Adopt the latest Channels/Runtime canaries and add the Open session footer

[jerelvelarde]

What this is

The OpenTag half of remote sessions — the Claude-Code-style affordance where a Slack reply carries a link that opens the same conversation somewhere with more room.

Draft for review of the approach; the footer works today and needs nothing from the platform.

Changes

chore(deps) — the latest canary pair. channels@0.4.1-canary.1785477663 / runtime@1.64.2-canary.1785477663. Stable latest (0.4.0 / 1.64.1) is behind this canary line on both packages, so pinning stable would roll back managed image delivery and SDK-owned tool rendering.

test(channel) — adapt to the 0.4 turn contract. Two breaking changes reach these tests:

• IncomingTurn now requires operation: MessageOperation — revision identity (kind, logicalMessageId, revisionId, mentioned).
• Every turn runs on a clone (isolateAgentInstance), and FakeAgent.clone() builds new FakeAgent(...), discarding subclass overrides. The recording double therefore never saw a run, and assertions reading FakeAgent's own counters saw nothing either, since those accumulate on the clone rather than the instance the test holds.

Recording now happens outside the instance, and clone() re-applies the subclass prototype. Assertions describe what OpenTag sent rather than the test double's post-clone internals.

feat(session) — the footer. app/session-link.ts + the wiring in channel.tsx. After a managed run, the reply carries <…|Open session> · gpt-5.5[low].

How the thread id is resolved

Worth stating explicitly, because it is easy to get wrong: the managed delivery adapter sets conversationKey: delivery.canonicalThreadId, and channelPreparedDeliverySchema has always carried canonicalThreadId. So Thread.conversationKey is the canonical thread id on the managed path — no platform change required.

Local adapters put a provider-shaped key in the same field (Slack's ${teamId}:${channel}:${threadTs}). That is not a thread id, and linking to it would produce a console URL for a thread that does not exist. The key is therefore accepted only when it has canonical UUID shape, and the footer is omitted otherwise.

Other deliberate choices:

• Config is all-or-nothing — a partial console location disables the footer rather than emitting a broken link — and rejects unresolved env templates, non-http schemes, and slugs that would escape the URL path.
• pathSegment selects the console view, so this targets the existing /threads/:id today and can move to a dedicated /sessions/:id with no code change.
• The model badge mirrors the agent's own env resolution, and a test reads agent/agent.py and fails if those defaults drift — otherwise the badge would quietly claim the wrong model answered.
• A failed footer post is reported and swallowed. The answer has already landed; a cosmetic addition must never turn a completed turn into a user-facing failure.

Verification

• pnpm check-types — clean
• pnpm test — 162/162, including an integration assertion that a managed turn's reply contains the footer and a non-managed turn's does not
• Runtime connects and declares against dev Intelligence (RUNTIME: Connected)

Not verified live in Slack. The probe set (image out, image in, tool rows, duplicate delivery, thinking status) never ran: the Kite (Dev) Slack app is not currently delivering into the test project, and the Channels surface went dark on a dev redeploy. So the claim here is "typechecks, tests green, runtime declares" — not "verified end to end."

One upstream nit

FakeAgent.clone() discarding subclasses looks like a bug worth a one-line fix — Object.create(Object.getPrototypeOf(this)) or a protected createClone() hook. Subclassing FakeAgent is the documented way to observe runs, and the SDK's own error message tells you to override clone(), which cannot help while the base discards your prototype.

Not in this PR

Planning docs (roadmap, PRDs, boundary analysis) are deliberately excluded — they reference private-repo internals and this repo is public. They live in Notion.