A completed background run's GET status returned a bare COMPLETED with an
empty `outputs` and a null `output`. The COMPLETED branch reconstructs
from `vertex_builds` keyed by job_id, which the durable path does not
write, so reconstruction raised ValueError and fell through to an empty
response; the result was only retrievable via a /events re-attach.
The runner now captures the terminal `output` events (the langflow
adapter's normalized ComponentOutput payloads) into `Job.result`, and the
status COMPLETED branch rebuilds the `outputs` map and resolved `output`
from them via `workflow_response_from_output_events`, matching the sync
response. agui-protocol runs emit no `output` events, so their status
stays result-less (the result remains on the /events log).
The v2 workflows endpoint applied `tweaks` only on mode=sync. The stream
and background paths build the graph via the v1 build-vertex loop
(`generate_flow_events`), which never received the tweaks, so they were
silently dropped. The confusing symptom: a model passed via tweaks
surfaced as "A model selection is required", and any per-component
override was ignored on non-sync runs.
Thread `parsed.tweaks` into `generate_flow_events` and apply them to the
built graph via `vertex.update_raw_params`. We do not use the lfx
`process_tweaks_on_graph` helper because it only sets `vertex.params`,
which does not persist to runtime (the same bug
`lfx.base.tools.run_flow._process_tweaks_on_graph` works around). No-tweaks
runs are unchanged (guarded by `if tweaks`). Adds a streaming regression
test that overrides ChatInput via tweaks and asserts the value drives the run.
The durable background-execution rewrite of workflow.py reverted
`side_channel_events` to its pre-"end" form, dropping the "end" event
from the AG-UI side-channel. That event carries `build_duration` to the
playground chat-view, and the message metadata badge only renders when
`hasDuration || hasTokens`. With build_duration gone the badge vanished,
failing the token-usage and shareable-playground "Finished In" regression
tests. Re-add "end" so the streaming playground path delivers it again.
The durable background-execution work added six Settings fields
(background_max_concurrency, background_job_timeout, background_lease_ttl_s,
background_heartbeat_interval_s, background_watchdog_interval_s, test_redis_url)
without updating EXPECTED_FIELDS, so test_field_count_unchanged failed 152 != 146.
These are intentional bg-exec config; add them to the gate.
As a module-level runtime value, FrameSourceFactory = Callable[..., Any] is a
GenericAlias that passes isinstance(obj, type) but makes issubclass(obj, Service)
raise on Python 3.10/3.14. The service factory scans this module for Service
subclasses (services/factory.py:90), so the runtime alias crashed service
initialization on those interpreters with 'issubclass() arg 1 must be a class'
-> 'Could not initialize services', erroring out dozens of unrelated tests at
setup (3.13 was unaffected, which is why local runs passed).
The alias is only referenced in a lazy annotation (from __future__ import
annotations), so moving it + the Callable import under TYPE_CHECKING removes it
from the runtime namespace with no behavior change.
Verified: alias absent from runtime module namespace, factory scan finds
BackgroundExecutionService cleanly, durable service tests 26/26 pass.
These 3 tests targeted the removed queue-service stop helper
(_cancel_workflow_queue_job / get_queue_service), inherited via the
agui->bg-default merge and failing with AttributeError across the stack:
- test_stop_workflow_success: adapted to the durable stop path
(revoke_task -> stop_job -> update_job_status(CANCELLED)).
- test_stop_workflow_allowed_for_legacy_job_with_no_user_id (IDOR): adapted to
the durable mechanism; still asserts the ownership check does not block a
legacy user_id=None row.
- test_stop_workflow_returns_503_when_queue_cancel_cannot_be_confirmed: dropped
— the durable stop writes a best-effort STOP signal and always finalizes
CANCELLED; the queue-service 'cannot confirm -> 503' path no longer exists.
test_workflow.py now passes 24/24.
A v2 background run created TWO JobType.WORKFLOW rows for one flow execution:
the durable row (submit()'s job_id, owned by JobRunner) plus an orphan keyed by
the run_id generate_flow_events mints, because the build pipeline's
track_job_status defaults True and the durable frame source never passed False.
The flow ran once (double bookkeeping), but every background run left a phantom
WORKFLOW row + job_events and double-fired the memory-base hook, skewing metrics.
Thread track_job_status through _stream_event_frames; pass False only from the
background frame source (the durable runner already owns the row + fires the
hook with the durable job_id). Stream/public paths keep default True. Also gate
build.py's memory-base hook fire behind track_job_status so background doesn't
double-fire. Adds a regression test (RED before fix: found 2 rows).
Resolves the workflow.py / test_workflow_agui.py conflict in favor of the
durable background-execution design: the in-memory _BackgroundRun buffer,
_BACKGROUND_RUNS registry, and their tests are dropped (replaced by the durable
job_events store + InMemoryLiveBus + JobRunner, which already port the
load-bearing guarantees: replay-then-tail, Last-Event-ID resume, exactly-once
non-resurrecting cancellation, FAILED-vs-COMPLETED, cross-worker stop,
privacy-404). The base's AG-UI translator fixes (inactive-vertex deltas +
dedupe) and the jest downlevelIteration fix merge in cleanly.
