feat: discover OSS MLflow capabilities and model metadata

[dgokeeffe]

Issue and stack

Closes #332.

Depends on #217, #333, #223, #239, and #243. This fork-based PR must target main; review only the capability-discovery commit relative to the pushed integrated consolidation base:

dgokeeffe/ucode@review/integrated-consolidation-base...review/oss-capability-discovery

Equivalent immutable range: 168038b...347f81d. Do not merge before all prerequisites. The incremental range contains no proxy or cache-diagnostic changes.

---

Unit 05 — OSS/MLflow capability discovery and model metadata

Objective and user-visible behavior

Dynamically discover MLflow chat-completions-only models, persist reasoning/context/output metadata, project capabilities onto UC IDs, and render bounded Pi/OpenCode model entries. Unknown or partial metadata degrades conservatively; static GLM/Kimi behavior remains intact.

Candidate: review/oss-capability-discovery (347f81d), incremental base integrated dev (168038b). This is the capability follow-up for existing PR #217; it does not include the repair proxy.

Exact scope

Production: src/ucode/databricks.py, src/ucode/cli.py, src/ucode/agents/pi.py, src/ucode/agents/opencode.py.
Tests/fixtures: tests/conftest.py, tests/test_databricks.py, tests/test_cli.py, tests/test_agent_pi.py, tests/test_agent_opencode.py.
Non-goals: SSE repair/lifecycle, cache affinity/retention, Pi home isolation, GPT off-thinking.

Before / after reproduction

git show dev:src/ucode/databricks.py | rg 'discover_oss_model_specs' || echo '<absent>'
# <absent>

git show review/oss-capability-discovery:src/ucode/databricks.py \
  | rg 'def discover_oss_model_specs|_OSS_MAX_OUTPUT_TOKENS'
# structured discovery function and bounded output table are present

Focused validation:

uv run --frozen pytest tests/test_agent_pi.py tests/test_agent_opencode.py tests/test_cli.py tests/test_databricks.py -q
# 585 passed
uv run --frozen ruff check .
# All checks passed

Full suite:

uv run --frozen pytest -q
# 2 failed, 1763 passed, 36 skipped

Failures: installed Claude capture and the pre-existing Pi preferred-model state assertion. Both reproduce on integrated dev with exact focused commands; this unit does not touch Claude capture or state serialization/default selection.

Live Databricks evidence (sanitized)

A live configured workspace call to discover_oss_model_specs completed without printing workspace, token, or model IDs:

{"live_evidence":"completed","spec_count":14,"all_ids_nonempty":true,"all_reasoning_boolean":true,"context_metadata_count":8,"output_limit_count":13,"reason_redacted":null}

Deterministic tests cover malformed payloads, mismatched entity capabilities, native-API exclusion, duplicate IDs, context-description parsing, UC ID projection, metadata outage fallback, conservative limit completion, and state-to-agent propagation.

Impact map

• Agents: Pi and OpenCode.
• Provider: MLflow chat completions only; native Claude/Gemini/GPT Responses endpoints are excluded.
• State/config: adds oss_model_specs; Pi/OpenCode consume it for reasoning and limits.
• Compatibility: broad dynamic models require live validation; static GLM/Kimi survive metadata outages; unknown limits get conservative complete pairs.
• Managed config/MCP/skills: no registration changes; existing state without specs retains static fallback.

Rollback and residual risk

Revert 347f81d; the original validated GLM/Kimi consolidation remains on dev. Residual risks: free-text context metadata can be absent; output ceilings are probed policy data and require maintenance as models evolve. The fallback direction caps conservatively rather than sending uncapped requests.

Hygiene

git diff --check dev..347f81d passes. Nine-file range; no proxy file, diagnostic scripts, generated files, credentials, uv.lock, .pi-subagents/, or goal.md; no merge markers or unresolved index entries.