[origin-review] anomalyco/opencode#40142 — CHECK #28

Why this matters

When a turn finishes length, the session loop treated it like stop — subagents exit silently mid-work (empty or reasoning-only) with no error.

Evidence

Tip fae62505cdf0f137ea583d233579ebfcfc38ad55 (base dev):

• Leaf: session loop entry gate + in-loop branch — one continuation if usable text/tools and prior turn was not also length; else OutputLengthError (existing schema/ACP max_tokens mapping)
• Tests claimed: three RED→GREEN cases in packages/opencode/test/session/prompt.test.ts (recoverable / reasoning-only / bound)
• Evidence class: SOURCE + REPORTED (deepseek V4 flash stall). Live suite on tip this pass: UNKNOWN (run not finished in env)
• Thread already has TUI copy suggestion for MessageOutputLengthError — optional leaf, not re-filed here

Invariant: truncation is either one bounded continue or an explicit output-length failure — never a silent successful stop.

Ask (design, light) — prospective; non-blocking

1. Is “exactly one continuation keyed off previous finish” the permanent truncation policy, or should recoverable length share a durable retry budget with other stream recoveries (core: recover gracefully from truncated provider streams anomalyco/opencode#38747 class)?
2. For never-again: should every finish reason that is not a chosen stop be forced through an error-or-continue matrix so new reasons cannot fall through as success?

• tip: fae62505cdf0f137ea583d233579ebfcfc38ad55
• evidence: SOURCE on length→continue / OutputLengthError branch + entry-gate exclude; REPORTED (Raynos deepseek stall); live prompt truncation suite PENDING/UNKNOWN this pass
• pack: pack-02 draft 008
• origin: fix(opencode): surface truncated turns instead of ending the loop anomalyco/opencode#40142

Fork-only archive of a sealed review draft. Not posted on origin.