Skip to content
Switches and arming order

Switches and arming order

PULSE (Prescriptive Unified Lead & Social Engine) by Runink ships with every outward action switched off. This page lists each switch in one place: what it defaults to, what it does, and the order to turn them on safely. The detail behind each one is on Publishing controls and Users and access.

Set these on TIDE’s PulseControlPlane resource or its overlay, never on the live Deployment. The operator puts back anything changed on the Deployment.

Yes and no values

The keys do not all parse yes and no the same way, and the difference matters:

  • PULSE_APPLY_ENABLED arms only on the exact value true. yes, on, 1, and TRUE leave it off, and the server warns with PUBLISHCFG-AMBIGUOUS.
  • PULSE_PUBLISH_<CHANNEL>, PULSE_PUBLISH_AUTOMATIONS, PULSE_AUDIT_APPLY, PULSE_OUTBOUND_CALLS and PULSE_AI_CALLS accept 1, t, true, y, yes, on and 0, f, false, n, no, off, in any case. Anything else, including empty, counts as unset, which for the last four means off.
  • AUTH_GOOGLE_ONLY is on only for the exact value true. PULSE_GRPC_REFLECTION is on for true in any case.
  • SELF_HEAL_AUTONOMOUS and PULSE_OPPORTUNITIES_DISCOVERY default to on, so they read the other way round: they take the same tokens, and unset or unparseable leaves them on. Only an explicit no turns them off.
  • PULSE_JUDGEMENT is on unless set to off (any case).
  • LINKEDIN_COMPANY_PAGES is on for 1, t, T, true, TRUE or True; anything else, including yes, leaves it off.
  • PULSE_WHATSAPP_AUTOREPLY takes the same tokens and is off unless set to an explicit yes.

