Running GLM — set-glm

The hard part is not switching the model — it is the context window. The Claude Code CLI does not know GLM's window, so it cuts to a conservative value (~200k) — and a long prompt either fails loudly or runs in a silent compact loop. This skill exists so that nobody has to measure that again.

⚠ Switching is always a USER decision. No automatics and no silent fallback: on missing config the runner stops, it does not fall back to Claude. A silent fallback is the worst outcome because the run would still finish — just inside the other framework, with nothing telling you which one.

Usage

set-glm --check                     # config + LIVE probe call — start here
set-glm -p "prompt"                 # one-shot call
set-glm -p "…" --output-format json # structured output (with token- and cost-fields)
set-glm                             # interactive session on GLM
set-glm --print-env                 # what it sets (token masked)

Every other flag passes through to claude unchanged (--allowedTools, --json-schema, --append-system-prompt, …). --model and --autocompact are only added when the caller did not supply them.

Configuration — ONE central file (since 2026-08-29)

~/.config/set-core/providers.json, mode 0600. This is the only place the framework reads a provider credential from; set-glm and the fleet's agent-owner call the same resolver, so the measured launch parameters live in one place and do not drift apart.

set-providers path        # where it is expected
set-providers show        # what is in it — token masked
set-providers migrate     # carry the old glm.env over, ONE command

Precedence has three levels: machine-level default → project override → the request itself. The model is decided per field; the credential and its endpoint are ONE block — a level either provides both or is rejected. A key is issued for one endpoint, and taking the two from different levels is a combination nobody has described: best case a 401, worst case the other bill.

The old glm.env — one release of grace, then gone

what it was	what became of it
./.env in the repo (GLM_* lines)	DISCONTINUED, no longer read — the error says it was discontinued, not merely that no credential was found
~/.config/set-core/glm.env	still works for one release, with a warning on every resolution

The warning names the old file, the new one, and the command. migrate is explicit, never a side effect of a read: a read writes nothing, and even the migration leaves glm.env in place. An existing providers.json is not overwritten silently.

Only the GLM_-prefixed lines are read from the old file — deliberately. A source .env would also pull in ANTHROPIC_API_KEY — exactly the key that would silently redirect the call to the platform bill.

The measured parameters (2026-08-29) — do not re-measure

what	value	env
endpoint	https://api.z.ai/api/anthropic	GLM_BASE_URL
model	glm-5.3-flash — no prefix	GLM_MODEL
context window	900000 (CLAUDE_CODE_MAX_CONTEXT_TOKENS)	GLM_CONTEXT_TOKENS
auto-compact	700k	GLM_AUTOCOMPACT

Measured capabilities: 800,016 input tokens accepted (model_context_window_exceeded at 1.05M, so the window is ~1M) · max output ≥131,072 · the prompt cache works without cache_control (2nd identical call: input=16, cache_read=120,000) · --json-schema, --allowedTools and the Write/Edit tool all work. Returns 200 for: glm-4.6, glm-5.3, glm-5.3-flash, glm-4.5, glm-4.5-air.

Three traps — each one a MEASURED failure

1. ⚠ The model name must not carry a gateway prefix. A zai/glm-5.3-flash-shaped name gets [1214][modelCode: does not exist] 400 back — that is the OpenRouter/LiteLLM format. The runner therefore rejects a name containing / BEFORE starting: otherwise a night loop fails on its first call.

2. ⚠⚠ --autocompact alone is not enough. The CLI clips the threshold to the window it assumes for an unknown model. The same ~250k-token prompt, four variants:

   setting	result
   only --autocompact 700k	Prompt is too long, input=0
   MAX_CONTEXT_TOKENS + DISABLE_…_ENFORCEMENT	OK, input_tokens=524,731
   only CLAUDE_CODE_MAX_CONTEXT_TOKENS	OK — necessary AND sufficient
   only …_DISABLE_UNKNOWN_MODEL_WINDOW_ENFORCEMENT	Prompt is too long

   The second env var is therefore NOT set: the redundant env only hides which one works.

   A compact is NOT an error message — from the outside it looks like a "slow model". The measured failure: compact at 180,174 tokens (227 s), again half a minute later, finally the CLI's own message — "Autocompact is thrashing: the context refilled to the limit within 3 turns of the previous compact, 3 times in a row" — 9.5 minutes, zero lines written. If a GLM run is suspiciously silent for a long time, look for compact_boundary and status: compacting events in the --output-format stream-json log; nothing among the tool events shows it.

