[schickling-assistant]

Problem

Automatic replication and peer-relayed client/PTY reads chose Tailscale before Fabric even when the member's fleet.toml listed Fabric first. The relay also ignored the local enabled-transport list. Preferring Fabric required explicit per-peer overrides.

Goal

Use the local member's ordered transports for automatic peer routes in both paths, with one command to prefer Fabric.

st fleet transports add fabric --prefer

Before: transports = ["fabric", "tailscale"] still dialed Tailscale first.
After: automatic routes try Fabric, then Tailscale, then loopback.

Decisions

• The dialing member's list controls local automatic-route preference and eligibility; remote peers need no new wire fields or upgrade to advertise usable endpoints.
• add fabric --prefer enables Fabric first or moves existing Fabric first without re-exposing it or changing its protocol. Ordinary add preserves order.
• Explicit --peer routes retain highest-priority override semantics and are not filtered by transports. The list governs automatic member endpoints only.
• Config-only legacy fleets without a fleet file, and old fleet files omitting transports, keep Tailscale-first defaults. An explicit empty list disables both automatic remote carriers. Loopback remains last.
• No new schema, protocol, environment variable, dependency, native Fabric client endpoint, or rekey operation.

Verification

Exact head: 663c5bee0bf61ffb89d1acdc01ba610ee014b87c.

• Independent correctness GO, including corrected manual restart semantics and omitted-field compatibility.
• Workspace CI run 39747738165: success. Hosted Linux compilation/tests, Clippy, fleet compatibility, and web/client lanes completed.
• Public repository run 39747742452: success.
• Tests added: configured_transports_filter_availability_and_preserve_local_order, routes_follow_local_order_and_filter_disabled_transports_with_loopback_last, dial_targets_follow_local_transport_order_with_mixed_legacy_endpoints, explicit_peer_routes_precede_local_order_and_bypass_disabled_transports, relay_next_hops_follow_local_transport_order_and_removal_with_legacy_peers, fabric_transport_preference_preserves_other_transports_and_existing_exposure, omitted_transport_list_defaults_to_tailscale_then_fabric, and explicit_transport_lists_preserve_empty_disabled_and_reordered_values.
• One local focused Nix/Cargo attempt failed before tests when sccache reset its connection compiling num-integer; no local passing test claim. Sources changed while that attempt was queued/running, so it is not exact-head evidence. No second heavy local build was run. Hosted exact-head execution is the authoritative proof. ST_SKIP_EVIDENCE=1 was not used.

Complexity

One shared LocalTransports::configured policy, one preference bit for the two supported remote carriers, and existing routing/restart paths. No new dependency.

Concerns

• Preference is local and takes effect through the existing service refresh/restart behavior; --no-service requires restarting both st3 daemon and st3 replication-worker. The relay, not just replication, must load the saved order.
• Explicit overrides intentionally win over automatic policy, including removed carriers. Remove the override to make automatic selection authoritative.

Open questions

• No blocking design question remains. Should a future command support preferring Tailscale directly rather than editing the ordered list? This PR only adds the requested Fabric preference affordance.

Friction & bottlenecks

branchy new could not fetch because the moving edge tag conflicted with the local tag. An isolated worktree was created from freshly fetched origin/main without modifying the canonical worktree or the tag. No performance claim is made.

Follow-ups

Native --endpoint fabric:// client support and Fabric rekey are explicitly out of scope.

References

Local fleet transport ordering decision q33; follows the existing fleet transport add/remove surface.

Posted on behalf of @schickling

field	value
agent_identity	agent/dotfiles/fabric-mesh
session	01a1ecd5
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.9.1
agent_runtime	OMP 18.9.1
agent_model	anthropic/claude-opus-5-5
worktree	unknown

[schickling-assistant]

Independent correctness review: GO at exact head 663c5be (reviewer agent). No remaining concrete blockers. Reviewed shared transport policy, replication and relay routes, explicit override precedence, legacy/omitted-list compatibility, and daemon-plus-worker restart semantics. Regression coverage reconstructs the relay from saved policy after restart. This is a static source-review GO; hosted execution proof is tracked separately.

Posted on behalf of @schickling

field	value
agent_identity	agent/dotfiles/fabric-mesh
session	01a1ecd5
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.9.1
agent_runtime	OMP 18.9.1
agent_model	anthropic/claude-opus-5-5
worktree	unknown

[schickling-assistant]

Compounding Manager review at exact head 663c5be: GO (static review against the merge base with main; no findings).

• The transport order from fleet.toml reaches both replication and peer-relay route selection.
• Explicit --peer overrides keep the highest priority and are not filtered.
• Without a transport list, the order stays Tailscale-first, so existing fleet.toml files behave as before. Disabled automatic carriers are dropped, and loopback stays last.
• Tests cover ordering, availability, disabled carriers, loopback-last, add fabric --prefer and the restart instructions for --no-service.
• Exactly one release note; no real hostnames in tests.
• Not executed by this review: builds, tests, live transport fallback and the service-manager restart. Hosted Workspace CI 39747738165 passed at this head.

Posted on behalf of @schickling

field	value
agent_identity	agent/compoundingtech/compounding-manager
session	01a1d992
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.10.2
agent_runtime	OMP 18.10.2
agent_model	anthropic/claude-opus-5-5
worktree	unknown

[schickling-assistant]

Intake: please give a FINAL GO at exact head 663c5be.

• Review: Compounding Manager (agent/compoundingtech/compounding-manager) GO at this head, with no findings.
• CI: hosted workspace run https://github.com/compoundingtech/smalltalk/actions/runs/39747738165 succeeded at this head. The public-repository guard (run 39747742452) also succeeded.
• Context: Johannes decided q33 on 2026-10-22. Fabric becomes the preferred st carrier through the fleet.toml transports order instead of hand-written --peer fabric:// drop-ins. The draft PR Clarify Fabric rekey boundary and compare identity cutover designs #3644 (re-key) stacks on this PR.

(Fabric-mesh, agent/dotfiles/fabric-mesh)

[myobie]

Intake A FINAL GO #3642 (feat(fleet): honor local transport order for replication and peer relay) at exact head 663c5be (2 commits e19b5af, 663c5be; not draft).

Read in source: smallclaims fleet/transport.rs LocalTransports::configured, fleet/file.rs default_transports, fractal-peer peer.rs, sync/worker.rs, fleet_cli.rs (st fleet transports add fabric --prefer) and the 8 tests. Explicit --peer routes keep precedence; loopback last; omitted list keeps Tailscale-first; explicit list omitting a route stops dialing it; no wire, schema or store change (worker.rs only passes the list; no store.rs).
CI: workspace runs 39747738165 and 39751015667 on this head, both attempt 1, success; public-repo guard 39747742452 success; all required checks green, both linux-tests shards nextest flaky 0, no TRY/FLAKY. No rerun was needed.
My earlier belt hold was a Nathan question on the --prefer flag plus the changed routing. Released because Johannes decided q33 (Fabric preferred through the fleet.toml transports list) on 2026-10-22, the flag is additive, and Nathan's 2026-10-22 delegation (recommendation YES = answered) covers it. The release note states the change.
Queue ALONE (routing behaviour change in replication and relay). It is 330 behind main; the merge queue re-runs. Not covered by me: the --no-service restart path (both daemon and replication worker must restart to load the saved order). A head change voids this.