feat(cli,mcp): take the requester id from the agent session

[V3RON]

Closes #158

Without --agent-id or SIMLOCK_AGENT_ID, the CLI and simlock mcp now use CLAUDE_CODE_SESSION_ID as claude-code:<id> or CODEX_SESSION_ID as codex:<id>, then the pid as before. One resolveRequesterId in the new src/agent-identity replaces the two copies of the rule. Docs updated in docs/CLI.md and README.md.

Beyond the three comments the spec named, two more comments and one docs/CLI.md sentence were fixed because the change made them false.

Completion conditions

• Two simlock lease in one Claude Code session, and one Codex session: same requester, second gets REQUESTER_ALREADY_LEASED naming the first lease. Checked with a throwaway e2e run against the fake driver.
• Against a gateway, lease --detach then simctl with no --lease runs on that device; another session's id is refused. Same throwaway e2e run.
• status and list --leases show claude-code:<id> / codex:<id>. Same run.
• simlock mcp leases under the CLI's id; a CLI lease after it gets REQUESTER_ALREADY_LEASED. Same run, plus the unit test "the CLI and simlock mcp resolve the same session-derived requester id".
• SIMLOCK_AGENT_ID beats the session id; --agent-id beats both. Unit tests and the e2e run.
• No ids at all: pid-derived, unchanged. New CLI and MCP fallback tests.
• Empty session variable skipped. Unit test, and the e2e run fell back to the pid.
• Docs: "Agent identity", "simlock mcp", README "MCP integration".

Checks: typecheck, typecheck:e2e, lint, format:check, unit (1872 passed), e2e (60 passed). Every new path was broken and failed on a named assertion.

Review

Spec review: 4 findings, 1 fixed. Code review: 9 findings, 5 fixed.

Rejected:

• Extra comment and doc edits beyond the three the spec named (raised twice) — architecture rule 13 requires them; the change made them false.
• Extra test pinning Claude Code over Codex when both are set — the spec says "the first table row"; the code review showed swapping rows stayed green without it.
• Parallel sub-agents and simlock mcp in one session now share one requester and one lease (raised twice) — this is the specified outcome.
• Empty SIMLOCK_AGENT_ID resolves to "" and no test pins it — unchanged behaviour; the spec keeps "defined wins, as today".
• The e2e harness passes process.env through, so under an agent tool e2e runs as the session id — out of scope here; the suite passes with and without the variable.

Written by an agent.

🤖 Generated with Claude Code

https://claude.ai/code/session_01E7hp6xYT1cA3BXRYiUukby

---

Generated by Claude Code

[V3RON]

Manual verification (real iOS simulators, inside a Claude Code session)

All four checks pass on feature/158 (88678bb), using a throwaway SIMLOCK_HOME.

Check	Result
1. Session id becomes the requester	Pass. The grant, list --leases and status all show claude-code:<CLAUDE_CODE_SESSION_ID>.
2. Second lease from the same session	Pass. Exit 13, REQUESTER_ALREADY_LEASED, and the message names the first lease id.
3. simlock mcp resolves the same id	Pass. lease_simulator was refused with REQUESTER_ALREADY_LEASED naming the CLI's lease.
4. Precedence and fallbacks	Pass. SIMLOCK_AGENT_ID gave explicit-agent, --agent-id gave flag-agent, CODEX_SESSION_ID gave codex:codex-thread-1, and an empty CLAUDE_CODE_SESSION_ID gave a bare pid-derived number.

Not covered: the gateway passthrough case (simlock simctl with no --lease against a gateway). The e2e run exercises it.

[V3RON]

Known limitation: simlock mcp under Codex

Found while verifying on real simulators. Under Codex, the CLI gets codex:<id> but simlock mcp gets a pid-derived mcp:<pid>.

Codex starts MCP servers without its own CODEX_* variables. I checked this with a stub MCP server that printed its CODEX_* environment under Codex 0.159.0, and it was empty. Claude Code does pass its session variable through, and that path passes. MCP itself gives stdio servers only a limited environment and has no session id for stdio, so this is client behaviour we cannot rely on.

Effect: an agent that leases through both the CLI and MCP in one Codex session is two requesters and can hold two leases. Each side still enforces one lease per requester.

Decision: leave MCP as it is. The PR now says this in docs/CLI.md, the README and KNOWN-PITFALLS.md. If it matters later, the fix is an optional agentId argument on lease_simulator, filled by the agent from its shell's session variable.