v0.90: make agent observability rich and capability-aware

[KAVentures]

What this changes

This upgrades agent observation across the supported integrations instead of special-casing Claude Code.

Claude Code

• Combines existing safe lifecycle hooks with Claude Code's stable OTel log events.
• Correlates hook prompt_id with OTel prompt.id so one user request becomes one structural execution.
• Adds observed model calls, model identity, token usage, human decision outcomes, retry exhaustion and subagent handoffs.
• SubagentStart now creates an explicit parent handoff plus child run boundary.
• Adds a dedicated write-only /agent-ingest/v1/claude-otel endpoint with a strict structural allowlist.
• One-click Connect now installs hooks + logs-only OTel configuration while explicitly leaving prompt/response/tool-content logging disabled.
• Existing foreign telemetry settings are never overwritten; Disconnect removes only values that still exactly match what OWG wrote.

Codex

• One-click setup now exports both Codex structural OTel logs and traces to the same write-only endpoint; prompt, agent-response and guardian-rationale content logging remains disabled.
• Uses turn IDs when available, extracts model/token usage from safe structural attributes when emitted and maps multi-agent spawn communication to handoffs.
• Handles current Codex telemetry where semantic event names can be replaced by tracing call-site names using only low-cardinality structural fields.
• Detects older OWG trace-only managed blocks as partial so Connect upgrades them rather than falsely reporting full coverage.

OpenAI Agents SDK

• Adds a stable adapter sensor identity.
• Response spans now contribute model identity and token usage without serializing response/input content.
• Existing tool, hierarchy and native handoff support is retained.

Generic OTel / custom agents

• Adds capability metadata instead of pretending every integration can observe every signal.
• Portable GenAI model/tool/usage/timing remains supported; provider-specific handoffs and approvals are explicitly unavailable unless a native adapter supplies them.

Dashboard + MCP

• 0 now means an observable signal was observed zero times.
• — means the active integration cannot observe that signal (or has not declared the capability).
• Rich reports include model calls, tools, handoffs, approval requests, token totals, model identity and telemetry depth without productivity scoring.
• Claude hook + OTel duplicate tool evidence is de-duplicated at read time.
• Compact MCP get_agent_runs exposes the same capability-aware summaries instead of dropping the richer signals.

Privacy

No prompts, assistant responses, reasoning, tool arguments/results, transcript paths, raw log bodies or arbitrary OTel attributes are added to canonical evidence. Provider-native payloads are treated as untrusted input and pass through strict allowlists plus the existing canonical content-field rejection.

Verification added

• Claude OTel privacy + structural projection
• Claude hook/OTel turn correlation and de-duplication
• Claude one-click telemetry conflict/reversibility behavior
• Codex logs+trace configuration and trace-only managed-block upgrade
• Codex turn/model/token + multi-agent spawn projection
• Codex semantic fallback without reading message content
• OpenAI Agents response model/usage projection
• write-only Claude OTel route auth/privacy
• capability-aware dashboard syntax/semantics

Version bumped consistently to 0.90.0.