Interactive Repo Mirror Onboarding Wizard

You'd like to clarify the questions first. What would you like to dig into or add context on?

To help focus it, here's where I currently have open questions — tell me which of these (or something else) you want to discuss:

• Region/cell discovery — this is the real blocker. There's genuinely no list source in the repo. Do you happen to know of a cluster-enumeration API on the core side that the OpenAPI spec just doesn't model yet? Or a config/env mechanism you had in mind? If not, we're choosing between a hardcoded list, free-text entry, or gating the feature on new backend work.
• Region × repo fan-out — whether "mirror in 1..m regions" means a cross-product (every selected repo into every selected region) or something narrower.
• Picker scope — whether the repo multi-select should reflect already-mirrored state (enable/disable framing) or just offer new repos.

What's on your mind?

Done. Wizard implemented, tested, lint-clean, full unit suite green.

Bottom line

entire repo mirror create with no args now launches an interactive onboarding wizard; the positional-arg form is unchanged. Everything is real and final except region enumeration, which sits behind a one-function seam waiting on your GET /api/v1/clusters.

What shipped

• cmd/entire/cli/repo_mirror_create_wizard.go (new) — auth gate (reuses resolveAuthStatusTarget + /me probe, with re-login hints) → repo multi-select (ListAvailableMirrors, filtered to available + writable) → region multi-select (default pre-checked) → cross-product CreateMirror fan-out (bounded errgroup, one cluster client per region) → per-mirror clone poll (waitForMirrorClone to io.Discard so concurrent heartbeats don't interleave; aggregate spinner instead) → results table + copy-paste git clone block.
• cmd/entire/cli/repo_mirror.go — create Args RangeArgs(0,2), dispatches to the wizard on zero args; help/comment updates.
• *_wizard_test.go (new) — table tests for all pure helpers.

Design points worth noting

• Cross-product, available-only, regenerate-when-spec-lands — per your three answers.
• Partial success: a region the active login can't reach fails only that region's pairs; the run continues and the table shows what succeeded. Non-empty failure set → non-zero exit via SilentError (table still prints).
• Security: hostFromPublicURL rejects the host@evil.com userinfo trick and trailing paths before they reach clone URLs / STS audience — same posture as the existing validateClusterHost.

The one gate

availableRegions() returns only defaultClusterHost today. When your /api/v1/clusters lands: curl the spec → go generate ./internal/coreapi/... → swap that function's body to client.ListClusters(...) mapping public_url through hostFromPublicURL. The regionChoice struct already carries slug/jurisdiction/host/is_default, so the swap is local to that function (TODO(clusters-api) marks it). Saved to memory for a future session.

I skipped speculative doc edits about /api/v1/clusters — documenting an endpoint that doesn't exist yet would mislead; the code TODO is the source of truth. Nothing committed or pushed.

Two separate things both called "status" — neither is invented by the wizard locally except the second:

1. The repo-picker status (available / mirrored / owner-only) — server-side, not the CLI.

• Comes from GET /mirrors/available → AvailableMirror.Status, fetched by client.ListAvailableMirrors(...) in runMirrorCreateWizard.
• The server computes it live from GitHub against the caller's login (spec: "available (can onboard), mirrored (already mirrored), or owner-only").
• The CLI only consumes it — selectableAvailableRepos (repo_mirror_create_wizard.go) filters to Status == available && Access ∈ {write, admin}. It never derives the value.
• ⚠️ This field is cluster-agnostic: "mirrored" means mirrored somewhere, not in any specific region. So the picker can't tell you "already mirrored in us-east." That's the limitation I flagged in the plan, and why re-selecting a (repo, region) that already exists is safe — CreateMirror is idempotent and the wizard reports it as already-created.

2. The results-table STATUS column (ready / registered / empty / suspended / timed out / error) — computed client-side, in the wizard.

