packed-cohort is red on main: sandbox 0.19.3 needs a harness enum agent-eval's pinned interface does not have

[drewstone]

What is broken

packed-cohort fails on every PR, including ones that touch nothing related. Reproduced on fix/supervise-root-driver-retry (run 31304816006); main's last green CI was 2026-08-06, before the triggering release.

Error: BACKEND_TYPE declares runners the canonical harness enum does not define: prime.
Reconcile them into @tangle-network/agent-interface.
  at @tangle-network/sandbox@0.19.3/dist/runtime-api-BQii48NT.js:49

The chain, each link verified

link	fact	evidence
1	@tangle-network/sandbox@0.19.3 declares PRIME: "prime" in BACKEND_TYPE and binds that table to harnessTypeSchema at import time, throwing when the canonical enum lacks it	npm pack @tangle-network/sandbox@0.19.3, dist/runtime-api-BQii48NT.js:13-49
2	prime was added to the canonical enum in agent-sdk d8020a5 (2026-08-06) and ships in @tangle-network/agent-interface@0.45.0	npm pack @tangle-network/agent-interface@0.45.0, dist/harness.js:10
3	@tangle-network/agent-eval@0.144.4 — the latest published — depends on @tangle-network/agent-interface@0.43.1 as an exact pin, not a range	npm view @tangle-network/agent-eval@latest dependencies
4	this repo peers agent-interface >=0.43.1 <0.44.0 and sandbox >=0.19.1 <0.20.0, so the cohort installer is free to take sandbox 0.19.3 beside interface 0.43.1	package.json peerDependencies

So the published cohort has no consistent solution: sandbox 0.19.3 requires interface >= 0.44, and the newest agent-eval forbids anything but 0.43.1.

Confirmed by bumping this repo's catalog to interface 0.45.0 / sandbox 0.19.3 locally: tsc then reports 19 errors, and every one is the duplicate-package shape (Two different types with this name exist) — agent-eval drags 0.43.1 in beside it. It is not an API migration; it is two copies of the enum.

The fix, in order

1. Republish @tangle-network/agent-eval against @tangle-network/agent-interface@0.45.0 (the pin is exact, so no consumer can resolve past it).
2. Bump this repo's catalog + peer ranges to interface >=0.45.0 <0.46.0 and sandbox >=0.19.3 <0.20.0.
3. packed-cohort goes green without any source change here.

Until step 1 lands, the honest alternative is to narrow this repo's sandbox peer to <0.19.3 — refusing a release that is genuinely incompatible with the rest of the cohort, rather than admitting it and failing at import.

Why it matters now

The check is a merge gate, so every PR in this repo currently carries one red check for a reason unrelated to its own diff — which is exactly how a real failure gets waved through later.

[drewstone]

Braid independently reproduces the incompatible latest cohort on 2026-08-09. With agent-runtime 0.128.0, agent-eval 0.144.4, agent-interface 0.45.0, sandbox 0.19.3, CLI Bridge provider 0.4.1, and Tangle provider 0.5.1, pnpm peers check reports runtime still requires agent-interface >=0.43.0 <0.44.0 and sandbox >=0.17.2 <0.18.0. The installed graph also contains agent-interface 0.43.0, 0.43.1, and 0.45.0 through agent-knowledge/eval/runtime. Runtime embeds agent-profile-materialize 0.10.2 even though 0.12.0 is published for interface 0.45.0. Braid compiles after the intentional systemPrompt capability migration and is running its full product checks on the latest direct dependencies, but the peer warnings and duplicate canonical type packages remain an upstream release-cohort defect.

[drewstone]

Root-caused and closed by #749: the wall was the packed-cohort job's PINNED sibling refs (interface 0.43.1 / eval 0.144.4 / knowledge 7.0.11) meeting a floating sandbox 0.19.3, not a missing agent-eval republish — eval 0.144.6 → interface 0.46.1 had already shipped. #749 repins both workflows to the published set, moves the catalog/peers with them, migrates the interface 0.44-0.46 changes, and restores the candidate system-prompt guard the materialize bump had silently disabled. Local run of the exact CI script with the new pins: Packed package cohort verified.