Model registry cleanup + engine reliability fixes

engines/mock_worker.py

@@ -13,6 +13,10 @@ def has_fail_directive(spec: str) -> bool:
     return any(line.strip() == "MOCK_FAIL" for line in spec.splitlines())
 
 
+def has_engine_down_directive(spec: str) -> bool:
+    return any(line.strip() == "MOCK_ENGINE_DOWN" for line in spec.splitlines())
+
+
 def parse_blocks(spec: str) -> list[tuple[str, str]]:
     lines = spec.splitlines()
     blocks: list[tuple[str, str]] = []
@@ -73,6 +77,9 @@ def main(argv: list[str]) -> int:
         return 2
 
     spec = argv[-1]
+    if has_engine_down_directive(spec):
+        print("mock-worker: 402 Payment Required: usage balance exhausted", file=sys.stderr)
+        return 1
     if has_fail_directive(spec):
         print("mock-worker: simulated failure")
         return 1

engines/opencode-sandboxed.sh

@@ -5,7 +5,9 @@
 # flag (required for headless runs) disables ALL of its interactive approval
 # prompts. This wrapper supplies the real containment: full network and reads,
 # writes confined to the task dir, a per-run scratch/cache dir, and OpenCode's
-# own state dirs.
+# own state dirs. The scratch root also carries a private per-run XDG_DATA_HOME,
+# which is what keeps concurrent workers off one shared session database (see the
+# comment on that export below — it is a correctness fix for parallel runs).
 #
 # Usage (as a ringer engine bin):
 #   opencode-sandboxed.sh <taskdir> [--no-sandbox] <opencode args...>
@@ -28,6 +30,9 @@ if ! OPENCODE_BIN="$(command -v opencode)" || [ -z "$OPENCODE_BIN" ]; then
 fi
 
 if [ "$SANDBOX" = "0" ]; then
+  # Full-access mode keeps OpenCode's real ~/.local/share/opencode, so it also
+  # keeps the shared-session-DB contention documented below. Fine for a lone
+  # full-access task; do not fan out several at once.
   exec "$OPENCODE_BIN" "$@" < /dev/null
 fi
 
@@ -59,7 +64,8 @@ cat > "$PROFILE" <<'SBEOF'
   (subpath (param "SCRATCH"))
   (subpath (param "OC_SHARE"))
   (subpath (param "OC_STATE"))
-  (subpath (param "OC_CONFIG")))
+  (subpath (param "OC_CONFIG"))
+  (subpath (param "OC_BASE")))
 ; /dev is needed for /dev/null, /dev/urandom, etc.; writes there can't create
 ; persistent files without root, so a few literals are allowed rather than via param.
 (allow file-write-data
@@ -72,6 +78,25 @@ export TMPDIR="$SCRATCH"
 export XDG_CACHE_HOME="$SCRATCH/cache"
 mkdir -p "$XDG_CACHE_HOME"
 
+# Per-run DATA root — this is a concurrency fix, not just containment. OpenCode
+# keeps its session store at $XDG_DATA_HOME/opencode/opencode.db, defaulting to
+# ~/.local/share/opencode: one file, shared by every worker, opened for write,
+# growing without bound. On 2026-07-27 it stood at 1.78 GB and two of three
+# workers launched simultaneously died before their first tool call with
+# "Error: Unexpected error / database is locked" — recorded as model failures
+# though no model had run. Since every non-Codex model routes through this
+# engine, shared-DB contention penalises exactly the cheap tier, and worsens as
+# parallelism rises. A private store per worker removes the contention entirely.
+# Credentials live beside the DB, so seed the fresh root with auth.json — copy,
+# never symlink: the profile denies writes outside SCRATCH, and a symlink would
+# resolve straight back to the shared file this exists to avoid.
+export XDG_DATA_HOME="$SCRATCH/share"
+mkdir -p "$XDG_DATA_HOME/opencode"
+if [ -f "$HOME/.local/share/opencode/auth.json" ]; then
+  cp "$HOME/.local/share/opencode/auth.json" "$XDG_DATA_HOME/opencode/auth.json"
+  chmod 600 "$XDG_DATA_HOME/opencode/auth.json"
+fi
+
 # Run as a child (not exec) so the EXIT trap fires and cleans up the profile +
 # scratch dir even on the success path; propagate the child's exit status.
 set +e
@@ -81,6 +106,7 @@ set +e
   -D "OC_SHARE=$HOME/.local/share/opencode" \
   -D "OC_STATE=$HOME/.local/state/opencode" \
   -D "OC_CONFIG=$HOME/.config/opencode" \
+  -D "OC_BASE=$HOME/.opencode" \
   -f "$PROFILE" "$OPENCODE_BIN" "$@" < /dev/null
 status=$?
 set -e

registry/model-identity.toml

@@ -64,6 +64,16 @@ confidence = "verified"     # Cursor (Anysphere) model, served through xAI's Gro
 source = "https://cursor.com/blog/composer-2-5"
 last_verified = 2026-07-10
 