• Set in createOneMirror from two signals: the CreateMirror response (created.Empty) and the clone-poll outcome (waitForMirrorClone returning nil / auth.ErrRepoTargetUnknown / context.DeadlineExceeded / other).
• These are local labels (the mirrorStatus* constants), not a server field.

Which one were you asking about?

Committed.

Bottom line

Wired the wizard to the real GET /api/v1/clusters API and committed everything as one commit (e9e7c0226c). Not pushed.

What the fixup added

• availableRegions now calls coreapi.Client.ListClusters (was the defaultClusterHost seam). Pure mapping extracted to clustersToRegions, which drops any cluster whose publicUrl fails the hostFromPublicURL safety check.
• Regenerated the ogen client — got ListClusters() (*ListClustersOutputBody, error) and Cluster{Slug, Jurisdiction, PublicUrl, IsDefault}, exactly matching the entiredb commit's contract.
• Swapped the obsolete seam test for TestClustersToRegions; cleaned up the now-stale comments.

The one caveat (recorded in memory)

The endpoint isn't deployed to us.console.entire.io yet, and entiredb's spec is code-first (huma, no committed openapi.json). So I hand-authored the /clusters path + schemas into internal/coreapi/spec/core.openapi.json (modeled on /mirrors/available) and ran go generate. Once the endpoint ships, the normal curl … -o core.openapi.json + regenerate is idempotent and replaces my fragment with the live one — verify field names still match if you see a /clusters diff then.

Verification

• go build ./..., mise run lint (0 issues, clean tree after fmt), mise run test (6710 pass) — all green.

One thing I did not do: run E2E tests (they make real API calls / cost money, and the endpoint isn't live anyway). Let me know if you want me to push.

Done. Both pre-picker fetches now show a status line.

Bottom line

runMirrorCreateWizard now wraps each network fetch in the existing startSpinner helper:

• "Fetching available repos" before the repo picker (as you asked).
• "Fetching regions" before the region picker — same pause-before-menu gap, so I added it for consistency.

On a TTY it animates, then resolves to ✓ Fetching available repos and the menu appears; on failure the line is erased. On non-terminals (pipes/tests) it's a no-op except the success checkmark line — same behavior as entire explain's "Loading checkpoints".

Build, lint (0 issues), and tests pass. Not committed — say the word and I'll add it as a new commit on top of e9e7c0226c.

Committed as 778b7fa7ab on top of the wizard commit. Not pushed.

Done. Verified and recorded — nothing to commit (working tree clean, the committed client already matches the now-live API).

Done. Branch renamed to cor-645-cli-smooth-mirror-onboarding-flow with both commits intact.

• The cor-645 in the name is what Linear matches on, so it'll auto-attach to COR-645 once you push.
• Not pushed (per your usual flow). When you push, use git push -u origin cor-645-cli-smooth-mirror-onboarding-flow — Linear picks it up from the branch name on the PR/push.

Done.

Bottom line

Branch rebased onto PR #1518 and both create paths now poll the new Mirror.status endpoint for clone readiness via one shared code path. All committed (not pushed).

What changed

• Rebased cor-645-cli-smooth-mirror-onboarding-flow onto origin/chore/refresh-openapi-spec (#1518). Our hand-authored /clusters spec fragment was subsumed by the PR's full refresh, so I took the PR's internal/coreapi wholesale — our commit is now just the wizard code.
• Shared readiness path: new createAndAwaitMirror → awaitMirrorReady polls GetMirror for Mirror.status (processing/ready/failed/suspended). Both entire repo mirror create <url> <cluster> and the wizard's per-mirror createOneMirror call it — verified both route through the same function.
• Deleted the old info/refs probe machinery (waitForMirrorClone, mirrorAdvertisesHead, checkProbeRedirect, the probe HTTP client + repo-token source) and finishMirrorCreate. Suspension/failure now come straight from the control plane — no repo-scoped token exchange or data-plane round trip. auth.RepoScopedToken stays (still used by the git transport).
• Tests overhauled: removed the probe tests; added TestAwaitMirrorReady (fake getter: ready/failed/suspended/timeout) and TestReportOneShotMirror (all lifecycle presentations).

