test: add omp (oh-my-pi) backend to skills-evals - #31

The first manual eval runs (workflows 26556757739 / 26556758520) had
the driver's `_extract_text` looking for either `type: "assistant"`
(Anthropic-shape) or `type: "text"` with a `part` envelope
(OpenCode-shape) — neither of which omp emits. omp's `--mode json`
wire emits the consolidated final assistant message as
`type: "message_end"` with `message.role: "assistant"` and the full
`message.content[]` array, while streaming `text_delta`/`thinking_delta`
events come wrapped in `message_update.assistantMessageEvent`.

Effect on the failing runs: the auto-yes detector got an empty string
every turn, so even when the agent emitted a 'Reply yes' gate the
driver short-circuited as `completed=True`. That alone wouldn't have
broken the passing tests (smoke + plan_extends both produced their
plans without needing auto-yes; the CI_DIRECTIVE keeps the agent
moving), but it masked a separate failure mode: when the model call
returned 402 / 429 / 500, omp's print-mode subprocess exited 0 and
emitted a final assistant message_end with `stopReason: "error"` and
empty `content[]`. Driver reported `completed=True, exit_code=0`, the
scenario test passed the early subprocess assertions, and the run
finally failed with the cryptic `expected 2 plan files, got 0` —
upstream cause invisible.

Fix:

  - `_extract_text` now reads `type: "message_end"` with role
    "assistant", iterates `message.content[]`, and pulls `text` items
    (filtering out `thinking` and `toolCall`).
  - `_summarize_event` learns the omp event types — message_end,
    message_update wrappers around assistantMessageEvent, turn/session
    boundaries — so CI logs are readable without re-paging the
    transcript.
  - New `_detect_stream_error` scans for the last assistant
    message_end and reports `stopReason: "error"` with the captured
    errorMessage. The driver loop runs it after every turn and flips
    `run.exit_code = 1` when fired, so the scenario test's
    `assert run.exit_code == 0` reports the LLM error directly.

No event-shape changes were needed for the smoke / plan_extends paths
that already passed — both happen to complete without needing the
auto-yes loop, so the prior empty-text bug never bit them.

Signed-off-by: Tanzim Saqib <[email protected]>

Adds omp (oh-my-pi, omp.sh / can1357/oh-my-pi) as a fifth row in the
agentTargets registry. `x-x init --agents omp` is now a real subcommand
flag value — the binary recognizes the key, the interactive picker
lists it, and the install path is bundled / unmerged through the same
loop as the other rows. No special-case branches anywhere in the
install/remove code.

Install path is `.agents/skills` at both scopes, the cross-agent open-
spec path Codex and Copilot already use. Rationale:

  - Officially documented as omp's `agents` skill provider in
    oh-my-pi/docs/skills.md ("priority 70 group ... agents") and
    backed by packages/coding-agent/src/discovery/agents.ts, which
    walks `.agent/` and `.agents/` at every cwd ancestor up to
    repoRoot plus `$HOME/.agents/` at user scope.
  - Symmetric across scopes — no userSkillsRel override needed.
    Codex and Copilot use the same path; install is idempotent so
    `--agents codex,copilot,omp` collapses to one shared
    `.agents/skills/` directory.
  - Cross-platform: node:path on omp's side joins to `.agents\skills`
    on Windows automatically; the Go side composes via filepath.Join.
    Verified through scripts/e2e_test.ps1 mirror.
  - Stays out of omp's native `.omp/...` tree (priority 100), which
    is omp's private session-storage area (`~/.omp/agent/`, owned
    end-to-end by omp's own runtime).

Skills-only — omp's settings live at `~/.omp/config.yml` (Settings →
Memory tab) and its model registry at `~/.omp/agent/models.yml`. Both
are user-owned end-to-end and outside the x-x install scope, same
reasoning as OpenCode and Copilot.

Mirrored through the full constants-drift chain per AGENTS.md:

  - constants.go agentTargets row 4
  - constants_e2e_test.go expected-map adds OMP_SKILLS_REL +
    COPILOT_SKILLS_REL (the latter was already declared in the shell
    file but never locked into the drift check)
  - scripts/e2e_test.sh OMP_SKILLS_REL readonly + two new cases
    (project + user scope) asserting on `.agents/skills/` presence,
    other agents' exclusive paths + per-agent config files absent
  - scripts/e2e_test.ps1 mirror of both readonly and cases, Windows
    backslash separators

skills-evals/tests/conftest.py: flip
AGENT_INIT_VALUE_FOR_KEY["omp"] from "claude" to "omp" (the binary now
recognizes the key) and AGENT_USER_SKILLS_REL_FOR_KEY["omp"] from
`.claude/skills` to `.agents/skills` so the user-scope post-install
log points at the actual install location.

Signed-off-by: Tanzim Saqib <[email protected]>