frame codec. Incremental; reassembles
continuation frames; enforces "client frames MUST be masked" (§5.1);
caps a reassembled message at 8 MiB. This is the attacker-adjacent
parser, so it has its own test table.
- rpc/ws_handshake — HTTP upgrade parse, Sec-WebSocket-Accept
(SHA-1 + base64 via libcrypto), and the two non-negotiable checks:
an Origin header must be present and must be moz-extension:// (a page
cannot pair). Version must be 13.
- rpc/ws_server — per-connection Handshake -> Open state machine on the
shared EventLoop. Token gate: session.pair mints a token behind the
approver + rate limiter; session.hello must present a valid one;
every other method is -32002 until authed. Privileged methods are
refused -32003 by the generated dispatch(). Ping -> Pong; Close
echoed. session.hello major-version mismatch -> -32001.
- rpc/pairing — PairingApprover interface + EnvAutoApprover dev stub
(approves iff VELOX_PAIR_AUTO=1); PairingRateLimiter (5 failures / 60 s
per origin, then 60 s lockout -> -32014, survives reconnect); a
four-digit code generator.
- store/pairings — the pairings table: create() returns the plaintext
token once and stores only its SHA-256; find_active_by_token,
touch, revoke, list_active.
- util/crypto — sha1 / sha256_hex / base64 / random_token over libcrypto.
- store/sqlite — pin the DB file (and -wal/-shm) to 0600.
- runtime_dir — resolve_data_dir() for $XDG_DATA_HOME/velox (velox.db).
- main.cpp — opens + migrates velox.db, starts both transports; a WS
bind failure is logged, not fatal (capture must fail open, the Unix
socket still serves the GUI/CLI).
Real gap, flagged not hidden: the pairing prompt is EnvAutoApprover for
now — a GUI dialog / desktop notification is build step 7. Pairing
needs VELOX_PAIR_AUTO=1 until then.
Tests (ASan+UBSan and TSan clean): veloxd.ws_frame (codec + handshake
vectors incl. the RFC 6455 §1.3 accept sample), veloxd.pairings (token
create/find/revoke, hash-not-token, rate-limit window + lockout +
per-origin isolation + success reset), veloxd.ws_server (full flow: 101
handshake, -32002 gate, deny-then-approve pairing, hello-with-token,
-32003 privileged refusal, real download.list). 27 daemon/cli tests
green; full tree green.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
docs/05 §3: "Download with Velox" on a link (linkUrl) or a media element
(srcUrl) hands that one URL to download.add with the page as referrer; a
configurable command (Ctrl+Shift+U) does the same for the active tab. These
are explicit user requests, so they skip shouldCapture and the fail-open
path — a failure is surfaced to the user via notifications instead.
The page/selection "Download all links…" item waits on the content-script
link harvester (build-order step 7) and will be added there.
manifest: commands.velox-grab-current-tab. 9 tests. 110 total, lint clean.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012Y9RU58hD1BuwP82DySUHk
- CategoryPanel: the left tree — All Downloads / Unfinished / Finished,
then Categories and Queues populated from category.list / queue.list,
with per-node task counts. Selecting a node emits a TaskSelection.
- DownloadFilterProxy: QSortFilterProxyModel keyed off that selection.
Client-side for M1 (the whole list fits); asTaskFilter() exposes the
equivalent TaskFilter for a server-side download.list once paging lands.
- MainWindow: menu bar (Tasks / Downloads / View / Help) sharing QAction
objects with the toolbar; QSplitter [panel | table]; Delete with a
confirm; Resume/Pause All; View menu toggles the panel. Actions
disabled while offline.
- Counts are computed off a throttled 400 ms timer, not the 4 Hz progress
path. Fixed a debounce-vs-throttle bug found in the first screenshot:
restarting the timer on every progress tick meant it never fired and
the status bar sat at "0 of 0 downloads".
- First-run column widths that fit the content.
- tst_downloadfilterproxy: nodes filter to their own rows, the
Finished/Unfinished split is correct, and the filter is dynamic (a row
that finishes leaves the Unfinished node with no re-list). Verified
against mockd --tasks 400 (screenshot: tree counts 75/84/89/83/69 sum
to 400, status bar "400 of 400, 21 active").