+[engines.grok.models."grok-4.5"]
+# Task-manifest "model" field for grok-engine tasks logs "grok-4.5" directly
+# (not the "grok-build" default_model_key above) — this is the actual slug
+# ringer observes on grok-engine attempts. Same artifact as grok-build.
+display = "Grok 4.5"
+lab = "xAI"
+confidence = "verified"
+source = "https://docs.x.ai/developers/release-notes"
+last_verified = 2026-08-01
+
 [engines.opencode]
 harness = "OpenCode"
 access = "OpenRouter API"
@@ -98,3 +108,119 @@ lab = "Meta"
 confidence = "verified"
 source = "manifest model field; OpenRouter slug meta-llama/llama-3.3-70b-instruct:free"
 last_verified = 2026-07-10
+
+[engines.opencode.models."openrouter/anthropic/claude-sonnet-5"]
+display = "Claude Sonnet 5"
+lab = "Anthropic"
+confidence = "verified"
+source = "manifest model field; OpenRouter slug anthropic/claude-sonnet-5"
+last_verified = 2026-08-01
+
+[engines.opencode.models."openrouter/cohere/north-mini-code:free"]
+display = "North Mini Code (free)"
+lab = "Cohere"
+confidence = "verified"
+source = "manifest model field; OpenRouter slug cohere/north-mini-code:free"
+last_verified = 2026-08-01
+
+[engines.opencode.models."openrouter/deepseek/deepseek-v4-flash"]
+display = "DeepSeek V4 Flash"
+lab = "DeepSeek"
+confidence = "verified"
+source = "manifest model field; OpenRouter slug deepseek/deepseek-v4-flash"
+last_verified = 2026-08-01
+
+[engines.opencode.models."openrouter/deepseek/deepseek-v4-pro"]
+display = "DeepSeek V4 Pro"
+lab = "DeepSeek"
+confidence = "verified"
+source = "manifest model field; OpenRouter slug deepseek/deepseek-v4-pro"
+last_verified = 2026-08-01
+
+[engines.opencode.models."openrouter/inclusionai/ling-2.6-flash"]
+# InclusionAI is Ant Group's AGI/open-source publishing org (home of the Ant
+# Ling team) — see https://huggingface.co/inclusionAI.
+display = "Ling 2.6 Flash"
+lab = "InclusionAI (Ant Group)"
+confidence = "verified"
+source = "https://huggingface.co/inclusionAI"
+last_verified = 2026-08-01
+
+[engines.opencode.models."openrouter/inclusionai/ling-3.0-flash:free"]
+display = "Ling 3.0 Flash (free)"
+lab = "InclusionAI (Ant Group)"
+confidence = "verified"
+source = "https://huggingface.co/inclusionAI"
+last_verified = 2026-08-01
+
+[engines.opencode.models."openrouter/moonshotai/kimi-k2.5"]
+display = "Kimi K2.5"
+lab = "Moonshot AI"
+confidence = "verified"
+source = "manifest model field; OpenRouter slug moonshotai/kimi-k2.5"
+last_verified = 2026-08-01
+
+[engines.opencode.models."openrouter/moonshotai/kimi-k2.6"]
+display = "Kimi K2.6"
+lab = "Moonshot AI"
+confidence = "verified"
+source = "https://openrouter.ai/moonshotai/kimi-k2.6"
+last_verified = 2026-08-01
+
+[engines.opencode.models."openrouter/nvidia/nemotron-3-nano-30b-a3b:free"]
+display = "Nemotron 3 Nano 30B A3B (free)"
+lab = "NVIDIA"
+confidence = "verified"
+source = "manifest model field; OpenRouter slug nvidia/nemotron-3-nano-30b-a3b:free"
+last_verified = 2026-08-01
+
+[engines.opencode.models."openrouter/nvidia/nemotron-3-super-120b-a12b"]
+display = "Nemotron 3 Super 120B A12B"
+lab = "NVIDIA"
+confidence = "verified"
+source = "manifest model field; OpenRouter slug nvidia/nemotron-3-super-120b-a12b"
+last_verified = 2026-08-01
+
+[engines.opencode.models."openrouter/nvidia/nemotron-3-ultra-550b-a55b:free"]
+display = "Nemotron 3 Ultra 550B A55B (free)"
+lab = "NVIDIA"
+confidence = "verified"
+source = "manifest model field; OpenRouter slug nvidia/nemotron-3-ultra-550b-a55b:free"
+last_verified = 2026-08-01
+
+[engines.opencode.models."openrouter/qwen/qwen3-coder"]
+display = "Qwen3 Coder 480B A35B"
+lab = "Alibaba (Qwen)"
+confidence = "verified"
+source = "manifest model field; OpenRouter slug qwen/qwen3-coder"
+last_verified = 2026-08-01
+
+[engines.opencode.models."openrouter/thinkingmachines/inkling"]
+display = "Inkling"
+lab = "Thinking Machines Lab"
+confidence = "verified"
+source = "https://openrouter.ai/thinkingmachines/inkling"
+last_verified = 2026-08-01
+
+[engines.opencode.models."openrouter/poolside/laguna-m.1:free"]
+# Not a typo — OpenRouter carried this slug from 2026-07-09 until it was
+# delisted 2026-08-01T23:13:22Z (same day as this entry, hours before this
+# registry check ran). Confirmed via ~/.ringer/openrouter-catalog.changes.jsonl
+# added/removed events, catalog name "Poolside: Laguna M.1 (free)". Historical
+# attempts against this slug should still resolve to Poolside identity even
+# though the model is no longer orderable.
+display = "Laguna M.1 (free)"
+lab = "Poolside"
+confidence = "verified"
+source = "OpenRouter catalog history (~/.ringer/openrouter-catalog.changes.jsonl): added 2026-07-09, delisted 2026-08-01"
+last_verified = 2026-08-01
+
+[engines.opencode.models."openrouter/nousresearch/hermes-3-llama-3.1-405b:free"]
+# Not a typo — OpenRouter carried the free tier of this model from 2026-07-09
+# until it was delisted 2026-07-19; the paid nousresearch/hermes-3-llama-3.1-405b
+# slug remains live. Catalog name "Nous: Hermes 3 405B Instruct (free)".
+display = "Hermes 3 405B Instruct (free)"
+lab = "Nous Research"
+confidence = "verified"
+source = "OpenRouter catalog history (~/.ringer/openrouter-catalog.changes.jsonl): added 2026-07-09, delisted 2026-07-19"
+last_verified = 2026-08-01