Add long-term (OCG-backed) agent memory compilation to agentspan-server

[NicholasDCole]

Summary

Adds OCG-backed long-term memory to AgentSpan agents, with deterministic recall before execution, best-effort raw run capture after execution, and completed-run feedback plumbing. It also fixes MCP tool compilation/execution and improves agent execution observability in the OSS UI.

Ports and extends agentspan-ai/agentspan#298 for Conductor OSS.

What changed

Long-term memory and recall

• Adds longTermMemory configuration with OCG URL, server-side credential name, agent identity, and optional user identity.
• Compiles a root-only cg_search_memories call before the first model or sub-agent task.
• Normalizes, bounds, and injects recalled content as explicitly untrusted supporting context.
• Propagates root recall to compiled sub-agents without issuing duplicate child recalls.
• Keeps recall optional so OCG failures do not fail the agent run.

MCP tool compilation and execution fixes

• Fixes MCP config.tool_names being ignored, which previously caused a configured six-tool OCG allowlist to expose all 14 discovered tools.
• Resolves explicit OCG tools from bundled schemas at compile time, avoiding unnecessary LIST_MCP_TOOLS calls while preserving generic discovery when no explicit list is configured.
• Rejects unknown explicit OCG tool names instead of silently widening access.
• Preserves credential-header placeholders while tools pass through the model and resolves them only on the executable CALL_MCP_TOOL task.
• Stamps _agent_tool_name onto dynamic MCP tasks and preserves it through JOIN compaction, including agent-as-tool sub-workflow inputs, so results are attributed and merged back into the next model turn correctly.
• Restricts normal model-callable OCG tools to the configured query/graph operations; memory mutation, sharing, history, and cleanup tools are not exposed.

Run capture and feedback

• Adds a reusable OCG client with server-side credential resolution, outbound-origin validation, bounded requests, retries, and contained failures.
• Exports raw root execution history asynchronously for completed, failed, timed-out, and terminated runs; OCG remains responsible for summarization, folding, ranking, retention, and indexing.
• Adds GET/POST /api/agent/executions/{executionId}/feedback and completed-execution feedback controls in the UI.
• Fails feedback closed with OCG_FEEDBACK_CONTRACT_UNAVAILABLE until OCG provides the required API-key-authenticated turn feedback contract; no guessed memory key or browser-side OCG credential is used.

Agent execution UI

• Aggregates prompt, completion, and total tokens across the complete sub-workflow tree on the server.
• Loads agent execution details through one /full request instead of recursively fetching children in the browser.
• Displays semantic MCP labels such as MCP_TOOL_CG_QUERY while retaining CALL_MCP_TOOL as the underlying Conductor task type.

Compatibility and robustness

• Adds top-level-only agent execution filtering.
• Defaults workflow media input to an empty list, preventing null-media failures for webhook and non-media runs.
• Adds the implementation design and OCG memory configuration documentation.

Security and failure behavior

• Agent definitions contain credential names, never API key values.
• OCG credentials are resolved and used only by the server.
• Recall, capture, and feedback failures do not change the workflow result.
• Child workflows do not independently export or duplicate the root memory lifecycle.

Verification

• ./gradlew :conductor-agentspan:test
• ./gradlew spotlessApply
• pnpm test -- --run in ui-next (719 passed, 1 skipped)
• pnpm typecheck in ui-next
• ESLint and Prettier checks for the changed UI files
• Live verification of aggregate token counts and semantic MCP labels in the OSS execution UI

Pull request type

• Feature
• Bugfix
• Refactoring
• Build-related change