[Mongoose] Add shared Execution Traces for capability runs

[Nickel85]

Goal

Add shared Execution Traces for capability runs so Mongoose can debug, evaluate, replay, and eventually optimize capability workflows without each capability inventing a private trace format.

Execution traces support the future Loop Runtime but are useful before it exists. They should capture what happened during a capability execution while preserving the local-first, secret-safe runtime model.

Roadmap Position

This extends existing observability, job logging, shared memory, and Njord finance audit/replay work. It does not replace domain audit records. Domain capabilities still decide what their decisions mean; Mongoose provides a common trace envelope for decisions, tool usage, state updates, and outcomes.

This fits the stable local runtime/observability runway better than the v0.9 shared memory milestone. v0.9 may store memory records that traces later reference, but it should not be forced to deliver the complete cross-capability trace system.

Execution Trace Scope

Capability executions should log structured trace events for:

• selected agent and capability
• goal/request context
• loop definition or direct entrypoint used
• decisions and validation results
• tool calls and provider access, with secret redaction
• state reads and state updates, using references where raw data is sensitive
• prompt/template references and model/provider metadata where applicable
• exit criteria checks
• output contract validation
• final outcome, errors, warnings, and user decisions

Traces should support debugging, evaluation, future optimization, prompt-context retrieval, and replay-safe summaries.

Acceptance Criteria

• Defines a provider-neutral execution trace envelope for capability runs.
• Supports both future Loop Runtime executions and existing direct capability executions.
• Captures decisions, tool usage, state references, validation results, exit criteria, and outcomes.
• Redacts secrets and avoids storing unnecessary PII.
• Stores traces under local-first Mongoose state/logging contracts.
• Links trace records to job ids, capability ids, memory records, and domain audit records when available.
• Defines replay-safe summaries for future analysis and prompt context.
• Makes trace collection inspectable through existing or planned run/job/status observability surfaces.
• Includes tests or fixtures for deterministic, LLM-assisted, failed, and user-aborted executions.

Non-goals

• Do not make traces an auto-approval mechanism.
• Do not require traces to store full raw prompts, raw provider payloads, or sensitive external data.
• Do not replace Njord finance audit records, commerce decision logs, or other domain-specific audit models.
• Do not require the full shared Loop Runtime to exist before direct-entrypoint capabilities can emit traces.

Depends On

• [Mongoose] Define shared agent memory provider contract #164 shared agent memory provider contract

Related

• [Mongoose] Add shared Loop Runtime for capability execution #167 shared Loop Runtime
• [Njord] Add finance loop audit log and replay model before guarded execution #153 Njord finance loop audit/replay model
• [Mongoose] Add memory-backed prompt context retrieval for interactive agents #165 memory-backed prompt context retrieval
• [Njord] Add audit log and reconciliation for YNAB writes #21 Njord write audit and reconciliation
• [Commerce] Add commerce decision log for listings, ads, and fulfillment decisions #100 commerce decision log
• [Mongoose] Add Loop Tuner meta-capability recommendations #169 Loop Tuner meta-capability