Merge main into major/agentic-chat (catch the base, through #866) - #867
Merged
Conversation
Documents, uploads and links showed as pills under the chat header while an attached knowledge base showed as a full-width bar above the composer, and a folder — which scopes the chat and is sent on every message — showed nothing anywhere. All three answer the same question about a conversation, so they share one row. The pill row was already the deliberate merge of chat uploads and file-browser selection; the KB bar was never brought along. This finishes that. - Type is icon + text tag + tint, never tint alone: --highlight-color is deploy-customisable and colour alone fails a colourblind reader. The KB and Folder tags preserve what the bar's "Knowledge Base: " prefix said. - Scope chips (KBs, capped at 3; folders) render first and always; the unbounded document tail collapses behind "+N more" past six, so a heavy library selection cannot push the knowledge base out of view. - FileBrowser reports selected folder titles alongside uuids, mirroring the document path, so a folder chip can name itself. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LKDBCBvedFaffXibXnFM2g
A step's tasks all receive an identical copy of one payload and run in parallel, but input was configured per task — so a two-task step asked the same question twice and let the answers disagree. The step editor gains Basic Setup / Input / Output tabs. Input is set once and shared by every task; Output describes what the engine actually does with the tasks' results rather than offering settings that do not exist. The task editor collapses to a single Design surface, keeping only what is genuinely per-task: what the task does, and post-processing of its own output. Different-source-per-task survives as a buried Advanced disclosure, badged OWN INPUT in the step's task list. Existing workflows are untouched: the engine falls back to a task's own keys when its step defines no input, and the step's Input tab seeds from the tasks rather than defaulting. Lifting a shared answer off tasks that disagree pins the odd ones out to the override instead of repointing them. Also makes a step's combined output deterministic — results were collected in completion order, so the same two-task step could emit [A, B] or [B, A] depending on which model answered first, with field_sources padded positionally against that list. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011tbi5K9tPm3DNKC5VJyXyQ
…tracts nothing The challenge step said "Include fields for: ... budget categories (personnel, equipment, travel, indirect costs) ...", so users typed "budget categories" as a field name and it extracted blank — extraction has no grouping; each category must be its own field. The chips below the sentence already showed the right names; the sentence contradicted them. Reword the step and star criteria 2 to name each budget category as its own field, matching the chips exactly, and say outright that a single "budget categories" field would extract nothing. The module objective in modules.ts repeated the same phrase and is reworded the same way. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011tbi5K9tPm3DNKC5VJyXyQ
…the previous content Refreshing a 2 CFR 200 Subpart source fetched eCFR's JavaScript shell instead of the regulation. The shell carried none of the phrases the bot-challenge and boilerplate gates recognise, so it passed them, and the refresh replaced 208 chunks with 1 — marked "Refreshed", green check. Chat then answered §200.414 questions from general knowledge. Phrase lists recognise the shells they have seen. The one thing known about a page that no phrase list can miss is the size of what is already indexed, so a refresh now refuses a page under a quarter of the retained text as not having loaded, and takes the failed-fetch path: previous content kept and served, source shows "Refresh failed — previous content kept: …" with the numbers, retrieval date unmoved. Prevents recurrence; does not restore a source already overwritten. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LKDBCBvedFaffXibXnFM2g
…feedback)
Six changes from a sit-down with a DGA working through the course:
1. Module 0 opens with the big picture — what AI is, why it's in research
administration, what the course delivers — before any terminology, and
teaches structured vs. unstructured data as the course's central idea
with no under-the-hood mechanics.
2. Jargon is defined in RA terms: JSON as "the fill-in-the-form format you
read, never write", Token via attachment-size limits, and Module 1's
pipeline lesson drops chunking/embedding/ChromaDB for "it reads the text
and builds an index, like the back of a book".
3. Every module now carries a worked example — eleven new walkthroughs
grounded in the modules' own sample documents.
4. Time estimates show everywhere they were missing: journey-map total,
locked cards, and the open module's header ("your place is saved").
5. An explicit pop-out button opens the course in its own browser window
for a second monitor, and the five cryptic mode icons get descriptive
tooltips with the pin group visually separated.
6. The Module 1 lab now says the Run button stays greyed out until a
document is ticked in the file browser — the exact stuck point.
