Surface the run's artist_account_id in headless agent prompts (tasks lose their artist)

[sweetmantech]

Follow-up to chat#1833 (task-email hallucination — closed 2026-07-02). While re-benchmarking Steph@OneRPM's 3 scheduled tasks against a fully corrected artist record, we found that headless runs receive the run's artistId but never surface it to the agent — so any task prompt that doesn't hand-embed the artist id sends the agent roster-guessing, where it nondeterministically lands on the org-scoped roster and can't find the artist at all.

Decision (2026-07-02, @sweetmantech): dropped the skills-side roster hardening (skills#71 closed unmerged). api#742's injected context already tells agents to use the given artist id; org and personal rosters are legitimately different sets (neither is authoritative), so "check personal before concluding absence" was a wrong assumption; and the API layer already resolves artist ids across both scopes (checkAccountArtistAccess: personal account_artist_ids OR shared org — no org_id needed for /api/artists/{id}/*).

Goal

Every headless agent run knows which artist it's running for. customer-prompt-task → POST /api/chat/runs already carries artistId on every scheduled run; the agent's system prompt must expose it (plus account_id) the same way the interactive chat does — so implicit-artist prompts ("Check if there are any new YouTube videos…") resolve deterministically instead of depending on prompt authors remembering to paste UUIDs. Surfaces: app/lib/workflows/runAgentStep.ts / lib/chat/buildAgentSystemPrompt.ts (headless prompt), lib/chat/buildRunAgentInput.ts (input threading), and the recoup-platform-api-access skill's roster-discovery guidance.

PRs (updated 2026-07-02)

PR	Item	Base	State
api#742	Thread artistId into the headless agent prompt — buildRunAgentInput + RunAgentWorkflowInput + runAgentStep → buildAgentSystemPrompt emits an IMPORTANT CONTEXT VALUES block (artist_account_id, account_id) like the interactive chat	main	✅ merged 2026-07-02 — see Done · primary fix
skills#71	Harden roster discovery in recoup-platform-api-access — prefer the run's artist context when present; never conclude "artist not found" from the org-scoped list alone (query the personal roster GET /api/artists before giving up)	main (skills)	❌ closed 2026-07-02 unmerged — superseded by api#742; see decision below

Merge sequencing: the api prompt-injection fix is primary and independent; the skills hardening is defense-in-depth for runs with no artist context. Either can merge first.

Done

• api#742 — surface the run's artist_account_id in the agent system prompt.
  ✅ Merged 2026-07-02 (squash to main, commit 80b1d500). buildAgentSystemPrompt emits an IMPORTANT CONTEXT VALUES block (artist_account_id + account_id) whenever the run carries them; threaded buildRunAgentInput → RunAgentWorkflowInput → runAgentStep; headless passes the validated artistId, interactive passes session.artist_id.
  Verified on preview (results): the implicit-artist Cosculluela prompt that went 0/2 on main passed first-run — 3 direct calls with the injected id, 0 org-scoped lookups, 0 "Rostrum Pacific", fresh scrape of the real 8.28M-sub channel, honest conditional no-op ("No new video in the last 24 hours").

Open — thread artistId into the headless agent prompt

• Surface artist_account_id in the headless prompt — ✅ shipped api#742, see Done.

Open — roster-discovery hardening (skills)

• recoup-platform-api-access: never conclude "artist not found" from the org roster alone. Dropped 2026-07-02 — superseded by api#742 (see decision callout above); skills#71 closed unmerged. The API already authorizes artist ids across personal + org scopes, and the injected context removes the discovery step entirely for runs that carry an artist.

Source references

• Parent: chat#1833 (closed) — the grounding/persistence fleet this rides on.
• Benchmark evidence: this issue's first comment (task prompts + round-1/round-2 email outcomes).
• Task→run plumbing: tasks/src/tasks/customerPromptTask.ts (passes artistId from taskConfig on every scheduled run).

