Daily dogfood: fast-dash

[dkedar7]

Tracking issue for the nightly dogfood routine — a clean-room adopter that installs fast-dash from PyPI, follows the README + docs.fastdash.app verbatim (web UI and the MCP agent surface), and files anything a real user would hit.

One heartbeat comment per run: date + bug #N / feature #N / already-known #N / clean run.

[dkedar7]

2026-06-24: bug #99 (blocker — backend="fastapi" never starts a server via the documented @fastdash/FastDash.run() script path; web UI + /mcp both dead), bug #100 (major — agent can't read back inputs it set; dash://components/get_dash_component don't reflect set_input, contradicting docs' "current values"; + minor set_inputs values vs invoke inputs naming nit and DynamicDash(port=...) kwarg leak). Clean-room install (fast_dash 0.3.0 / dash 4.3.0 / py 3.11.15) and the default Flask backend quickstart + full MCP surface (set_input/set_inputs/invoke/get_invocation/list_component_types/get_dash_component, dash://layout+components, DynamicDash set_form, DataFrame return) all worked.

---

Generated by Claude Code

[dkedar7]

2026-06-25: bug #102 — MCP agent-facing tool schemas (invoke/set_inputs/set_input) are generic untyped objects; type hints don't build the agent schemas, contradicting the ai_agents.md claim. Clean install (fast_dash 0.3.0, dash 4.3.0, mcp 1.28.0). Verified working this run: quickstart boot, MCP drive surface (set_input/set_inputs/invoke/get_invocation/list_component_types/set_form + helpful known_ids errors + atomic invoke validation), DynamicDash set_form→invoke flow, browser parity both directions (agent→browser ~0.5s drain; human edit→agent invoke), input type-hint table, set_form across all advertised types, stream=True on Flask. Re-confirmed already-known: #99 (fastapi backend), #100 (headless read-back).

---

Generated by Claude Code

[dkedar7]

2026-06-26: bug #106 (major — describe_app() doesn't reflect a DynamicDash form built by the set_form tool; the agent-generated form is drivable via set_inputs/invoke but undiscoverable through the documented "Start here" contract tool — state.pending_specs is ignored, only state.inputs is mirrored). Clean-room install fast_dash 0.3.1 / dash 4.3.0 / mcp 1.28.0 / py 3.11. Verified working this run: quickstart boot (GET / 200, Textarea renders), full MCP drive surface (describe_app w/ type+default+current_value, set_input/set_inputs/invoke/get_invocation/list_component_types, helpful known_ids error on bad id), DataFrame return over invoke (no crash, session stays alive), native dash://layout+dash://components+get_dash_component. Re-confirmed FIXED: #99 (backend="fastapi" boots as a script, /mcp serves + drives over the ASGI backend) and #100 (describe_app reflects set_input headlessly). No cp1252 risk in fast_dash's own print/log strings. Browser UI driver not exercised (headless HTTP/MCP path used).

---

Generated by Claude Code

[dkedar7]

2026-06-27: bug #110 (major — single-select str-with-list-default dropdown seeds its MCP current_value as the whole options list: describe_app self-contradicts (type: string vs current_value: [..]), and invoke() with defaults passes a list to the str param while the browser holds null — breaks the contract-correctness + human↔agent parity promises; same root cause visible for list/dict defaults). Clean-room install fast_dash 0.3.2 / dash 4.3.0 / mcp 1.28.1 / py 3.11. Verified working this run: quickstart boot (GET / 200, Dash app), full MCP drive surface (describe_app w/ type+default+options+current_value, set_input/set_inputs/invoke/get_invocation/list_component_types, helpful known_ids error on bad id), describe_app now surfaces options for Literal/dropdown, DataFrame return over invoke (no crash, session alive), native dash://layout+dash://components+get_dash_component, GET /mcp 200 (no 405/hang). Documented limitation holds: multi-function + mcp_server=True emits a clear UserWarning and the /mcp MCP client connect fails (surface effectively absent). No cp1252 risk in fast_dash's own banner/log strings. Browser UI driver not exercised (headless HTTP/MCP + _dash-layout truth used for parity).

---

Generated by the nightly dogfood routine.