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_016Ne28kx4VreeBWZv82Nksd
Downloads that never reach the header hook (form POST results, service-worker
responses, clicks Firefox routes straight to its downloader) surface here.
If one looks like the daemon's, offer it FIRST and only cancel + erase
Firefox's copy on {action:"take"} — a failed or slow offer can never leave
the user with nothing. blob:/data: downloads are left to Firefox (the daemon
can't fetch a blob URL).
- offered-urls.ts: short-lived, bounded TTL set of URLs the header hook has
already offered; the safety net checks it (via wasOffered) so nothing is
double-handled. Hook gains an onOffered hook to populate it.
- background/index.ts: both paths share one offer(), getCookies, rules
mirror, and OfferedUrls instance.
13 new tests incl. fail-open (offer rejects -> 'error', ignore ->
'offer_declined', cancel() throwing after take still returns 'taken').
101 tests green; web-ext lint clean.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012Y9RU58hD1BuwP82DySUHk
The download entry point the AGENT-CORE brief asked for on day one and
that slipped. DAEMON has an RPC surface and a store and, until this is
agreed, nothing in velox::core to call.
vdm/task/download.hpp — DownloadSpec (the resolved subset DAEMON hands in:
absolute save_path, verbatim browser headers, requested segments/buffer,
optional probe_hint / checksum / auth, allow_resume), EngineState (the
CORE-owned subset of the wire TaskState), Progress / SegmentProgress,
DownloadCallbacks (on_progress <=4 Hz, on_state for every transition
incl. auto-pauses, on_auth_required, on_decision_needed, on_finished
last), DownloadHandle (pause/resume/cancel — idempotent per the ADR 0013
signature — plus provide_auth / decide / refresh_url, and synchronous
state()/progress() snapshots).
vdm/engine.hpp — Engine: start(spec, callbacks) -> handle, segment_budget()
(DAEMON's sched/ admission surface, ADR 0011), live connection.* setters,
a standalone probe() on the pool outside the segment budget.
core/docs/engine-api-m1.md — the review doc: field semantics, the state
machine, threading/lifetime rules (which thread callbacks arrive on, what
is legal from inside one, handle/engine lifetime), the shared-`paused`
idempotency contract as a signature, and five open questions for DAEMON.
Value types compile and are covered by api_compiles_test; Engine /
DownloadHandle bodies land in stage 8, built against whatever DAEMON
signs off here.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HPPSGhiArbvQgwC2DNiURS
For a response shouldCapture() likes: gather cookies, offer to the daemon
under a hard 750 ms budget, and return {cancel:true} ONLY on an explicit
{action:"take"}. Every other outcome — shouldCapture says no, daemon down /
slow / erroring, cookies fail, anything throws — resolves to {} and Firefox
downloads normally.
Fail-open tests written first (tests/capture/hook.test.ts): offer() rejects,
offer() never settles (resolves within budget), timeout-shaped rejection,
getRules() throws, malformed details. Plus the happy paths and a check that
the capture.offer payload is well-formed from details + stash + headers.
background/index.ts: wire the stash + hook onto browser.webRequest, mirror
capture rules via capture.getRules on connect and on a capture.* settings
change; DEFAULT_CAPTURE_RULES (enabled:false) until the first mirror lands.
84 tests green; web-ext lint clean.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012Y9RU58hD1BuwP82DySUHk
Six lanes pick ADR numbers with no allocator. This lane dodged one collision
by taking 0012 while DAEMON drafted 0011, and just hit a real one — two 0014s
in the same integration round. Codify what already happened in practice: take
the next free number in main's docs/adr/, and on collision whoever merges
second renumbers and fixes cross-refs rather than round-tripping.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012fgjnqFCS5h5L7gZTZo3rV
PKG landed docs/adr/0014-conformance-runs-through-ctest.md (dad88fc) and this
lane's 0014-generated-binding-changes-and-versioning.md landed in the same
integration round, both as 0014. Renumbered this one to 0015 — second merger
renumbers. Updated the two cross-references (contracts/README.md rule 4,
proto-answers-daemon-m1.md P1) and the in-file header.