[sweetmantech]

Evidence exhibit — Steph@OneRPM's 3 scheduled tasks: prompts + benchmark email outcomes

All three tasks run under account 94d3f7e5… (model: null → Sonnet-5 default per tasks#154). Benchmarked twice on 2026-07-02 under a staff account (recipients rewritten to a staff inbox; artistId passed on every run): round 1 against the then-dirty Apache record, round 2 after the artist correction (verified socials + image + KB — see chat#1833 closure).

Task 1 — "Daily YouTube Video Performance Report" (c1ecfdbb, daily 0 12 * * *, artist Cosculluela 87360f64…)

Check if there are any new YouTube videos posted in the last 24 hours. If a new video is found, analyze its performance metrics (views, likes, comments, watch time if available) and send a detailed report via email to stephanie.guerrero@onerpm.com with the subject 'New YouTube Video Performance Report'.

Artist: implicit (never named). Outcomes: round 1 (4fa700b5) ❌ no email — agent resolved the roster as "one artist: LV8R" and found no YouTube; round 2 (fd8e8b26) ❌ no email — same trap via the org-scoped lookup ("Roster scoped to this org → Rostrum Pacific → only LV8R"). Cosculluela's YouTube (@cosculluelaelprincipe, 8.28M subscribers, social linked) was never touched in either round.

Task 2 — "Weekly Apache Artist Report" (80c34563, Mondays 0 10 * * 1, artist Apache ebae4bb9…)

Generate a comprehensive weekly report for Apache (artist_account_id: ebae4bb9-e38f-4763-b3c8-ff30e99f5d01) covering recent fan data, streaming performance, social engagement metrics, and emerging trends. Format this as a professional email and send it to stephanie.guerrero@onerpm.com with the subject 'Apache: Weekly Artist Report'.

Artist: explicit (id embedded). Outcomes: round 1 (253d0ff0) ✅ sent — 12 Research + 4 fans calls, real data, but opened with an "ACCOUNT VERIFICATION FLAG" (the roster then mixed the Venezuelan Apache with German rapper Apache 207 + the Apache Software Foundation). Round 2 (3e73ec51) ✅ sent — flag gone; reports the corrected reality (TikTok @apache_apa 184,300 · Facebook apacheoficial 980,622), zero wrong-Apache mentions. Worked both rounds — the only task that embeds its artist id.

Task 3 — "Weekly YouTube CPM Report" (70956e7e, Wednesdays 0 9 * * 3, artist Apache ebae4bb9…)

Generate a comprehensive CPM analysis report for Apache's YouTube channel. Include: 1) Top 5 videos by CPM, 2) Performance trends, 3) Brand affinity data showing which advertisers are most aligned with the channel, 4) Interactive HTML visuals showing CPM by content type, audience demographics, and ad category. Format as an interactive HTML email that can be forwarded to potential brand partners. Email to stephanie.guerrero@onerpm.com with a clear subject line 'Apache YouTube CPM Weekly Report'.

Artist: implicit (named "Apache" but no id). Outcomes: round 1 (149596f9) ✅ sent — happened to find Apache via the personal roster; honest report (12.5KB HTML, zero fabricated CPM $, cites sources). Round 2 (aa4eb91f) ❌ no email — took the org-scoped path this time: "No 'Apache' artist in this Recoup account." Same prompt, opposite outcomes → coin-flip roster discovery.

Summary matrix

task	artist in prompt?	round 1	round 2	delivered value
1 · Cosculluela daily video	❌ implicit	no email (wrong-roster no-op)	no email (org-roster trap)	0/2
2 · Apache weekly	✅ artist_account_id embedded	✅ sent (+ data-quality flag)	✅ sent, clean + corrected numbers	2/2
3 · Apache CPM	⚠️ name only	✅ sent (honest, lucky path)	no email ("no Apache here")	1/2

The correlation is exact: prompt-embedded artist id → 2/2; implicit artist → 1/4 across rounds, despite artistId being passed on every single run. Injecting the run's artist context into the headless prompt (this issue's primary fix) turns the 1/4 class into the 2/2 class without touching any customer's prompt.