Verified: 209 backend tests pass (services/background_execution, v2
test_workflow_agui, test_agui_translator, test_converters).
build.py keeps reporting a conditionally-excluded vertex in
inactivated_vertices on every subsequent end_vertex (the excluded set
persists until the ConditionalRouter clears it), so the translator was
putting the same inactive STATE_DELTA on the wire once per remaining
vertex. Track emitted inactive nodes and skip re-emitting; drop a node
from the set when it actually runs again (build_start/end_vertex) so a
loop re-activation can still re-emit inactive later.
A branch component (If-Else, Conditional Router) reports its not-taken
vertices in build_data.inactivated_vertices, but the AG-UI translator only
emitted the branch node's own success/error status and dropped that list. The
canvas seeds every planned node as pending from vertices_sorted; skipped
vertices then get no build_start/end_vertex, so they stayed stuck on pending
instead of rendering as inactive (the v1 build path marked them INACTIVE).
The translator now appends an inactive STATE_DELTA op per inactivated vertex,
and the frontend bridge maps the new inactive status to BuildStatus.INACTIVE
and tears its edges down like a completed node. Fixes the If-Else regression
in general-bugs-reset-flow-run.spec.ts.
ts-jest compiles with target es5; without downlevelIteration, [...set] and
for...of over a Set/Map emit ES5 that yields nothing. That silently broke the
AG-UI bridge tests: runningNodeIds spread, markRunningNodesFailed, and
restoreOriginalBuildStatuses all iterated empty. Production (Vite/SWC, modern
target) was never affected; only the ts-jest harness was. Fixes the 3 failing
jest tests on this branch with no other suite changes (4994/4994 pass).
A partial add_message re-fire (the agent path emits these at tool
start/end for a message it is still streaming) was treated as the
finalizer: it closed and tombstoned the id, so the post-tool answer
was dropped. Only a non-partial add_message finalizes now; state
defaults to complete, so payloads without properties are unchanged.
remove_message now purges a buffered message and tombstones its id,
so text the backend retracted is not flushed to the client later.
Parallel components stream tokens for different message ids interleaved.
The translator tracked a single open message: the first foreign token
closed the open message and tombstoned its id, so every later event for
it was dropped and its remaining text never reached the client.
Tokens for a message that cannot take the wire now buffer until the open
message genuinely ends (its add_message finalizer), then flush in arrival
order; complete messages landing mid-stream buffer the same way instead
of interleaving a second START. end/error drain all buffers before the
terminal event. The wire still carries at most one open text message, so
the stream stays AG-UI-conformant.
Port the 6 background-execution settings and background_backend_is_scaled into 1.11.0's RuntimeSettings group mixin; union the BACKGROUND_EXECUTION_SERVICE / TELEMETRY_WRITER_SERVICE additions in schema.py and deps.py.
Adopt 1.11.0's composed group-mixin Settings for lfx settings/base.py (all v2 settings already present in the new groups); port the Python 3.14 cors_origins ['*']->'*' fix into groups/security.py.
Integration tests failed twice in nightly run 27260425158 with pyleak
EventLoopBlockError - first Integration Tests 3.14, then 3.12 on the
rerun, each time in a different test. The blocking stack points at
refresh_models_dev_periodically: every app boot unconditionally starts
a lifespan task that immediately fetches https://models.dev/api.json,
so the request lands mid-test in whatever test happens to be running.
Under pyleak's asyncio debug instrumentation the fetch blocked the loop
0.797s against a 0.2s threshold. Whichever test draws the short straw
flakes - which is why it looked transient and moved between versions.
Add a LANGFLOW_MODELS_DEV_REFRESH env gate (default unchanged: enabled)
and disable it session-wide in the backend test conftest. Tests fall
back to the bundled static model lists, which is also deterministic.
Verified: with the gate set, app boot makes zero models.dev requests;
the previously failing integration test passes.
The test spawns a child via multiprocessing spawn context, which
cold-imports the full langflow package (plus coverage's multiprocessing
hooks in CI) before appending a single event. The 10s join timeout is
routinely exceeded on a loaded CI runner sharing 4 vCPUs with a second
xdist worker: in nightly run 27253229568 (Unit Tests - Python 3.12 -
Group 5) the test failed all 12 executions (5 reruns x 2 step attempts),
each rerun exactly 10s apart - the join deadline, not a product bug.
Raise the liveness bound to 60s (join returns immediately when the
child exits, so the passing case is unaffected) and kill the child on
timeout so a hung spawn can't leak into later tests.
Unit Tests - Python 3.12 - Group 3 has been failing at 98-99% on the
nightly: the job hits the nick-fields/retry 40-minute per-attempt timeout
and pytest is SIGTERM'd mid-test, so it looks like one flaky last test
when it is actually a deterministic timeout (and the internal retry plus
run-level retries can never succeed).
Root cause is twofold:
1. .test_durations was last regenerated ~May 2025 and covered only 2,219
of ~9,500 current unit tests, so pytest-split weighted 77% of the
suite at the 0.73s average. The expensive client-fixture tests
(test_webhook.py, test_login.py - each pays a full create_app +
lifespan boot per test, 60-120s late in a CI run) clustered into
group 3's tail, making it ~8-10 minutes slower than its siblings
(33:15 on 3.13, >40min on 3.12 which is additionally slowed by
astrapy disabling SSL connection reuse on Python 3.12.0-11).
