Classify agents on every command invocation

[gtsiolis]

Tags every command invocation in CLI telemetry so we can answer "are agents using lstk?". Detection is environment-based (the strongest signal) with a TTY fallback for humans, and is attached to every telemetry event. A detected agent is also forwarded into the emulator container as AI_AGENT.

Agent and CI are tracked as independent axes — an agent can run inside CI, so each is recorded separately rather than collapsed into one mutually exclusive value. Every event carries is_ci plus agent_identity / ci_identity when detected, and a derived caller_type (agent/ci/human) for single-label segmentation. This mirrors the legacy CLI (localstack/localstack-cli-standalone#19), where is_ci is its own field independent of agent detection.

See PRO-276 for more context.

Notes

• Orthogonal axes — an agent running inside CI is recorded on both (agent_identity and is_ci/ci_identity); caller_type derives to agent (agent > ci > human) so segmentation stays single-label without discarding the CI signal.
• is_ci is sourced from the single caller classifier (14 known CI providers), not a separate bare CI env lookup.
• AI_AGENT on the container is session-origin — it reflects who started the container, not who makes later calls into it, and is forwarded whenever an agent is detected (including inside CI).
• Detection-only, no free-text override, so no user-supplied value reaches telemetry and no sanitization is needed.

[gtsiolis]

Cc @carole-lavillonniere @anisaoshafi because #58, #127, etc.
Cc @andrewrjones @purcell @whummer because https://github.com/localstack/localstack-cli-standalone/pull/19

[carole-lavillonniere]

question: why do we have a type with mutually exclusive values (ci can also be agent but we can't represent this here), when instead we could have 2 binary fields: agent=yes/no ci=yes/no? Data model could become something like:

  is_agent / agent_identity   (cursor, claude-code, …)
  is_ci    / ci_identity      (github-actions, jenkins, …)
  detection_method

Then we could also remove the other lookup in internal/telemetry/client.go line 77 but instead use the value from internal/caller/caller.go

[gtsiolis]

Nice catch, @carole-lavillonniere! ⚾

I was trying to mirror the approach from localstack/localstack-cli-standalone#19, but you're right! Made some changes in 27b8021, and updated PR description. Could you take another look and take it over from here? 🙏