Release map: next minor — stable migrated platform

[onsails]

Destination

Next minor release: all three agents (him, right, riskoff) stable on microsandbox — auth works, DB healthy, toolchain provisioned, no unexplained prod errors, upgrade machinery actually upgrades. Nothing left to decide before cutting the release.

Notes

• Domain: right-agent platform on microsandbox (post-OpenShell migration, completed 2026-08-22).
• Tracker: GitHub Issues on onsails/right-agent; map membership via "Part of #" in ticket bodies (repo convention; no sub-issue wiring used).
• Existing stabilisation map for riskoff: Stabilise riskoff for release: eliminate unexplained errors #182 (all execution tickets landed; only the 24h soak gate remains).
• Wayfinder skill applies; one ticket per session, claim before work.

Tickets

• Run the 24h production soak + curated error allowlist — STARTED after Heal WAL desync loop in him/right data.db + harden right_db recovery (#197) #203 closure; release remains incomplete until its contiguous 24h gate passes.

Decisions so far

• Replace OpenShell with microsandbox as the Sandbox Backend — code + migration complete (all 3 agents on msb, verified); pending close.

• Stabilise riskoff for release — all error classes fixed; remaining: 24h soak + curated allowlist.

• Verify cross-provider credential substitution under msb (#92) — NO LEAK: per-SNI host gate in pinned SDK 0.6.10 blocks wrong-host placeholders (BlockAndLog); Provider credentials are injected by env-var name, not scoped to the provider's hosts (cross-provider exfiltration) #92 was OpenShell-era, closed; docs corrected.

• Fix silent auth-token load failures (#196) — root cause was 429-blind keepalive probe + silent-None token load, NOT a missing token; fixed 429-aware probe + stdout-tail diagnostics + WARN instrumentation (3c8617a).

• Heal WAL desync loop + harden recovery (#197) — COMPLETE. Deployed single-owner standard-local topology keeps the Aggregator as the sole live database owner. db-repair provides coordinated offline forensic recovery only when corruption is proven. Healthy production databases were not rewritten; accepted windows and the clean coordinated restart had zero WAL/lock/owner recurrence.

• Fix in-guest claude upgrade no-op (#199) — not a no-op: upgrades happened but logging showed only the report's first line and up-to-date matched a substring present in every run; fixed classification + last-line logging (88467ed).

• Migration bookkeeping: #172 closed + #181 refreshed — done.

• Actualize docs and site for the microsandbox era — stale Provider credentials are injected by env-var name, not scoped to the provider's hosts (cross-provider exfiltration) #92 exfiltration notice replaced with per-SNI gating; rest of surface verified clean (3e690be).

• Cleanup dead memory-server + connection Arc (#127) — deleted dead stdio server/codegen path; owned Connection at every caller (3569975).

• Fix curator idle gate restart reset (#135) — curator reads MAX archived-message timestamp; shared delivery IdleTimestamp unchanged (736cd7d).

Not yet specified

• After Run the 24h production soak + curated error allowlist (release gate, from #182) #205's soak runs clean: the actual release cut mechanics (version bump, release-plz run, changelog review).

Out of scope

• Turso cloud chain (Design Turso cloud configuration model #74, Add embedded replica and sync mode #75, Design cloud backup and restore UX #76, Add sync health checks and doctor output #77, Plan migration from local data.db to cloud-backed storage #78) — separate epic.
• Learning/curator v2 (Future work: prefilter recalls Hindsight memory and/or chat history for 'done-before' check #101, Proactive cron prompt-evolution nudge (v2 of cron↔skill linking) #128, Curator: evaluate & tune LLM consolidation quality (after outcome telemetry) #132, Curator: recover archived skills back to active #134) — separate effort.
• Features (feat: Agent Output Canvas — rich HTML documents via Mini App #58 Agent Output Canvas, TG Managed Bots #52 TG Managed Bots, site: replace stylized control-plane card with a real Mini App screenshot #131 site screenshot) — post-release.
• Switch release-plz workflow back to upstream action once PR #2789 merges #45 release-plz upstream switch — blocked on external PR release-plz#2789.
• Provider credentials are injected by env-var name, not scoped to the provider's hosts (cross-provider exfiltration) #92 full endpoint-scoped credential injection — research ticket only (verify msb semantics); full fix only if the research proves a leak.