Also while in contracts/README.md: the file-tree comment and the "Method
surface" heading still said v1.2.0; both now v1.4.0 to match VERSION.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012fgjnqFCS5h5L7gZTZo3rV
Rebased onto main at 1.4.0. The regenerated Dispatcher returns
HandlerResult<T> = expected<T, HandlerError{code, message, data}>
(ADR 0014); the covariant-return break on all 39 overrides is the swap
predicted in daemon/docs/proto-requests-m1.md P1.
- dispatcher.hpp/.cpp: Result<T> -> HandlerResult<T> on every override;
not_implemented() now returns HandlerError{InternalError, ...} rather
than a ParseError forwarded as -32603.
- download.get: returns -32010 TaskNotFound with data.taskId. Not a
placeholder — with no store, every id is genuinely not-found, which
is the real answer for contracts/ fixture download.get.not-found. It
becomes a store lookup when store/ is wired in.
- uds_roundtrip: the -32603-collapse guard is now a -32010 + data.taskId
assertion, the regression guard the P1 note promised.
download.add (-32011) and download.probe (-32013) stay InternalError
until they have real bodies (canonicalization / probe); they get their
fixture codes when that logic lands.
All 24 tests green; uds_roundtrip TSan-clean.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
The daemon's persistent state. SQLite in WAL mode, foreign keys on,
5 s busy timeout so a writer waits rather than SQLITE_BUSY under the
RPC loop.
- store/sqlite — RAII Db/Stmt over the C API; errors returned as
DbResult<T> (std::expected), never thrown — the RPC loop must not
unwind. transaction() helper: BEGIN / fn / COMMIT, ROLLBACK on error.
- store/migrations/0001_initial.sql — the eight tables from the brief:
settings, categories, queues, tasks, segments, rules, history,
pairings. Notable choices:
* tasks columns project onto proto TaskSummary with no computation;
requested vs effective segments/buffer split per ADR 0010/0012;
pause_reason column per ADR 0013.
* segments end_byte is NOT constrained >= 0 so a whole-file
zero-length download is one row with end_byte = -1 (ADR 0010 B3a).
* pairings stores only token_sha256 — the plaintext token is
returned once from session.pair and never persisted (CLAUDE.md §4).
* indices on tasks(state), (category_id), (queue_id, queue_position),
(created_at), (completed_at) for the "1000 tasks, download.list
under 50 ms" DoD.
* six built-in categories + a Main queue seeded.
- store/migrations — runs every embedded migration past PRAGMA
user_version, each in its own transaction, forward-only. SQL files
are embedded at build time by cmake/embed_migrations.cmake.
Test veloxd.store_migrations (ASan+UBSan and TSan clean): fresh DB ->
head, all tables present, seed rows, FK cascade (segment orphan
rejected, task delete cascades), the end_byte=-1 zero-length case,
idempotent re-run, and forward-only from every released user_version.
Also: daemon/docs/proto-requests-m1.md — P1 marked landed on lane/proto
as 1.4.0 (HandlerError/HandlerResult), to be adopted in rpc/ once that
merges to main; P2 resolved.
Not linked into the running daemon yet — the store is wired to the
dispatcher when download.add/list/get get real bodies, next.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
The scriptable client, built now rather than last: it is how the daemon
gets exercised before the GUI is pointed at it (AGENT-DAEMON.md).
- src/client — synchronous blocking RPC over the Unix socket: resolve
$XDG_RUNTIME_DIR/velox/velox.sock, connect, session.hello, one framed
request/reply per call. Distinguishes transport failure (exit 3) from
a daemon-returned error (exit 1).
- src/main — subcommands add/ls/pause/resume/rm; --json prints the raw
JSON-RPC result or error; --dir/--out/--segments on add;
--delete-file on rm. Usage errors exit 2.
ls works end to end against veloxd today (empty table). add and the
bulk verbs reach the daemon and surface its "not implemented" (-32603)
cleanly until the store lands — the plumbing is done, the commands
light up as handlers do.
Test velox.client: the real Client against an in-process UdsServer —
no-daemon path, session.hello, download.list, and a not-implemented
method surfacing as an RPC error rather than a transport error.
