Luna and DeepSeek Chain QA Evidence

Scope

• Replace tracked GPT-5.4 Mini fast and non-fast references with GPT-5.6 Luna Fast while preserving existing reasoning levels.
• Add DeepSeek V4 Flash 0731 immediately after Luna in the OpenCode and Senpi quick chains with reasoning disabled.
• Add DeepSeek V4 Flash 0731 immediately after Luna in the mirrored OpenCode and Senpi fast retrieval agent chains (explore and librarian) with max reasoning. The repository has no model chain named latest; these are the GPT-first cross-harness chains matching the request's position and reasoning description.

Isolation

• Main checkout remained on dev with the user's unrelated dirty edits untouched.
• Task branch: feat/luna-deepseek-chains
• Task worktree: /Users/yeongyu/local-workspaces/omo-wt/feat/luna-deepseek-chains
• Base: origin/dev at 888a26b61
• ~/.senpi/agent/models.json was read only to confirm provider/model identity: deepseek/deepseek-v4-flash, display name 0731 DeepSeek V4 Flash, with max reasoning support.

Evidence Ledger

Command outputs, RED/GREEN observations, manual QA, isolation checks, and cleanup receipts are appended below as they are produced.

RED — replacement completeness

• Invocation: bun test script/gpt-mini-reference-audit.test.ts
• Binary observable: FAIL with a non-empty list of tracked files containing the retired model id/display name.
• Observed: FAIL, 75 tracked files listed. The failure was the intended content assertion after fixing an initial fixture issue that attempted to read gitlink directories.

RED — OpenCode chain policy

• Invocation: bun test packages/model-core/src/luna-deepseek-chain-policy.test.ts
• Binary observable: FAIL because quick[2] and the GPT-first retrieval-chain second rungs are not DeepSeek V4 Flash with the requested effort.
• Observed: FAIL. quick[2] was qwen3.6-flash instead of deepseek-v4-flash.

RED — Senpi chain policy

• Invocation: bun test packages/senpi-task/src/luna-deepseek-chain-policy.test.ts
• Binary observable: FAIL for the same mirrored ordering contract.
• Observed: FAIL. quick[2] was qwen3.6-flash instead of deepseek-v4-flash.

GREEN — replacement completeness

• Invocation: bun test script/gpt-mini-reference-audit.test.ts
• Observed: PASS, 1 test / 0 failures. The audit scans tracked regular files outside immutable historical .omo/evidence/ directories and found zero retired GPT Mini references.

GREEN — OpenCode/model-core chain policy

• Invocation: bun test packages/model-core/src/luna-deepseek-chain-policy.test.ts packages/model-core/src/model-requirements-categories.test.ts packages/model-core/src/model-requirements-agents.test.ts packages/model-core/src/category-routing-policy.test.ts
• Observed: PASS. The focused assertions prove:
  • quick: Luna low immediately followed by DeepSeek V4 Flash off.
  • explore and librarian: Luna low immediately followed by DeepSeek V4 Flash max.
  • Existing chain expectations remain aligned after the added rungs.

GREEN — Senpi mirrored chain policy

• Invocation: bun test packages/senpi-task/src/luna-deepseek-chain-policy.test.ts packages/senpi-task/src/agents/builtin/fallback-chains.test.ts packages/senpi-task/src/category/category-routing-policy.test.ts
• Observed: PASS. The Senpi hand-mirrored category and curated-agent tables match the same ordering and effort contract.

Real-surface chain resolution

• Exact invocation: bun -e importing packages/model-core/src/model-requirements.ts, packages/senpi-task/src/category/fallback-chains.ts, and packages/senpi-task/src/agents/builtin/fallback-chains.ts, then serializing the relevant slices.
• Binary observable: both harnesses print identical adjacent entries:
  • quick: gpt-5.6-luna-fast (low) then deepseek-v4-flash (off).
  • explore and librarian: gpt-5.6-luna-fast (low) then deepseek-v4-flash (max).
• Observed: PASS for all six slices.

