feat(agent-cli): add ori code harness for terminal-bench

[devin-ai-integration]

TL;DR

Adds code to ORI_AGENTS so terminal-bench can drive Ori's native coding agent (ori code) instead of only the wrapped third-party agents (pi, claude, prime-agent, omp).

What changed?

• ORI_CODE_HARNESS in src/benchmarks/agent-cli/harness.ts:
  • run script: ori code --model "$TB_MODEL" --reasoning-effort <effort> --approvals self-drive --output jsonl --prompt-file /instruction.md, stdin from /dev/null, reasoning.delta/tool.output.delta events filtered before tee to /logs/agent/ori-code.txt
  • image steps install only curl/ca-certificates/git; the bootstrap installs the ori CLI (oriChannel controls the version) and checks ori --version once. agentPackage other than the "ori" sentinel throws, since ori code ships inside the CLI.
  • parseOriCodeStream reads the {kind:"event"|"result"} envelopes: assistant text from assistant.text.delta, tool calls from tool.started, usage/cost/generationIds from the turn.succeeded payload, errors from turn.failed/session.failed/runtime.error and a non-ok result line. inputTokens is reported as-is (ori's inputTokens already includes cached tokens).
  • systemPrompt/appendSystemPrompt/allowedTools/disallowedTools exit 2 before launch — ori code has no flags for them. isolateAgentConfig is a no-op (the sandbox has no user config).
• ORI_AGENTS gains "code"; fake sandbox helper recognises ori code as the agent command.

Why?

We want terminal-bench numbers for the actual ori code path (requested by Louis in #agents-ori), and no existing harness invokes it.

How to test

Event shape was captured from a real headless run against meta/muse-spark-1.3 (bun framework/cli/bin/ori.ts code --model meta/muse-spark-1.3 --reasoning-effort medium --approvals self-drive --output jsonl --prompt-file instruction.md </dev/null), exit 0, 21 NDJSON lines, turn.succeeded carrying usage.{inputTokens,outputTokens,costUsd,generationIds,model} and a final {"kind":"result","ok":true,"sessionId":...} line; the test fixture mirrors it.

Live smoke:

bun run bench -- --benchmark terminal_bench --model meta/muse-spark-1.3 --reasoning-effort medium --limit 1 --solver-config '{"agent":"code"}'

Expect the agent log at /logs/agent/ori-code.txt in the sandbox to start with {"event":{"audit":... and metadata agent: "code", agentTurns: 1, non-empty generationIds.

Benchmark impact

New agent option only; existing agents' scripts and parsers are untouched.

Reviewer focus

• ori code writes .ori/ under the task cwd (/app); tasks whose verifier inspects the directory listing could be affected. Same class of risk as other agents' scratch files, flagging in case we want a different cwd.
• Headless ori code exits non-zero with interaction_pending if the model asks a question; we parse that as an error. ori#2557 removes that path upstream.

Checklist

• Tests cover changed behavior
• Public API or configuration changes are backward compatible, or the break is documented
• Benchmark changes document dataset provenance and licensing (n/a)
• No credentials, private results, or restricted dataset contents are included
• Documentation is updated where needed (no README agent list exists)

Link to Devin session: https://openrouter.devinenterprise.com/sessions/c8c55d9cfa2e43569688382fc7deb436
Open in Devin Desktop: https://openrouter.devinenterprise.com/desktop/session/c8c55d9cfa2e43569688382fc7deb436?variant=devin
Requested by: @louisgv

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR that start with 'DevinAI' or '@devin'.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

---

Original prompt from Louis

SYSTEM:
<latest_message>
Louis Vichy (U0ACWETHNLT) [ts=1788844310.775339]: @Devin with ori cli against terminal bench on it, using the latest muse model -- i.e make sure we're running ori code!
</latest_message>

=== BEGIN THREAD HISTORY (in #agents-ori) ===
Louis Vichy (U0ACWETHNLT) [ts=1788844310.775339]: @Devin with ori cli against terminal bench on it, using the latest muse model -- i.e make sure we're running ori code!
=== END THREAD HISTORY ===
Channel ID: C0BS37QUK7T
Thread URL: https://openrouter.slack.com/archives/C0BS37QUK7T/p1788844310775339?thread_ts=1788844310.775339&amp;cid=C0BS37QUK7T

The <latest_message> is the message that you should use to guide your goals + task for this session, and you should use the rest of the slack thread as context.
A [ts=...] marker on a Slack message is that message's timestamp. To act on a specific message with the slack tool (e.g. adding an emoji reaction via the reaction command), pass that value as timestamp along with the Channel ID — no extra lookup call is needed.