The weekly store_pytest_durations workflow that should refresh the
file has been dying at the 6-hour GitHub job limit every week (serial
full-suite run no longer fits), so the file silently froze.
This regenerates the file from real per-test wall-clock measured in
the passing Python 3.13 nightly jobs (run 27246228762 attempt 1, all
5 groups, parsed from the -vv xdist logs: per-worker start-to-start
deltas). 9,508 tests now have measured durations; old entries are
kept where no new measurement exists. Simulated least_duration
split goes from one outlier group to 5 even groups (~24.6 min each)
with the >50s tests spread 1-2 per group instead of 7 in one.
2. timeout_minutes 40 -> 50 gives headroom for matrix-cell variance
(3.12 runs ~20% slower than 3.13) so a slightly slow cell degrades
gracefully instead of burning 2x40min and failing the whole run.
Follow-ups (not in this PR): fix store_pytest_durations to run with
xdist or split groups so it fits the 6h limit; investigate the in-worker
degradation that makes client-fixture boots cost 8-12s early in a run
but 60-120s after ~30 minutes.
The first 1.11.0.dev0 nightly failed its "Test Langflow Main CLI" step:
the fork bump's sync_bundle_lfx_pin.py re-synced every bundle floor to
lfx>=1.11.0, and PEP 440 sorts the nightly's 1.11.0.dev0 BELOW 1.11.0,
so the workspace-built bundles (whose metadata shadows the satisfiable
PyPI lfx-arxiv 0.1.1 during `uv pip install dist/*.whl`) reject the
branch's own lfx while langflow-base pins it exactly — unresolvable.
Floor at lfx>=X.Y.0.dev0,<(X+1).0.0 instead: X.Y.0.dev0 is the lowest
version PEP 440 admits in the minor line, so every devN / rcN / final
satisfies the floor while older lines and the next major stay excluded.
This closes the NIGHTLY.md activation-gate hole structurally — future
minor forks re-sync to a floor their own nightlies already satisfy.
- sync_bundle_lfx_pin.py: lfx_floor_spec emits the .dev0 floor
- port_bundle.py: mirrored _current_lfx_floor kept in step
- bundle pyprojects restamped via the script (arxiv/docling/duckduckgo/ibm)
- test_bundle_lfx_pin.py expectations updated (20/20 passing)
- NIGHTLY.md gate section annotated with the post-activation fix
uv.lock is unaffected (workspace lfx is recorded as an editable source
with no specifier — which is also why uv sync/lock passed in the same
job). The release.yml RC floor-relax sed still matches the new form;
now redundant but harmless. No bundle version bump needed: published
0.1.1 floors >=1.10.0.rc0, satisfiable by the whole 1.11 line.
The release-1.11.0 fork bumped lfx to 1.11.0 but left component_index.json's version field at 1.10.0. _read_component_index fails closed on exact version mismatch, so the bundled registry loads as None and the upgrade-gate tests fail (UpgradeFlowError: 'bundled component registry is empty or missing') across the LFX test matrix. That failure blocks the nightly's release-nightly-build job, so no canonical 1.11.0.devN is published and check-nightly-status blocks CI everywhere.
Surgical restamp: version -> 1.11.0 and recomputed sha256 (same orjson OPT_SORT_KEYS hashing and OPT_SORT_KEYS|OPT_INDENT_2 serialization as scripts/build_component_index.py). Entries unchanged. Verified the reader validation passes and the three failing tests go green.
* docs: record nightly→stable bundle cutover plan (gated on lfx 1.10.0)
Add src/bundles/NIGHTLY.md documenting why langflow-nightly currently
renames the bundles (lfx and lfx-nightly ship the same lfx/ import, so a
stable bundle would co-install both and collide) and the deferred cutover
(Approach A: canonical pre-releases; B: lfx as a bundle extra), gated on
stable lfx 1.10.0 being published to PyPI.
Also expand two docstrings in scripts/ci/update_lfx_version.py to state
the deeper install-conflict reason, not just the resolve failure. No
behavior change.
* feat(ci): nightly Approach A — canonical pre-releases, drop nightly bundles [DRAFT/gated] (#13529)
feat(ci): nightly Approach A — canonical pre-releases, drop nightly bundles
DRAFT reference implementation of the nightly→stable-bundle cutover documented
in src/bundles/NIGHTLY.md. Publishes the nightly under CANONICAL package names as
.devN pre-releases instead of separate *-nightly distributions, so the stable
lfx-* bundles resolve against a single canonical lfx (no dual-lfx install
collision) and no nightly bundle packages are produced.