3. ⚠ A .env is often git-tracked. In a worktree, a git reset --hard takes the GLM_* lines with it, and the run stops with "GLM_TOKEN is missing". This is one reason the in-repo .env tier was discontinued: the credential's home is the machine-level ~/.config/set-core/providers.json, and one repo's cleanup must not take the others' with it.

Subagents on GLM — the provider comes from the PARENT, not the agent definition

Measured 2026-08-29, three attempts. A subagent inherits the parent session's ENDPOINT (the spawn carries the parent's env forward); the agent definition's model: field only chooses within that endpoint.

parent	the agent's model: field	what happens
Anthropic	glm-5.3-flash	⛔ model_not_found (HTTP 404) — the subagent starts, then falls over
GLM (z.ai env)	glm-5.3-flash	✅ works, modelUsage = glm-5.3-flash only
GLM (z.ai env)	sonnet	⚠ runs, but the call goes to z.ai and is billed as claude-sonnet-5

The recipe is therefore: start the PARENT on GLM. There is no way to switch provider per agent — set-glm is the parent session, and its subagents carry it with them.

⚠ The third row is the trap, and it must be said out loud. Under a GLM parent, a model: sonnet agent does not fall over — z.ai accepts the name — and the modelUsage shows two rows: glm-5.3-flash (costBasis: unknown) and claude-sonnet-5 (costBasis: list, ctx=200000). So the screen shows an Anthropic model name while the bill is z.ai's, and the env holds no Anthropic auth at all. This is the "other bill, same look" defect class — the same one the ledger's provider field closes from the other side.

And it is not relabelling: REAL Claude comes back. Measured — asked "which model family are you?" through the z.ai endpoint with --model sonnet, the answer was "Claude"; the same question with glm-5.3-flash answered "GLM"; on the Anthropic endpoint with sonnet, "Claude" again. So z.ai forwards the Claude call — onto its own bill. (A model's self-report alone is not proof, but the GLM control self-reports as GLM, so the signal is worth something.)

This sharpens what the real risk is: you do not get something other than what you asked for — you get exactly that, on a different bill, and nothing signals it. A model: sonnet agent under a GLM parent is therefore not "lower quality" — it is invisibly billed elsewhere.

Therefore: if you want an agent definition to run under a GLM parent, either leave out the model: field (it inherits the parent's model, and there is nothing to misread) or write a GLM model into it. A model: sonnet under a GLM parent is misleading: it promises what it does not deliver.

Which GLM model for review — measured on five cases

Five code-review cases (counter-race · VAT rounding · date-UTC · cron idempotency · one deliberately clean file to measure the false-positive rate):

model	hits	avg time	avg output
glm-5.3-flash	5/5	12.7 s	222 tokens
sonnet (for comparison)	5/5	5.4 s	74 tokens
glm-4.6	2/2 (short trial)	15.7 s	604 tokens
glm-4.5-air	⛔ 0/2	11.0 s	878 tokens

glm-4.5-air is excluded: twice, on two different defect classes, it missed — once it said "NO BUG" on a real one — and burned 8× more tokens than flash to do it. For a review agent that is the worst outcome: silence sounds reassuring.

glm-5.3-flash is suitable — it also said "NO BUG" correctly on the clean file (it does not manufacture false positives). In exchange it is 2.4× slower and 3× more verbose than sonnet.

⚠ The sample was easy: five self-contained, 5–8-line, textbook defects. A real review is harder — a large file, project context, rules in tension — and this measurement says nothing about that.

What the cost field LIES about

The --output-format json total_cost_usd field computes with Anthropic pricing, so for a GLM run it is not a real bill. For comparing two models, wall-clock and the token counts are the honest measures, dollars are not — and if you write it into a ledger, state next to it that it is price-equivalent.

If you run a loop with it (A/B, night loop)

• Same prompt, same task — otherwise it is not a measurement, it is an anecdote.
• The gate line is part of the measurement. The measured GLM run's work was good, but a lint gate rejected the commit because the model tried to silence a rule with a non-existent ignore directive instead of the repo's own precedent. Only the gate caught it — the test was green. Whoever compares models counts the commits that went through the gate, not the lines written.
• Log events, not just output (--output-format stream-json): otherwise the compact is invisible, and you will blame the model for the slowdown.

Related

• bin/set-glm — the runner; the measured rationale is in the file header, in comments
• bin/claude-local — the same for a local Ollama model (the sibling case)