ASan+UBSan clean; full tree (21 tests, incl. conformance) green.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
First real code in daemon/. veloxd now listens on
$XDG_RUNTIME_DIR/velox/velox.sock (0600, SO_PEERCRED same-UID check),
frames NDJSON, and routes every method through the generated
velox::proto::dispatch(). The CLI and GUI have a server to talk to.
Modules:
- rpc/ndjson.hpp — newline-delimited framing, 8 MiB frame cap, CRLF-
tolerant, partial-tail buffering. Header-only, tested.
- rpc/event_loop — single-threaded poll(2) reactor; never blocks the
loop. stop()/wake() are async-signal-safe (eventfd).
- rpc/runtime_dir — $XDG_RUNTIME_DIR/velox resolution, 0700, owner-checked;
refuses an insecure fallback rather than using /tmp.
- rpc/uds_server — listener + non-blocking per-conn read/write with
backpressure; handles session.hello (protocol-major
check -> -32001, sessionId, transport=uds) and
session.subscribe in the server layer; routes the
rest through dispatch().
- rpc/dispatcher — VeloxDispatcher : proto::Dispatcher, all 39 methods.
download.list answers an empty table; the rest return
"not implemented" (-> -32603) until the store lands.
- main.cpp — abstract-namespace single-instance lock, signal ->
clean shutdown, socket unlinked on exit.
Tests (ASan+UBSan and TSan clean):
- veloxd.ndjson — framing edge cases
- veloxd.uds_roundtrip — real socket: hello ok / version mismatch / empty
list / -32601 / -32700 / pipelined requests, and a
guard on the -32603 collapse documented in P1.
Known gap, filed not worked around: daemon/docs/proto-requests-m1.md P1 —
the generated Dispatcher has no error channel below -32603, so handlers
cannot yet return -32010/-32011/-32013 with their data payloads. The
server layer handles -32001/-32002/-32003 around dispatch(); genuine
in-handler errors collapse to -32603 until PROTO gives handlers a real
error return. Three error fixtures are non-conformant until then.
Not in this drop: rpc/ws_server (next; needs the store for hashed pairing
tokens), store/, sched/, cli/. WS reuses this event loop.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Upd9WhG9oppieig5nRDLig
Records the rule 1.4.0 applied, so the next generated-binding retype cites it
instead of relitigating README rule 4's "retype → major + ADR" from scratch.
The rule: VERSION tracks the wire protocol, not any binding's API or ABI. A
change that leaves the wire byte-identical but breaks a generated binding's
source API (a C++ virtual's return type, a struct name) is a minor bump plus
a migration note — major would make session.hello refuse a client whose wire
behaviour is unchanged, which is worse than the problem. An ADR is still
required when the change encodes a design decision; "only one lane consumes
it" is not a reason to skip that, since GUI already links velox::proto and
the next such change starts with more than one consumer.
Rule 4 in contracts/README.md now points here so its "major + ADR" line
isn't read in isolation. proto-answers-daemon-m1.md's P1 writeup references
it as the durable home for the reasoning that was otherwise only in a commit
message.
Also names the nlohmann brace-init hazard the 1.4.0 work hit: json{nullptr}
is the array [null], not JSON null, so HandlerError::data is `= nullptr`. The
kind of thing a regeneration reintroduces; caught here only by an end-to-end
assertion on dispatch() output, which is called out to keep.
Docs only — no schema, VERSION, or generated-code change.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012fgjnqFCS5h5L7gZTZo3rV
vdm/ids.hpp — TaskId, an opaque engine handle (DAEMON keeps the wire
UUID <-> TaskId map; the engine never sees the UUID).
segment/segmenter — per-download range management (docs/04 §3). Initial
lazy split; assign_slot() splits the largest remaining range when the
budget grants a slot; on_complete(may_steal) either *steals* the second
half of the largest remaining range for the same worker (slot-neutral) or
returns nullopt so the caller *yields* the slot (ADR 0011 A1); on_failed()
returns requeue only on the 3rd consecutive connection error with a mirror
present — the remaining range is orphaned and re-split. Non-resumable or
unknown-size => exactly 1 segment; never split below min_segment_bytes
(1 MiB). Resume ctor rebuilds from a persisted table (falls back to a
fresh layout if it doesn't tile [0,total)). One mutex == "the task lock";
segment fields are std::atomic and the store is a std::deque so a steal's
append never moves a worker's record.