Point 7 of the feedback (wrong duplicate-name message on import) is a
backend bug, fixed separately.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011tbi5K9tPm3DNKC5VJyXyQ
An Explore import told a user she "already had a workflow with the same name" while her personal library showed no such row — and she was right to disbelieve it. The uniqueness scope counts more than the personal library: a teammate's team-shared workflow, her own workflow filed under the Team tab, and a workflow whose library bookmark was removed while the object (and its name) lives on. The flat "already exists in your library" message claimed all of these were sitting in front of her. The 409 now resolves the conflicting workflow and names the case — team library (hers or a teammate's), or existing-but-unlisted — and quotes the stored name's exact capitalization, since the match is case-insensitive and "budget analyzer" is invisible to someone scanning for "Budget Analyzer". Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011tbi5K9tPm3DNKC5VJyXyQ
…don't offer a File view that 404s
A KB built on the NSF PAPPG answered from text that held the Introduction,
Chapter II and Chapter XII and none of I, III, IV or V, behind a green
check; the inspector's File view answered {"detail":"File not found"}.
The document pipeline already records a partial conversion on the document
(ingestion_warnings: partial_ocr, sparse_text) and chat and the file list
say so. The KB did not: not the source row, the inspector, the export, or
source health — a fifth of the KB quality score — which scored a partial
document as fully healthy.
- Source row: amber incomplete icon + a line with the pipeline's reason,
the treatment a truncated URL source already gets.
- Inspector: banner above the text; export: ingestion_warnings per document
source; source health: "partial", not counted healthy.
- Read live from the document, so a successful re-extraction clears it.
- Inspector detail reports document_file: available / no_access / missing,
decided the way the download route decides it; the File view is offered
only when available, with the reason in the header otherwise.
Surfaces what the pipeline detected; does not detect more. sparse_text is a
40-chars-per-page floor and partial_ocr needs the converter to report it, so
a document missing half its chapters can still pass both.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LKDBCBvedFaffXibXnFM2g
The recipe called `build`/`update` without `--repo`, so the tool discovered the repository from the working directory. Under `git worktree` that is the worktree, not the checkout holding the graph, so a run from one built a second index there containing only the files that worktree had touched -- and then answered every query from it reporting `"status": "ok"`. Measured on a working copy: four worktrees carried indexes of 7, 27, 26 and 6 files against a real 967. A partial index returns 0 for anything it never parsed and a true 0 is indistinguishable, so `callers_of` on a live symbol reports no callers, which reads as "safe to change". The first `git worktree list` entry is always the main working tree, so one graph now serves every worktree. ORIG_HEAD is read from that same tree, since that is the tree being indexed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KBMRynZ8iqpC7JKvwdBw9m
The page noted that semantic search "needs an extra" but not what happens without it. It does not fail: `search` returns `"status": "ok"` with `"search_mode": "fts"` -- keyword matching under a semantic name. A query whose terms appear literally still looks right, so the degradation is invisible until a conceptual query quietly returns nothing, which reads as "there is nothing there". The install line now takes `[embeddings]` and the setup adds the separate `embed` step, since building the graph does not populate vectors. Three `search_mode` values are documented -- `semantic`, `hybrid` (also correct) and the degraded `fts`. Also documents two traps that belong to the tool rather than the recipe: `detect-changes` reads its diff from the same `--repo` path it reads the graph from, so from a worktree it reports on main; and a symbol the graph never indexed is answered exactly like one with no callers, which only 2.3.8+ annotate with a `confidence` field. The network note now separates the local model -- downloaded once from Hugging Face, then run on-device, sending no repository content anywhere -- from the optional remote backends, which stay off for the reason already given. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KBMRynZ8iqpC7JKvwdBw9m
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KBMRynZ8iqpC7JKvwdBw9m
