[JacobStephens2]

Summary

Adds research notes on whether Temporal could carry thirdshift: running Harness sessions on the user's own subscriptions, and running a coding process.

 docs/research/
+└── temporal-for-thirdshift.md   # primitives → thirdshift mapping, AI offerings, Rust SDK, cost, assessment

How Temporal's primitives map onto thirdshift:

Workflow            → one per Command (Run, Spec run, Pickup run, …)
Activity            → each Harness session, each git/gh step
Child Workflow      → Spec run Tickets
Schedule            → crontab Passes (Overlap Policy ≈ Pass lock)
Signal / Update     → Day shift decisions
Worker with 2 slots → ~2 building Runs cap
Workflow Id         → one Run per issue

Findings:

• Subscriptions: Temporal bills only Actions, storage and support. A self-hosted Worker can exec claude -p or codex exec under the user's login. Temporal's agent-framework integrations call model APIs with API keys instead.
• AI offerings: none of them turns an issue into a pull request. temporalio/pi-temporal is the closest analogue, and it uses API keys.
• Rust SDK: GA since 1.0.0 on 2026-09-04.
• Assessment: don't adopt now. Crash recovery and a UI don't yet outweigh the costs:
  • a long-lived Worker daemon, which reverses ADR 0009
  • a server to run, or Cloud
  • determinism and Workflow-versioning rules
  • rewriting the Delivery and Repair loop

Evidence

A docs-only change. Every fact links to its primary source (docs.temporal.io, Temporal source and release notes, the pricing page) and says how it was confirmed. The temporal CLI wasn't installed or run, so the dev server's footprint on the 4-core/8 GB host is unmeasured. The doc says so.

Merge Danger

Door: two-way

Blast Radius: none

A new Markdown file under docs/research/. No code or prompts change.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MrZceMizNVdvpLgvQ1P3PE