The job receipt

Read this when a job is finished (or nearly), when the user asks what a video cost or how many rounds it took, or when a gallery, site or post wants to show a job's receipt.

Every job ends with three things, written by showtime receipt [job]:

• receipt.md: for people. Fixed headings, below.
• receipt.json: the stable machine format (schema 1, below). A gallery or site reads this one.
• one line in share.txt, between --- Receipt (optional, delete if you like) --- and --- end receipt --- (replaced on every run, the rest of share.txt is left alone).

showtime job note <job> --stage deliver, showtime deliver exports and the MCP receipt tool write it; showtime qa of the job's latest render keeps receipt.md and receipt.json current (not share.txt). Run showtime receipt <job> last, after share.txt is written, and it is always safe to run again.

Essentials#

• Run showtime receipt <job> last, after share.txt is written (safe to re-run): it writes receipt.md, receipt.json and one line in share.txt between the receipt markers (§ What goes in it)
• Record the request verbatim: showtime job init <slug> --request "<verbatim>" (§ What goes in it)
• The receipt counts only what was made or recorded; "images made for looking" is not "images looked at"; nothing is estimated silently (§ What goes in it)
• Tokens and cost come only from the named log of the current session (--transcript <file>), numbers only; otherwise "not reported by this agent"; the cost is a floor, not a bill (§ Tokens and cost)
• showtime receipt <job> --card prints the delivery card (modes.md §5) with the facts filled in (files, length, loudness, QA, review, findings and waivers, cost) and the agent's lines (Look, Assumed, Next) as prompts; job note --stage deliver prints it too (§ What goes in it)

What goes in it, and where it comes from#

Honesty rules: showtime counts what it made or recorded, never what the agent did in its head. "Images made for looking" is not "images looked at": whether each was opened is not known to showtime, and the receipt says so. Nothing is estimated silently. A clean removes images; the receipt keeps the larger count it had.

Tokens and cost#

Only from a session log that is named, only the current session, and only its numbers (model name, time stamp, token counts). No prompt, reply, tool call or file content is read out, kept or copied; the log's path is kept in <job>/work/logs/receipt-source.json for the next regeneration and never appears in the receipt.

• Claude Code. The plugin's Stop hook (hooks/hooks.json) hands the session's transcript_path to showtime receipt --hook at the end of a turn; without the plugin: showtime receipt <job> --transcript <file> (or SHOWTIME_TRANSCRIPT). Usage per model message (input, output, cache read, cache write; sub-agent logs beside the session are added; a message logged twice counts once). Window: the job's start to its last update plus 30 minutes; --whole-session counts everything. Cost is API-equivalent: the same tokens at public list prices (lib/st/job/usage.py PRICES, each read on its provider's pricing page, dated and sourced in the table: Opus 5.5, Sonnet 5.5, Sonnet 5, Fable 5.1, Haiku 4.5, GPT-6.1 Sol, GPT-6 Sol, GPT-6 Astra, Grok 4.7). A plan does not pay per video. A model with no price is listed with its tokens and no cost, and the total says it is a lower bound. Work the host keeps outside the session log is not counted, so treat it as a floor, not a bill.
• Codex. --transcript <rollout.jsonl>: tokens (input, cached, output, reasoning) from the cumulative token_count events in the window. Codex reports tokens, not dollars: cost says "not reported by this agent".
• Devin. Under Devin (or --transcript <sessions.db> --host devin) the CLI's session database (~/.local/share/devin/cli/sessions.db) is opened read-only; only each request's model id, time and token metrics are selected, from the sessions whose working folder holds the job. Devin reports tokens, not dollars: the cost is an estimate at the same API list prices (PRICES; Devin's ids drop their effort suffix, and a request past a model's long-context threshold is priced at that price) and is labelled "est.". SWE-2 has no list price there, so its tokens are listed unpriced and the total is a lower bound. A Fusion session lists the lead and the sidekick model with their share of the tokens.
• Everything else, and the MCP path (Claude Desktop, Cursor, ...): the server cannot see a session log, so tokens and cost say "not reported by this agent". The receipt still names the client through MCP.

receipt.json (schema 1)#

Keys never change meaning within a schema; new keys may be added.

schema, job, showtime, generated
request      {text, source}            text is masked (home paths, keys, e-mail addresses)
mode, platform, agent
assumptions  {not_checked[], checked[], decisions[{question, answer}], open_questions[]}
review       {packed, with_findings, pairwise, findings_files, self_reviewed, lead_review,
              rounds[{round, kind, findings, findings_files, self_review}],
              findings: null (no critic answered) | {raised, blockers, should_fix,
                        fixed[{id, severity, text, by, note}], waived[{id, severity, text, by, reason}],
                        open[{id, severity, text}]}}
delivery     {final, kind, poster?, share?, credits?, captions?, exports[], duration, size, fps, lufs, true_peak,
              target_lufs, qa, qa_fail, qa_warn}       file names only
qa           {runs, verdict, fail, warn, file}
renders      {full, preview, partial, spliced, seconds, source, list[{kind, file, seconds, span, spliced_from}]}
images       {total, by_kind{review, qa, snap, check, views, looks}, note}
time         {start, end, ends_at, wall_seconds, render_seconds, note}
usage        {status: "reported" | "not_reported", host, note?,
              tokens{input, output, cache_read, write_5m, write_1h, total}   (Codex: input, cached_input, output, reasoning)
              cost_usd (number | null), cost_note, prices_as_of, window,
              subagents?, models{<model>: {input, output, cache_read, write_5m, write_1h, messages, cost_usd, token_share?}}}

View or edit this page on GitHub