The switches

KeyDefaultEffect
PULSE_APPLY_ENABLEDoffThe global kill switch for every outward channel, the first half of the audit-apply gate, and what starts the daily site-audit engine. While it is off, nothing publishes, and approving a draft does not queue it at its planned date.
PULSE_PUBLISH_<CHANNEL>follows the kill switchNarrows one channel: BLOG, PODCAST, AUDIOBOOK, YOUTUBE, INSTAGRAM, SPOTIFY, LINKEDIN, EMAIL, WHATSAPP, SMS. Can only turn a channel off.
PULSE_PUBLISH_INSTAGRAMfollows the kill switchApproving (or a schedule or Autopilot reaching) an Instagram piece publishes it by what it is: a post (one photo, with caption, hashtags and alt text), a carousel (its rendered 1080×1350 slides), a Reel (the owner’s own video, attached in Studio; MP4/MOV, 3 s to 15 min, at most 300 MB and 1920 px wide, index at the start) or Stories (each attached photo or video, 3 to 60 s; no caption, and stickers cannot be posted through the API). A Reel or Story with no video of the owner’s gets one made by PULSE at approval (or on download), and so does a YouTube Short (9:16, at most 3 min, uploaded through PULSE_PUBLISH_YOUTUBE like any video; the owner’s own video goes first): the script’s lines drawn as frames in the account’s brand colours with burned-in captions; Instagram’s API publishes no stickers, so a Story’s stickers are drawn as still cards and listed on a checklist in caption.txt to add by hand (except a Mention sticker: the API mentions accounts without one), read aloud by ttsd (TTSD_URL; none = a silent video that says “no voiceover: speech service unavailable”), assembled by ffmpeg (H.264 + AAC, 30 fps, one at a time, 2 threads, 2 min per encode; Reels at most 90 s, Stories split into parts of at most 60 s); a server without ffmpeg says “video is not available on this server”. A piece without its media is refused for scheduling and withheld at approval with “needs a video / photo: upload one in Studio”. Instagram fetches every file from PULSE_PUBLIC_BASE_URL at an unguessable, signed address that expires within two hours; without that URL the channel is not ready. Before publishing PULSE reads the account’s content_publishing_limit; at the limit the post moves to after the 24-hour window with the reason. Mentions go in the caption as @name; each one is looked up again with business discovery right before posting and only a found (business or creator) account is sent in user_tags (a photo at its centre, a Reel or the first Story by username); one that cannot be looked up is not tagged. Needs the Instagram professional account connected with instagram_basic and instagram_content_publish (Meta app review); a missing permission is recorded as “reconnect Instagram to allow …”.
META_APP_ID, META_APP_SECRETunsetThe instance’s own Meta app (Facebook Login for Business). With both set, Settings › Integrations shows Connect with Instagram: the person signs in at Facebook, PULSE stores a sealed long-lived token, finds the Instagram professional account through their Page, and renews the token in its last ten days (a failed renewal is shown on the card). The app must allow the redirect PULSE_PUBLIC_BASE_URL/oauth/social/callback. Unset: the card says sign-in is not set up and accepts a pasted token instead.
LINKEDIN_CLIENT_ID, LINKEDIN_CLIENT_SECRETunsetThe instance’s own LinkedIn app. With both set (and an https PULSE_PUBLIC_BASE_URL), the LinkedIn card shows Connect with LinkedIn: the person signs in at LinkedIn (openid profile w_member_social), PULSE stores the sealed token and the member, and approved LinkedIn posts use that token. LinkedIn tokens last 60 days; PULSE renews them only when LinkedIn issued a refresh token, otherwise the card shows “reconnect before” the date. The app must allow the redirect PULSE_PUBLIC_BASE_URL/oauth/social/callback. Unset: the card accepts a pasted token instead. PULSE_PUBLISH_LINKEDIN and PULSE_PUBLISHER_USER_IDS still decide whether anything is posted.
LINKEDIN_COMPANY_PAGESoffSet true only once LinkedIn approved the app for the Community Management API. The card then offers Connect company pages, which also asks for r_organization_social, w_organization_social and rw_organization_admin, and lists the pages the member administers to post as.
PULSE_PUBLISH_EMAILfollows the kill switchSend on a lead’s email draft (SendOutreachEmail): one email, sent when the user taps Send, from that user’s own connected mailbox (Gmail, Microsoft 365 or SMTP, see Connections). There is no shared sender, so PULSE_PUBLISHER_USER_IDS does not apply. Off, a send is recorded withheld and the sheet keeps Copy and Open in email. Editors and admins only.
PULSE_PUBLISH_WHATSAPP, PULSE_PUBLISH_SMSfollows the kill switchA WhatsApp message or SMS through the shared Twilio carrier: Do it on a lead’s next step under Leads › To do today, and Send on a result. Each send needs the person to confirm its preview. Recipients are the contacts an admin added under Settings › Messaging, plus, for SMS only, the person’s own open leads whose phone number has a recorded source (see below). Off, the send is refused as “not configured” with the switch named; nothing is sent. Calls are not affected.