- tag scripts (pypi/lfx/sdk_nightly_tag.py): count .devN against the canonical
PyPI histories instead of the *-nightly projects
- update scripts: stop renaming to *-nightly; set .devN versions; re-pin
inter-package deps to exact canonical dev versions; delete the bundle
rename/repin (update_lfx_dep_in_bundles, rename_bundles_for_nightly)
- release_nightly.yml: publish canonical pre-releases; remove bundle build,
dist-nightly-bundles artifact, publish-nightly-bundles job + its gate; verify
canonical names; main wheel glob dist/langflow-*.whl
- nightly_build.yml: drop the bundle git-add in the tag commit
- NIGHTLY.md: Approach A marked implemented + activation gate + A1/A2 follow-ups
Held as DRAFT: do not activate until stable lfx 1.10.0 is published AND the
nightly base is the next minor (release-1.11.0). A 1.10.0.devN core sorts below
1.10.0 and would fail the bundles' >=1.10.0 floor. Stacked on #13528.
(.secrets.baseline: incidental line-number shifts for the two workflows + prune
of pre-existing stale Pokédex Agent.json entries.)
* docs(ci): drop internal 'Approach A' label from nightly cutover comments
Comment- and docstring-only change across scripts/ci/* and the two nightly
workflows; no logic change. The two workflow edits stay single-line so
.secrets.baseline line numbers are unaffected. src/bundles/NIGHTLY.md keeps
its A/B decision-record framing intentionally.
* feat(ci): make nightly consumers work with canonical pre-releases
Follow-ups from the nightly cutover that are part of its blast radius (the
nightly now publishes canonical `.devN` pre-releases, not `*-nightly`
distributions):
- version.py: derive the "Nightly" label from the `.dev` version marker, since
the canonical `langflow`/`langflow-base` distribution matches first in the
lookup. Keeps the startup banner and telemetry `package` field identifying
nightlies. Adds a canonical-dev test; updates the base-dev assertion.
- ci.yml check-nightly-status: query the canonical `langflow` project and pick
the latest `.devN` release date instead of `langflow-nightly`'s `.info.version`.
- db-migration-validation.yml: install the nightly as `langflow[postgresql]==<dev>`
(pre-release) instead of `langflow-nightly[...]`; verify via version("langflow").
- src/lfx/README.md: nightly install is `uv pip install --pre lfx`.
- NIGHTLY.md: rewrite the follow-ups section (these are addressed; Docker image,
A2 meta-package, and website docs remain deferred by design).
The `langflowai/langflow-nightly` Docker image name is intentionally unchanged.
* fix(ci): correct nightly verify uv tree parsing + stale base-dep regex
Addresses review of #13528:
- release_nightly.yml LFX verify: `uv tree | grep lfx | head -n1` matches the
bundle `lfx-ibm` first → 'Name lfx-ibm does not match lfx'. Root the tree with
`uv tree --package lfx` so the first line is the lfx package itself.
- release_nightly.yml base verify: under canonical naming `langflow-base` prints
as a top-level `langflow-base v<ver>` line, so the old $2/$3 field parse read
name="v0.10.0" and version="". Use `uv tree --package langflow-base` and $1/$2.
- update_lf_base_dependency.py update_base_dep: regex only accepted ~=/==, so its
CLI entry point couldn't match the current root dep `langflow-base[complete]>=0.10.0`.
Add >= (parity with update_uv_dependency.py). The active nightly path uses
update_uv_dependency.py and was unaffected.
* docs(nightly): point NIGHTLY.md status at the activation gate, not draft state
Per review of #13528: this file ships inside #13528 (#13529 was folded in), so
the 'stacked on prep #13528' + 'held as a draft' framing is stale and misleading.
Reword the status block to state the real guard is the activation gate (stable
lfx 1.10.0 published AND next-minor base), not merge/draft state. Also reword the
follow-ups heading 'decide before un-drafting' -> 'decide before activating'.
A completed background job's GET status 500'd with 'No vertex builds found
for job_id'. The background build path differed from the sync path twice:
1. generate_flow_events minted a fresh run_id instead of using job_id, so
vertex builds were keyed by an id the status query never uses. Thread
run_id through _stream_event_frames -> generate_flow_events and pass
job_id from the background buffer so graph.run_id == job_id (the sync
path already does graph.set_run_id(job_id)).
2. The SSE build loop (build_vertices) only persisted builds when log_builds
was set and never passed job_id. Tie log_builds to job-tracked runs
(run_id present) and pass job_id=graph.run_id on the persist call.
Job-tracked runs also persist streaming terminal vertices so
reconstruction is complete; the live build path (run_id is None) keeps
its original behavior, so the v1 build path is unchanged.
Test: a real background run polled to completion, then GET status asserts a
reconstructed 200 (verified RED: 500 'No vertex builds found' before the
fix). Covers the non-streaming flow. v1 build path unchanged (35 build
tests pass); AG-UI suite 46 pass.
LE-1389
The v2 public endpoint only ran validate_flow_for_current_settings and
skipped validate_public_flow_no_code_execution, which the v1
build_public_tmp path applies. A public flow containing a Python
interpreter/REPL (or the legacy Python Code Structured tool, Smart
Transform lambda) was therefore an unauthenticated server-side
code-execution primitive (report H1-3754930).
Mirror v1: import the validator and call it right after the
public-access gate. PublicFlowValidationError subclasses
CustomComponentValidationError, so the existing handler already
sanitizes it to a 400 'This flow cannot be executed.' without leaking
the blocked component class names.