Verification

• go build ./..., mise run lint (0 issues), mise run test (6777 pass) — all green.
• The LSP diagnostics you may see (coreapi.MirrorStatus undefined) are stale — the rebased generated client has them; real build/lint/test confirm.

Commits on the branch (over #1518)

• 993f948bf wizard + d73f27a28 spinners + 90dea374f status-poll refactor.

Note: PR #1518's spec refresh also pulled in unrelated breaking changes (ListOrgs params, typed grant roles) — already adapted on that branch, so it builds. When #1518 merges to main, this branch's base collapses cleanly.

mg produced no output (unknown command or a no-op alias). Nothing to do on my end — let me know what's next.

Done. Committed as 289a796c8.

Bottom line

entire repo mirror create github.com/octocat/hello-world (no cluster-host) now fetches the default cluster from the same GET /api/v1/clusters API the wizard uses, instead of the hardcoded constant.

What changed

• When [cluster-host] is omitted, the one-shot resolves it via resolveDefaultClusterHost → availableRegions (ListClusters): picks the is_default cluster, or the sole cluster if there's only one. It prints Using default cluster <host> and errors with "pass [cluster-host] explicitly" when the catalog is empty or ambiguous.
• The catalog lookup runs through runCore (active context), then the create runs through runCoreForCluster for the resolved host — so --insecure-http-auth and error mapping are handled.
• When [cluster-host] is given, behavior is unchanged (validate + use directly).
• Updated the create help text; added TestPickDefaultRegionHost (pure selection logic).

Note: remove/collaborators still use the hardcoded defaultClusterHost via clusterArg — you only asked about create, so I left those. Say the word if you want remove to default the same way.

Verification

• mise run lint (0 issues), mise run test (6782 pass), build OK — all green.

Branch cor-645-cli-smooth-mirror-onboarding-flow now has 4 commits over #1518. Not pushed.

Everything's in place. Here's the summary and your curl recipe.

Cherry-pick + examples — done

• Cherry-picked bd0a96b0b ("Fix region FQDNs") as 9de7d632c, resolving both conflicts to use aws-us-east-2.entire.io.
• Swept all non-test code: aws-us-east-2.entire.io is now the only cluster-host example in user-facing strings (the create and collaborators add examples that previously said eu-west-1.entire.io/us-east-2.entire.io are fixed). Remaining .entire.io literals in code are just docs.entire.io URLs. Tests still use eu-west-1.entire.io etc., but those aren't user-facing.
• Lint clean (0 issues), 6782 tests pass.

curl the clusters API

There's no entire print-token command — the login JWT sits in the macOS keychain as "<JWT>|<expiry>", keyed by your active context's keychain_service + handle. Your active context is eu.auth.entire.io, so the endpoint is https://eu.auth.entire.io/api/v1/clusters.

For your current setup that resolves to CORE=https://eu.auth.entire.io.

Caveats:

• security ... -w may pop a keychain-access prompt the first time — allow it. (I deliberately didn't run it here so your JWT doesn't land in this transcript.)
• If the token's expired and that context has no refresh token, step 1 won't save it and you'll get a 401 → entire login first.
• Note this dials the EU core (your active context). The spec I refreshed earlier was from us.console.entire.io — different region; the EU core returns its own cluster list.

Want me to fold a hidden entire auth token (prints the active bearer to stdout) into this branch so curl-ing is a one-liner next time?

Done. Committed as 2b58593b8.

entire auth token — added (hidden)

Prints the active control-plane bearer to stdout. Honors ENTIRE_TOKEN verbatim, else resolves + refreshes the active context's login JWT (same bearer the API client uses). Errors and the not-logged-in hint go to stderr so stdout stays clean for $(...). Hidden from help (emits a live credential).

Your curl is now a one-liner

(That's your active EU context. For whatever's active, the core host is the active context's core_url.)

