Iterion

Appearance

ADR-043 — Provider/model support tiers: Anthropic + OpenAI first-class, z.ai/GLM as an opt-in fallback lane ​

Status: accepted (2026-06-23)

Context ​

The 2026-06-23 GLM-5.2 dogfood campaign ran the bot catalogue on a mixed stack — claude_code nodes routed to z.ai/GLM-5.2 (1M context, billed on a separate z.ai credit) with Anthropic/opus + OpenAI/gpt-5.5 forfait as the other lane — to save Anthropic quota and compare. See 2026-06-23-glm-dogfood-campaign.md and /tmp/opus-vs-glm52.md for the run-level detail.

Observed, live:

  • GLM-5.2 strengths: 1M context (digested sec-audit's large triage input where gpt-5.5/forfait stalled ~18 min); separate credit pool (parallel-provider capacity, Anthropic savings); drop-in via the claude_code backend (zai hint).
  • GLM-5.2 weaknesses: structured-output reliability < opus on complex schemas (feature-gap-fill's reviewer looped on missing required field on GLM-5.2 even with the engine missing-field retry, then completed on opus; sec-audit's voters needed retry + an explicit OUTPUT CONTRACT to tolerate GLM); survey verbosity / slow convergence (adr-cartograph ~900 tool-calls); occasional transient throttle
    • the z.ai 5h rate-limit cap tripped mid-campaign.
  • Anthropic/opus: reliable structured output (completed what GLM missed), tighter loop convergence; smaller standard context than GLM; forfait quota to husband.

Decision ​

Anthropic and OpenAI are the first-class model/provider tiers. Bots' default models, schemas, and reliability guarantees target opus/gpt. z.ai/GLM is a supported but second-tier, opt-in lane — best used as the primary of a per-node provider+model failover chain with an Anthropic fallback (provider: "zai:glm-5.2,anthropic:claude-opus-4-8", ADR-004), where its 1M context and cost help while opus catches its structured-output misses.

We are not investing now in making GLM first-class (the work below). It would add real complexity for a second-tier provider; the failover chain already gives GLM's upside with opus's reliability as the safety net.

Follow-ups required IF we later promote GLM to first-class (NOT doing now) ​

  1. Generalize the per-provider+model failover chain to every review-loop bot's claude_code reviewers (feature-dev, feature-gap-fill, branch-improve-loop, whole-improve-loop) — today only sec-audit-source uses it; the others fail hard on a GLM structured-output miss instead of falling to opus.
  2. Schema-robustness pass: bots that consume strict structured output must tolerate a transient missing/short field (bounded retry + defensive coercion) rather than hard-fail — done for sec-audit-source's voters + the engine validateAndRetry missing-field retry; extend the pattern to the other strict-schema nodes.
  3. z.ai rate-limit ergonomics: surface the 5h-cap window + make the failover automatic+seamless (the chain handles per-call fallback; a global "z.ai capped, prefer anthropic for N hours" hint would avoid per-call retries during a cap).
  4. Model-spec registry: GLM coverage in the dynamic registry (ADR-042) is via the curated fallback until models.dev indexes new GLM releases; a z.ai-specific spec source would remove the lag.
  5. GLM survey/convergence prompt tuning (verbosity) if GLM becomes a default.

Consequences ​

  • Bots stay reliable on the first-class lane by default; GLM is available for cost/context wins via explicit opt-in (env model overrides + the failover chain).
  • The dogfood-built features (per-provider+model failover ADR-004, dynamic model-spec registry ADR-042) remain valuable regardless — they make the multi-provider story work without GLM being first-class.