Add a non-mocking test that builds a public flow with a real
PythonREPLComponent and asserts the sanitized 400 (verified RED: returns
200 without the gate).
LE-1389
* fix(security): remove the disabled Python Code Structured tool component
Follow-up to #13538, which neutered PythonCodeStructuredTool to a
non-executable stub "for one release cycle, full removal later." This
completes that removal.
- Delete the component and its registration in lfx.components.tools.
- Drop its entry from the component index (num_components 355 -> 354,
sha256 recomputed surgically) and from stable_hash_history.json.
- Remove its 18 i18n keys from every locale file.
- Replace the dedicated stub unit test with a removal test in
test_dynamic_import_integration.py.
- Add a regressions entry to regressions/1.10.x.yaml.
The unauthenticated public-build RCE fix (report H1-3754930) is
unaffected: PythonCodeStructuredTool stays in
CODE_EXECUTION_COMPONENT_TYPES, so build_public_tmp still rejects any
saved or crafted flow that carries the type. instantiate_class execs the
node's stored `code` field regardless of whether the class still exists,
so the type-name block -- not the class -- is what closes the path.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* docs: document removal
* docs: typo
* Apply suggestion from @mendonk
Co-authored-by: Mendon Kissling <59585235+mendonk@users.noreply.github.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Mendon Kissling <59585235+mendonk@users.noreply.github.com>
* fix(bundles): bump lfx-* bundles to 0.1.1 to republish with corrected lfx pin
The published 0.1.0 artifacts on PyPI carry stale lfx pins from before the lfx 0.5.0->1.10.0 version realignment:
- lfx-arxiv, lfx-duckduckgo: lfx>=0.5.0,<0.6.0 (hard-broken; the <0.6.0 cap can never resolve lfx 1.10.0)
- lfx-docling, lfx-ibm: lfx>=0.5.0 (uncapped floor/cap mismatch)
The source pin was already corrected to lfx>=1.10.0,<2.0.0 in #13516, but the bundles were never re-published. PyPI versions are immutable, so shipping the fix requires a version bump. Bump all four to 0.1.1 so the Release Bundles workflow cuts fresh wheels carrying the correct pin. Root pyproject keeps its >=0.1.0 floors (satisfied by 0.1.1); only the bundle dist versions and their uv.lock stamps change.
* fix(release): relax bundle lfx floor for pre-release builds + idempotent bundle publish
The RC pre-release run builds lfx as 1.10.0rc0, but bundles floor lfx at >=1.10.0. Under PEP 440 a pre-release sorts below the final, so 1.10.0rc0 fails >=1.10.0 and the cross-platform install test cannot resolve the bundle wheels against the RC lfx wheel.
build-base/build-main/build-lfx already rewrite their inter-package deps to the pre-release version when pre_release=true; build-bundles was the only release artifact missing that step. Add it: when pre_release=true, rewrite each bundle's lfx floor to the exact pre-release version, keeping the wide <2.0.0 BUNDLE_API cap. Stable source stays at >=1.10.0 -- only RC wheels are relaxed, at build time, so no source churn or re-tag.
Also make publish-bundles tolerate 'already exists' duplicate wheels on rerun, matching release_bundles.yml.
* fix(security): block code-execution components on unauthenticated public flow builds (H1-3754930)
PythonCodeStructuredTool exec()'d its attacker-controlled `tool_code` field at
flow-build time. Because a PUBLIC flow can be built with no authentication via
POST /api/v1/build_public_tmp/{flow_id}/flow, that sink was reachable as an
unauthenticated server-side RCE — and other components (Python Interpreter/REPL,
Smart Transform) execute code on the same path, so removing one component alone
would not close the gap.
- Harden the public build path: build_public_tmp now rejects flows containing
code-execution components via validate_public_flow_no_code_execution(). The
check keys on the node `type`, so it holds regardless of the stored component
`code`, and is enforced ONLY on the unauthenticated public path —
authenticated /build is unchanged.
- Neuter PythonCodeStructuredTool to a non-executable compatibility stub: all
exec()/eval() sinks removed; build_tool returns a tool that raises a
deprecation error. The component stays registered with identical
display_name/inputs so saved flows still load and locale keys don't change.
It will be fully removed in a future release.
component_index.json (code_hash) is regenerated by autofix.ci.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* [autofix.ci] apply automated fixes
* fix: validate_public_flow_no_code_execution
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
* fix: serialize DataFrames in tool mode to prevent pandas truncation
When Memory Base or Knowledge components are wired as agent tools, the DataFrame
result was returned directly without serialization. LangChain would then stringify
this for the agent observation using pandas' default repr, which truncates all
cells to 50 characters (display.max_colwidth=50), inserting "...".
This breaks the F3 agent use case where agents need to reliably recall facts from
memory. The agent receives truncated content and cannot see the full text.
The fix serializes DataFrames through the existing serialize() function, which
converts them to list[dict] format with full, untruncated content. This maintains
consistency with how other result types (Message, Data, etc.) are handled in tool mode.
