[ralyodio]

With this PR, resumable chat turns survive a deploy. Until now, an answer still streaming when c0mpute.com redeployed was lost. The process that generated it was killed (after Docker's 10 s stop grace), and so was the standby container that served the deploy window.

How it works

• Durable runs. The chat_runs and chat_run_chunks tables (SQLite Cloud) hold each run's request, its output and a heartbeat.
  • The request holds the prompts, so it is sealed with the conversation key.
  • The owning server appends output and refreshes the heartbeat about once a second.
• Takeover. On boot (instrumentation.ts → runManager()) and every 15 s, each server claims runs whose owner hasn't written for 20 s. The claim is a conditional UPDATE on the old owner, so only one server wins.
  • Plain answer: continued from the saved partial text. The request is history + {assistant: partial} + a "continue exactly where it stopped" instruction.
  • Plugin turn: runs again, without its approvals, so an approved destructive tool call is never executed twice; it asks again.
  • After 3 deaths, or past 20 minutes, the run ends with a clear error instead of looping.
• Works from any server. A reattach or DELETE that lands on a server that doesn't own the run reads it from the database (polled). A cancel is recorded there, and the owner stops at its next tick. A server that finds its run taken over stops quietly.
• Finished runs are deleted after an hour.
• Without a database or a key, everything stays in memory, as before.
• The client contract is unchanged: same headers, events and routes. Web, TUI and CLI already reattach on a dropped stream, so they pick up the continued answer with no client changes.

Tests

Web src/lib/chat: 119 pass. The only failure is chat music > node progress…, which fails on master too.

New durable tests, run with two simulated servers on one SQLite database and a shared clock:

• A server SIGKILLed mid-answer: another takes the run over after it goes stale and continues from the saved partial. The replay reads as one answer, and the stored request has no plain-text prompt.
• Only one of two racing servers wins a takeover.
• A server that doesn't own a run follows it from the database to the end.
• A cancel on another server stops the owner.
• A plugin turn re-runs without approvals.
• A run that keeps dying is given up with a message.

tsc is clean in every touched file, including instrumentation.ts. Version bumped to v0.3.59.

🤖 Generated with Claude Code

[github-actions]

ThreatCrush Security Scan

16 finding(s)

HIGH/CRITICAL: 1 | MEDIUM: 11 | LOW: 4

Severity	Rule	Location
HIGH	js-ssrf-outbound-request	apps/web/public/sw.js:24
MEDIUM	js-unescaped-html-sink	apps/web/src/app/blog/[slug]/page.tsx:53
MEDIUM	js-unescaped-html-sink	apps/web/src/app/layout.tsx:109
MEDIUM	sql-template-interpolation	apps/web/src/lib/chat/convo-store.ts:116
MEDIUM	sql-template-interpolation	apps/web/src/lib/chat/convo-store.ts:117
MEDIUM	sql-template-interpolation	apps/web/src/lib/chat/push.ts:476
MEDIUM	sh-remote-script-execution	scripts/dev-setup.sh:25
MEDIUM	sh-eval-expansion	scripts/dev-setup.sh:35
MEDIUM	sh-remote-script-execution	scripts/install.sh:185
MEDIUM	sh-remote-script-execution	scripts/install.sh:210
MEDIUM	sh-remote-script-execution	scripts/install.sh:408
MEDIUM	sh-remote-script-execution	scripts/install.sh:450
LOW	py-ssrf-outbound-request	examples/business-strategist-agent/chat.py:37
LOW	py-ssrf-outbound-request	examples/business-strategist-agent/train.py:54
LOW	secret-generic-credential	node/crates/c0mpute-core/src/webhook.rs:1171
LOW	py-path-traversal	node/crates/c0mpute-core/tests/fixtures/fake-infernet-text:29

Snippets are redacted; ThreatCrush never prints matched credential material.