Admin

Add a new person with name, initials, email, optional password, multiple roles, internal/external type, and optional talent flag/gender. Password defaults to firstname_temp123 if left blank, with mustChangePassword = true. No welcome email is sent — use "Send password reset email" to give them access.

Add an internal on-camera talent account. Always assigned only the Talent role. No email is sent on creation.

Grant a member any combination of roles. External members are blocked from elevated roles (Admin, Power User, etc.) server-side.

Mark a member as on-camera talent and set gender (male / female / either) for TAL character matching.

Switch a member between Internal (sees all clients) and External (restricted to assigned clients). Switching to External requires at least one assigned client.

For external members, pick one or more clients they are scoped to. All other content is hidden server-side.

Updates the member's display name and initials throughout the app.

Revokes or restores login access without deleting records. Deactivated members cannot log in.

Sets mustChangePassword = true — the member must set a new password on their next login.

Generates a single-use 1-hour reset token and emails the member: "Reset your Helicon Flow password." Prior unused tokens are invalidated.

Hard-deletes a member with no associated records. Returns a 409 error with a "Deactivate instead" suggestion if any scripts, reviews, production, or audit records exist.

Opens a settings modal to toggle a member's global notifications on/off, plus per-event in-app/email preferences. Events include "Talent availability break" (on by default for Producers, opt-in for everyone else) alongside the qual, script-status, video-status and comment events.

Checks whether an email already belongs to a team member before creating a new talent account. Returns existence, name, active status, and existing roles.

Creates a Talent member with a temporary password and a 72-hour setup link.

If the email collides with an existing staff member, adds the Talent role to their existing account instead of creating a duplicate.

Edit name, active status, gender, mustChangePassword, or email.

Generates a fresh 72-hour setup token and resends the setup email.

Add a client with name, initials, service type, production track, contact info, posting channel, allowed script formats, and notes.

Update any client field. Allowed script formats control what writers can choose for that client.

Soft-delete a client (hiding it from active lists) or reactivate an archived one.

Download all records for Scripts, Live Production, or AI Production as re-importable CSVs.

Bulk-create scripts from a CSV. Auto-detects format, maps status, groups scene rows by script name, skips duplicates and invalid client/writer references. Returns created count and per-row errors.

Bulk-create live production records, resolving client/producer/editor by name. Optionally links a script by name.

Bulk-create AI production records. Requires a matching existing script by name.

Download blank CSVs with correct headers for each type/format.

Lists recent field-change audit entries: date/time, entity type, field changed, old → new value, changed by. Loaded newest-first, capped at 500.

Narrow the log to one entity type (script, live production, AI production, client, team member, scheduler).

Limit entries to a from/to date window.

One list feeds two rules: reaction phrases (a Line ≤6 words containing one makes the row a reaction scene) and stage directions (a Line ≤4 words with no sentence punctuation matching one is non-speaking). Add a phrase with its kind — Reaction + direction, Reaction phrase, or Stage direction — or remove one. A row containing real dialogue is never affected. VO rows are always speaking, and the "VO:" tag itself is not treated as a spoken word. A Line with no letters or digits at all ("...", "…", "—") is always a reaction beat — no phrase needed.

Auto-build quiet period (minutes), head and tail pads (seconds), reaction take cap (seconds), XML timebase, the rough-cut join offset (frames), voice cross (frames), reaction join offset (frames) and completeness share (0–1), the audio channel count assumed for a file whose properties could not be read, the source channel that carries the voice, and the default clip path-prefix template (blank = bare file names; {file} stands for the file name). Per-project prefixes are set through the API setting path_prefix_by_client. Rough Cut v2: the placement mode (takemap — the model reads every clip, the default; state — the on-set cue state machine; legacy — the older dialogue-first matcher), the take-map model id (e.g. claude-opus-5; changing it needs no code), the take-map spend cap per build (USD; a clip over the cap falls back to the state machine), the pinned sequence width and height (1080 × 1920 by default), the whole-take cap (seconds; a longer whole-take fallback becomes a labelled hole), and whether the download also includes the 02 ALL TAKES sequence (off by default). Rough Cut v3: the join style — straight (the default: every join in the rough cut is a straight cut) or jl (the J-cut / L-cut sound offsets chosen by listening to the handles). Rough Cut v3b: the eye's span check mode — corroborate (the default: a take nobody is mouthing becomes a director feed only when its character is not the clip's bound on-camera character), strict (any such take — the earlier rule) or off. Rough Cut v4: the judge — strict (the default: the words check is the gate, nothing changes) or judge (one model pass reads every real delivery in the story and the room and may place a variant, verbatim words and reason shown, one-click revert). Rough Cut v6: two more small model passes run in both modes inside the same per-build cap — the script read (reaction beats and VO carriers) and the amended script (the set's cut / amended / added decisions with quotes); no new setting. Values are validated: caps and sizes must be non-negative, sizes whole pixels between 16 and 16384, the cap at most $500, the mode one of the three, the join style one of the two, the span check mode one of the three, the judge one of the two, the model a plain model id. Visual index: on/off (off = zero vision calls), the vision model id (claude-haiku-4-5 by default), one frame every N seconds (0.5–60), the spend cap per clip and the spend cap per build for on-demand indexing (each at most $500).