PULSE_PUBLISH_LINKEDIN_REPLIES, PULSE_PUBLISH_INSTAGRAM_REPLIES, PULSE_PUBLISH_YOUTUBE_REPLIESfollows the kill switchReplies PULSE sends from the Autopilot reply queue (Campaigns › Approvals): an answer to a comment (and, on Instagram, to a post that tags the account). A reply needs this switch and the channel’s own PULSE_PUBLISH_<CHANNEL>, so false here keeps posting on and answering off. Off, an approved reply is recorded withheld with the switch named; nothing is sent.
PULSE_PUBLISH_INSTAGRAM_DMSfollows the kill switchAnswers to Instagram direct messages from the reply queue. Needs PULSE_PUBLISH_INSTAGRAM too. Independently of every switch, a message is answered only within 24 hours of the person’s last message (Meta’s window); after that the reply is withheld with that reason.
PULSE_PUBLISH_GBPfollows the kill switchApproving a Google Business Profile post (update, offer, event) publishes it on the managed location it names, with the account’s own Google sign-in. Not connected, or a location PULSE no longer manages, is withheld with the reason.
PULSE_PUBLISH_GBP_REPLIESfollows the kill switchA reply to a Google review, posted only when a person approves it (a negative review also needs the person to confirm they read it). Needs PULSE_PUBLISH_GBP too, so false here keeps posts on and review replies off. Off, the approval is recorded withheld with the switch named.
PULSE_EMAIL_DAILY_LIMIT50Emails one user may send in any 24 hours, to protect their mailbox’s reputation. A positive whole number; anything else means 50.
GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRETunsetThe instance’s own Google OAuth client. Also used for sign-in and token refresh. Needed for Connect Gmail, which asks for gmail.send and, to see replies, gmail.readonly. Also Connect with YouTube (asks for youtube.upload, youtube.force-ssl, yt-analytics.readonly; redirect PULSE_PUBLIC_BASE_URL/oauth/social/callback must be allowed on the client). Unset: the YouTube card accepts a pasted token instead. The same client connects Google Search Console (read-only webmasters.readonly, redirect PULSE_PUBLIC_BASE_URL/oauth/search-console/callback); see Google.
MICROSOFT_CLIENT_ID, MICROSOFT_CLIENT_SECRET, MICROSOFT_TENANT_IDunset, unset, commonThe instance’s own Microsoft Entra app registration, for Connect Microsoft 365 (delegated Mail.Send, Mail.Read for replies, User.Read, offline_access). Set the secret from the instance’s configuration, never from a repository secret.
PULSE_PUBLISH_CONFIG_PATH, PULSE_PUBLISH_CONFIG_JSON/etc/runink/pulse-publishing.json, noneThe same per-channel opinions as a file, or inline JSON. An environment key wins over the file.
PULSE_PUBLISHER_USER_IDSempty, which means nobodyUser ids (the users.id column, not emails) allowed to publish through the channels that post as the deployment: LinkedIn, blog, podcast, and audiobook. Anyone else’s dispatch is recorded withheld.
Settings › Integrations › Active (per user, not an env var)every channel active until the user savesA channel its owner switched off gets no daily drafts, and its queued posts are recorded withheld: channel not active. It can only narrow: it never arms a channel the switches above leave off.
PULSE_PUBLISH_LINKEDINfollows the kill switchKeep this false to keep LinkedIn drafts-only. See the daily LinkedIn draft.
Settings › Integrations › Autopilot (per account, not an env var)off for every accountAfter a person approves a post, PULSE queues it at the account’s best time instead of sending it at once, and the publishing worker sends it through the same gate as any scheduled post: the kill switch, PULSE_PUBLISH_<CHANNEL>, PULSE_PUBLISHER_USER_IDS and the Active switch above all still decide at send time. Autopilot never queues content nobody approved, and never queues on a disarmed channel (the approval then sends, and is recorded withheld, as before). It pauses the account by itself on an expired connection, a rate limit, or three failed sends in a row, and says so under Leads › To do today with a Reconnect link; transient failures are retried with backoff (5, 20, 80 minutes). Turning Autopilot off withholds its queued posts. See Autopilot and agencies.
Autopilot › Answer simple comments by itself (per account)offReplies to comments PULSE reads (LinkedIn, Instagram and YouTube; on Instagram also mentions and direct messages) are drafted into Campaigns › Approvals. A person may let three low-risk kinds go out by themselves: thanks, an answer taken from approved FAQ content, and a link to the business’s own site. Each one must pass OpenBias and the plan gate (a concur verdict; with PULSE_JUDGEMENT=off nothing replies by itself) and is logged on its row. A reply is an outbound post on the channel, so it also needs the kill switch, PULSE_PUBLISH_<CHANNEL>, the channel’s reply switch (PULSE_PUBLISH_<CHANNEL>_REPLIES, or PULSE_PUBLISH_INSTAGRAM_DMS for a direct message), the publisher allowlist and the Active switch.
PULSE_LINKEDIN_CTA_URLhttps://runink.orgThe address the daily LinkedIn draft ends with.
PULSE_AUDIT_APPLYoffThe second half of the audit-apply gate. With the kill switch, it allows the apply, commit, and pr modes of ApplyAuditActions (audit fixes), and Propose as pull request on the website repository, for administrators only. Off (or the kill switch off) refuses both before anything is read from or sent to GitHub.