- Affects: Memory Base and Knowledge components used as agent tools
- Does not affect: Component-to-component wiring or normal (non-tool) execution
- Testing: Added tests verifying DataFrames serialize to list[dict] with complete content
Improved MB GUI component description
* chore: auto-bake note keys and regenerate backend locales/en.json [skip ci]
* [autofix.ci] apply automated fixes
---------
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
The hard_proof marker name was a vibe word that said nothing about what the
tests need. Rename it to real_services everywhere: the pytest marker
registration, the *_hard_proof.py test files, the Makefile target
(real_services_tests), the -m selector in migration-validation.yml, and the
CI job. real_services says what these tests require: real Postgres + Redis +
worker subprocesses. (integration was already taken for the external-API
suite under tests/integration.)
* refactor: remove getenvvar component and update locales
* test: add test for GetEnvVar component removal and update component index
---------
Co-authored-by: Janardan S Kavia <janardanskavia@Janardans-MacBook-Pro.local>
fix: use shadcn canvas colors for playground code blocks and chat-bg for the container
Standardize the playground tool-call surfaces on shadcn theme tokens so
they swap correctly between light and dark mode:
- Code blocks (SimplifiedCodeTabComponent) use bg-canvas: light grey
(#F4F4F5) / dark black (#000) so they stand out as the code surface.
- The 'Called tool' container (ContentBlockDisplay) uses bg-background so
it matches the chat background in both themes instead of inverting to
solid black in dark mode (it was bg-primary-foreground).
Previously the container and code blocks were both bg-primary-foreground,
which rendered the whole section black in dark mode with no contrast.
* perf(telemetry): batched off-pool writer for transactions + vertex_builds
Adds TelemetryWriterService that buffers transaction and vertex_build rows
in memory and drains them in batched INSERTs via a dedicated AsyncEngine
(pool_size=1 for SQLite, 2 for Postgres, max_overflow=0). Retention is
amortized in a 60s sweeper instead of running on every insert.
Producers (log_transaction, log_vertex_build) enqueue instead of opening
a DB session, so telemetry traffic no longer competes with the
request-handling pool. Falls back to the legacy direct-write path via
LANGFLOW_TELEMETRY_WRITER_ENABLED=false.
Durability: in-flight rows spill to a diskcache.Deque per PID on shutdown
and are restored on next startup; orphan PID directories left by crashed
workers are adopted.
* perf(telemetry): tighten error handling and add coverage
Review feedback from pr-review-toolkit + silent-failure-hunter:
- Retention sweep snapshots the dirty-flow sets before commit and only
clears them after it lands; a crashed sweep no longer drops the flows on
the floor, so per-flow caps cannot drift unboundedly.
- Producer fall-through is no longer silent. transaction_service and
log_vertex_build each log a one-shot WARNING when telemetry_writer_enabled
is True but the writer is not running.
- Lifespan startup failure now logs ERROR instead of WARNING — if the user
opted into the writer and it didn't come up, that's an error.
- Shutdown drain timeout no longer suppressed silently: logs a WARNING with
pending row count + a hint to raise telemetry_writer_shutdown_drain_s.
- Writer's flush loop catches asyncio.CancelledError separately and
re-prepends the in-flight batch to the buffer so teardown's disk spill
catches it.
- After 6 consecutive batch failures the writer emits a loud ERROR with
buffer depths so operators see sustained data-loss risk.
Added tests:
- test_sanitization_survives_writer_round_trip
- test_retention_failure_preserves_dirty_flows
- test_in_flight_batch_returned_on_cancel
* perf(telemetry): address copilot review
- _restore_from_disk + _adopt_orphan_outboxes now route through _enqueue so a
large disk-spilled or orphan outbox can't bypass telemetry_writer_max_queue
and OOM the process. Oldest rows are dropped and counted via the existing
dropped_transactions / dropped_vertex_builds counters.
- chmod 0o700 the outbox root + per-PID directory so sanitized-but-still-
sensitive payloads aren't exposed cross-user on multi-tenant hosts.
Suppressed on platforms where chmod is a no-op (Windows).
- Added test_adopt_orphan_outboxes_honors_max_queue.
The remaining two copilot notes (private API access to DatabaseService and
diskcache.Cache.close) were also flagged by the in-tree review; tracking
separately. The "diskcache not declared" note is a false positive — the
dependency is at src/backend/base/pyproject.toml:76.
* perf(telemetry): address coderabbit review
- Sweeper hands off dirty sets via capture-and-clear so concurrent
flushes during a retention pass aren't wiped by the post-commit
subtract; failure path restores the snapshot.
- Stress README uses a concrete Postgres DSN matching the docker
example instead of an unset env var.
- Test PID-probe loops bounded via a shared helper with pytest.fail
fallback.
* perf(telemetry): swap diskcache outbox for stdlib sqlite (CVE-2025-69872)
Replaces the diskcache.Deque-backed spill outbox with a stdlib sqlite3
outbox (WAL mode, JSON payloads in TEXT). diskcache 5.6.3 has
CVE-2025-69872 (pickle deserialization RCE for an attacker with write
access to the cache dir), no fixed version released, and was never
declared as a dependency — import failed on Python 3.10.
The replacement is encapsulated in a small _Outbox helper that owns the
connection, schema, JSON codec, and lifecycle. The codec uses tagged
wrappers for datetime and UUID so SQLAlchemy's typed columns accept
restored payloads on the way back out.
