Sequence enrollment

A sequence is Oxygen's outbound journey: LinkedIn, email, WhatsApp, call, and CRM-task steps dispatched with rate limits, reply-stop, and a unibox. Enrollment puts leads into one. It is always free and never sends anything. Only starting the sequence dispatches real actions, behind its own approval gate.

Prerequisites

Enrollment itself doesn't touch a provider or spend anything — it only needs the sequence, and for --from-table, its bound table, to exist. Sender and mailbox health matter once you start the sequence live, not before.

Connect a sender with oxygen senders connect --country DE — use the account owner's normal LinkedIn login country, not the campaign target. It mints a shareable hosted-auth link (valid 10 minutes, one account per link) you can open yourself or send to the account owner; they sign in to their own LinkedIn and need no Oxygen login. The connected account appears under Connections and in oxygen senders list.

Three enrollment paths

Web — the Leads stage

The Leads stage has one button, Import leads. It opens a source chooser with three cards: Upload a file (CSV, JSON, JSONL, or XLSX), Copy from a table (rows from another workspace table; disabled until a leads table is bound), and HubSpot list (keeps a saved list synced through a hosted DNC-first workflow). None of them is an enroll button: the two direct paths wait for their durable ingestion run to reach a terminal state, then enroll the sequence's bound table on your behalf, paging until nothing is left. A toast reports the enrolled count and, for a LinkedIn sequence, how many were skipped (no LinkedIn URL). A sequence with no bound table yet gets one created and bound by the import. Leads land as pending; nothing sends until you start the sequence.

CLI/MCP — bound table (--from-table)

The one-command path for a sequence already bound to a source table:

oxygen sequences enroll <sequence> --from-table --json

In MCP, call oxygen_sequences_enroll with from_table: true instead of a leads array.

The response's from_table block reports the batch:

At most 500 rows enroll per call. When has_more is true, run the identical command again — already-enrolled rows are skipped automatically, so the enrolled set is its own cursor; there's no offset or page token to track. Once every row is enrolled, re-running returns enrolled: 0 with the note "All table rows are already enrolled in this sequence." — success, not an error.

A sequence with no bound source table fails with sequence_no_source_table. Create the sequence with --table <table-id>, or use an explicit leads list instead.

CLI/MCP — explicit leads

For leads outside the bound table, or a sequence with no table at all:

oxygen sequences enroll <sequence> --leads-file leads.json --json

{
  "leads": [
    { "lead_name": "Ada Lovelace", "row_values": { "email": "ada@acme.com", "first_name": "Ada" } },
    { "lead_profile_url": "https://www.linkedin.com/in/example" }
  ]
}

MCP's oxygen_sequences_enroll accepts the same list under leads. Each entry identifies its lead one of these ways:

LinkedIn URL resolution

A LinkedIn sequence no longer needs a pre-resolved provider id on every lead. Oxygen resolves the LinkedIn member id at send time from, in order: an explicit lead_provider_id, an explicit lead_profile_url, or the enrolled row's LinkedIn URL column — the sequence's configured linkedin_url_column_key, falling back to linkedin_url, linkedinUrl, linkedin, profile_url, or contact_linkedin.

Rows with no resolvable /in/ URL are not enrolled — they're dropped and counted in the enroll response's unresolved_linkedin_leads. Fix the cell (or point linkedin_url_column_key at the right column) and re-run --from-table; already-resolved rows are skipped, so only the fixed rows enroll.

Idempotency and skips

Enrollment is idempotent per (sequence, table_row_id) — re-running any enrollment path only adds what's new. Enrolling by email/phone with no table dedupes per email/phone instead.

Every enroll response reports skipped with a skipped_by_reason breakdown:

unresolved_linkedin_leads is reported separately from skipped_by_reason since it isn't an exclusion decision — the row just has nothing to dispatch to yet.

Nothing sends until you start

Enrollment — any path — never dispatches anything. Enrollments land pending; if the sequence is already active they activate for the next dispatch tick, but dispatch itself only fires for steps a live start has authorized (see Modes for dry_run vs live).

oxygen sequences enrollments <sequence> --json                            # inspect enrollment status
oxygen sequences start <sequence> --json                                  # preview — no sends
oxygen sequences start <sequence> --dry-run --json                        # simulate — no sends, no credits
oxygen sequences start <sequence> --approved --max-credits <n> --max-live-sends <n> --json

--approved is the only flag that makes a start live; without it you get a preview with rendered recipient samples, character counts, copy blockers, sender/mailbox capacity, and CRM readiness.

