chore(deps): bump agent-runtime ^0.50 + agent-eval ^0.91

[drewstone]

Pure dependency bump of the Tangle substrate pins to the canonical fleet versions.

• @tangle-network/agent-runtime: ^0.36.0 → ^0.50.0
• @tangle-network/agent-eval: ^0.70.0 → ^0.91.0

Widest skew in the fleet (14 minors on runtime). The repo's evals import surface is the canonical pair {createOpenAICompatibleBackend, runAgentTaskStream} + AgentTaskSpec — non-breaking across this jump.

Verification: npm run typecheck:evals passes clean (exit 0) under the bumped pins. The one-shot judge divergence in evals/src/profiles/types.ts (EvalProfile) is intentional and left untouched. Merges cleanly into main.

[tangletools]

✅ Auto-approved PR — d41e69ad

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-14T00:53:09Z}

[tangletools]

🟠 Value Audit — better-approach-exists

Verdict	better-approach-exists
Concerns	1 (1 strong-concern)
Heuristic	0.0s
Duplication	0.0s
Interrogation	143.6s (2 bridge agents)
Total	143.6s

💰 Value — sound

Bumps the root evals package to canonical Tangle substrate versions (^0.91 agent-eval, ^0.50 agent-runtime) and keeps the existing single-point runtime abstraction intact; typecheck:evals passes.

• What it does: Updates package.json:38-41 and package-lock.json to pin @tangle-network/agent-eval to ^0.91.0 (from ^0.70.0) and @tangle-network/agent-runtime to ^0.50.0 (from ^0.36.0). The lockfile now resolves agent-eval 0.91.0 and agent-runtime 0.50.0, matching the newer peer-dependency ranges agent-runtime declares (@tangle-network/agent-eval >=0.83.0 <1.0.0, `@tangle-network/sandbox >=0.1.2 <0.
• Goals it achieves: Closes the 14-minor fleet skew on agent-runtime and brings the eval harness onto the current substrate line so it receives upstream runtime/eval fixes, features, and compatible peer ranges. The repo's eval import surface (createOpenAICompatibleBackend, runAgentTaskStream, AgentTaskSpec from agent-runtime; campaign/trace APIs from agent-eval) is preserved as non-breaking across this jump.
• Assessment: Good change. It is scoped exactly to the package that consumes these dependencies (only ./package.json references them; arena/package.json, sdk-ts/package.json, and the CJS tool package do not). The evals already centralize agent-runtime usage behind evals/src/sim/llm-call.ts:33-34, so the bump touches a single abstraction boundary rather than scattered call sites. Verification held: after `npm
• Better / existing approach: none — this is the right approach. Searched the workspace for package.json files and agent-runtime/agent-eval imports: only the root package depends on these libraries, and only evals/src imports them. There is no duplicate dependency set to consolidate and no alternative abstraction already present that should absorb this bump. A workspace-wide bump would be wrong because the other packages do no

🎯 Usefulness — better-approach-exists

Bumps eval substrate to current runtime/eval and typechecks, but leaves the provisioned trading-agent sandbox pinned to the old versions in activate.rs, so the fleet skew the PR aims to close remains for deployed agents.

• Integration: The bumped runtime/eval is reachable from the eval substrate: evals/src/sim/llm-call.ts:33-34 imports createOpenAICompatibleBackend, runAgentTaskStream, and AgentTaskSpec from @tangle-network/agent-runtime, and many eval modules import @tangle-network/agent-eval (e.g. evals/src/product/chat-sandbox-runner.ts:5, evals/src/trading/lifecycle-runner.ts:3, `evals/src/analysis/rlm-analys
• Fit with existing patterns: It fits the established eval pattern: all judge/user-sim LLM calls route through the single llm-call.ts helper that wraps runtime primitives (evals/src/sim/llm-call.ts:1-34). It does not compete with any existing pattern. The local EvalProfile divergence in evals/src/profiles/types.ts:36-53 is intentional and untouched, which is consistent with the documented design.
• Real-world viability: The runtime 0.50 surface used by this repo (createOpenAICompatibleBackend, runAgentTaskStream, AgentTaskSpec) is stable and the new optional peer deps (playwright, @tangle-network/sandbox) are optional, so install-time breakage is unlikely. Error-path handling in evals/src/sim/llm-call.ts:139-159 (backend_error events, AbortController timeout, catch-and-return) is unchanged and will co

🎯 Usefulness Audit

🔴 Root deps bumped but provisioned trading-agent sandbox stays on old runtime/eval [integration] ``

package.json:39,41 now pins ^0.91.0 / ^0.50.0, but trading-blueprint-lib/src/jobs/activate.rs:36-38 still emits ^0.70.0 / ^0.36.0 for the generated trading-agent package (activate.rs:72-94). The test trading_agent_substrate_versions_match_root_package at activate.rs:1619-1635 explicitly enforces that the generated sandbox matches root package.json; because the constants were not updated, that invariant is broken. Update TRADING_AGENT_AGENT_EVAL_VERSION to ^0.91.0 and `T

---

What this audit checks

It judges the change on its merits — not whether it was tasked out in an issue. Unticketed, fast-moving work is fine; the question is whether the change is good and whether a better or existing approach should be used instead.

Pass	What it asks
Heuristic	Vague title? Whitespace-only or cruft-bearing diff? (content signals only)
Duplication	Do added function/class names already exist elsewhere in the repo?
Value Audit	What does it do? What goal does it achieve? Is it good? Better architecture or already-exists?
Usefulness Audit	Does it integrate and fit? Will it hold up in real use and actually get used?

Findings are concerns, not blocks — the human reviewer decides what to do with them.

_{value-audit · 20260614T005715Z}