A green play and a flask said nothing about the one thing that separates them:
whether you answer the flow's chat and decision steps, or an LLM answers them for
you. The flask was actively misleading - "experiment" is the bias experiments'
word, so it pointed at a different feature.
Both buttons now keep the play, and a small badge carries the difference: a
person, or a robot. A gear was the first thought and is the wrong glyph - it is
the settings icon everywhere else in the product, so on a run button it reads as
"configure this run" rather than "a machine runs it".
They also sit next to each other now. Stop and resume used to separate them, so
each had to be understood alone - which is precisely the job a 17px badge cannot
do. Side by side they are one choice, and each is what makes the other legible.
The tooltips stop repeating the button and say who answers: "Run - you answer the
human steps", "Run simulated - an LLM answers the human steps". That is the
sentence the icon can only gesture at.
557 frontend tests green.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Starting a fresh run means retyping values you already typed. The panel now
offers the other runs of this execution's own group - the group is keyed by
source flow, so a run of a different flow is never offered: matching its inputs
by name would be a coincidence, not a copy.
The copy fills the panel's pending edits rather than writing anything. That is
the whole trick: the values arrive with the unsaved-change styling already on
them, the user reviews them, and the single Save sends them in one bulk request.
No new write path, no new error handling.
It is deliberately partial, and says so. A file input holds a temp file on the
server, which the backend copies by reference when it reruns an execution; from
here there is only a path, and copying it would point this run at another run's
upload. Credentials are not copyable at all - the vault decrypts them
server-side. Both are reported in the summary rather than dropped quietly,
because a copy that silently leaves gaps is worse than one that names them.
Worth knowing before reaching for this: a rerun, bias rerun included, already
arrives fully populated - createBiasRerun calls copyReusableInputs, which carries
over every node input, the global descriptors, the authorizations and the
simulation descriptor. This is for the case that copies nothing: a new run
created from the flow.
An input already matching the source is not offered as a change, since it would
join the unsaved count and invite a pointless save.
555 frontend tests green; initial bundle unchanged at 2.88 kB over budget.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The Bias impact tab is empty because no experiment has ever run - zero rows in
bias_impact_report_entity - and nothing in the UI said how one is started.
The empty state now explains what a report is, says how many nodes carry a probe
that can be activated on this run, and offers the action. When something is in
the way it states that instead: a subflow cannot be a baseline, an unfinished run
cannot be compared against a rerun, a flow with no activatable probe has nothing
to measure.
"Create biased rerun" no longer returns silently when no node qualifies - a
button that does nothing and explains nothing is indistinguishable from a broken
one. It now names the reason.
The per-node measure control is shown on interactive blocks instead of hidden.
ChatInteraction cannot be replayed in isolation, so the isolated experiment is
genuinely unavailable there - but hiding the control made an annotated node look
identical to an unannotated one. It is rendered disabled, and the tooltip points
at the full-flow route that does work. That is a rule of the domain, worth
stating rather than concealing.
Splitting the node scan out of biasRerunCandidates keeps this free: the count is
computed from the flow snapshot, while only the capability check needs the
network.
506 frontend tests green; each new assertion fails when its behaviour is undone.
Initial bundle now 2.86 kB over budget, up from 2.21 kB - reported, not raised.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The globals go in a single request, so a failure belongs to the batch and not to
any field in it. Repeating the same text on every input said one thing N times
and implied N separate problems.
One message now sits above the fields, and it can say what the per-field version
could not: the request was all-or-nothing, so none of the values were saved. It
clears on the next edit, so a stale notice cannot outlive the state it described.
Node inputs keep their per-input errors: each is its own request, so a failure
there is attributable and partial success is real. File uploads likewise.
489 frontend tests green.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The tally said "0 of 3" while ignoring the manual inputs below it, which are just
as required: the backend only reaches READY when every step is ready, so a step
missing its manual input blocks the start exactly as an unsatisfied global does.
It now counts the whole panel.