`token_safety_margin` gains a rung between the tiktoken-is-exact check and `DEFAULT_TOKEN_SAFETY_MARGIN`. When a provider has counted the request, the margin becomes the measured ratio of its count to our local one, rather than the 1.5 that covers the worst content anyone measured. Deliberately a ratio and not a total. The margin is applied per component, `BudgetPlan.total_input_tokens` is a computed property summing those parts, and the planner recounts on mutated text inside its trim loop. A recorded total stops being true the moment anything is trimmed; a ratio stays approximately valid. Clamped at 1.0: a provider counting below tiktoken may only reduce over-inflation, never reclaim window. Rejected outright on a model-name mismatch, a zero baseline, a non-positive count, or a non-finite ratio, each of which falls through to the existing ladder — so an absent or broken measurement is the 1.5 guess, by construction rather than by a second code path remembering. The "estimated, not exact" warning stays quiet only when the measurement actually supplied the margin. Inert until a caller passes one. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HucxUJ7gaA6iKveF8aruW2
`_sized_for` re-derived the current model's safety margin from its name and config in order to restate a request in a candidate's units. That is correct only while every count comes from the same estimate-plus-default path. A provider-counted request carries a much tighter margin, and `token_safety_margin` cannot know that happened — it still answers 1.5. Dividing an already-tight number by a factor never applied understates the request by the whole difference, at the exact point the router is deciding whether a candidate can hold it. That is the #648 defect — an estimate that reads low, so the request hard-fails — relocated to the routing boundary, which this function's own docstring says it exists to prevent. Callers may now state the margin their number was measured with. `choose_document_model` and `suggest_document_model` pass it through; the candidate's margin stays derived, because the candidate has not been counted. Omitted, every caller keeps today's arithmetic to the token. A supplied margin below 1.0, non-finite, or not a number at all is refused in favour of the honest derivation, matching how `context_budget._configured_margin` refuses the same thing. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HucxUJ7gaA6iKveF8aruW2
New `native_token_count`: fetches a provider's own count for a prepared request, or reports plainly that it could not. Two properties, and the second is the load-bearing one. A returned figure is the provider's own, unmodified. Anything going wrong returns nothing at all — the caller drops its safety margin on the strength of a usable result, so a count that comes back zero, short, or from a provider that counted half the request is worse than no count, converting a conservative over-estimate into a confident under-estimate. So: gated before any I/O on a protocol that can actually count, since `OpenAIModel.count_tokens` raises and every OpenAI-compatible model would otherwise decrypt a key and build a client per chat turn to learn that. Wrapped in a five-second timeout that is deliberately not the request's own settings — `build_thinking_model_settings` sets 120s and the shared client retries a 429 six times honouring Retry-After up to 60s, which would put a two-minute stall in front of every message. Blanket `except` by necessity: Anthropic raises ModelHTTPError, Google's genai errors are unwrapped, and a misconfigured protocol reaches OpenAIModel; there is no complete and stable union of those. CancelledError passes through, being the caller going away rather than a failure. Google's count omits the system instruction — pydantic-ai attaches `system_instruction` only for providers other than `google-gla`, and an api-key GoogleProvider is `google-gla`. `covers_system_prompt` says so rather than letting a caller silently drop a multi-kilobyte grounding preamble from its budget. Nothing imports this yet. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HucxUJ7gaA6iKveF8aruW2
Wires the native count into `chat_stream`. Before the request is sized, the provider is asked what the assembled prompt actually costs; the same payload is counted locally; the ratio becomes this turn's safety margin in place of the 1.5 guess. The payload counted is the whole user prompt, documents included. `_build_chat_prompt` is the existing assembly extracted verbatim so the pre-flight counts the request chat is about to make rather than a reconstruction of it — in document chat the documents are the payload, and they are also the digit-dense content the 1.5 was sized for. Sampling the question and history alone would measure prose and apply the answer to a budget table. The baseline covers exactly the components the provider counted, which differ by provider: Anthropic's count includes the system prompt, Google's does not for api-key providers. Counting it on one side and not the other would depress the ratio for a reason unrelated to tokenizer divergence, and a depressed ratio clamps to a margin of 1.0 over a tiktoken figure that itself under-counts these models — the hard-fail direction. Routing is given the margin the number was measured with, rather than re-deriving it. That value survives a model switch on purpose: it describes how `requested_input_tokens` was measured, and that does not change when the model does. A count for one model is not passed to the planner for another. `_native_margin` already rejects on a name mismatch, but relying on that silently is how the next person learns it the hard way. Anything unusable — ineligible protocol, timeout, 429, malformed response — leaves the stream exactly as it is today. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HucxUJ7gaA6iKveF8aruW2