segment/budget — the global allocator (ADR 0011). Owns exactly one
ceiling (maxActiveSegments) and min-1-before-seconds fairness: a two-pass
allocation (guarantee pass gives every wanting task 1 slot in DAEMON's
priority order, then a growth pass round-robins the rest up to each
task's effective cap = min(per_task_cap, host cap, 1 if non-resumable)),
recomputed from scratch on every edge so a live set_max_active_segments
cut naturally yields the excess lowest-priority-first, never a
mid-segment kill. DAEMON-facing surface exactly as promised in
daemon/docs/core-requests-m1.md / ADR 0011: budget(), segments_active(),
starved_tasks(), starved_since(), set_max_active_segments (drain),
set_host_segment_cap, set_task_order, on_budget_changed (a jthread
coalesces at <=4 Hz; the tasks_starved 0<->nonzero edge fires
immediately). Callbacks are copied out and run after the lock is
dropped.
Tests: segmenter split/steal/requeue/resume math + a concurrent
steal-and-advance run; budget min-1 under a tight budget, round-robin
growth, host-cap and non-resumable clamps, live-lower shedding
lowest-priority-first, starvation below the task count, starved-edge
notification, and a concurrent set_want hammer. Green under ASan/UBSan;
the steal path and the budget green under TSan.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HPPSGhiArbvQgwC2DNiURS
DAEMON's daemon/docs/proto-requests-m1.md P1: velox::proto::Dispatcher's
on_* methods returned Result<T> = expected<T, ParseError>, and dispatch()
mapped every handler error to -32603 InternalError. A handler had no way to
return -32010 (download.get not-found), -32011 (download.add invalid-path)
or -32013 (probe-failed) with their data payloads -- three error fixtures a
conformant server must satisfy were unreachable, blocking DAEMON's
"conformance as a server" M1 DoD.
Two error channels now, kept separate on purpose:
- parse: Result<T> / ParseError -- dispatch() failing to turn the wire into
typed params. Always -32602, always structural.
- handler: HandlerResult<T> / HandlerError -- a handler deciding the request
can't be fulfilled. Carries any ErrorCode + message + free-form data.
struct HandlerError {
ErrorCode code{ErrorCode::InternalError}; // bare {} is a valid -32603
std::string message;
nlohmann::json data = nullptr; // straight into the error's data
};
template <class T> using HandlerResult = std::expected<T, HandlerError>;
dispatch()'s handler branch is now
make_error(id, r.error().code, r.error().message, r.error().data)
instead of a hard-coded InternalError. -32001/-32002/-32003 stay the server
layer's to raise around dispatch(), as DAEMON already does.
Verified end to end against the real dispatch() path: a handler returning
TaskNotFound/InvalidPath/ProbeFailed produces -32010/-32011/-32013 with the
data object intact, and a bare HandlerError{} still yields a clean -32603
with no data field. The `= nullptr` on the member (not `{nullptr}`) matters:
brace-init of nlohmann::json from nullptr is the array [null], not JSON null.
FixtureDispatcher regenerated to HandlerResult; conformance_main.cpp only
inspects dispatch()'s JSON and needed no change. TS side is untouched beyond
the version string -- no server Dispatcher is generated there.
P2 also handled: session.hello.version-mismatch's data.expected was a stale
"1.0.0"; now $any, with a note that the error-fixture compare is on `code`
only so a server echoing kProtocolVersion there is fine.
