fix: split Harness name and version in persona analytics (CLI-1771) - #7162

Summary

• persona.agent fused Claude Code's name, version, and role into one string (claude-code_2-1-233_agent), fragmenting it into ~101 distinct values in analytics and making "how many Claude Code sessions" impossible to query. This splits name and version into two fields.
• Replaces the hand-ported detect-agent signature table (cliv2/internal/persona/agent) with the upstream github.com/vercel/detect-agent module — grows coverage 13 → 20 harnesses and fixes a live bug where plain-CLI OpenCode was undetected (our port matched the launcher's env var, not the one OpenCode itself sets).
• Adds a harness-agnostic regex that splits a version-shaped suffix off any detected name (name[_/]version[_/role]), and canonicalises the remaining fragment against detect-agent's own KnownAgents vocabulary rather than a hand-written alias list.
• New persona.agent_version is written only when a version is found; persona.agent always resolves to claude_code for Claude Code regardless of whether it was detected via AI_AGENT or via environment signature.
• cursor-cli, augment-cli, and every other already-clean harness are unaffected — no version-shaped suffix, so the regex doesn't match and the value passes through unchanged.

Signature renames, flagging ahead of release per the spec: dropping the hand-ported table for upstream detect-agent's own vocabulary changes the canonical string several harnesses report. None of these are remapped back to the old values — intentional, current analytics data for these harnesses is already messed up, so this is a good time to adopt vercel's naming instead of preserving a value we hand-picked ourselves:

This lands in the next release; will also be called out as a Jira comment on CLI-1771.

Datadog: where the new field lands

Confirmed against live analytics-service logs (service:analytics-service "analytics payload"). Today:

custom.analytics-service.interaction.extension.persona.agent = "claude-code_2-1-220_harness"

After this change, agent_version is a new sibling key at the same nesting level, queryable as @analytics-service.interaction.extension.persona.agent_version:

persona:
    agent: claude_code
    agent_version: "2.1.220"
    interactive: false
    interactive_mode: 0

agent_version is omitted from the payload entirely (not an empty string) whenever no version is parsed out.

License bundling — fixed

github.com/vercel/detect-agent has no LICENSE file in its repo (confirmed via the GitHub API), despite tagged releases and external contributors. This hard-failed CI and any local build.

Every third-party Go dependency in cliv2 ships its license text embedded in the CLI binary: cliv2/scripts/prepare_licenses.go runs go-licenses save ./... and writes one LICENSE (or NOTICE/COPYING) file per module path under cliv2/internal/embedded/_data/licenses/, which cliv2/internal/embedded/file.go pulls into the binary via //go:embed _data and surfaces to end users as the OSS notices. This step runs as part of make configure in cliv2/Makefile, ahead of make build.

Reproduced locally on this branch, matches CI exactly:

E... library.go:122] Failed to find license for github.com/vercel/detect-agent: cannot find a known open source license for ".../github.com/vercel/detect-agent@v1.2.0" ...
F... main.go:77] one or more libraries have an incompatible/unknown license: map["unknown":["github.com/vercel/detect-agent"]]
Error running go-licenses save: exit status 1

So make configure / make build fail outright with this dependency in place — this is not a merge that could accidentally ship unnoticed; the build gate stops it. go test ./... still passes (it doesn't go through make configure), which is why the test-plan checkboxes below are green despite this blocker.

The script already has a manual-override mechanism for packages go-licenses can't auto-detect (github.com/davecgh/go-spew, github.com/alexbrainman/sspi, github.com/pmezard/go-difflib, go.dev are each fetched from a hardcoded URL) — but go-licenses save fatal-exits before reaching that step for a package it finds zero license-like files for at all, so the override alone wasn't reachable. Fixed with two changes in cliv2/scripts/prepare_licenses.go:

1. --ignore github.com/vercel/detect-agent on the go-licenses save invocation, so it no longer requires a verdict on that package (dependencies of the ignored package would still be checked, but it has none).
2. A new manual-download entry supplying the Apache-2.0 license text directly (https://www.apache.org/licenses/LICENSE-2.0.txt), matching github.com/vercel/detect-agent's own package.json declaration ("license": "Apache-2.0") — its repo just never committed the file.

Verified both changes are needed: removing the --ignore and keeping only the manual-download entry still reproduces the exact same fatal error, since go-licenses save never reaches the manual-download loop. With both in place, go run scripts/prepare_licenses.go exits 0 and embeds a real Apache-2.0 LICENSE file for github.com/vercel/detect-agent.

Note this asserts Apache-2.0 based on package.json's declaration, not on a license file the copyright holder actually published — still worth an explicit legal nod given that.

Test plan

• go test ./... passes in cliv2 and cliv2-private
• go vet ./... and golangci-lint run clean on changed packages
• go mod tidy -diff clean in both cliv2 and cliv2-private
• New test coverage in cliv2/internal/persona/persona_test.go (TestReport_Agent) at the single seam where persona values are recorded onto analytics — extends the existing fake-analytics-recorder test rather than a new harness, using t.Setenv per case (the upstream module reads the process environment directly, no injection point). Covers every real-world shape observed in production: fused underscore (both _agent and _harness roles), legacy slash format normalising to the same bucket, bare signature detection, an already-clean harness, a fused identifier with no version-shaped segment, an identifier that fails to canonicalise, a redacted (***) identifier, and a version-shaped run with no trailing role.
• Verified against 14 days of production analytics-service logs: no harness other than Claude Code currently reports a version-shaped persona.agent value at any real volume, confirming the ticket's Out-of-Scope call.
• Manual: after merge, re-run the distinct-persona.agent-value count and confirm Claude Code's ~101 buckets collapse to one, with persona.agent_version populated on those events.

Ref: CLI-1771