The two ceilings bound different things. --max-credits caps the LinkedIn track and is optional — omit it and the LinkedIn track runs unbounded. --max-live-sends caps total real external actions and is required whenever the sequence has email, WhatsApp, or crm_task steps, because those actions cost 0 Oxygen credits and a credit cap cannot bound them. Its effective value must also exceed the sequence's live_sends_used, or the start is refused.

Replies land in your CRM

When someone replies, Oxygen files them as a person in your CRM and moves their lead stage:

Auto-replies, wrong-person and bounced conversations are skipped — those aren't people to add. Stage moves are forward-only, so a reply can never move someone backwards out of a closed stage (a customer stays a customer).

This is one automation for the whole workspace, not a per-campaign setting. It switches on the first time you start a sequence live and then covers every sequence you create afterwards — you never set it up again.

oxygen crm automation rules --json                                  # is it armed?
oxygen crm automation set crm-lead-stage-router --disarmed --live   # turn it off

crm automation set defaults to dry-run; --live is what writes the change. Turning it off is a decision Oxygen remembers: a disarmed automation is never switched back on automatically.

Want different behaviour — adding everyone you email rather than only the people who reply, or filing them in a nightly batch instead of instantly? That's a fork, not a setting; see Customizing an existing automation.

Recovering terminal failures proven to have no effect

Current provider-capacity pressure is a deferral, not a terminal failure: Oxygen records the reason as provider_capacity, waits without sending, and retries when capacity returns. Historical Sequence actions created before that deferral path can instead be left terminal even though the provider call definitively never happened. The same recovery seam also recognizes narrowly allowlisted Oxygen request-format defects when the provider's validation response proves the action was not applied—for example, the historical WhatsApp cold-chat request that sent bare digits instead of a provider-ready JID.

Preview the exact legacy scope for one Sequence:

oxygen sequences recover-capacity <sequence> --json

The preview is read-only, costs 0 credits, makes 0 provider calls, and writes nothing externally. scanned_count is the number of terminal candidates inspected; eligible_count is the exact safe recovery scope; excluded_by_reason explains everything refused. A zero scanned_count across a Sequence means there is no candidate for this recovery path—it does not mean failures are unmeasured.

Sequence action errors use the canonical failure_class values capacity, recipient, authorization_scope, funding, validation, provider, internal, effect_unknown, and unknown. provider_capacity is a deferral reason, while provider_capacity_deferred is its error code; neither is a separate failure class. Recovery accepts a terminal capacity failure only with definitive no-effect evidence, and a non-capacity failure only when its exact historical signature is explicitly allowlisted as an Oxygen request defect with a provider-confirmed no-effect outcome. It refuses ambiguous effects, ordinary recipient failures, later delivery evidence, suppression, and any changed execution scope.

Applying recovery is deliberately separate from preview. Pause the Sequence, inspect the returned scope and exclusions, then use the preview-bound approval token only if that exact local repair is intended. Applying restores local action/enrollment state; it never sends, calls a provider, spends credits, or resumes dispatch, and the Sequence remains paused.

Track performance and organize

oxygen sequences list --stats --json          # lifetime funnel per sequence: enrolled, sent, replies, reply rate, email opens/clicks, opportunities
oxygen sequences analytics --range 30d --json # all campaigns + cross-campaign mailbox/domain attribution
oxygen sequences variants <sequence>          # one campaign: generic action/base-variant, mailbox, domain, reply type, and bounce tables
oxygen sequences events <sequence> --json     # per-lead activity with mailbox/domain provenance where known
oxygen sequences list --tag q3-outbound --json  # only sequences carrying a workspace tag
oxygen sequences update <sequence> --tags q3-outbound,saas-founders  # replaces the set; "" clears; works on archived sequences too

--stats costs a stats pass per sequence — skip it for a quick name lookup. Tags link the sequence to its lead table, workflow, and campaign-learnings wiki page; the /sequencer web list shows the same stats as sortable columns with per-row start/pause/duplicate/delete actions.

Mailbox/domain attribution is built from Oxygen's durable send, conversation, reply-signal, and hard-bounce provenance. An event that preserved the campaign but not the sending identity stays under unattributed rather than being assigned to a likely inbox. Provider-owned campaigns do not enter this report merely because their integration is connected; their telemetry must first be normalized into Oxygen's native action/message ledgers.

Related

• Approvals, Spend caps — the gate a live start enforces.
• Runs — the adjacent waiting state for current provider-capacity pressure outside Sequences.
• Modes — dry_run vs live for sequences start.
• Tables — the bound source table --from-table reads.
• Web app — the sequence's Leads stage.