Hardening informed by a survey of OTel collector, Fluent Bit, Vector,
Prometheus remote_write, Datadog agent, Logstash, and Filebeat:
- Per-PID outbox dirs now stamp an owner.json (hostname + Linux boot_id
or time()-monotonic() proxy). Adoption only proceeds when host+boot
match; cross-host or pre-owner-file dirs are logged and skipped so a
recycled PID after a container restart cannot pull in a stranger's
spill data.
- PRAGMA synchronous=FULL on the spill connection so the shutdown
commit hits the platter (NORMAL only fsyncs on WAL checkpoint, which
may never run if the process exits immediately after commit).
- Spill honors telemetry_writer_max_queue with drop-oldest, matching
the producer-side overflow policy so a backlogged buffer at shutdown
can neither stall teardown nor fill disk.
- append_all encodes payloads up front and only clears the deque after
the transaction commits — no more partial-drain on mid-flight SQLite
failure. drain collapses to a single DELETE FROM outbox after the
SELECT.
- Exception handlers at the spill/restore/adopt boundaries narrowed to
(sqlite3.Error, OSError) so genuinely unexpected exceptions
propagate rather than being silently logged.
Tests cover: cross-host orphan skipped, missing-owner orphan skipped,
spill cap drops oldest, and a realistic UUID+datetime payload
surviving spill → restore → SQLAlchemy INSERT.
* [autofix.ci] apply automated fixes
* fix(telemetry): name wait_for inner tasks so pyleak can filter
Integration tests using @pyleak_marker were flagging an inner asyncio
task created by the telemetry writer's `wait_for(Event.wait(), ...)`
loops. The wrapper task is auto-named (Task-N) and lands mid-await
across the test boundary, so pyleak counts it as leaked. The writer
itself shuts down cleanly via teardown; the "leak" is a pattern
mismatch with pyleak's per-test snapshot model, not a real lifecycle
bug.
Wrap Event.wait() in an explicitly named task (`telemetry-writer-tick`,
`telemetry-writer-backoff`, `telemetry-sweeper-tick`) via a small
_wait_or_shutdown helper, and extend pyleak_marker with a task_name_filter
that excludes `telemetry-*` from leak detection. CI was green on earlier
commits in this branch only because the diskcache import error
prevented the writer from starting at all — fixing the import surfaced
this latent pyleak collision.
* [autofix.ci] apply automated fixes
* [autofix.ci] apply automated fixes (attempt 2/3)
* test(telemetry): disable writer in tests so reads see writes synchronously
* perf(telemetry): age out cross-host orphan outboxes on shared volumes
A pod on a shared PV (NFS/RWX) cannot adopt rows from a dead pod on a
different host because the dead pod's hostname doesn't match. Without a
janitor those directories accumulate forever. Add an owner-file mtime
heartbeat from the sweeper plus a prune pass that deletes cross-host
outboxes whose owner file hasn't been touched within
telemetry_writer_orphan_max_age_s (default 1h). Same-host orphans still
flow through the existing adoption path.
* perf(telemetry): add byte-aware flush + drop strategy
Today the writer bounds memory by row count only, so a worker logging fat
vertex_build artifacts can hold tens of MB per row and silently dwarf the
configured max_queue. Add a size_strategy switch ('count' | 'bytes' |
'either', default 'count' for parity) with two byte thresholds:
- batch_size_bytes (256KB) caps per-flush INSERT size when bytes apply
- max_queue_bytes (200MB) drops oldest by bytes when bytes apply
Parallel sizes deques mirror the payload buffers so accounting stays
consistent across drain, spill, restore, and the cancel/retry rollback.
Single rows above the byte budget are still emitted (no row is refused);
operators get dropped_*_bytes counters alongside dropped_*.
* test(telemetry): tighten byte-strategy assertions and cover end-to-end paths
Address gaps in the byte-aware strategy tests:
- pin exact dropped counts and remaining buffer state for the drop-by-bytes
case (was 'greater than zero')
- compute exact drain count from encoded size (was a 1-4 range)
- round-trip bytes through spill+restore to verify size deque is rebuilt
- round-trip bytes through orphan adoption for the same reason
- exercise the sweeper loop end-to-end to confirm heartbeat + prune are
wired in (previously only unit-tested in isolation)
Clarify in the settings docs that the byte caps measure encoded JSON size,
not Python in-memory footprint.