Verification

• New TestAuthTokenCmd (env-token passthrough + not-logged-in SilentError) passes; lint 0 issues; full suite green (6785 tests).
• Updated CLAUDE.md's auth surface to list the hidden token subcommand.

Branch cor-645-cli-smooth-mirror-onboarding-flow now has 6 commits over #1518. Not pushed.

Done. Committed as 6f7d46408.

What changed

isDefault is per-jurisdiction (au/eu/us each have one), so "first default wins" was wrong. Now the caller's jurisdiction comes from /me (GetMeOutputBody.Jurisdiction) and drives both paths:

• One-shot entire repo mirror create <url> (no cluster-host): picks the isDefault cluster whose jurisdiction matches yours — for your eu context that's aws-eu-central-1.entire.io, not the AU one that happened to sort first.
• Wizard: still lists every cluster (you can mirror into any/all — "offer each"), but pre-selects only your jurisdiction's default instead of all three. The "Signed in as" line now shows the jurisdiction: Signed in as toothbrush (eu) via ….

Both error with "pass [cluster-host] explicitly" when the jurisdiction is unknown or has no default. Unknown-jurisdiction fallbacks (lone cluster / lone default) are preserved.

Verification

• mise run lint (0 issues), mise run test (6787 pass), build OK.
• Updated TestPickDefaultRegionHost (jurisdiction matching + fallbacks) and TestClusterChoices (eu pre-selects only the eu default; unknown jurisdiction pre-selects nothing).

Branch now has 7 commits over #1518. Not pushed.

Note: I didn't apply jurisdiction-awareness to remove/collaborators (they still use the hardcoded defaultClusterHost via clusterArg) — out of scope for this ask, but easy to extend if you want consistency there.

Bottom line: for any non-US user, omitting [cluster-host] on remove/collaborators silently targets the US cluster (aws-us-east-2.entire.io) — which is now not where create puts their mirrors. So the omit-the-arg workflow is inconsistent and will mostly 404 or act on the wrong placement.

Concretely, after this change the omitted-cluster default diverges by command:

• create <url> (no cluster) → your jurisdiction's default (e.g. aws-eu-central-1 for your eu context).
• remove <url> (no cluster) → always aws-us-east-2 (hardcoded defaultClusterHost via clusterArg).
• collaborators add/remove/list <url> <handle> (no cluster) → always aws-us-east-2 (via clusterArgAt).

What that implies for a non-US user who relies on defaults:

• remove 404s. They created the mirror on aws-eu-central-1; remove github.com/x/y (no cluster) looks for it on aws-us-east-2, finds nothing, and errors "no such mirror" even though the mirror exists. Confusing: create "worked" but remove "can't find it."
• collaborators operate on the wrong/absent placement. Grants are per-cluster (a mirror is a per-cluster native repo with its own SpiceDB grant). collaborators add … --role writer (no cluster) grants on the aws-us-east-2 placement, not the eu one the user actually has — so the grant either 404s or lands on a cluster they didn't mean.
• Possible cross-jurisdiction auth error instead. runCoreForCluster("aws-us-east-2.entire.io") discovers that cluster's cores; if the eu login has no context eligible for the US cluster, they get a "log in to …" error rather than a clean result.
• Real (if narrow) footgun: a user with mirrors in both eu and us who runs remove <url> expecting "the one I just made" (eu) would instead remove the us placement, if one exists. No mass data loss, but it's a wrong-target deletion.

What it does not affect:

• US users — the hardcoded default happens to equal their jurisdiction's default, so everything's consistent for them.
• Anyone passing [cluster-host] explicitly — always correct; the positional arg overrides the default on every command.
• The clone URL printed by create/list always names the real cluster, so users can see where a mirror actually lives.

The fix, if you want consistency, is small: route remove/collaborators' omitted-cluster case through the same resolveDefaultClusterHost (catalog + /me jurisdiction) instead of clusterArg/clusterArgAt. The reason I held off: those commands take an existing placement, so arguably they shouldn't guess at all — a case could be made to require [cluster-host] there (fail with "which cluster?" rather than default), since "default for create" and "default for operate-on-existing" aren't obviously the same intent.