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HucxUJ7gaA6iKveF8aruW2
`org_name` was doing two incompatible jobs. It is the word users read in conversation — the chat welcome, the input placeholder, "I'm your X assistant", the browser tab — and it is also an institutional field that leaves the deployment: `verification_service` stamps it as the institution in a creator credit built so attribution survives catalog export into other installs, and `export_import_service` writes it as `deployment_org`. No single value is correct for both. Set it to the institution and the assistant introduces itself as a university. Set it to the product and every catalog item the deployment contributes is credited to an institution that does not exist, somewhere its readers cannot tell it was a branding setting. `app_name` resolves as `app_name or org_name or "Vandalizer"`. Conversational surfaces read it; the footer, email, team naming, issued credits and every logo `alt` keep reading `org_name`. The fallback is the compatibility story: an install that already set `org_name` and leaves `app_name` empty is byte-for-byte unchanged. `isCustomized` now also trips on a customised `app_name`. That flag keeps the "Powered by Vandalizer" credit and the NSF acknowledgement visible under GPL v3, and without this someone could rename every conversational surface while never touching the field it watched. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HucxUJ7gaA6iKveF8aruW2
The sign-in, password-reset and invite-acceptance pages hardcoded Vandalizer's gold instead of reading the configured brand colour. A white-labelled install therefore greeted new users with its own logo above a button in someone else's brand — on the first screen anyone sees. Landing.tsx was migrated for exactly this reason; these surfaces were missed. Doing it safely needed a token that did not exist. `getAccessibleOnLight` darkens the brand colour until it passes contrast on white; nothing did the same for the near-black auth and footer surfaces. Without it, migrating a deployment branded to anything dark would trade a legible gold for an unreadable version of its own colour: RWU navy #163A64 measures 1.72:1 on #0a0a0a, against 4.79:1 for the lightened variant, hue preserved. Seventeen buttons paired the gold background with hardcoded `text-black`. Those move to the runtime-computed contrast colour in the same edit — changing the background alone is how black text ends up on navy. Deliberately left: the footer's "Powered by Vandalizer" credit, GPL link and NSF acknowledgement (the obligation is the words; only colour classes changed), two SSO buttons whose black text sits on permanent white, and one `hover:text-[#d49e00]` on a dark surface — `--highlight-hover` is a *darkened* brand colour and would read worse there than what it replaces, so a correct fix needs an on-dark hover variant that does not exist yet. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HucxUJ7gaA6iKveF8aruW2
`--color-panel-dark` was a fixed #191919 that nothing derived from anything, so a deployment branded to navy got navy accents on neutral-black chrome — the workspace tab strip, the document-panel header, the knowledge panels, the left rail. Derived now from the brand's hue and saturation with the lightness pinned. That pin is what makes it safe rather than reckless: contrast against the white chrome text stays predictable whatever an admin picks, where tinting neutrals freely makes contrast a runtime property of an arbitrary input. Navy #163A64 gives #0f1d2e at 16.99:1 on white, against #191919's 17.58:1. Saturation is capped so a saturated brand does not read muddy; a greyscale brand comes back neutral rather than faintly tinted. Thirteen hardcoded #191919 backgrounds move onto the token, so the chrome cannot end up half-branded — including LeftPanel's "Black header bar", which is the surface that prompted this. Unbranded deployments are unchanged. When highlight_color is still the shipped default the variable is *removed* rather than set, so the :root neutral wins. removeProperty and not an early skip: applyTheme runs for the cached theme and again for the server's, so an admin resetting a brand back to the default has to clear the value this same call previously wrote. Mirrors the _THEME_DEFAULT_COLOR guard in email_service.py, which protects branded email for the same reason. Left alone: both certification-badge gradients (the second stop is out of scope, and tinting one end gives a hue break), four text colours sharing the hex, and one 8-digit alpha form needing color-mix() rather than a swap. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HucxUJ7gaA6iKveF8aruW2
