fix: centralize Claude and GPT context capabilities

[dgokeeffe]

Issue and stack

Closes #222.

Depends on #217. GitHub must target main for fork PRs; review the concern-isolated incremental range:

dgokeeffe/ucode@036572c...f299e50

Do not merge before #217. After #217 merges, the visible PR diff collapses to this six-file context-policy unit.

---

Unit 03 — shared Claude/GPT context capability policy

Objective and user-visible behavior

Centralize Claude family/version capabilities and GPT token windows so Pi does not truncate custom gateway models to small defaults and Claude Code applies the 1M selector consistently.

Candidate: review/shared-context-policy (f299e50), mandatory PR target review/oss-base (036572c), existing PR #223 after unit 02 is separated.

Exact scope

Production: src/ucode/databricks.py, src/ucode/agents/claude.py, src/ucode/agents/pi.py.
Tests: tests/test_databricks.py, tests/test_agent_claude.py, tests/test_agent_pi.py.
Non-goals: GPT off-thinking mapping, OpenCode routing, dynamic OSS discovery, proxying, cache retention.

Before / after reproduction

git show 036572c:src/ucode/agents/claude.py | rg 'claude_model_supports_1m'
# no match

git show review/shared-context-policy:src/ucode/agents/claude.py | rg 'claude_model_supports_1m'
# shared helper import and call are present

Focused validation:

uv run --frozen pytest tests/test_databricks.py tests/test_agent_claude.py tests/test_agent_pi.py -q
# 371 passed
uv run --frozen ruff check .
# All checks passed

Full suite:

uv run --frozen pytest -q
# 2 failed, 1614 passed, 36 skipped

Only installed Claude/Pi capture tests fail on this older stack. Unit 01/integrated dev resolves the Pi path isolation case; integrated dev independently reproduces the Claude capture failure.

Capability evidence

Table-driven tests cover dotted, system.ai.*, databricks-*, provider-qualified, [1m]-suffixed, malformed, and future-major IDs. Verified policy includes GPT-5.6 Sol/Terra/Luna 1.05M windows, model-specific GPT fallbacks, Claude Opus 4.6+, Sonnet 4.5+, Fable 5, Haiku, and conservative unknown-model behavior. No credentials or endpoint identifiers are recorded.

Impact map

• Agents: Claude Code and Pi.
• Providers: Anthropic and OpenAI Responses metadata only.
• Config: Pi model entries gain explicit contextWindow, maxTokens, reasoning/image metadata; Claude selector may gain [1m] only from shared policy.
• Compatibility: conservative 200k Claude fallback and bounded GPT fallback reduce unknown-ID risk.
• Managed config/MCP/skills: model IDs flow through existing paths; no state or registration schema changes.

Rollback and residual risk

Revert f299e50; this restores duplicated policy and Pi's smaller custom-model defaults. Residual risk is provider capability evolution; update the centralized table and table-driven tests together.

Hygiene

git diff --check review/oss-base..review/shared-context-policy passes. PR #223 must target the OSS base; its target-relative diff is the review contract. Six-file range; no generated files, credentials, uv.lock, .pi-subagents/, or goal.md; no merge markers or unresolved index entries.