---

Generated by Claude Code

[dkedar7]

2026-06-28: bug #116 (major — describe_app() leaks a raw <fast_dash.utils.depends_on object at 0x...> repr as the dependent input's default for the README cascading-inputs pattern, and never surfaces the dependent dropdown's options (stays null even after the parent is set), so a headless agent following only the documented "Start here" contract can't drive a depends_on input; same contract-correctness class as #110/#106; bundled lower-pri notes: dict-default multiselect keys not surfaced as options, and set_input accepts out-of-options values silently). Clean-room install fast_dash 0.3.3 / dash 4.3.0 / mcp 1.28.1 / py 3.11.15, real streamablehttp_client to /mcp. Verified working this run: quickstart boot (GET / 200, Dash app), full MCP drive surface (describe_app w/ type+default+options+current_value, set_input/set_inputs/invoke/get_invocation/list_component_types, helpful known_ids error on bad id, atomic invoke rejects bad key without mutating state, partial set_inputs errors), DataFrame return over invoke (no crash, session alive), Plotly Figure return, GET /mcp 200 (no 405/hang). Re-confirmed FIXED: #110 (str/list/dict/range defaults seed null current_value, slider browser value is also null so invoke-with-defaults matches a UI Run — parity holds). Re-confirmed FIXED: #99 (backend="fastapi" boots as a script via FastDash(...).run(), server: uvicorn, GET / and /mcp both 200, MCP drive surface works on ASGI). DynamicDash set_form→describe_app→set_inputs→invoke flow works (#106 fix holds). Minor: describe_app type field is the Python/JSON type for static FastDash ("integer") but the component name for DynamicDash ("Slider"). No cp1252 risk in fast_dash's own banner/log strings. Browser UI driver not exercised (headless HTTP/MCP + dash://layout truth used for parity).

---

Generated by the nightly dogfood routine.

---

Generated by Claude Code

[dkedar7]

2026-06-29: bug #120 (moderate — describe_app() omits a Slider's min/max/step for static FastDash apps, so the contract is indistinguishable from an unbounded number box; an agent can't discover a slider's range and set_input/invoke accept out-of-range values (99999, -50) the UI slider physically can't produce — a human↔agent parity gap in the same class as the fixed #110/#116/0.3.4 options-validation. Asymmetric: DynamicDash sliders do surface props:{min,max}, static ones don't). Clean-room install fast_dash 0.3.4 / dash 4.3.0 / mcp 1.28.1 / py 3.11.15, real streamablehttp_client to /mcp. Verified working this run: quickstart boot (GET / 200, Dash app); full MCP drive surface (describe_app, set_input/set_inputs/invoke/get_invocation/list_component_types, get_dash_component, dash://layout+dash://components); 0.3.4 options-validation rejects out-of-options values with allowed options error; unknown-id error returns known_ids; Plotly Figure return over invoke (no crash); full README type-hint table renders correct widgets (Literal→Select, Enum→Select, list→Select, list-of-str→MultiSelect, dict→MultiSelect, date→DatePickerSingle, bool→Checkbox) and describe_app surfaces their options; DynamicDash set_form→describe_app→set_inputs→invoke flow; GET /mcp 200 (no 405/hang). Re-confirmed FIXED: #99 (backend="fastapi" boots as a script, GET / 200 server: uvicorn, MCP invoke works over ASGI). Minor (noted, not filed): static-app Enum input reports default: null while Literal reports its default. No cp1252 risk in fast_dash's own banner/log strings. Browser UI driver not exercised (headless HTTP/MCP + dash://layout truth used for parity).

---

Generated by the nightly dogfood routine.

---

Generated by Claude Code

[dkedar7]

2026-06-30: bug #126 (minor — describe_app() mis-reports Enum input defaults: a plain enum.Enum param reports default: null (should be the member's .value, e.g. "choco"), while the equivalent Literal and the str/IntEnum subclasses report theirs correctly; and an IntEnum reports default/options as ints (2, [1,2]) but current_value as the string "2" — the value the live dmc.Select actually emits — so the contract self-contradicts and an agent building invoke from default/options passes an int where a UI Run passes a string. Same contract-correctness class as the fixed #110/#116/#120; root cause: _describe_static_inputs only fills default for str/bool/int/float, so an Enum member falls through, and the str/IntEnum coincidence skips the str(member.value) coercion the component applies). Clean-room install fast_dash 0.3.5 / dash 4.3.0 / mcp 1.28.1 / py 3.11.15, real streamablehttp_client to /mcp. Verified working this run: quickstart boot (GET / 200 Dash app, GET /mcp 200 no 405/hang); full MCP drive surface (describe_app, set_input/set_inputs/invoke/get_invocation/list_component_types, get_dash_component, dash://layout+dash://components); options-validation rejects strawberry with allowed options; unknown id returns known_ids; Literal default correct; DynamicDash set_form→describe_app(reflects form props)→set_inputs→invoke; DataFrame return over invoke (no crash, session alive). Re-confirmed FIXED in 0.3.5: #120 (static Slider props:{min,max,step}). No cp1252 risk in fast_dash's own banner/log strings. Browser UI driver not exercised (headless HTTP/MCP + get_dash_component reflection used for parity).

---

Generated by the nightly dogfood routine.

