Morph Migrate

Morph Migrate

I migrate this app onto a Morph open model. Morph serves GLM-5.2 (morph-glm52-744b) on an OpenAI- and Anthropic-compatible API at https://api.morphllm.com, so the migration is usually a base URL + model swap, not a rewrite. I never change the app's behavior beyond the model call, and I keep a one-flag rollback the whole way.

I work in four phases and I stop for approval before I touch any code.

Phase 1 — Explore

Map how this app calls an LLM before proposing anything. Fan out:

• Find LLM call sites and the client construction (SDK, base URL, model id). Search for openai, anthropic, chat.completions, messages.create, baseURL/base_url, model:/model=.
• Identify the current provider and model, and where the API key comes from.
• Check whether Morph is already wired up: look for a MORPH_API_KEY in the environment/.env without printing its value (grep for the name only).
• Note the language/framework and how the app builds (so I can verify later).

Report the call sites, the incumbent provider+model, and the config surface as a short list. No edits in this phase.

Phase 2 — Choose & plan

Ask exactly:

Full migration to GLM-5.2, or a 5% production trial?

Then write a concrete plan for the chosen path — the exact files and lines I'll touch, the env flag that flips it back, and how I'll verify. Show the plan and get explicit approval before editing. Do not edit in this phase.

Phase 3 — Execute

Full migration

• Repoint the LLM client at Morph: baseURL → https://api.morphllm.com, model → morph-glm52-744b, key → MORPH_API_KEY.
• Keep the old path behind a flag so a revert is one env var, e.g. LLM_PROVIDER=morph|incumbent (default to what preserves current behavior if the flag is unset, so an un-set deploy is safe).
• Change nothing else — no prompt edits, no control-flow changes.

5% production trial (canary)

• Add a minimal traffic split keyed off MORPH_CANARY_PERCENT (default 5): that percentage of calls route to Morph (morph-glm52-744b), the rest stay on the incumbent, unchanged. Decide per-request (a hash or RNG at the call site); don't reshape the surrounding code.
• Score both arms with Reflex so the comparison is apples-to-apples. In the app's own code, after each completed turn (Morph arm and incumbent arm), POST https://api.morphllm.com/v1/reflex/predict with header Authorization: Bearer $MORPH_API_KEY and body {"model":"<reflex>","text":"<the turn's text>"}, where <reflex> is a failure mode worth watching (stuck-in-a-loop is a good default; pick one or two that fit this app). Read the predicted label from the classes array in the response, and record it next to which arm served the turn (morph | incumbent) — a log line, a metrics counter, whatever this app already uses.
• This scoring runs in the app at request time, so it's a plain HTTPS call, not the coding-agent reflex_predict MCP tool. The model field is the reflex name, not the LLM; the arm is your own tag. Reflex scoring is observational — it must never gate or alter a response.

Then verify the build compiles. Max 2 fix attempts; if it still fails, stop and report rather than thrashing.

Phase 4 — Verify

• Trigger exactly one request and confirm it reached Morph (assert the response, or check the base URL/model actually used).
• Confirm the Reflex scoring is recording: check that per-arm labels are landing in the log/counter you wired in Phase 3. Then tally the recorded labels by arm and show the failure-rate comparison directly (Morph vs incumbent) from where the app logged them. These are live predicts, so read them from the app's own sink — the Morph dashboard's reflex_summary only rolls up traced turns, not raw /v1/reflex/predict calls, so don't rely on it for the trial's numbers.
• End with the upgrade script for the user:

  When the comparison looks clean, tell me "go to 100%". I'll flip MORPH_CANARY_PERCENT=100 (or set LLM_PROVIDER=morph) and drop the split.

Guardrails

• Observational-safe: never let Reflex scoring change a response or block a call.
• Never change control flow beyond the traffic split. No prompt or tool changes.
• Keep the one-flag rollback (LLM_PROVIDER / MORPH_CANARY_PERCENT) intact.
• Never print the API key. Reference MORPH_API_KEY; check for its presence by name, never echo its value.
• Codex apps: Codex speaks only the OpenAI Responses API, and Morph's chat models serve Chat Completions/Messages, not Responses. Do not fake an env swap. Front Morph with a Responses gateway (LiteLLM) and point Codex's base_url at it — see docs/sdk/components/coding-agents.mdx.