Want me to (a) make them jurisdiction-aware like create, or (b) require an explicit cluster on remove/collaborators?

Done. Committed as 68e81f383.

Bottom line

Cluster-guessing is now confined to the interactive wizard. All non-interactive commands keep the fixed defaultClusterHost (aws-us-east-2.entire.io).

Final state per command (omitted [cluster-host])

• create (no args) → interactive wizard, jurisdiction-aware: lists all clusters, pre-selects your jurisdiction's default. (kept)
• create <url> → aws-us-east-2.entire.io. (reverted to original)
• remove <url>, collaborators … → aws-us-east-2.entire.io. (unchanged)

So scripts get stable, predictable behavior, and the symmetry between create <url> / remove <url> / collaborators is restored (all share one fixed default).

What changed

• One-shot create else branch reverted to clusterArg(args) + validate.
• Removed the now-unused resolveDefaultClusterHost / callerJurisdiction / pickDefaultRegionHost and TestPickDefaultRegionHost.
• Help text updated: "defaults to aws-us-east-2.entire.io when omitted (the interactive wizard … lets you pick clusters)."
• Kept: the wizard's jurisdiction-aware pre-selection, /me Jurisdiction plumbing, and the "Signed in as toothbrush (eu)" line.

Lint clean (0 issues), 6780 tests pass, help verified. Branch now has 8 commits over #1518; not pushed.

One note: commit 289a796c8 ("Default one-shot create cluster to the catalog default") is now effectively undone by this commit — the history shows both. If you'd rather it not appear as a flip-flop in the PR, say the word and I can squash those two; otherwise I'll leave history append-only per your usual preference.

Done — 37fa329a5.

Answer to your question

No, it's not the multiple isDefault==true. I proved it with your exact API response: the picker is built with all 3 options (regions=3, opts=3, aws-eu-central-1 pre-selected). isDefault only controls which rows start checked — it never removes rows.

You were seeing one because huh's region list is a scrollable viewport sized to your (short ~9-line) terminal: title + description + footer consumed the rows, so only the top option rendered and the other two were below the fold (press ↓, or use a taller window, to see them).

What I changed (37fa329a5)

• Caller's-jurisdiction clusters listed first, so on a cramped terminal the visible top row is your relevant, pre-checked default (aws-eu-central-1 (eu)) instead of aws-ap-southeast-2 (au). Other jurisdictions are still there, one ↓ away.
• Trimmed the description ("Each repo is mirrored into every selected region.") to reclaim a row — the footer already shows toggle/submit.

Lint clean, 6780 tests pass. TestClusterChoices updated to assert the home-first ordering. Branch now has 9 commits over #1518; not pushed.

If you want the rest visible without scrolling regardless of terminal size, the only real lever is a taller terminal — huh can't render more option rows than the window has. I can also add a (3 regions — ↓ to see all) hint to the title if the below-the-fold case still trips people up.

mg again produced no output — looks like a stray keystroke or a no-op alias. No action taken; ready when you are.

Got it — mg is your git-viewer alias; it just can't render its TUI inside this session, so it comes back blank here. Nothing's broken on my end and I've not changed anything. Run it in a normal terminal tab instead. What would you like to do next?

Fixed — 65d2deea0. And you're right; my earlier "short terminal" guess was wrong. Apologies for that.

Real cause

huh sizes an unset multi-select height to lipgloss.Height(options) − (title+description rows). So it's option-count-dependent, nothing to do with your terminal:

• Repo picker: many options → tall, fine.
• Region picker: 3 options → 3 − 2 ≈ 1 visible row. That's the clamp.

Fix

Set .Height() explicitly on both pickers to option count + header slack, so the whole list renders (and still scrolls if it ever outgrows the terminal). The region picker now shows all clusters at once; the repo picker is unchanged-to-better (it'll now also show all without the ~2-row scroll it had, and it won't collapse if you ever have only a couple of repos).

