v2 session.prompt dropped system and noReply #48345

Activity

1. github-actions

   assigned

   neriousy

   on Sep 10, 2026
2. opencode-agent

   added

   2.0

   on Sep 10, 2026
3. remorses commented on Sep 10, 2026

   

   remorses

   on Sep 10, 2026

   Last edited by remorses

   ContributorAuthor

   More actions

   while waiting for system on prompt/command, i tried session.instructions.entry.put as the replacement.

   that api is capped at 8kb (InstructionEntry.MaxValueBytes in https://github.com/anomalyco/opencode/blob/v2/packages/schema/src/instruction-entry.ts, from #36254). over that it 413s with InstructionEntryValueTooLargeError. no truncate.

   that limit does not make sense as a stand-in for v1 system.

   v1 promptAsync({ system }) took a full session system prompt. kimaki's is tens of kb: discord routing, slash commands, shell extra fields, session ids. AGENTS.md / skill files already inject large instruction text through discovery. hashing and blob storage in #36254 already handle arbitrary json values. an 8kb wall only on the public PUT is an artificial client limit, not a storage one.

   please remove MaxValueBytes (or raise it to something that can actually hold a session system prompt). otherwise entry.put cannot replace system, and every harness has to invent a side channel (file, sqlite, plugin hook) to do what v1 did in one field.

   still want system / noReply on session.prompt and session.command as the real fix. dropping the 8kb cap is the workaround so entry.put after session.create can hold the same text until then.
4. opencode-agent commented on Sep 10, 2026

   

   opencode-agentBot

   on Sep 10, 2026 – with OpenCode Agent

   Contributor

   More actions

   These behaviors are available through separate V2 APIs:

   // Admit a normal user message without starting the agent loop.
   await client.session.prompt({
     sessionID,
     text: userText,
     resume: false,
   })
   
   // Admit model-visible synthetic context without starting the agent loop.
   await client.session.synthetic({
     sessionID,
     text: context,
     resume: false,
   })
   
   // Attach durable, session-scoped system instructions.
   await client.session.instructions.entry.put({
     sessionID,
     key: "kimaki",
     value: system,
   })

   resume: false is the V2 equivalent of admitting input without requesting a reply. Instruction-entry changes apply at the next step boundary. The 8 KiB limit is per entry; larger instruction sets can be split across stable keys such as kimaki.000, kimaki.001, etc.

   session.command immediately executes a slash-command callback rather than admitting a prompt, so prompt admission and system context are handled through the APIs above.

   Closing as covered by the current V2 API.
5. opencode-agent

   closed this as not plannedon Sep 10, 2026
6. nexxeln

   reopened this on Sep 10, 2026
7. nexxeln commented on Sep 10, 2026

   

   nexxeln

   on Sep 10, 2026

   Contributor

   More actions

   looking into the limit thing
8. nexxeln commented on Sep 10, 2026

   

   nexxeln

   on Sep 10, 2026

   Contributor

   More actions

   raised the limit!

   👍React with 👍1Reacted by Tommy D. Rossi
9. nexxeln

   closed this as completedon Sep 10, 2026