…odel requests A ValidationRun's model field recorded the requested model, but a model pinned in system or per-set extraction config silently won inside the engine — so a run could be labeled with a model that never executed, and the regression suite's "run under model X" measured the pinned model instead. Workflow and KB validation runs recorded no model at all. - ExtractionEngine.effective_model_info() resolves what extract() will run (model, provenance, configured temperature) without running it. - Extraction validation forces an explicitly requested model into the engine config (including two-pass per-pass models) so it actually runs, and persists the resolved model plus a model_settings snapshot. - Workflow validation attributes the models its graded WorkflowResults snapshotted at dispatch — a single label only when they agree. - KB validation records the answer-generation model (distinct from the judge), takes an explicit answer model that beats a KB config override, and pins the no-KB baseline to the same model so lift measures the KB, not a model swap. - ValidationRun gains model_settings so later edits to a model's live config can't rewrite what a historical score measured. Wiring tests verified red against the old behavior before the fix. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018f8shnU7x3BV6PiMZV928r
…omparison The suite ran every verified item synchronously inside the HTTP request (a catalog-wide sweep of LLM calls no gateway timeout survives) and returned rows into component state where a reload or a second run destroyed them. It computed no catalog-wide mean, never echoed which model produced a table, validated an item once per user who had added it, ignored the model parameter for KBs, and labeled rows with raw ids (LibraryItem has no .name). - RegressionSuiteRun document + tasks.passive.regression_suite Celery task; progress (completed/total, per-item rows) lands on the doc as the sweep advances. - POST /quality/regression-suite dispatches and returns run_uuid; new GET runs list/detail endpoints; endpoint map regenerated. - run_regression_suite dedupes (kind, item_id), forwards the model to KB validation (workflow validation grades historical executions, so no model can be forced onto it), resolves real item titles, and computes the mean over successfully validated items only. - Quality tab polls the running job, lists past runs as reopenable chips, and puts two completed runs side by side per item with a mean-vs-mean header - model A vs model B over the verified catalog. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018f8shnU7x3BV6PiMZV928r
…dation_runs indexes Fleet-wide, validation results could not be grouped by model at all -- the only per-model data was the single-item model_comparison drill-down. - get_quality_by_model(days) + GET /api/admin/quality/by-model: avg score, run count, items covered, per-kind split, grouped by the model that executed each run; runs with model=None reported as their own "(unattributed)" row rather than hidden. - Admin Quality tab renders it as a Model Performance panel, refreshed when a regression sweep completes. - QualityTimeline gains a "model changed" chip symmetric to the existing judge-model warning, so a score dip from swapping the model under test stops reading as a content regression (the history payload already carried the model; the component dropped it). - First indexes on validation_runs: (item_kind, item_id, created_at) and (model, created_at); the mgmt API model filter was a full collection scan. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018f8shnU7x3BV6PiMZV928r
…y what they measured Drift monitoring compared the pinned official_baseline_score against the item's latest (possibly unrelated) validation score; the frozen official_baseline dict was inert payload nothing executed, and last_drift_score's docstring described a re-run that did not exist. - reexecute_official_baseline(): replays an extraction baseline's frozen test cases against the current config via run_validation_v2 - source text recovered from the stored ExtractionTestCase or the examiner's inline addition, expected values from the frozen rows, the baseline's own num_runs replicated so the sample-size discount matches (fewer replicates would read as false drift). - Gated behind quality_config.monitoring.baseline_reexecution (default off - a re-run is real LLM spend per pinned item per day). Workflow and KB baselines cannot be replayed from the frozen dict alone and keep the proxy. - Every check records last_drift_basis (baseline_reexecution | latest_validation_proxy); the drift alert message names its basis; the Catalog coverage tab marks proxy deltas "(proxy)"; the model field's documentation now tells the truth for both modes, and the false "full re-run happens in drift monitoring" comment on pin_retroactive_baseline is corrected. - Metadata is reloaded after a re-execution before drift fields are saved, so the save doesn't clobber the quality scores the re-run itself just refreshed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018f8shnU7x3BV6PiMZV928r
…inned baselines The bundled catalog's gold/silver tiers are hand-typed literals sitting almost exactly on the items with no evaluation data (the 11 gold workflows have zero checks; the two workflows with checks are "fair" or untiered), and the badge rendered a hand-typed word identically to a measured score. - quality_asserted (tier present, no measured quality_score) exposed on the catalog listing, library listing, and coverage report. - QualityContractBadge renders an asserted tier in the neutral style as "Gold (asserted)" with a tooltip naming the provenance; the flag clears itself the moment a real validation run lands. - export_validated_seeds.py also exports a pinned official_baseline (+score) into seed files; seed_catalog.upsert_verified_metadata pins a shipped baseline as "catalog-seed" on install, refreshing its own prior pin but never overwriting one a local examiner pinned. - Deliberately does NOT author eval content - test cases and expected answers must come from measured runs, not the seed-file author. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018f8shnU7x3BV6PiMZV928r