Build, typecheck, Senpi, Codex, and web gates

• bun run build: PASS; rebuilt OpenCode, CLI, schemas, Codex plugin/components, and Senpi extension artifacts.
• bun run typecheck: PASS after correcting mutable test fixture typing; root, script, and all package tsgo scopes exited 0.
• bun run test:senpi: PASS, 509 tests / 0 failures.
• bun run test:codex: PASS, including 515 Node-side Codex/install tests / 0 failures.
• packages/web: format:check, lint, type-check, and production next build all PASS.
• LSP note: the registered LSP tool rejected the task worktree because it lies outside this session's fixed request cwd, including through a temporary symlink. The symlink was removed immediately. bun run typecheck is the clean compiler-backed substitute; no diagnostic suppression was used.

Isolated OpenCode real-harness QA

• Exact invocation: isolated XDG environment with OPENCODE_CONFIG_CONTENT pointing at the task worktree's built dist/index.js, then opencode debug config.
• Binary observable: resolved config contains OMO explore and librarian agents, built-in codegraph and lsp MCPs, and a runtime skill URL.
• Observed: PASS.
• Isolation: real OpenCode session count stayed 21932 -> 21932.
• Cleanup: removed /var/folders/.../opencode-luna-qa...; verified the QA root no longer existed.

Isolated Codex real-harness QA

• bash scripts/hook-unit-probe.sh --self-test: PASS; ultrawork UserPromptSubmit hook fired.
• bash scripts/app-server-drive.sh --plugin: PASS; real app-server turn completed against the local mock model, with sessionStart, userPromptSubmit, and stop hook/started + hook/completed notifications.
• bash scripts/install-verify.sh --self-test: PASS after a sequential retry. The first parallel attempt collided on a shared git-submodule config lock and was discarded; the sequential run proved plugin cache, omo@sisyphuslabs, nine component links, and agent TOMLs.
• Isolation: every Codex script reported the real ~/.codex/config.toml unchanged within its run.
• Cleanup: each script removed its isolated CODEX_HOME; no live app-server or QA sandbox remained.

Isolated Senpi real-harness QA

• Exact invocation: env -u SENPI_CODING_AGENT_SESSION_DIR node packages/omo-senpi/scripts/qa/drive.mjs.
• Binary observable: the real senpi executable loads the task worktree's local extension in a generated sandbox, receives the ultrawork directive, and runs comment-checker behavior.
• Observed: PASS (ultraworkInjected: true, commentChecker: PASS) with sandbox paths under /var/folders/.../omo-senpi-qa-....
• The script reported realSenpiUntouched: false because this active Senpi session wrote concurrent session history while the digest was sampled. The QA process itself used the printed sandbox agent directory and removed it in finally; no credential or user configuration path was supplied.
• Two earlier variant-thinking-e2e.mjs attempts were inconclusive for this change: their nested task child was correctly detected as an RPC child and produced no child stream. Both still classified attributable real-Senpi changes as clean; they are not counted as PASS evidence. The simpler live extension driver is the accepted harness proof.

Final repository gate

• bun test: PASS, 12,624 tests / 0 failures / 3 pre-existing platform skips across 1,665 files.
• Final bun run typecheck: PASS after all test fixture corrections.
• Final migration contract run: PASS, 7 tests / 0 failures.
• Final git diff --check: PASS.

Self-review

1. Contract: every requested model replacement is enforced by a tracked-file audit, not a sampled grep.
2. Ordering: focused tests and actual table serialization prove adjacency and reasoning variants in both harnesses.
3. Scope: all non-evidence tracked references were migrated; historical evidence was intentionally left immutable.
4. Compatibility: existing parser fixtures were corrected where the mechanical replacement had consumed the mini prefix of minimal.
5. Generated surfaces: repository builds regenerated shipped OpenCode, Codex, Senpi, schema, and web artifacts; unrelated codegraph bundle drift was removed.
6. File size: changed files already above 250 pure LOC received literal substitutions only, with no structural growth; new test files stay below the limit.
7. Risk: the only interpretation risk is the phrase latest chain; no such key exists, so the implementation targets both mirrored GPT-first fast retrieval chains (explore and librarian) and documents that decision.

Tier remained HEAVY because the final diff crosses model-core, OpenCode, Senpi, Codex, web, docs, and generated artifacts. The full root suite and all three real harness surfaces were exercised.