Deepline Plays

Getting Started With Deepline Plays

Deepline GTM work is paid uncertainty reduction. The play is the notebook, stable ids are the cache, datasets are durable row state, getters are the interface, and CSV is an input/output boundary. Start from live Deepline contracts, then run exactly or build a small durable wrapper.

Default rule: every user request is fresh unless the user names a project, play file, run id, saved play, or asks to continue prior work. Do not inspect old local .play.ts, README, or CSV files for route evidence.

Run one Deepline command at a time when you need its output. Avoid jq, shell parsing, background jobs, raw curl, env spelunking, or local provider probes around Deepline commands. JSON output is for reading directly.

If CLI V2 seems broken, check deepline switch status; use deepline switch sdk to switch to SDK mode or deepline switch python to roll back to the Python CLI.

Core Loop

1. Preflight: check auth/health/balance when spend or cloud execution is likely. 2. Describe before spend: plays search -> plays describe; for tools, tools search -> tools describe. 3. Choose direct vs build: direct-run only when the described contract exactly matches input, output, export, freshness, and pricing. Otherwise bootstrap/wrap/fork. 4. Check before run: plays describe is the gate for prebuilts; plays check <file> is mandatory for local, bootstrapped, or forked plays. 5. Pilot before scale: run 1-3 rows or a small sample, then inspect/export. 6. Report reality: run id, export path, charged credits or why not visible, executed/reused/failed counts when available, and repair class.

Safe planning-only commands: auth/health/balance, plays search, plays describe, tools search, tools describe, plays check, plays bootstrap --help, and local scaffolding. Do not call plays run or provider execution in planning-only mode.

Which Piece?

Find And Describe

Names are hints. Live search and describe are source of truth.

deepline plays search "<job words>" --json
deepline plays describe prebuilt/<candidate> --json
deepline tools search "<provider need>" --categories <category> --json
deepline tools describe <tool-id> --json

Rules:

• plays search defaults to trusted Deepline-managed prebuilts. Use --all only when the user explicitly names or asks for their own/user/org play.
• Search results are candidates. Describe before using them.
• Do not use tools describe for a prebuilt play.
• Job-change/email/phone/LinkedIn/contact workflows are live play searches, not job docs.
• If the task is account sourcing, contacts, provider strategy, signals, or TAM, load references/find-companies-contacts-tam.md.

Contract checklist from plays describe / tools describe:

• canonical reference and namespace
• owner/editability/cloneability
• scalar, CSV, or API input modes
• required fields and accepted aliases
• output object and row dataset shape
• nested paths and flattened export headers
• pricing and billing mode
• freshness/staleness behavior
• direct-run vs bootstrap/wrap/fork constraints
• run/export/clone starter commands
• declared getters such as result.extractedValues.email?.get()

Direct Prebuilt Run

Direct-run only when exact:

• described scalar/CSV/API input matches the user input
• no CSV mapping or semantic repair is needed
• output schema includes the requested result
• export dataset path is known
• freshness/caching behavior is acceptable
• pricing mode and likely scale are acceptable

Typical flow:

deepline plays describe prebuilt/<name> --json
deepline plays run prebuilt/<name> --input '{"field":"value"}' --watch
deepline runs get <run-id> --full --json
deepline runs export <run-id> --dataset result.rows --out rows.csv

For CSV prebuilts, compare required headers to actual headers. If aliases are unsupported or output projection is custom, bootstrap a wrapper instead of editing the prebuilt.

Bootstrap, Wrap, Fork

Bootstrap is the composition tool. It is not anti-prebuilt.