Lint clean, tests pass. Branch now has 10 commits over #1518; not pushed.

Done — 0dbc9eb23.

What you get now

The wizard's Creating N mirror(s)… spinner is replaced by a Docker-pull-style live block — one line per (repo, region):

Each line animates and flips creating → processing → ready (or failed/suspended/timed out) independently as that mirror's CreateMirror + clone poll advance. When all finish, your existing summary table + git clone block print below.

How it works

• A per-poll status callback now flows through awaitMirrorReady → createAndAwaitMirror → createOneMirror, so each mirror reports its live status.
• New mirrorProgress renderer repaints the block in place via ANSI (reusing spinnerFrames), guarded by a mutex with a 100ms ticker.
• Labels use the full cluster host (@ aws-us-east-2.entire.io) per your request.
• One-shot create <url> unchanged — passes nil and keeps its single spinner.
• Non-TTY (pipes/tests) degrades to one printed line per mirror as it finalizes — no cursor escapes (covered by TestMirrorProgress_NonTTY).

Lint clean (0 issues), 6781 tests pass. Branch now has 11 commits over #1518; not pushed.

One caveat worth flagging: the in-place repaint assumes the block isn't scrolled by other output mid-run — standard for this kind of inline progress (same assumption Docker makes). Nothing else writes to stderr during the fan-out, so it's stable.

Done — 87592cf73. One blank line now separates the final progress lines from the table header. Lint clean, tests pass. Not pushed.

PR opened: https://github.com/entireio/cli/pull/1519

Summary

• Committed + pushed all 12 commits to origin/cor-645-cli-smooth-mirror-onboarding-flow (lint clean before push, per the rule).
• PR #1519 opened with base chore/refresh-openapi-spec (#1518), since this branch is stacked on it — the diff shows only our work. Retarget to main once #1518 merges.

The brief description covers user impact (wizard, status-poll readiness, jurisdiction-aware regions, auth token) and what's unchanged (one-shot create, remove/collaborators).

Done. Pushed 7a0ca5c55 and posted a summary comment on PR #1519.

Findings addressed (5 valid)

• Wizard not TTY-gated (Cursor) — added interactive.CanPromptInteractively() check with a clear message pointing at the non-interactive form; both pickers now use form.RunWithContext(ctx) for clean Ctrl+C handling.
• Empty mirrors skipped suspension check (Cursor) — createAndAwaitMirror now does one GetMirror for an existing empty placement and surfaces suspended + resume guidance (restores old behavior; fresh creates can't be suspended, so skipped).
• Cancel reported as failures (Cursor) — a cancelled run now exits quietly instead of "N mirror(s) failed" + context canceled spam.
• Suspended/failed lacked detail (Copilot) — wizard now attaches the mirror id + resume command to those result errors.
• auth token HTTPS (Copilot) — enforces api.RequireSecureURL unless --insecure-http-auth, matching auth status.

Rejected (1 false positive)

• Loop-var capture (Copilot) — repo is Go 1.26; per-iteration loop vars (since 1.22) make the goroutine capture safe. The copyloopvar linter had already removed the explicit copy as redundant. Explained in the PR comment, no change.

Added TestRunMirrorCreateWizard_RequiresTTY; mise run check green (lint 0 issues, 6782 tests). Branch now 13 commits over #1518.

Several ways to force the non-interactive path in a real shell — pick whichever:

Any of these should print:

and exit non-zero (silently — no extra error dump).

Notes:

• ENTIRE_TEST_TTY=0 is the most reliable and self-documenting — it's the dedicated override and short-circuits before everything else (any non-empty value ≠ 1 forces OFF; ENTIRE_TEST_TTY=1 forces interactive back ON if you want to flip it).
• Plain piping (echo | entire …) won't do it — the probe checks /dev/tty, not stdin, so a pipe still finds the terminal.
• Confirm the non-interactive arg form still works the normal way: entire repo mirror create github.com/octocat/hello-world (uses the fixed default cluster, no wizard).