… regression suite - The "Intelligence" bar beside every configured model (admin model list + chat model picker) was computed from the admin's own tier dropdown and thinking checkbox, rendered like a measured figure. It is now labeled "Tier" with a tooltip stating it reflects configured settings, not measured capability, and pointing at Admin - Quality - Model Performance for measured numbers. - The regression suite's model dropdown was fed by the superadmin-only system config endpoint, leaving staff-level admins with only "Default Model" and no way to target a model. It now uses the same non-privileged /api/config/models listing the chat picker uses. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018f8shnU7x3BV6PiMZV928r
Adds a workflow-level default model applied to every step that has no Model Override of its own. run_workflow / run_workflow_batch fall back to input_config.default_model before the user default; a per-step override still wins, and "Automatic (system default)" (blank) keeps prior behaviour. - backend: fallback in run_workflow and run_workflow_batch - frontend: "Model — all steps" selector on the Design canvas + type field - test: 5 unit tests covering the resolution precedence Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…on catch it Extractions and workflows failed against a vLLM model registered under its HuggingFace repo name — "Native structured output is not supported by this model." on every run — while chat worked and the same weights under a bare name were fine. JSON-schema output is a property of the server: vLLM enforces `response_format` via guided decoding for anything it serves. The profile answered for the model family's own hosted API instead — `Qwen/Qwen3-32B` resolved through OpenRouter's family map to a profile leaving `supports_json_schema_output` at its False default, and pydantic-ai refused the request. VLLMProvider now declares the capability for every model it serves, keeping each family's own schema transformer. The per-model "supports structured output" toggle now does something. It was written by the model editor and read by nothing on the extraction path, so switching it off — the obvious escape hatch here — changed nothing. The reason none of this was visible: the admin Test button ran one free-text completion, so it went green for the entire outage. It now makes a second, schema-constrained round trip using the same output-mode decision the extraction engine makes (shared as one function, so the diagnostic cannot pass on a configuration a real run fails on), and reports a model that chats but cannot do that as failing. "Connected" and "usable" are different claims. Two supporting fixes from the same incident: * An unsupported output mode is classified with its two real remedies instead of falling through to "read the raw error". * The Endpoint step reports the URL actually dialed, not the one typed — the vLLM and Ollama providers append `/v1` to a stored endpoint and the OpenAI one does not, so changing the protocol dropdown silently changed the URL. Separately, found while building a model to test that: the external-OpenAI branch passed `openai_client=` to a constructor taking only provider/profile/ settings, so any model added through the "OpenAI" or "Custom" setup preset raised TypeError before a request left the server. No test built a model. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PJ3BAa7MWo8CySegCumLEC
# Conflicts: # CHANGELOG.md
# Conflicts: # CHANGELOG.md
… own workflow bookmark
LibraryItem.find_one({item_id}) matched any bookmark of that object,
including a teammate's team-library row, so a workflow the user had
removed from their library could still be reported as listed in it.
Filter by kind and by who added it.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qk88vvuDRJz5Nft2pRNpeY
…tes up front mkdir -p runs before the tool's first build, so the tool may never write its own .gitignore into the directory; without this the flock lock file shows up as untracked. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Qk88vvuDRJz5Nft2pRNpeY
…tually breaks Only steps that route through the extraction engine use NativeOutput; LLM prompt steps answer in free text and keep working. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Qk88vvuDRJz5Nft2pRNpeY
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Qk88vvuDRJz5Nft2pRNpeY
…r Unreleased, not v4.12.0 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Qk88vvuDRJz5Nft2pRNpeY
getPanelDark on an empty or short highlight_color produced a non-colour, and var(--panel-dark, #191919) does not fall back for an invalid value, so the whole app shell painted transparent. Treat anything that is not #rrggbb like the default. Also files the CHANGELOG entry under Changed: the chrome following the brand is new behaviour, not a bug fix. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Qk88vvuDRJz5Nft2pRNpeY
