Pre-companion v0.5 feature preservation#

This audit compares the candidate with the actual pre-companion v0.5 source, preserved in the development baseline commit cc88d4d and the separately hashed coordinating snapshot. The public v0.4.1 tag is an additional compatibility baseline, not a replacement for implemented v0.5 behavior. Unreleased status does not authorize removing a capability.

retained means the supported path still works; refactored means behavior is preserved through an integrated path; intentionally gated means an existing explicit opt-in, resource, evidence or supervisor control remains. The audit found a companion integration gap: it discarded the full configuration and called a single-model Bridge directly. Existing MCP/workflow paths were retained, but companion jobs bypassed them. The adapter now repairs that gap. No duplicate workflow engine, legacy registry UI or process implementation was restored.

Preservation matrix#

Paths below are relative to this source tree. Tests use deterministic provider responses unless explicitly identified as live observations below. A test of transport or routing is not a model-quality evaluation.

No requested capability remains classified as missing or broken after these repairs. This does not imply every model performs every advisory role well. Some role/skill pairs need an explicit custom stage; being a reusable catalog definition does not mean every built-in workflow invokes it automatically.

Exact Qwen3 14B preservation#

The installed loopback inventory observed during this audit contained qwen3:14b-q4_K_M, digest bdbd181c33f2ed1b31c972991882db3cf4d192569092138a7d29e973cd9debe8, 9,276,198,565 bytes, advertised 14.8B / Q4_K_M, completion/tools/thinking. The exact name was selected through the candidate inventory parser; no alias or different model was substituted. Advertised capabilities are not quality scores.

The actual pre-companion configuration enables Qwen for general/coding/reasoning, strength 4, minimum tier Medium, weight estimate 8.67 GiB, thinking false. Reasoning and second_opinion bind to that exact name; consensus roles are coding=GPT-OSS and reasoning=Qwen. GPT-OSS remains the substantive preference (strength 5, Large, thinking low), with Qwen3 4B for bounded lightweight work. The quality companion example preserves these enabled routes without copying private paths or changing the user's config.

The Medium-cap fallback test actually reaches the Qwen transport with its exact name and think=false. Fixed consensus separately pins both role models; it never lets a preference replace the independent Qwen opinion. A live bounded consensus probe verifies transport usability only; it is not a new broad model benchmark. No model was downloaded or deleted.

Limits and evidence interpretation#

The original seven MCP tools remain independently usable and, as before, do not enforce project workflow policy. Use the workflow task interface when that policy is required. Exact mode intentionally selects one model; it does not masquerade as adaptive workflow execution. The dock is a queue/status client, not a visual workflow editor. It links decisions to explicit CLI actions.

Completed-step restart is supported. A process dying between the workflow ledger commit and companion result commit can leave missing handoff content. Spent reservations and result digests remain; automatic generation replay is forbidden. Inspect workflow-status, retain the interrupted evidence, and explicitly scope new work if content cannot be recovered. This is not token-position resumption.

The contribution ledger identifies direct Bridge calls, actual multi-stage workflow contributions, durable workflow execution, and fixed sequential consensus separately. Authenticated callbacks remain unsent. Windows tests, mocked providers, live local calls and untested release gates are reported separately in readiness.