agent: route interactive wakes through sessions

[ikma-ricon]

Summary

• route interactive shimmer agent through sessions new + foreground sessions wake
• remove ambient pi / AGENT_HARNESS resolution from the interactive path
• preserve optional session resume, optional initial message, caller cwd, selected-agent identity, and runtime environment scrubbing
• document that both interactive and headless wakes require a provider-qualified model

Interactive execution now follows the intended dependency direction: shimmer -> sessions -> pi. Sessions owns Pi resolution and keeps the foreground wake attached to the caller's terminal.

Scope

This is the interactive-runtime slice of #757. It does not implement the dedicated OS-user defaults or --shared-os-user behavior described in that broader issue.

Validation

• mise run test test/agent/agent.bats (40/40)
• mise run test (181/181)
• codebase lint "$PWD" (9/9)
• readme build --check
• git diff --check

A live interactive wake was intentionally not run because this delegated desk was not authorized to mutate live agent sessions. The Shimmer tests prove delegation/arguments/environment behavior, while sessions v0.4.11 owns and tests the foreground Pi/TTY execution path.

[rho-ricon]

Adversarial pass found no blocking issues.

I checked the updated shimmer agent path against the actual diff and sessions v0.4.11 wake/new behavior. The branch now resolves sessions after shimmer_prepare_agent_child_env, creates interactive sessions through sessions new, resumes by passing the provided session ID/name straight to sessions wake, forwards optional interactive --message, preserves selected-agent identity while scrubbing mise/caller env, and uses foreground exec sessions wake rather than resolving Pi directly in shimmer.

Validation I ran:

• mise run test agent passed, 40/40. The installed-sessions presence made the legacy missing-PATH skip fire, but the post-cleanup missing-sessions boundary test ran and passed.
• mise run test passed, 181/181.
• gh pr checks 796 --watch=false shows Ubuntu and macOS checks passing.

One degraded check: I did not launch a real interactive Pi wake because that would create/live-run an agent session, which the review packet explicitly disallowed. I instead verified the boundary through the task diff, the BATS mocks, and sessions v0.4.11 task implementation.

[ikma-ricon]

Manual interactive smoke passed

Or exercised the PR branch end-to-end in a real terminal.

Fresh launch:

SHIMMER_CALLER_PWD=/tmp/shimmer-796-smoke \
  mise -C /Users/rikonor/desks/ikma-shimmer-sessions-20260710/shimmer \
  run agent --model openai-codex/gpt-5.5

Observed:

• created tracked session bed8adeb-aa10-4efa-9d30-405f8e1bd9c5 (ikma-interactive-1783956902);
• opened Pi interactively with no initial message;
• runtime was sessions-owned Pi 0.80.6-kkl.1, provider openai-codex, model gpt-5.5;
• cwd was /tmp/shimmer-796-smoke;
• interactive input and response worked normally.

Resume launch:

SHIMMER_CALLER_PWD=/tmp/shimmer-796-smoke \
  mise -C /Users/rikonor/desks/ikma-shimmer-sessions-20260710/shimmer \
  run agent \
  --session bed8adeb-aa10-4efa-9d30-405f8e1bd9c5 \
  --model openai-codex/gpt-5.5 \
  "Reply exactly: resumed successfully"

The prior conversation was present, the initial message produced resumed successfully, and Pi remained interactive afterward. This closes the previously noted degraded TTY/end-to-end check.

Setup friction unrelated to the branch behavior: invoking the global shimmer shim from a bare /tmp directory initially had no selected shiv:shimmer version. Establishing identity from Ikma's home repo succeeded, after which the branch-direct commands above worked.