todo2code

Ticket 074: Route semantic LLM calls through SubLLM

• ID: ticket-074
• Owner: unresolved:human
• Status: DONE
• Workflow state: DONE
• Created: 2026-08-12

Goal and scope

Route Subactor-owned semantic executions of todo2code through the central Python subllm package instead of requiring a separately configured direct OpenRouter client. Add the stable todo2code/semantic route, preserve the standalone deterministic mode, and keep provider credentials out of command output, persisted audits and logs.

The todo2code implementation is limited to the existing internal LLM boundary and its focused tests. Central policy and the Platform intent-gate caller are changed in their owning repositories. No new npm runtime dependency or public todo2code API is introduced.

Acceptance criteria

• AC-01: A human owner approves this exact scope and transition to EDIT.
• AC-02: SubLLM resolves todo2code/semantic to direct Z.AI GLM 5.2 when the shared Z.AI credential is valid, with OpenRouter ordered second.
• AC-03: The Node LLM boundary consumes SubLLM’s resolved public route and the selected credential without printing or persisting the credential.
• AC-04: Z.AI requests carry user_id=todo2code, a bounded unique request_id, model=glm-5.2, and no OpenRouter-only fields.
• AC-05: Existing explicit standalone OpenRouter behavior remains available outside the Subactor integration; an explicitly requested but unavailable SubLLM route fails closed rather than silently bypassing policy.
• AC-06: Focused tests, full todo2code verification, governance checks, Docker checks and SubLLM verification pass.
• AC-07: One minimal structured-output request through the production todo2code LLM boundary returns HTTP 200 from direct Z.AI and reports glm-5.2, without exposing the credential.

Participants

• Human participant: unresolved; no user-* file was created by this script.
• Agent participant: ai-codex.md

Approval gate

The human owner approved continuation on 2026-08-12 after reviewing the bounded plan. A required-LLM validation run proved the route and exposed provider-stale audit configuration and parser labels. The bounded repair completed in IN_PROGRESS / VALIDATION; protected exact-head review was subsequently obtained before publication. The approval permitted the scoped implementation; it did not authorize secret disclosure.

Verification evidence

• SubLLM verification passed 43 tests plus Ruff, package build and policy checks. Its public route resolves todo2code/semantic to direct zai, logical and wire model glm-5.2, priority 10; OpenRouter remains the second configured candidate.
• Focused todo2code bridge tests pass 3/3, including direct Z.AI request shape, secret-safe shared credential loading and fail-closed missing-package behavior. The final full host verification passes 428 tests with one expected local JDK skip; an earlier unrelated watch timing flake passed both its isolated retry and the final full rerun.
• Governance passes with 0 errors and 0 warnings. Docker image build, health and doctor smoke pass.
• Local Lizard analysis reports no complexity threshold violations after the first Koru run identified CC=22; isolating subprocess/credential handling reduced the highest reported function complexity to 13 without changing the transport contract.
• A required-LLM pipeline run produced 11 audited responses, all through subllm -> zai -> glm-5.2, using 449,820 total tokens. One documentation response reached the 6,000-token output limit and was partial; the provider route and dependency resolution did not fail. The repaired audit now records the effective SubLLM application, function, provider, model, priority and API base, and response parsing names Z.AI rather than OpenRouter.
• A minimal live structured request through production OpenRouterClient resolved todo2code/semantic via SubLLM to https://api.z.ai/api/coding/paas/v4, returned provider zai, model glm-5.2, response ID 20260812200120928ace40448f43fc and 157 total tokens. No credential or response content was recorded.

Publication evidence

Protected PR #88 passed Koru, hosted verify, required JDK, Docker smoke and review-triggered governance. Independent validator-agent run 31596459305 reviewed exact head 4c35747ae4a8c7ca1fca1187ef56ab0d1675b03e in five chunks; the deterministic authority and advisory openrouter/z-ai/glm-5.2 both returned APPROVE, with no blocking findings. The PR was squash-merged as main@ceae6963279e6dc70fac901c4ee9e7b93cf68b94, and post-merge CI run 31596778339 passed verify, Docker smoke, governance and required Java.

Non-goals

• No automatic replay of a failed paid Z.AI request through OpenRouter.
• No provider or model policy copied into todo2code.
• No new npm runtime dependency or public SDK contract.
• No changes to deterministic extraction semantics.
• No commit of .env, API key IDs, secrets or response payloads containing sensitive input.

This site is open source. Improve this page.