Settings › Integrations › Website repository (per account, not an env var)not connectedThe customer’s own site repository on GitHub, with a fine-grained token limited to that repository (Contents and Pull requests: read and write), sealed with the envelope key. An administrator connects it; the card shows connected, missing permission, token rejected, branch not found or unreachable. A proposal is one branch pulse/<audit>-<finding> and one pull request through the GitHub API: PULSE never pushes to the base branch, never merges, never edits .github/workflows/ or .git/, refuses .. paths and links, and caps a proposal at 20 files, 256 KiB each and 1 MiB in all. The plan gate judges the edit before the pull request is opened; a dissent opens nothing.
PULSE_APPLY_ROOTunset, which refuses every modeThe only directory an audit fix’s repo_path may resolve into. Applies to plan too.
PULSE_PUBLISH_AUTOMATIONSoffLets automations (the automation card a person opens on a lead or a campaign item) do outbound work: schedule a post, write to HubSpot, send email from the owner’s mailbox, and place AI calls (which also need PULSE_OUTBOUND_CALLS, PULSE_AI_CALLS and the lead’s consent record). Under the kill switch: both must be on, and each step still needs its own channel switch and the publisher allowlist. Each automation must also be armed by its owner; a dry run never sends.
PULSE_OPPORTUNITIES_DISCOVERYonThe daily read of open procurement data (EU TED, UK Find a Tender and Contracts Finder, Brazil PNCP, CanadaBuys, and US SAM.gov for accounts that added their own key) and of requests for proposals found by PULSE’s crawler, for every account with an offering to match. Reading only: nothing is sent or submitted, and the kill switch does not apply. Off stops the daily read; Find → Public tenders: look now on Leads still reads for the person who presses it. Drafting a response (on request, or the opt-in auto-draft) spends model time on the account’s own inference.
PULSE_OUTBOUND_CALLSoffOne of the switches an AI call needs (with PULSE_AI_CALLS and the kill switch). The caller must also be an editor or admin.
PULSE_AI_CALLSoffAllows AI voice calls to leads (AICallService.PlaceAICall, AI calls). Needs PULSE_APPLY_ENABLED=true and PULSE_OUTBOUND_CALLS as well, an editor or admin, Twilio voice settings, PULSE_PUBLIC_BASE_URL (https), and for each lead a recorded consent, a time zone and a drafted call script.
PULSE_SITE_CHAT_SITE_MESSAGES_PER_MINUTE, PULSE_SITE_CHAT_SITE_BURST, PULSE_SITE_CHAT_VISITOR_MESSAGES_PER_MINUTE, PULSE_SITE_CHAT_DAILY_MESSAGESunset, which turns the website chat’s visitor side off (503)The website chat (Settings › Website chat): the PULSE assistant answering visitors of the account’s own site from its public facts only (offerings, approved FAQ content, pages of its own site), on PULSE’s own inference. Each one a positive number: the sustained rate and burst per site key, the rate per visitor address (held in memory only) and the visitor messages one site may send per UTC day (a contact or ask-a-person form counts as one). All four must be set, and PULSE_PUBLIC_BASE_URL too for the snippet; Settings names what is missing. The site key in the snippet is public, works only from the addresses the owner listed, and is revoked or replaced there. Nothing is sent from it: a lead it makes (only with the visitor’s consent) waits in Leads › Inbound like a website-form lead.
PULSE_SITE_CHAT_RETENTION_HOURS168 (7 days), at most 720How long a visitor’s chat is kept before the hourly sweep deletes it. The widget states it to the visitor.
PULSE_PUBLIC_BASE_URLunsetThe public origin for absolute podcast-feed links, and the base of the mailbox OAuth callback <PULSE_PUBLIC_BASE_URL>/oauth/mail/callback. Without it, an armed feed answers 503 instead of rendering links nobody can fetch, and Gmail / Microsoft 365 cannot be connected (SMTP still can). It must be https for Connect with Instagram, YouTube, LinkedIn and Search Console to be offered, and for Instagram to fetch media; the TIDE edge must route /oauth/social/callback and /oauth/search-console/callback to PULSE. See Public address and redirects.
PULSE_JUDGEMENTonoff turns the plan gate off: proposals are no longer judged before they are saved, and Autopilot answers nothing by itself (an auto-reply needs a concur verdict). Leave it on.
TTSD_URLthe in-cluster ttsd ServiceTIDE’s shared text-to-speech service, as a full URL. Set to empty to turn voice output off: Reels and Stories PULSE makes are then silent and say so, and read-back and podcast audio are unavailable. Sends nothing outward by itself.
PULSE_ENVunsetproduction (any case) turns the startup security check from a warning into a refusal to start. TIDE’s k0s overlay sets it. See below.
PULSE_GRPC_REFLECTIONofftrue turns on gRPC reflection. For local work only; PULSE_ENV does not affect it.
AUTH_ALLOWED_EMAILSempty, which means nobodyWho may sign in with Google.
AUTH_GOOGLE_ONLYoffExactly true hides the password form.
AUTH_SEED_USERSunsetOperator-granted roles as email=persona,…, re-applied at every start. See Seeded roles.
SELF_HEAL_AUTONOMOUSonAn explicit no stops the self-healing loop applying anything; plans are still reasoned and recorded. See Operations.
PULSE_WHATSAPP_AUTOREPLYoffAn explicit yes lets PULSE answer inbound WhatsApp and SMS messages with a model-written reply. Each reply is a send, so it also needs PULSE_APPLY_ENABLED=true and PULSE_PUBLISH_WHATSAPP (or PULSE_PUBLISH_SMS for an SMS) not switched off. Off, messages are still received and nothing is answered. See Telephony and messaging.

