CODING AGENT · Profile revision 1

Claude Code / Claude Agent SDK

A coding-agent and SDK family with local session JSONL, subagent hierarchy, tool permission decisions, hooks, OTel spans/events, and W3C trace propagation.

coding-agentagent-sdkopentelemetryhuman-control

Why it is in the map

Anthropic documents unusually rich local and exported evidence plus an important service-credential versus application-end-user identity boundary.

Tracevity boundary

A documented field can support reconstruction, but its presence alone does not prove completeness, authenticity, authorization, or external settlement.

Profile findings

What the documentation can—and cannot—establish

Each row distinguishes source evidence from Tracevity's bounded interpretation. “Not documented” describes the reviewed sources; it does not prove an implementation cannot record the artifact.

01

System identity

Documented

Session, service version, span hierarchy, and subagent relationships provide strong execution correlation.

Tracevity interpretation

This does not create a universal versioned agent-definition record.

Limit

Agent-definition version and stable instance identity across every local and distributed artifact are not canonical.

Primary evidence (1)
02

Principal and delegation

Partially documented

Authenticated account or installation identity is documented, with an explicit boundary between service credential and application end user.

Tracevity interpretation

Account presence alone does not establish the initiating application principal or a delegation chain.

Limit

Delegated identity, credential scope/expiry, elevation, and application-user authenticity are not universal.

Primary evidence (1)
03

Instruction and context

Documented

Local sessions can preserve rich instruction, response, and tool context.

Tracevity interpretation

Conversation resumption is not evidence that all external context or filesystem state is unchanged.

Limit

Source fingerprints, complete retrieval provenance, missing-context representation, and exact JSONL field schema are not established.

Primary evidence (1)
04

Decision artifacts

Documented

Plans, refusals, permission decisions, and hook results can survive as explicit recorded artifacts.

Tracevity interpretation

These artifacts do not reveal hidden chain-of-thought or true internal motivation.

Limit

Selected alternative and calibrated confidence are not universal, and capture differs between local session and OTel paths.

Primary evidence (1)
05

Model activity

Documented

Model/API requests, responses or refusals, errors, usage, and latency are representable.

Tracevity interpretation

Complete request/response bodies and exact model version depend on capture controls and emitted attributes.

Limit

Raw bodies are capture-controlled; provider build/version and cache completeness are not guaranteed.

Primary evidence (1)
06

Tool and MCP activity

Documented

Tool/MCP identity, calls/results, errors, timing, permission wait, and subagent correlation are representable.

Tracevity interpretation

Tool execution evidence does not establish provider settlement or the authority behind credentials.

Limit

Retry and downstream provider correlation are not universally complete, and detailed arguments/results can be disabled.

Primary evidence (1)
07

Effects

Partially documented

Attempted tool execution and some local resulting content can be reconstructed.

Tracevity interpretation

Neither a tool result nor file snapshot universally proves provider acceptance or external settlement.

Limit

Canonical effect type, requested effect, acceptance, external state, and state delta are absent for arbitrary tools.

Primary evidence (1)
08

Human control

Documented

Approval/denial, blocking/modification, permission-mode change, interruption, and stop controls are documented.

Tracevity interpretation

Human control is reconstructable only when the selected local or telemetry path retains the relevant event and actor data.

Limit

Takeover, rollback request, compensation, and a complete reviewer-authority chain are not universal.

Primary evidence (1)
09

Outcome

Documented

Runtime success/failure, refusals, tool results, and subagent completion can be reconstructed.

Tracevity interpretation

These are recorded runtime outcomes, not universal external destination verification.

Limit

Externally verified success, ambiguous settlement, rollback, and compensation are not canonical across tools.

Primary evidence (1)
10

Trace integrity

Not documented

Local session history has ordered entries but no documented universal integrity guarantee.

Tracevity interpretation

Plaintext JSONL and optional append-only custom storage are not built-in signing or tamper evidence.

Limit

Clock provenance, immutability, signatures, tamper evidence, and independent verification are not documented.

Primary evidence (1)
11

Portability

Documented

Local JSONL/storage adapters, OTLP, and W3C trace propagation provide multiple portability surfaces.

Tracevity interpretation

These paths do not establish lossless mapping among local sessions, OTel events, and third-party backends.

Limit

Stable JSONL schema version, documented conversions, import destinations, and exact semantic loss are unknown.

Primary evidence (1)
12

Privacy

Documented

Sensitive local and remote capture, retention, persistence controls, and deployment-specific regimes are explicitly documented.

Tracevity interpretation

The regimes must remain separate; no one retention value applies to every Claude Code/SDK deployment.

Limit

This finding is public-doc-only: no private session or backend record was inspected, and redaction/filter coverage varies by capture path.

Primary evidence (1)

Reconstruction reading

Do not collapse these findings into one score.

This profile describes documented evidence surfaces. Whether they are sufficient depends on the reconstruction question and the other identity, authorization, tool, and destination records available.

Potentially useful evidence

  • Identity: Session, service version, span hierarchy, and subagent relationships provide strong execution correlation.
  • Principal: Authenticated account or installation identity is documented, with an explicit boundary between service credential and application end user.
  • Context: Local sessions can preserve rich instruction, response, and tool context.
  • Decisions: Plans, refusals, permission decisions, and hook results can survive as explicit recorded artifacts.
  • Models: Model/API requests, responses or refusals, errors, usage, and latency are representable.

Explicit gaps or uncertainty

  • Integrity: Clock provenance, immutability, signatures, tamper evidence, and independent verification are not documented.

Source register

7 reviewed primary sources

  1. Observability with OpenTelemetryAnthropic · OFFICIAL DOCUMENTATION · observed August 28, 2026 · current
    Open source ↗
  2. Monitor team usage with OpenTelemetryAnthropic · OFFICIAL DOCUMENTATION · observed August 28, 2026 · current
    Open source ↗
  3. HooksAnthropic · OFFICIAL DOCUMENTATION · observed August 28, 2026 · current
    Open source ↗
  4. SessionsAnthropic · OFFICIAL DOCUMENTATION · observed August 28, 2026 · current
    Open source ↗
  5. Session storageAnthropic · OFFICIAL DOCUMENTATION · observed August 28, 2026 · current
    Open source ↗
  6. Explore the .claude directoryAnthropic · OFFICIAL DOCUMENTATION · observed August 28, 2026 · current
    Open source ↗
  7. Data usageAnthropic · OFFICIAL POLICY · observed August 28, 2026 · current
    Open source ↗

Next question

What evidence does your use case require?

Move from documented field presence to an explicit reconstruction target, or examine selected directional format mappings without changing this system's documentation posture.

Trace Reconstruction Requirements Explore compatibility evidence