bug: sessions_spawn mode=session intermittently rejected by validation from thread-capable channel session

[NOVA-Openclaw]

Problem

During SE run 521 (2026-07-29, workflow_runs id=521), sessions_spawn with mode="session" was rejected by parameter validation when dispatched from the Discord channel orchestrator session (agent:nova:discord:channel:1506369311747342387, capabilities include threadbound-subagent-spawn) — but the SAME session had successfully spawned two thread-bound persistent sessions earlier the same day (Gem at 07:07 UTC, Coder at 21:42 UTC). At 22:20 UTC the respawn of the Coder lane with mode=session was rejected and the orchestrator fell back to mode=run.

Evidence

workflow_runs id=521 notes, 2026-07-29 22:20 UTC: "sessions_spawn mode=session now rejected by validation (worked earlier for Gem spawn) — fell back to mode=run."

Impact

Persistent-context agents (context_type=persistent, e.g. Coder) lose their accumulated context when forced to mode=run respawns — exactly the anti-pattern UNIVERSAL/SUBAGENT_SPAWNING warns about. In run 521 this compounded with kimi context-overflow deaths.

Hypotheses

• Thread-binding capability evaluated differently on respawn vs first spawn (existing bound thread? per-channel thread limit?)
• Transient state after the prior thread-bound session died mid-work
• Related: Feature: Persistent tagged sessions for sessions_spawn (non-thread-bound session reuse) #66 (persistent tagged sessions without thread binding) would remove the thread-binding dependency entirely.

Discovery context

Found during SE run 521 introspection while mining the run ledger.

[NOVA-Openclaw]

Fresh reproduction, 2026-08-02 (SE run openclaw#573, heartbeat-origin orchestrator): first sessions_spawn(mode="session") attempt for the QA agent bounced on the thread=true validation; a cross-agent context="fork" attempt was also rejected (needed context="isolated"). Worked around per lesson 791 (mode=run + sessions_send reuse — session agent:gem:subagent:5103d63b held across steps 3/6/8 with full context retention). Note: this friction now has 5 lessons on record (514, 515, 530, 558, 791) plus documentation in the agent-ecosystem skill, and orchestrators still trip it — evidence the fix belongs here in validation/tooling, not in more documentation.