free-tier: Free usage limit reached follows the session ID; forks work #59280

Summary

Free models return AI_APICallError: Free usage limit reached when I send a prompt to an existing session, but the same prompt to a fork of that session (new session ID, byte-identical history) succeeds in the same minute with the same model and account. New sessions always work. The rejection follows the session ID, not the account, model, directory or content.

Environment

• opencode version: CLI 1.19.4, desktop 2.1.4
• OS: Windows 11, 10.0.26300 (win32 x64)
• Terminal: OpenCode Desktop (Electron); CLI tests in xterm-256color
• Shell: C:\Windows\system32\cmd.exe
• Install/channel: npm global (opencode-ai@1.19.4) and desktop latest
• Active plugins: .opencode/plugins/graphify.js, @dietrichgebert/ponytail

Reproduction

1. Have an existing session that has been used for a while with a free model (I see it with opencode/deepseek-v4.1-flash-free and opencode/big-pickle).

2. Resume it:

   opencode run -s <sessionID> -m opencode/deepseek-v4.1-flash-free "Reply with exactly: TEST"

3. The stream fails immediately with Free usage limit reached; the assistant message is stored with tokens.output = 0 and no parts.

4. Fork the same session:

   opencode run -s <sessionID> --fork -m opencode/deepseek-v4.1-flash-free "Reply with exactly: TEST"

5. The new <title> (fork #1) session accepts the request and replies normally.

6. Resume the original again: it fails again, while the fork continues to work.

Expected Behavior

Resuming a session should behave exactly like resuming its fork: history loads and the request is accepted.

Actual Behavior

A/B run at the same time, same account, same model, same directory:

New sessions in the same directory with the same model work while the old session is rejected.

Things that do not change the result:

• Switching free models (both deepseek-v4.1-flash-free and big-pickle are rejected on the old ID).
• Rebuilding the session storage: opencode export <id> > s.json && opencode import s.json rewrites all message/part rows, and the same ID is still rejected.
• Changing the session directory.

Things that do change the result:

• --fork (new session ID, identical messages), or starting a new session.

Additional Context

Log lines (paths and IDs shortened):

timestamp=2026-10-26T10:03:40.509Z level=INFO  message=stream  providerID=opencode modelID=deepseek-v4.1-flash-free session.id=ses_0d73c9ee… small=false agent=build mode=primary
timestamp=2026-10-26T10:03:41.620Z level=ERROR message="stream error" providerID=opencode modelID=deepseek-v4.1-flash-free session.id=ses_0d73c9ee… error.error="AI_APICallError: Free usage limit reached. Free models work in OpenCode (https://opencode.ai); subscribe to OpenCode Go to keep going: https://opencode.ai/go"

timestamp=2026-10-26T09:56:13.658Z level=INFO message=process session.id=ses_e885c7aa… messageID=msg_…
timestamp=2026-10-26T09:56:13.662Z level=INFO message=stream  providerID=opencode modelID=deepseek-v4.1-flash-free session.id=ses_e885c7aa… small=false agent=build mode=primary
timestamp=2026-10-26T09:56:19.345Z level=INFO message=loop    session.id=ses_e885c7aa… step=1
timestamp=2026-10-26T09:56:19.359Z level=INFO message="exiting loop" session.id=ses_e885c7aa…

Other notes:

• The database is a restored backup, so it contains both storage generations (session_message and message/part). 438 of 440 old sessions already have rows in the current tables, and export/import converts the remaining two, so legacy storage is not the cause. The failure follows the session ID across both generations.
• The desktop shows this as "You've used today's free messages", which sent me looking at quota for days. The account was not out of free usage.
• Separately, opencode run on these older/forked sessions often does not exit after the reply is already written to the DB (the request completes, no output is printed, the process hangs). I can file that separately if it is not the same bug.

Related: #59252, #58284, #57540.

If the free lane actually keeps a per-session ledger, could the error be made explicit (e.g. "this conversation reached its free limit — fork or start a new session")? Right now it reads as an account-wide quota and is indistinguishable from a real one.