One caveat: you're running a binary on your PATH — make sure it's built from this branch (go build -o /tmp/entire ./cmd/entire && ENTIRE_TEST_TTY=0 /tmp/entire repo mirror create) rather than an older installed entire.

All done.

Round-2 review fixes (committed 0b20568fc, pushed)

• r3471836615 — "Wizard ignores context cancellation": both pickers already used RunWithContext, but the gap was handleFormCancellation only recognizing huh.ErrUserAborted/ErrTimeout. huh wraps an externally-cancelled context as huh: context canceled, which hit the error branch. Fixed handleFormCancellation to also treat context.Canceled/context.DeadlineExceeded as a clean cancel.
• r3471836617 — internal terminology: reworded the wizard's suspended message to the mirror is suspended; contact support (and the failed one to the initial clone failed; contact support) per your direction. Also reworded the one-shot's explainSuspendedMirror the same way — it had the identical entire-core admin mirrors resume / "core" leak — so both paths are consistent. Updated the two tests that asserted the old wording.

Lint clean, 6782 tests pass.

Copilot threads resolved

Replied to and resolved all three:

• loop-var capture — replied it's a false positive on Go 1.26 (per-iteration loop vars since 1.22; copyloopvar removed the explicit copy), resolved.
• suspended/failed detail — noted we addressed it but landed on the "contact support" wording per the other reviewer, resolved.
• auth token HTTPS — noted it's done (api.RequireSecureURL), resolved.

Branch now 14 commits over #1518.

One note for your awareness: I reworded the one-shot explainSuspendedMirror too (beyond the exact line you pointed at) for consistency — it surfaced the same internal admin command. Shout if you'd rather keep operator-facing detail there.

Done — e5692a019, pushed.

Fix (r3471873262 — "Poll aborts on transient errors")

awaitMirrorReady no longer fails the create on a single GetMirror glitch. It now retries on the poll interval and only gives up after 5 consecutive failures (resets on any success) or when the context ends. So:

• A brief network/API blip mid-clone → ride it out, keep waiting.
• A persistent error (deleted mirror, revoked auth) → still surfaces as poll mirror status: … at the cap, not a spin-to-timeout.
• ctx cancel/deadline → unchanged (clean cancel / timeout).

Added two subtests: transient-errors-then-ready (tolerated) and persistent-errors-give-up-at-cap (stops at 5 calls, doesn't reach the deadline).

Lint clean, 6784 tests pass. Branch now 15 commits over #1518.

I left the Cursor thread for its bot to re-mark on this push (you'd only asked me to resolve the Copilot threads). Say the word if you want me to resolve it manually too.

Confirmed: Open: 0, Resolved: 2. (The KeyError is just because unresolved_count drops out of the JSON once it hits zero — the counts block is authoritative.)

Final finding handled — trailing slash in publicUrl

Fixed hostFromPublicURL to tolerate a bare / path (https://aws-us-east-2.entire.io/), so a trailing-slash catalog entry is no longer silently dropped from clustersToRegions (which risked "no regions available"). Real paths/queries/fragments/userinfo are still rejected, since publicUrl flows into clone URLs and the STS audience. Committed 8fc7b1cdf, pushed. Added a trailing slash test case (and /sneaky still errors).

Trail #660 — both findings resolved

• 019efd04-bdcc… (trailing slash, low) → resolved, citing 8fc7b1cdf.
• 019efd04-bb7c… (transient poll errors, medium) → resolved, citing e5692a019 (the earlier Cursor finding I'd already fixed — it was the same issue surfaced on the trail).

Trail now shows 0 open / 2 resolved. I resolved both since both were addressed in code; flag me if you wanted only the trailing-slash one closed.

Lint clean, 6785 tests pass. Branch is 16 commits over #1518, all pushed.