…ntainers, trim the changelog claim compose.yaml carried two healthcheck keys on the celery service; docker compose config rejects the file outright (podman-compose's PyYAML kept the last key silently, which is why the live test passed). main already had a compose-level inspect-ping probe for celery, so the new block goes. compose_container used ps -a and took the newest match, so exec targets could land on an exited or one-off container; it now prefers a running one and falls back to any. The CHANGELOG entry no longer claims a latent Docker celery bug that main had already fixed, and notes that a stopped stack now counts as an existing deployment. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Qk88vvuDRJz5Nft2pRNpeY
's test block Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Qk88vvuDRJz5Nft2pRNpeY
feat(chat): one chip row for documents, folders and knowledge bases
Certification: start broad, cut jargon, worked examples, time estimates, pop-out window (DGA feedback)
…cation Duplicate-workflow-name error: say where the conflicting workflow actually is
…search-mode fix(review-graph): one shared graph across worktrees, and say that semantic search degrades silently
…d-config-probes fix(models): make vLLM structured output work, and make the Test button catch it
…nting feat(budget): measure the token margin against the provider instead of guessing 50% high
fix(branding): a white-labelled sign-in page shouldn't show our gold button
feat(branding): the dark app chrome follows the brand colour
fix: generic container-engine support (podman / podman-compose)
Brings the v5 branch up to origin/main: the 2026-09-08 PR batch (#795, #799, #800, #815, #857, #818, #821, #824, #866) plus the model-eval stack (#831-#838), #842, #859, #860, #796-#798, #801, #820. Eleven files conflicted; how each was resolved: - backend/app/database.py, chat_service.py: union of both sides' imports and document registrations. - LoginForm.tsx: the branch's "Need an account?" link to the v5 landing page, with main's --highlight-on-dark hover token. - ChatInput.tsx: the branch's document-aware placeholder, with main's appName instead of orgName. - LeftPanel.tsx: the branch's useWorkspace destructuring plus setSelectedFolderNames from #795. - modules.ts: main's broader Module 0 description (with "validate" kept), and both lesson sets at the two collision points — main's worked examples first, then the v5.0 insight lessons. - exercises.json, Landing.tsx: the branch's chat-driven lab instructions and v5 launch page (main's changes there were the orgName/appName rename, already covered). - AttachmentList.tsx: main's unified chip row (documents, folders, KBs), with the branch's processing/ready state on document chips ported on. - ChatPanel.tsx: the branch's agentic panel, with main's chip-row wiring (folder and KB props on AttachmentList), the old KB bar removed, and the unused Link2 import dropped. - WorkflowEditorPanel.tsx: main's step-level input/output restructure, with the branch's punctuation pass re-applied where the text survives (35 strings; 3 more that moved to StepInputSources.tsx). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Qk88vvuDRJz5Nft2pRNpeY
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Catch-the-base merge: brings
major/agentic-chatup toorigin/mainthrough #866 (22 first-parent merges since #856).What comes across
The 2026-09-08 batch — #795 (one chip row for documents, folders, KBs), #799 (certification rework), #800, #815, #857 (vLLM structured output + Test probe), #818 (provider-native token counting), #821/#824 (branding on dark surfaces and chrome), #866 (podman) — plus the model-eval stack #831–#838, #842 (workflow default model), #859 (dep bumps), #860 (CHANGELOG repair), #796 (step-level input/output), #797, #798, #801, #820 (app name vs org name).
Conflicts (11 files) and how each was resolved
backend/app/database.py,chat_service.pyLoginForm.tsx--highlight-on-darkhover token.ChatInput.tsxappName.LeftPanel.tsxuseWorkspacedestructuring plussetSelectedFolderNamesfrom #795.modules.tsexercises.json,Landing.tsxAttachmentList.tsxChatPanel.tsxAttachmentList), the old KB bar removed, unusedLink2import dropped.WorkflowEditorPanel.tsxStepInputSources.tsx).CHANGELOG merged clean:
[Unreleased]is byte-identical to main's (23 bullets), the5.0.0section is untouched.Verified on the merged tree
tsc -b: clean4997 passed, 161 skippedchat/,contexts/,certification/,auth/,workspace/: 278 passed; oneHomeExperiencetest hit the 5 s timeout on the first cold run of 42 files and passes on every re-run, so it's a cold-start flake rather than a merge effect.Owed after this lands
From the #839 review:
chat_tools.get_quality_info's per-itemmodel_comparisonnow duplicates main'squality_service.get_quality_item_detail(#833, same shape). Fold it into the shared helper, move thedemo_seedexclusion and(unattributed)label there, and addsource != "demo_seed"toget_quality_by_modelas #839 promised.🤖 Generated with Claude Code
https://claude.ai/code/session_01Qk88vvuDRJz5Nft2pRNpeY