PRISM V2 — Status Pointer

Read this first when resuming. Full plan: 00-manifesto.md. Execution map + Fable-5 handoff: 06-v2-execution-map.md (the build spine — read after this). Algolia-as-DB findings: 02-algolia-as-database.md. Claude Agent SDK executioner findings (live-probed 2026-07-05): 03-executioner-claude-agent-sdk-findings.md.

Where things stand (2026-07-06) — STRATEGIC PIVOT, foundation decisions CLOSED

Two foundation decisions were settled by Arijit this session. They resolve the long-standing contradiction between the manifesto's original "rebuild on Algolia tech" rationale and the later "best-of-breed" architecture lock:

1. North star = STANDALONE PRODUCT FIRST. V2 optimizes to become a sellable, domain-agnostic "Prospect Research Operating System." Algolia is now just the first domain module, not the platform. Consequence: the original Phase-1 driver — "rebuild on Algolia tech so it's an easy internal-Algolia sell" — is RETIRED. Internal-Algolia adoption is no longer the design constraint. (Phase 3 productization is now the thesis, not the speculative far-future.)
2. Stack = BEST-OF-BREED, confirmed and no longer open. Data/RAG = Postgres + pgvector. Executioner = Claude Agent SDK (self-hosted). This closes the executioner question — the 2026-07-05 "ADR still unwritten / evidence not decision" status below is now superseded: the decision is made, best-of-breed wins because the product is standalone, not Algolia-internal. Algolia-as-DB and Agent Studio-as-executioner are not the direction. (Agent Studio may still serve as an optional domain-connector for Algolia-flavored deployments later — not core.)

Other calls locked this session: 3 roles in scope — AE, BDR, Marketer (Sales Leader deferred). First visible milestone = all 3 role doors live (breadth-first). Design system source of truth = the SPA template (already embeds the premium language) → extract into a reusable design-system + Claude Design prompt. Deprioritized to Phase 2: HeyGen/LiveAvatar embodied avatar + Telegram notify (built + staged, parked — get base functionality working first).

Full ordered build plan, workstreams, open research, and the Fable-5 handoff prompt: 06-v2-execution-map.md.

Where things stood (2026-07-05) — superseded on the executioner decision, see above

• Executioner decision is now a live-evidence 5-way, not a 4-way guess. 03-executioner-claude-agent-sdk-findings.md live-probed the actual installed claude-agent-sdk package (v0.2.110, source-inspected, not recalled) against every non-negotiable executioner requirement from Phase 1. Verdict: real primitives exist for multi-step control, subagent dispatch, pluggable state persistence (SessionStore, Postgres-backable), status tracking, and mandatory gating (hooks with decision: "block") — the one real gap is durable/automatic retry, which is Temporal's strength. Also surfaced a 5th candidate this session: the Managed Agents API (/v1/agents + /v1/sessions), Anthropic-hosted rather than self-hosted, with native multiagent coordination + Outcomes (rubric-graded gate) + memory stores — found via docs only, NOT yet live-probed.
• Not yet done: an actual end-to-end proof-of-concept (needs a real ANTHROPIC_API_KEY, not available this session), and the same live-probe treatment for the Managed Agents API. The ADR is still unwritten — this is evidence, not the decision.

Where things stood (2026-07-04, prior)

• Manifesto written, all 3 phases captured. HARD RULE: this lives ONLY here, under docs/PRISM-V2/ — not the vault. An earlier vault mirror (Projects/PRISM/wiki/2026-07-04-v2-manifesto.md) was created before this rule existed and has been removed; do not recreate it.
• Phase 1 research:
• Agent Studio: substantial prior research located in Algolia-Central2 repo, migrated to vault (was a real gap — existed only in-repo before today). Key finding: Agent Studio has no native multi-agent/handoff/workflow/memory primitives (live-probed, confirmed 404 on /teams, /handoffs, /orchestrators, /workflows, /sessions, /memory etc.) — looks like a fit for the chat layer, not the executioner.
• Algolia-as-database: fresh research done, written up in 02-algolia-as-database.md. Mixed verdict: good fit for PRISM's content/search layer (documents, research corpus), weak fit for the executioner's own state-tracking (no transactions, no joins, async-by-design indexing).
• Executioner decision (Temporal vs Agent Studio vs other): NOT resolved, needs Arijit's explicit call. Important context found: Temporal was dropped once already (2026-06-28 ADR, Projects/PRISM/wiki/decisions/2026-06-28-prism-is-chowmes-prism-hermes-instance.md) but specifically because it was redundant alongside Hermes — that objection doesn't hold now that Hermes is being removed entirely. Temporal should be re-evaluated fresh, not treated as closed.
• Phase 2 and 3: captured as vision only, no design work started.

Next when resuming

1. Bring the executioner question (Temporal vs Agent Studio vs other) to Arijit as an explicit decision point — cite the "Temporal objection no longer applies" finding and the Algolia-as-DB split-architecture recommendation (content layer on Algolia, state-tracking layer elsewhere).
2. Once decided, write the ADR and update this manifesto.
3. Phase 2 brainstorming can start once Phase 1's executioner question is settled.