---

Generated by Claude Code

[dkedar7]

2026-07-01: bug #131 (minor–moderate — describe_app() type field is semantically inconsistent across app classes: JSON type ("integer"/"string") for static FastDash but the component name ("Slider", == tag) for DynamicDash, with no JSON type surfaced at all; entry key sets also differ. The docs' only describe_app example shows type as a JSON type, so generic agent code keyed on it silently fails to recognize a DynamicDash field. Same contract-correctness class as the fixed #110/#116/#120/#126; previously noted in heartbeats 06-28 & 06-30 but never tracked as its own issue). Clean-room install fast_dash 0.4.1 / dash 4.3.0 / mcp 1.28.1 / py 3.11, real streamablehttp_client to /mcp.

Verified working this run: quickstart boot (GET / 200 Dash app, GET /mcp 200 SSE, no 405/hang); full MCP drive surface (describe_app, set_input/set_inputs/invoke/get_invocation/list_component_types, get_dash_component, dash://layout+dash://components); bad-id → known_ids error; bool input checked wiring (0.4.0) round-trips over MCP and in the real browser; real headless-browser (Playwright/Chromium) human↔agent parity — an agent invoke() set the live browser state (checkbox + color + dropdown), a human UI Run read it back identically; multi-output Figure + DataFrame over invoke (summarized, session stays alive, no crash) on backend="fastapi" (uvicorn, GET / + /mcp 200 — #99 holds); DynamicDash set_form→describe_app(reflects props)→set_inputs→invoke, plus set_form validation (unknown type lists allowed types; missing name errors); 0.4.0 str-inference table verified (short→TextInput, >120→Textarea, multiline→Textarea, hex→ColorInput, list-default→Select); stream=True boots on both Flask and FastAPI backends; required-param (no default) → describe_app shows default:null and invoke returns a clean TypeError object without tearing down the session; steps mode + mcp_server=True warns loudly at run() ("supported only for single-function apps; ignoring…") and the /mcp surface is absent. Re-confirmed FIXED in 0.4.1: #126 (Enum defaults). No cp1252 risk in fast_dash's own banner/log strings.

---

Generated by the nightly dogfood routine.

---

Generated by Claude Code

[dkedar7]

2026-07-02: bug #132 — clean-room install of fast-dash 0.4.1 (PyPI; dash 4.3.0, mcp 1.28.1). Quickstart, static-FastDash MCP (describe_app/set_input/set_inputs/invoke/get_invocation/list_component_types + native get_dash_component/dash://layout/dash://components), DynamicDash set_form, backend="fastapi" extra, DataFrame/Figure returns, and all README snippets (mosaic tuple-return, Fastify, depends_on, steps, multi-function) all worked; multi-function warns clearly that MCP is skipped; wrong set_input id returns known_ids. New finding: describe_app() reports every Optional[T] input as type:"string" while the UI renders the real typed widget (Optional[int]→NumberInput, Optional[bool]→Checkbox) — a contract↔UI parity break in the same _json_type_name class as #126/#131 → filed #132.

---

Generated by Claude Code

[dkedar7]

2026-07-03: bug #134 — clean-room pip install fast-dash from PyPI → fast_dash 0.4.1 (dash 4.3.0, mcp 1.28.1, py 3.11.15, Linux). New finding: describe_app() reports default: null for a datetime.date input (a documented type-hint-table type, rendered as a date picker) despite a real default — the input's own current_value carries "2021-06-15" and the adjacent int/str inputs report their defaults, so the contract self-contradicts; datetime.datetime (tag Timestamp) has the same gap (and also drops the time component from its seed). Same _describe_static_inputs scalar-only-branch class as the fixed #126 (Enum) and open #132 (Optional[T]) → filed #134 (minor; drivable, contract-accuracy only).

Verified working this run (real streamablehttp_client to /mcp): verbatim README 30-second quickstart (@fastdash greet) boots — GET / 200, Dash app, banner strings pure ASCII (no cp1252 risk); full MCP drive surface (describe_app, set_input, invoke, get_invocation, tools list = describe_app/set_input/set_inputs/invoke/get_invocation/list_component_types/set_form/get_dash_component); date input drivable end-to-end (set_input/invoke round-trip a "2023-03-03" string, callback runs, session stays alive); depends_on cascade fully works over MCP (#116 holds) — after set_input(country=India), describe_app resolves the dependent state options to India's states, an invalid option (California) is rejected with allowed options, and invoke({}) produces "Karnataka, India"; GET /mcp DELETE→405 (session teardown, expected). Browser UI driver not exercised this run (headless HTTP/MCP + describe_app/current_value truth used for parity).

---

Generated by the nightly dogfood routine.

---

Generated by Claude Code