howells GLM review
GLM 5.3 Flash is the cheap second-opinion lane: a bounded contract check, conformance to stated criteria, UI copy, search relevance, or any concrete review that benefits from an independent read. What comes back is a set of leads. Verify each one before you act on it.
The question decides the lane rather than its importance. A check with a right answer comes here. A judgement call with no checkable answer goes to fable-review, which costs a great deal more.
The coding plan is the only route
Run the model as zai-coding-plan/glm-5.3-flash. The same weights are reachable through zai/, openrouter/z-ai/ and opencode-go/; those bill against different accounts and this skill uses the coding plan.
Confirm the lane before the first invocation in a session:
opencode models zai-coding-plan | grep -qx 'zai-coding-plan/glm-5.3-flash' && echo lane-ok
Model listing establishes availability, not successful authentication. Check the actual invocation outcome before reporting a completed review.
Without lane-ok, stop and name the cause: opencode missing from the path, Z.AI Coding Plan credentials absent (opencode auth list shows no Z.AI Coding Plan entry), or the lane authenticated but no longer serving that model. Swapping in another provider, another GLM size, or another model is a failed run.
Build a bounded brief
Investigate first, so GLM gets verified current facts rather than rediscovering the repository. A brief carries:
- the exact question or decision;
- the artifact: paths, relevant excerpts, diff or base commit, current status;
- constraints, settled decisions, and explicit non-goals;
- the review lenses that matter here;
- validation already run, and known uncertainty;
- the boundary, verbatim:
Read-only review. Do not edit files or run state-changing commands.
Ask for a verdict, findings with file:line evidence, disagreements with the proposed direction, and concrete changes.
The brief is done when someone with no access to this session could act on it alone. "Review this repo" fails that bar; narrow the artifact and the questions first.
Invoke
Run from the repository under review. Set REPO to its absolute path, SCOPE to a short label and PROMPT to the complete brief before using the example. Confirm the installed CLI supports the flags with opencode run --help.
opencode run \
--pure \
--dir "$REPO" \
--agent plan \
--model zai-coding-plan/glm-5.3-flash \
--title "GLM review: $SCOPE" \
--format default \
-- "$PROMPT"
Leave --auto off. Plan mode and --pure do not by themselves isolate MCP integrations or prove read-only access.
Before invoking the model, inspect the resolved configuration locally (opencode debug config and opencode debug agent plan); do not print credentials or unrelated configuration into the conversation. Apply a run-scoped configuration that denies every tool except read, glob and grep, denies shell/edit/subagent access, and disables each inherited MCP server by name. In V1, permissions use permission and MCP entries accept enabled: false; V2 uses a different permission schema. Use the installed version's schema and verify the resolved agent's effective permissions before running. Configuration files merge: an empty mcp object or a minimal OPENCODE_CONFIG file does not erase inherited servers or specific allow rules. Do not change the user's persistent configuration.
If the host cannot establish those restrictions, pass the relevant source and diff as the brief in a verified isolated environment, or report that the read-only lane is unavailable. A prose refusal probe is not a substitute for inspecting permissions. Current primary documentation: configuration, permissions and MCP servers.
Retry once on a transport or internal-server failure. Report an authentication, quota, or billing rejection and stop without retrying: those repeat.
Wait for actual output. A started process or a created session is not a review. Enforce a ten-minute wall-clock limit with the host’s process-tree timeout, or a runner using Python subprocess.Popen(start_new_session=True) and communicate(timeout=600) that terminates the process group on timeout. Short tool yields must preserve the running session. Do not assume GNU timeout exists on macOS. At the limit, terminate the run, say so, and stop. Output that arrives empty or cut off mid-findings is a failed run too, and is reported as one rather than salvaged into a partial verdict.
Verify, then report
Check every material claim against the source, the diff, the tests, or the running product. A claim you could not confirm stays labelled unconfirmed.
Return:
- the scope GLM received;
- findings you verified, most serious first;
- claims that failed verification, and what you found instead;
- recommended changes, kept separate from changes already made;
- any lane failure.
Pasting GLM's reply verbatim is not a report. Implement nothing unless implementation was already part of the user's request.