Skip to content

Garry's Opinionated OpenClaw/Hermes Agent Brain

TypeScript

docs(opencode): the quota abort leaves rate limits alone; T5 needs a gbrain tool_use event

garrytanec065a86b22b
diff --git a/docs/mcp/OPENCODE-CLI-PIN.md b/docs/mcp/OPENCODE-CLI-PIN.md
index eda123748c..c5fde172ec 100644
--- a/docs/mcp/OPENCODE-CLI-PIN.md
+++ b/docs/mcp/OPENCODE-CLI-PIN.md
@@ -184,8 +184,9 @@ bin-dir prepend).
and prints NOTHING on stdout, in either format; the only trace is a
`message="stream error" … small=false … error.error="…"` line in
`<XDG_DATA_HOME>/opencode/log/opencode.log`. `opencodeOneShotTurn` reads that
- log, ends the turn on a quota refusal, and reports every main-turn stream
- error.
+ log, ends the turn on a quota refusal (a provider rate limit, which
+ opencode retries through, does not end it), and reports every main-turn
+ stream error.
- **MCP tools fire in keyless run mode WITHOUT `--auto`** (verified: the free
model called `gbrain_recall` and returned a seeded per-run nonce with
`--auto` absent). `--auto` exists (`auto-approve permissions that are not
@@ -226,7 +227,8 @@ bin-dir prepend).
reports that no paid turn ran. The free-tier nonce SMOKE (T4) runs only
with `GBRAIN_REAL_OPENCODE_FREE_TIER_E2E=1`, which the non-gating canary sets.
- The separate credentialed lane carries the gating nonce SMOKE (T5, one
- `anthropic/claude-haiku-5-5` turn). The nightly schedule always runs it; on
+ `anthropic/claude-haiku-5-5` turn that, like T4, must show a `gbrain_*`
+ `tool_use` event, since opencode's own file tools could find the nonce). The nightly schedule always runs it; on
a dispatch, set `run_opencode_paid` or the repository variable
`GBRAIN_OPENCODE_PAID_E2E=1`. This authorizes paid provider usage, and
`GBRAIN_PAID_LANES=0` keeps it off everywhere. It requires