Bootstrap/wrap when:

• CSV headers need mapping, validation, or projection
• a prebuilt is a useful stage but not the whole answer
• company rows need people/contact/channel fanout
• provider source rows need durable row state
• final output needs flat user-facing columns
• row gates, fallback legs, miss reasons, or stale policy matter

Fork when internals need to change:

• provider/tool order
• internal stale policy
• getter metadata
• billing stage
• native prebuilt logic

Do not fork for simple CSV aliases or final formatting. Wrap instead.

deepline plays bootstrap <family> --from <source> --using play:prebuilt/<candidate> --limit 5 --out workflow.play.ts
deepline plays get prebuilt/<name> --source --out fork.play.ts
deepline plays check workflow.play.ts

Route families: people-list, company-list, people-email, people-phone, company-people, company-people-email, company-people-phone.

If bootstrap syntax fails, run deepline plays bootstrap --help or route help and retry with explicit stage flags such as --people, --email, or --phone.

Authoring Basics

Use the current V2 shape from generated references when exact syntax matters:

import { definePlay } from 'deepline';

type Input = { limit?: number };

export default definePlay(
  'gtm-play',
  async (ctx, input: Input = {}) => {
    return { ok: true, limit: input.limit ?? 5 };
  },
  { billing: { maxCreditsPerRun: 50 } },
);

Authoring rules:

• Prefer typed inline input. Import validators only if generated refs or bootstrap output prove they exist.
• Use ctx.csv, ctx.dataset, ctx.tools.execute, ctx.runPlay, ctx.step, ctx.fetch, and ctx.secrets.
• Do not use local fs, raw fetch, shell commands, env reads, Date.now, or Math.random inside play bodies.
• Use stable ids for paid work. Rename ids only to refresh wrong/stale provider data or changed semantics.
• Prefer one paid operation per dataset cell. A paid/tool/play-backed column should return the provider or child-play default output as directly as possible, such as email_lookup_raw.
• Put shaping, projection, status, miss_reason, display fields, and transformations in separate pure columns after the paid column.
• Good shape: email_lookup_raw: rowCtx.runPlay(...) or rowCtx.tools.execute(...), then resolved_email, email_status, and miss_reason as pure extraction/mapping columns.
• Avoid one paid column that calls a provider or child play, transforms the output, invents status, and hides the raw/default result at the same time.
• Return datasets for CSV/exportable outputs.
• Use declared getters. Do not parse raw payload paths when extractedValues.*.get() exists.
• Project to flat user-facing columns with status, miss_reason, evidence/source, and requested output fields.

Dynamic staleness pattern:

.withColumn("job_change", {
  run: async ({ row, ctx, previousCell }) => {
    if (previousCell?.value.status === "stale_contact") {
      return previousCell.value;
    }

    return await checkJobChange(row, ctx);
  },
  staleAfterSeconds: (value) =>
    value.status === "stale_contact" ? null : 30 * 24 * 60 * 60,
})

Semantics:

• value is the new value returned by run.
• null stale time means no next expiry.
• Existing cells with missing stale metadata may schedule once to backfill.
• The previousCell guard avoids paying again during metadata backfill.

Load References

Load only when the task needs it:

• references/find-companies-contacts-tam.md (Find Companies, Contacts, And TAM): account sourcing, contacts, TAM, portfolio/investor lists, hiring-qualified companies, signals, personas, provider playbooks, account-first strategy, evidence, or fanout economics.
• references/run-export-inspect-repair.md (Run, Export, Inspect, Repair): before scale; after every meaningful run; for billing, rerun, export, cached rows, failed rows, logs, suspicious output, partial repair, or UI/run mismatch.
• references/sdk-reference.md: exact current SDK signatures.
• references/api-reference.md: exact API/manual invocation, polling, streaming, stop, list, inspect/export, and artifact routes.

Do not load a separate reference for ordinary prebuilt search/describe/run, CSV contracts, bootstrap, wrapping, or forking. Those basics are here.

Finish Shape

When work ran, summarize:

• route and play reference
• run id
• rows requested and returned
• executed/reused/failed counts when visible
• charged credits or why credits are missing/zero
• export path and dataset path
• miss/failure classes
• next action: scale, rerun, repair, or stop

When no paid run happened, say so explicitly and list the safe commands used.