feat(harness): continuous conversation execution across providers

[erinepshovel-code]

Purpose

Complete the A0 conversation harness so the user does not have to carry provider switching or long-context continuity manually.

The conversation remains the durable identity. Provider/model changes are internal execution transformations with provenance.

Behavior

• persists bounded harness_state on the owning conversation
• preserves the first task anchor and latest 24 exact messages
• compacts older turns into a bounded historical checkpoint while retaining the full persisted transcript as authority
• routes ordinary single turns through work_harness.run_single_turn
• preserves explicit model pins exactly
• introduces string-compatible typed provider-failure markers so legacy callers still receive sanitized display text while the harness can distinguish a real provider failure from ordinary model prose
• in auto mode, falls back only while the run-scoped registry-tool execution count is zero
• records the first registry-tool dispatch as the replay boundary
• refuses automatic provider replay after that boundary; unresolved continuation is surfaced as hmmm instead of risking duplicate mutation
• records attempted models, fallback count, actual provider, and tool-execution count in usage provenance
• repairs Python ORM drift for max_tool_rounds and inference_mode
• adds harness_state to both declared schemas and the still-active production startup-DDL path

Architectural boundary

METAPAT gate: no root/canon change. Identity (conversation), state (harness_state), relation (provider routing), and transformation (provider execution change) remain separate.

The migration foundation has not yet cut production over to Alembic, so this PR uses the current idempotent startup-DDL authority for harness_state rather than pretending that migration cutover is complete.

Exact-head evidence

Head: aa0ed96108f6858d6a60b81fabb863f63c9c930a

• provider adapter contracts: PASS
• full Python test suite: PASS
• Python backend boot: PASS
• console-tab regression guard: PASS
• production container build: PASS
• clean frontend build / no @replit bundle references: PASS
• PR deployment jobs: skipped by workflow policy

hmmm

Provider quota APIs are not uniformly queryable before a call. Safe automatic continuation is therefore evidence-based: before the first registry-tool dispatch, provider failure may switch providers; after that boundary, A0 preserves the conversation and refuses automatic replay because the preceding tool may have mutated external state.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.