Every switch is also reported at startup. Look for the PUBLISHCFG lines, including PUBLISHCFG audit-apply: armed=…; PULSE_APPLY_ROOT set=…; callers must be admin, and for security config validated.

Safe arming order

Turn things on one at a time, and check the startup log after each restart.

Confirm the production gate passes

With PULSE_ENV=production, the server refuses to start on an insecure configuration. Before setting it, confirm that the current pod logs security config validated. See Operations.

Name who may publish as the deployment

Set PULSE_PUBLISHER_USER_IDS to the owner’s <user-id>, read from the id column of the users table. Do this before arming, so that the moment the kill switch goes on, only the owner’s content can leave on LinkedIn, the blog, or the podcast feed.

Narrow the channels you do not want

Set PULSE_PUBLISH_<CHANNEL>=false for each channel that must stay off, including PULSE_PUBLISH_LINKEDIN=false if LinkedIn is to stay drafts-only, PULSE_PUBLISH_EMAIL=false if users must not send email from PULSE, and PULSE_PUBLISH_WHATSAPP=false / PULSE_PUBLISH_SMS=false if nobody may message from PULSE. Narrowing first means arming the global turns on only what you meant.

Set up the sign-in apps, if people should connect with one click

Set an https PULSE_PUBLIC_BASE_URL routed by the TIDE edge, then the app clients: META_APP_ID / META_APP_SECRET, GOOGLE_CLIENT_ID / GOOGLE_CLIENT_SECRET, LINKEDIN_CLIENT_ID / LINKEDIN_CLIENT_SECRET. None of them publishes anything: connecting only stores a sealed token. What each developer console needs is on Setup. Leave LINKEDIN_COMPANY_PAGES off until LinkedIn has approved the Community Management API for the app.

Arm the kill switch

Set PULSE_APPLY_ENABLED=true. Check one PUBLISHCFG <channel>: armed=… line per channel, and that the audit-apply line still says armed=false.

Arm server writes, only if you need them

To let administrators apply audit fixes, set PULSE_APPLY_ROOT to the directory that holds the site checkouts, then PULSE_AUDIT_APPLY=true. Leave both unset to keep the server read-only. The same two switches arm Propose as pull request on a connected website repository (Settings › Integrations › Website repository), which needs no PULSE_APPLY_ROOT: it writes only to a new pulse/ branch on GitHub and opens a pull request for a person to merge.

Arm automations, only if you need them

Set PULSE_PUBLISH_AUTOMATIONS=true once the channels the automations use are armed and allowlisted. Each automation still runs by itself only after its owner arms it, and every step runs with the owner’s role.

Let accounts run on Autopilot, only if you want that

Autopilot is switched on per account by its owner, under Settings › Integrations › Autopilot. It adds no server key: it sends only through channels you armed above, so a channel you narrowed stays off for Autopilot too. To stop every Autopilot post at once, unset PULSE_APPLY_ENABLED (or the channel’s PULSE_PUBLISH_<CHANNEL>): queued posts are then recorded withheld when they come due and are never replayed when you re-arm.

