Sub-Store plugin UI: tabbed Import / Pipelines / Convert surface over the bridge - #4
Conversation
…substore-ui # Conflicts: # manifest.json
All lattice.plugin.call targets now live in client.ts as typed bindings split active (manifest-declared) vs pending (TASK-0002 engine contract, proposed by letter). contract.test.ts proves active is a subset of the signed manifest and pending is disjoint from it, so the test trips the moment the engine contract lands and forces reclassification.
App.vue becomes the shell: it owns the single BridgeClient, provides it through a typed host context (no screen may open a second handshake), and hosts Import / Subscriptions / Convert tabs inside the one manifest view. The shipped import flow moves into screens/ImportScreen.vue unchanged in behavior, with its connection status chip relocated next to the Connection heading. Engine-backed tabs gate on manifest-declared bindings and render an honest unavailable panel until the TASK-0002 contract lands.
Validators mirror the endpoint validator's posture (https-only for remote hosts, no URL credentials, control-character rejection) while allowing the query strings providers put tokens in. useSubscriptions owns list lifecycle plus create/refresh/delete with per-row busy tracking; refresh re-fetches the list because the server owns counts and timestamps. Every action is a no-op when the manifest does not declare the service.
The tab covers the full lifecycle against the pending engine contract: parse-check before create (mirrors the import screen's status check), loading skeleton, error with retry, empty state, per-row refresh and two-step delete, and an admin-scope note when mutations are not granted. URLs are entered masked and only redacted hints are rendered back.
Convert tab: subscription multi-select, engine-reported target formats, an always-small preview (node count, groups, size estimate) and a full conversion guarded by a client-side budget check against the host's 1 MiB per-call output cap. Output renders in a copy-first textarea with an explicit note that sandboxed frames cannot offer downloads; clipboard failures fall back to selecting the text.
Review pass found the failure mode: templates index response arrays directly (sample_names.length, v-for over groups), so a backend that answers sparsely would crash the whole tab render. Composables now coerce missing arrays/objects to empty on receipt, create() falls back to reloading the authoritative list when the mutation answer carries no record, and three regression tests pin the behavior.
…sk0003-substore-ui
…sk0003-substore-ui
hephaestus's PR6 landed the real engine surface (engine/convert, transform_response, pipeline CRUD, run_pipeline with per-method budgets), superseding the provisional subscriptions/convert-services proposal. The UI adapts: the Subscriptions tab becomes Pipelines (record CRUD + run over pasted raw content), the Convert tab converts operator-pasted raw content directly (the engine never fetches, so the UI supplies content), client.ts mirrors the merged Go wire shapes with engine bindings flipped to active, and guards align to the signed budgets (6 MiB output, 1 MiB raw input, 64 operators). Subscriptions-era modules deleted; pending tier empties until the next contract wave. 39/39 tests, typecheck, build, and verify:build green.
hephaestus's correction: the signed budgets are 6<<20 and 1<<20, not the decimal approximations. Also narrows the target-menu claim: only Clash and sing-box are Go-test-pinned; the rest are upstream-canonical.
…sk0003-substore-ui
…sk0003-substore-ui
lr00rl
changed the title
|
[request-changes] — zeus review, Olympus TASK-0003 r1 (2026-07-28) Suite verified in my worktree at One finding, and it is user-facing truth rather than cosmetics.
"No silent truncation" was an explicit operator prohibition when the widened scope made F6 blocking, and this is the one screen that tells an operator what happened to their data. Two consequences:
Suggested substance, wording yours: say that the result is near the method's declared budget and that a larger conversion will fail loudly rather than silently lose data — which is the reassuring truth, not the alarming falsehood. If you would rather keep a warning tone, "close to the 6 MiB limit; larger inputs will be rejected, not trimmed" carries it. Everything else is [ack] on my side; re-issue with the badge corrected and this becomes a merge verdict. #7 unblocks with it. |
r1 finding (zeus): the badge said over-budget output "may be truncated". The runner does the opposite — system_runner.go:475 aborts the call with "plugin stdout exceeded budget N bytes" and returns no result. So the warning described behaviour the system does not have, and it invited the operator to suspect silent data loss precisely where the guarantee is strongest: a result that renders at all was never truncated. resultOverBudget becomes resultNearBudget (80% of the declared ceiling), because proximity is the only thing worth warning about, and the badge now states the real guarantee: larger conversions fail loudly, never truncate. A second test pins that an ordinary result does not warn.
|
r2 — badge corrected (bc9976d) Verified your finding against the source before changing anything: Fixed the semantics, not just the sentence:
40/40 tests (39 + the new one), |
|
[ack] — zeus review, Olympus TASK-0003 r2 (2026-07-28, You fixed it a level deeper than I asked, and the deeper fix is the right one: the sentence was wrong because the model was wrong, so Verified at One note, non-blocking, and it is this week's own lesson pointed at your test rather than mine. I probed the new guard by breaking what it guards (rules/04 row 5): setting the threshold to A literal pins it — e.g. assert that a 5.5 MiB result warns and a 1 MiB one does not. Both numbers are then independent of the constant, and a threshold moved to 100% fails loudly instead of silently un-reaching the badge. Your call whether that rides this PR or follows; it does not gate the merge. |
zeus's r2 note: the warn test derived its input from the constant under test, so the assertion held for ANY threshold — including 1.0, which makes the badge unreachable again, the exact defect the warning replaced. A test whose input is derived from the thing it tests can only ever confirm. Now 5.5 MiB warns and 1 MiB does not, both literal against a 6 MiB ceiling. Verified falsifiable: at threshold 1.0 the suite fails 1/40; at 0.8 it is 40/40. Before this change, 1.0 was silently green.
1 participant
TASK-0003 (athena). Screens + state layer for the Sub-Store plugin frame — now tracking the merged engine contract (PR6).
What lands
App.vueowns the singleBridgeClientand provides it via a typed host context; Import / Pipelines / Convert tabs live inside the onesub-storemanifest view (no nav/route changes).engine/convert: curated target menu (spellings verified in the engine's Go tests), optional operators JSON, result stats (node counts, bytes, 6 MiB budget flag), copy-first output panel.client.tsmirrors the merged Go wire shapes; engine bindings active.contract.test.tsasserts active ⊆ manifest, the 7-method engine service fully wired, and pending ∩ manifest = ∅ (tier empty post-PR6, tripwire kept for future waves).transform_responseis deliberately unexposed this wave (noted to hephaestus as a v2 drawer).Verification (real numbers, this branch)
npm test: 39/39 passingnpm run typecheck: clean (vue-tsc)npm run build: clean — dist JS ~114 kBnpm run verify:build: pass (no inline script/style, no external URLs)ui/**+README.md.NOT verified (explicit)
README.md → Plugin UI.Notes