* ref: updates to telemetry writer PR (#13294)
* fix(telemetry-writer): harden error handling and add missing test coverage
Critical:
- teardown() now re-raises CancelledError after awaiting the writer task so the
asyncio cancellation chain propagates correctly on lifespan task kill
- suppress(OSError) on owner file write replaced with explicit error log so
operators know when disk-spilled rows will not be recoverable on restart
High / important:
- Escalation threshold check changed from == to >= so the error log fires on
every failure past the threshold, not just the 6th
- Dead `except OSError` on time.time() - time.monotonic() changed to
`except Exception` since time functions cannot raise OSError
- Removed redundant `from uuid import UUID` inside _run_retention_pass (already
imported at module top)
- Orphan directory cleanup: replaced blanket suppress(OSError) with per-child
suppress so ENOTEMPTY on an individual rmdir doesn't abort the whole loop and
leave the parent directory leaking silently; outer failure now logs at debug
Tests (3 new):
- test_retention_sweep_caps_vertex_builds_per_vertex: inserts 8 builds for a
single vertex with max_per_vertex=3 so the per-vertex DELETE subquery
actually executes (previous test used 8 distinct vertex IDs, bypassing it)
- test_either_strategy_trips_on_bytes_first: verifies bytes can be the first
trigger under 'either' strategy (previous test only covered the count-first path)
- test_writer_retries_on_batch_failure: injects 2 flush failures then success;
confirms failed_batches increments, rows are preserved in the buffer, and
flushed_rows reflects the final successful write
* ci: add stress-tests job to nightly build pipeline
Wires the stress-tests workflow into nightly so telemetry write stress
tests run automatically. Also gates release-nightly-build and
slack-notification on stress-tests result so a stress regression blocks
the nightly release and surfaces in Slack.
* ci: run stress tests in nightly without blocking release
Stress tests are informational for now — a failure is visible in the
workflow run and Slack but does not gate the nightly release or build.
* fix(telemetry-writer): address PR review on cancelled teardown and nightly stress tests
- Add the stress-tests.yml reusable workflow (was untracked, so the
nightly stress-tests job could never resolve)
- Notify Slack on stress-tests failure (add to slack-notification gate
and FAILED_JOB detection as non-blocking)
- Run sweeper cancel, disk spill, and engine dispose in a finally so a
cancelled teardown() still persists the in-memory buffer
- Add tests for the cancelled-teardown spill path and the >= escalation
threshold; drop the misleading wait-loop in the retry test
---------
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
Co-authored-by: Jordan Frazier <122494242+jordanrfrazier@users.noreply.github.com>
* fix: commit slider value before node selection consumes the interaction
Sliders inside React Flow nodes (e.g. the URL component's Depth field) lost
the value the user set when the node was not yet selected. React Flow selects
a node on click and pans/drags it on pointer down, while Radix drives the
slider with the same pointer events. The interactive SliderPrimitive.Root had
no React Flow isolation, so the first interaction on an unselected node was
consumed by node selection: the slider reacted visually but the chosen value
reverted or snapped to wherever the pointer landed.
Stop pointer/click propagation on the slider root and add React Flow's
nodrag/nopan/noflow/nowheel opt-out classes (matching the slider's value-text
input). Radix composes the handlers, so value setting is unaffected.
Adds a regression test asserting slider pointer-down/click do not bubble to
the node wrapper while unrelated children still do.
* Update src/frontend/src/components/core/parameterRenderComponent/components/sliderComponent/__tests__/slider-node-selection.test.tsx
Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>
* [autofix.ci] apply automated fixes
---------
Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
* fix(i18n): translate missing frontend strings found in QA
- API Keys page: replace hardcoded "Never" with t("settings.apiKeys.never")
- New Flow screen: migrate all WELCOME_* string constants in
flow-builder-welcome.tsx to t() calls; remove string constants from
flow-builder-welcome.constants.ts (keep WELCOME_MAX_INPUT_LENGTH)
- Get Started sidebar card: add useTranslation and replace 5 hardcoded
strings (All Set, Get started, Star repo, Join the community, Create a
flow) with existing sidebar.* i18n keys
- en.json: add settings.apiKeys.never + 10 flowBuilderWelcome.* keys
- All non-English locale files (ja, fr, es, de, pt, zh-Hans) updated via
GP upload/download
* fix(i18n): translate additional QA-identified untranslated strings
- Translate canvas assistant banner ("Try the new Langflow Assistant!", "New" pill)
- Translate search placeholder flow type name (was showing raw "flows"/"mcp" in mixed-language)
- Translate UNNAMED tool badge in ToolsComponent and ToolsTable
- Extract all 61 flow default descriptions to i18n keys (flow.defaultDescription.0-60)
so new flows get a localized random description instead of always English
- Fix backend ja.json: "Vector Store RAG" → "ベクターストア RAG" for consistency
with the frontend welcome screen translation
- Run GP upload+download to populate fr, ja, es, de, pt, zh-Hans locales
* [autofix.ci] apply automated fixes
* fix(i18n): fix CI failures — Biome import order and updated test constants
- Fix Biome import sort in flow-builder-welcome.tsx (lucide-react after react)
- Update flow-builder-welcome test to use inline English strings instead of
importing the deleted WELCOME_* constants from flow-builder-welcome.constants
* fix(i18n): truncate long checklist labels with tooltip in get-started sidebar
Long translations (e.g. Japanese) were wrapping to two lines in the narrow
sidebar. Now each label truncates with ellipsis and shows the full text in a
tooltip on hover. Icons also get shrink-0 to stay fixed-size.
* [autofix.ci] apply automated fixes
* fix(i18n): align assistant banner translation with panel title in ja, es, pt
GP inconsistently kept "Assistant" in English in the banner string while
translating it natively in the panel title. Patched directly:
- ja: "Assistant" → "アシスタント"
- es: "Assistant" → "Asistente"
- pt: "Assistant" → "Assistente"
fr, de, zh-Hans were already consistent.
---------
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