Arm outbound calls, only if you need them

Set PULSE_OUTBOUND_CALLS=true and PULSE_AI_CALLS=true once the Twilio voice settings are in place and the compliance work for calling real numbers is done. See Telephony and messaging.

To disarm, reverse the order, or simply unset PULSE_APPLY_ENABLED: that stops every channel, the audit-apply modes and the daily site-audit engine at once. It also stops every automation’s outbound steps. It does not stop outbound calls, WhatsApp and SMS replies, or self-healing, each of which has its own key.

PULSE_ENV=production

TIDE’s k0s overlay sets PULSE_ENV=production on the PulseControlPlane, so it takes effect when that resource is applied. With it set, the server refuses to start, logging refusing to start in production, when:

  • the session key (PULSE_JWT_SECRET) is missing or is one the platform refuses, or
  • CORS_ALLOWED_ORIGINS is unset or contains *. The backend serves no web bundle, so there is no same-origin exemption: pin the origins in the deployment before this change rolls out.

A missing or refused session key stops startup whatever PULSE_ENV says; what production adds is that the CORS problem stops it too, instead of warning.

PULSE_ENV also changes one TLS log line. It does nothing else: no billing, usage counting, or plan behavior reads it, and it does not turn reflection on or off.

Image build arguments

The backend image (grpc/Dockerfile) takes its base images through one build argument, each base pinned by digest. Point it at a registry mirror and the same digests come from there.

Build argumentDefaultBases
BASE_REGISTRYdocker.io/librarygolang:1.26-bookworm (build) and debian:bookworm-slim (runtime)

BASE_REGISTRY defaults to the upstream registry. There is no Flutter stage in the image: PULSE ships as an Android and iOS app and the backend serves no web bundle.

Messaging a lead at their own number

Do it on a lead’s WhatsApp or SMS step can text the lead’s own number, not only a contact an admin added. Every rule below is enforced on the server:

  • SMS only, after a confirm tap. The person sees the exact text and the masked number, and taps to send. It still needs PULSE_APPLY_ENABLED=true and PULSE_PUBLISH_SMS not switched off.
  • Only a number with a recorded source. A phone number counts only when PULSE recorded where it came from: the lead’s own web page (the audit), a person typing it into PULSE, your website form, or HubSpot. The Leads page shows the source. A number without a source is never texted.
  • STOP is honoured. A reply of STOP (also STOPALL, UNSUBSCRIBE, CANCEL, END, QUIT, PARAR, SAIR, BAJA, ARRET) is recorded for that number. After that, the lead’s step no longer offers WhatsApp or SMS, and a message that was already previewed is refused when confirmed. START, YES or UNSTOP lifts the block. When Twilio itself refuses a send because the number unsubscribed (error 21610), PULSE records the same block. For the replies to reach PULSE, set the Twilio number’s “A message comes in” webhook to <PULSE_PUBLIC_BASE_URL>/comms/twilio/inbound. PULSE checks Twilio’s signature on every request, and the webhook refuses everything (503) while no auth token is set.
  • WhatsApp needs an approved template. Meta only lets a business start a WhatsApp conversation with a template it has approved. The shared carrier sends plain text only, so PULSE never starts a WhatsApp conversation with a lead. Do it says so plainly and offers SMS or email instead. WhatsApp to a contact an admin added works as before.

Checked against the code

Every key on this page is read by the backend (grpc/cmd, grpc/internal) or, for AUTH_ALLOWED_EMAILS and PULSE_JWT_SECRET, by the shared security library it uses. The backend also reads these keys, which arm nothing and are listed on the configuration reference: PULSE_SKILLS_MODELS, PULSE_SEARCH_MIN_GAP_MS, LINKEDIN_API_BASE, LINKEDIN_API_VERSION, LINKEDIN_ACCESS_TOKEN, LINKEDIN_AUTHOR_URN, GOOGLE_TOKEN_ENDPOINT, MESH_ISSUER_URL, MESH_ISSUER_TOKEN_FILE, FFMPEG_PATH and the rest of the storage, inference, telephony and billing settings.