"Provided" is decided per input and passed down, rather than inferred in the
panel from a globals-only list. A global uses the backend's own
missingGlobalInputKeys; a node input is judged on its stored value, ignoring
unsaved edits - otherwise typing would make an input look satisfied before it was
sent. An empty list, or a list of blanks, does not count as supplied.
"Flow inputs" and "Manual inputs" are now "Global inputs" and "Node inputs",
which is what they are called everywhere else in the codebase and in the API.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Each input opened with a centred gradient badge whose largest text read "Flow" -
the same word on every global input, so the most visual weight carried the least
information - followed by a "Type:" line and a four-row textarea. A one-word
positionTitle got the same box as a CV, and three short answers filled the panel.
An input is now one line of chrome - name, a small type chip, and a dot saying
whether it is still required - over a field that starts small. Removing an item
from a list is an icon rather than a full-width "Remove" button, which at this
width used to push the field out of the panel.
Saving is one action for the panel. Per-input buttons meant a click and a round
trip each, with nothing to say how much was still unsaved; a sticky bar now
reports the pending count and saves them together. Each one still goes through
the same single-input request, so a failure is still reported against its own
input.
The panel also stops ignoring missingGlobalInputKeys, which the backend has been
sending all along: it now shows "2 of 5 provided" and marks exactly the inputs
that block the start.
The editor is one template instead of four near-identical copies - global and
node, each single and multiple. That was not the goal here, but writing the same
change four times is how those four drifted apart in the first place. First tests
for this component.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The direction badge always said "Bias". It derived the direction from
biasExecutionContext.activeBiasProbes, and that field is not on the bias context
at all - probes belong to a step view - so the lookup was always undefined and
every variant fell through to the bias branch. My own change one commit ago.
A persisted snapshot settles what the payload really is: bias and mitigation are
tracked in four separate collections - annotations per node and activated
subflows per container, one pair for each direction. The direction is which of
them are non-empty, so a variant can now correctly read as Bias, Mitigation, or
Bias + mitigation, including when an intervention was switched on for a whole
container subflow rather than per annotation. A variant with nothing recorded
says "unspecified" instead of guessing.
The model was wrong in a second way, with a second victim:
activeAnnotationIdsByNode is not sent either, so the bias highlighting on graph
nodes read an absent field and never lit anything up. It now combines both
directions.
The model is corrected to the real shape and the two absent fields are gone from
it, so nothing can quietly read undefined again. The derivation lives in pure,
tested helpers rather than inline, and the dev fake now splits activations by
direction like the real payload - flattening them hid this very difference in
development.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Nothing in the run history said whether a run was a bias variant. Worse, the
viewer header claimed "Bias variant" for *every* execution: it tested for the
presence of biasExecutionContext, which the backend sets on all of them,
defaulting to NORMAL. Only the mode distinguishes them. The same faulty test also
offered the bias comparison action on any plain rerun.
Each history row now carries a kind badge - Run, Rerun, Bias, Mitigation, or Bias
+ mitigation - coloured and with a tooltip saying what it means for the result. A
bias variant reads as a variant even when it is also a rerun, because carrying
probes is what changes how its output should be read.
The rows also stop showing raw uuids. A rerun names the run it came from by
number ("from #1") instead of repeating a 36-character id, the execution id is
gone from the row entirely - the viewer header owns it - and the group header no
longer prints the source flow id. "1 runs" reads "1 run".
The viewer header keeps only what changes how a result should be read: Simulated,
the bias variant, Subflow. Execution id, simulator descriptor, experiment id,
baseline and probe internals moved behind a Details toggle, closed by default.
Two Italian strings in that block are now English, like the rest of the app.
Cards are flat here too, matching the flows list: no gradients, no lift on hover,
a left accent for the selected run.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
An outcome payload is often a generated document - the local example is a full
rejection email - which pushed the execution graph down the page. The section now
collapses.