Version: minor, 1.3.0 -> 1.4.0. Wire is byte-identical (no schema, fixture,
or OpenRPC change) but every Dispatcher implementer must swap Result ->
HandlerResult on regen, and the bump is how lanes are told to. Not an ADR:
one lane consumes this binding, it's the one that asked, and the shape is
the one they proposed. Answered in contracts/proto-answers-daemon-m1.md.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012fgjnqFCS5h5L7gZTZo3rV
R3: the three GUI DoD gates (10k rows at 60fps, flat RSS over 10 min,
--slow/--flaky/--drop-connection recovery) have nowhere to run. GUI owns the
harness, PKG/QA owns the job. tests/integration/README.md records the wiring
contract — driver invocation, exit-code and --json semantics — and carries
the pre-drafted per-PR and nightly job stanzas with TODO(GUI) markers for the
harness path. Wire for real when GUI files the follow-up.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_0143aKiohmDiyefJBwHDJJqw
PROTO wired the suite into ctest (label "conformance": the end-to-end
`conformance` test that shells to run.sh, plus the native `conformance_cpp`);
the CI job called run.sh directly. Two entry points, and the required M0 gate
exercised only one of them, so the ctest registration could rot.
The conformance job now configures, builds velox_conformance_cpp, and runs
`ctest --preset dev -L conformance`. The dev test preset's
noTestsAction: error is the rot guard: an empty label match exits non-zero
instead of the old silent pass. build/sanitizers exclude the heavy e2e test
with -E '^conformance$' (the dedicated job owns that run; conformance_cpp
still runs under every sanitizer). ADR 0014 records the decision.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_0143aKiohmDiyefJBwHDJJqw
libFuzzer writes crash-* / oom-* / leak-* / timeout-* into CWD on a find and
each holds the crashing input verbatim. None were ignored; CORE caught two by
hand before committing. Prevention, not cleanup.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_0143aKiohmDiyefJBwHDJJqw
--check verified outcomes — binaries on PATH, pkg-config modules — on a box
that already installed everything. It never looked at the APT_* names, which
is the only part that breaks on a clean machine of the wrong release, as R1
just showed. Add a loop over the assembled PKGS array that fails on any name
with no installable candidate (apt-cache policy; no root, no network).
All-missing is treated as stale lists (warn), not 36 bad names.
The bootstrap-script job runs on a 24.04 runner, where the bad name still
resolves, so name validation there checks the wrong archive. Add
bootstrap-script-2604: a real --with-clang install in an ubuntu:26.04
container — the release the project ships on, and the first time the
fuzz-toolchain half of bootstrap is exercised anywhere (CORE had been running
clang++-21 directly). Also pass --with-clang/--packaging to the 24.04 --check
so the optional and M6 names can't rot unnoticed. BRANCH_PROTECTION.md gains
the 2604 row and the stale "when X merges" rows are corrected to "now".
gui/docs/pkg-qa-requests-m1.md R2 + the --with-clang note.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_0143aKiohmDiyefJBwHDJJqw
`apt-cache policy libqt6svg6-dev` is "Candidate: (none)" on 26.04; the package
that carries the Svg headers and Qt6SvgConfig.cmake is qt6-svg-dev. Since
gui/CMakeLists.txt landed on main the root build's
find_package(Qt6 ... Svg REQUIRED) is live, so a clean 26.04 box could not
configure the project at all. Same name in README.md and AGENT-PKG-QA.md.
Reported in gui/docs/pkg-qa-requests-m1.md R1.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_0143aKiohmDiyefJBwHDJJqw
core/docs/proto-requests-conformance-cmake.md: the runner's header comment
said it links libveloxproto, but it actually compiled
core/generated/velox_proto.cpp straight into the executable and found
nlohmann_json itself — the only option while ADR 0009's target didn't exist.
It exists now on main as velox::proto (core/CMakeLists.txt, PUBLIC generated
include dir, PUBLIC nlohmann_json).
if(TARGET velox::proto): link it. else: fall back to compiling the generated
.cpp directly, for a configure with no core/ in the tree. Both paths verified
— full tree links libveloxproto.a (compiled once, by veloxproto's own
target); with core/CMakeLists.txt hidden the fallback compiles the .cpp and
finds nlohmann itself. conformance_cpp passes either way.
Beyond tidiness: once GUI links velox::proto too, the conformance runner
linking the same target is what guarantees the suite and the clients exercise
byte-identical generated code, rather than two compiles of one .cpp under two
warning configs — the exact skew a conformance suite exists to catch.
run.sh's own direct g++ compile is unaffected and stays independent by
design; this only changes the ctest-driven path CI and lanes use.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012fgjnqFCS5h5L7gZTZo3rV
samiandClaude Sonnet 5