It starts open, because this is the flow's answer and it was invisible until a
moment ago. Collapsed, the header keeps the outcome codes, so the conclusion is
never hidden entirely: only the payload folds away.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A flow result is built only from *unconnected* outputs, so wiring the last block
into an End node moves its value out of the result and into the outcome payload -
which nothing in the UI read. The run then looked like it produced nothing at all.
There is already an execution in a local database whose outcome payload is a full
generated rejection email that was invisible for exactly this reason.
The run view now has an Outcomes section above the graph, listing each End the run
passed through: its code, its label, the step it came from, and its payload. An
End reached with no value says so, rather than showing an empty box - the two
cases mean different things.
A text payload renders as text rather than through the JSON tree. The tree would
have kept the line breaks but wraps strings in quotes, and the common case here is
a generated document, not a data structure.
This is the smallest of the options for the underlying gap: the data was already
persisted and already in the API payload, so nothing in the engine changed. The
gap itself remains - End cannot be attached as a pure ordering dependency
(canDependOnOtherNodes is false), so a single-exit flow still has to choose
between a labelled end and a value in `result`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
- Lower rete editor min zoom (0.35 -> 0.1) so wide flows fit on screen.
- Remove the Export action from the read-only subflow preview dialog,
since it's only ever opened from the execution view.
- Add a fullscreen toggle to the flow editor and task execution viewer,
placed next to the title; disables the "flows" sidebar tab and hides
the right assistant/validation panel while the flow editor is
fullscreen, and collapses the left sidebar if it was open.
- Allow dragging nodes in the read-only execution graph and persist
custom positions client-side in sessionStorage, keyed by execution id,
restored on reload/refresh.
An execution cannot start until every requiredAuthorizations entry is
satisfied. A key of the form LLMProvider::<provider>::authorization is
answered with a vault secret id chosen from a picker filtered by that
provider, or from a credential created on the spot; anything else keeps
the literal-value panel it had.
A provider key never falls through to the literal-value panel, whatever
the provider catalog says or fails to say, so the API key itself can no
longer be pasted as the authorization value. The catalog is therefore no
longer part of the gate: an outstanding requirement blocks the start on
its own, and a failed catalog read is a notice with a retry.
The authorizations PUT answers with the recomputed execution, so its
response replaces the execution in the store rather than triggering a
list refresh, and the backend's 400 message reaches the user. A settled
requirement stays on screen with the credential label and a change
action, and a banner above the graph names the missing providers when
the context aside is collapsed.
The gate itself is now pure - buildAuthorizationGate and
isExecutionStartable in execution-viewer.utils.ts - and covered by
tests, since the component cannot be instantiated under the test
environment.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Finishes the bias impact experiments plan (docs/bias-impact-experiments-plan.md
steps 7-12):
- Side-effect policy selector: reuse the .llm-warning visual language for the
external side-effects banner, add a REQUIRE_CONFIRMATION note.
- Full-flow compare ("Compare with baseline"): new bias-compare-dialog
(service + host) triggering compareBiasExecutions and opening the shared
report viewer; inline errors read from errors[].message/detail.
- Persisted reports: new bias-impact-report-list (list + detail in one view)
wired into a new "Bias impact reports" tab in task-execution-viewer;
404/403 on report detail show the same inline message on purpose.
- Canvas: annotation badge on generic-node (count, executable-probe
indicator, severity from the backend catalog); new
BiasComparisonViewStateService driving bias-active / downstream-changed /
routing-change highlighting on task-step-node and custom-connection, fed by
a highlightOnCanvas event from bias-impact-report-viewer wired in all three
places that render it; legend + "back to normal view" action in the canvas
toolbar.
- Fixed a bug where the bias variant context badge only rendered for
simulated executions.
- Fixed "Measure bias impact" to stay visible-but-disabled with an
explanatory tooltip while the baseline hasn't reached a final state,
instead of being hidden outright, per the §12 checklist.
- Added a Retry action to the compare dialog's error state for parity with
the report list.
- Added the end-to-end facade flow test (annotation -> capability -> isolated
experiment -> report) plus coverage for all new components/services.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Lucio Lelii