E2E Report — Built-in Skill Migration (T3)

Date: 2026-04-23 Runner: e2e-tester (aionui-builtin-skill-migration team) Plan: ../plans/2026-04-23-builtin-skill-migration-plan.md §Task 3 Backend SHA (aionui-backend): run 1 0ab877f → run 2 04f1537 (H1 fix: camelCase audit on skill.rs) Frontend SHA (AionUi): 69585b28b (feat/backend-migration-builtin-skills — unchanged across runs; test file only) Verdict: CLEAN on run 2 — 8/8 green, 13.5s wall time. Run 1 surfaced 2 Class F/D backend contract defects (D1 ReadBuiltinResourceRequest, D2 SkillListItemResponse missing camelCase rename); both fixed in H1. All three previously-failing scenarios now pass. T3 unblocks T4.

Environment

Commands run:

cd /Users/zhoukai/Documents/github/AionUi
git fetch origin feat/backend-migration-builtin-skills
git reset --hard origin/feat/backend-migration-builtin-skills
bunx electron-vite build
bun run test:e2e tests/e2e/features/builtin-skill-migration/

Scope decision: singleton fixture vs cold-restart scenarios

Identical pattern to the assistant-user-data pilot's T5:

S6 (orphan sweep) seeds {tmp}/agent-skills/<id>/ before the sibling backend starts, then asserts the startup task removes those dirs because the empty conversations table has no matching rows. S8 verifies the backend does NOT touch {data_dir}/builtin-skills/ (that cleanup is Electron-main-process territory) and annotates whether any legacy dirs survive under the live cache — authoritative assertion deferred to T4's packaging smoke.

Per-scenario matrix (run 2 — after H1 fix)

Run 2 total wall clock: 13.5s (down from 38s — no retries, fewer ~10s Electron boots needed).

Test file

• tests/e2e/features/builtin-skill-migration/builtin-skill-migration.e2e.ts

Defects from run 1 (resolved in H1 / 04f1537)

All three backend contract defects below were fixed by backend-dev under task H1 (fix(api-types): enforce camelCase on wire for all skill.rs public types) — an audit-level pass over aionui-api-types/src/skill.rs that added #[serde(rename_all = "camelCase")] to every request/response type that still lacked it. Run 2 verified all three paths live.

D1 — ReadBuiltinResourceRequest missing camelCase rename (Class F) — RESOLVED

Location: aionui-backend/crates/aionui-api-types/src/skill.rs:170-173

#[derive(Debug, Clone, Deserialize)]
pub struct ReadBuiltinResourceRequest {
    pub file_name: String,
}

Problem: Lacks #[serde(rename_all = "camelCase")]. Every other skill-side request type (MaterializeSkillsRequest, BuiltinAutoSkillResponse, ExternalSkillSourceResponse etc.) is camelCased. The frontend (src/common/adapter/ipcBridge.ts:341 and AcpSkillManager.ts:341) invokes this endpoint with { fileName }. Result: every POST /api/skills/builtin-skill call 400s with missing field file_name.

Live probe:

$ curl -sX POST http://127.0.0.1:25904/api/skills/builtin-skill \
    -H 'Content-Type: application/json' -d '{"fileName":"auto-inject/cron/SKILL.md"}'
{"success":false,"error":"Bad request: Failed to deserialize the JSON body into the target type: missing field `file_name` at line 1 column 40","code":"BAD_REQUEST"}

$ curl -sX POST http://127.0.0.1:25904/api/skills/builtin-skill \
    -H 'Content-Type: application/json' -d '{"file_name":"auto-inject/cron/SKILL.md"}'
{"success":true,"data":"---\nname: cron\n..."}

Impact: ACP conversations cannot load any builtin skill body — the auto-inject flow degrades silently (AcpSkillManager.loadSkillBody catches and returns ''), so conversations start without the builtin skill injection that is the whole point of this migration.

Fix: add #[serde(rename_all = "camelCase")] above the struct. Same change likely also needed on ReadBuiltinResourceRequest consumers — a quick audit of all aionui-api-types/src/skill.rs request structs against the frontend's camelCase wire format is warranted:

#[derive(Debug, Clone, Deserialize)]
#[serde(rename_all = "camelCase")]
pub struct ReadBuiltinResourceRequest {
    pub file_name: String,
}

D2 — SkillListItemResponse missing camelCase rename (Class D) — RESOLVED

Location: aionui-backend/crates/aionui-api-types/src/skill.rs:27-36

#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
pub struct SkillListItemResponse {
    pub name: String,
    pub description: String,
    pub location: String,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub relative_location: Option<String>,
    pub is_custom: bool,
    pub source: SkillSourceResponse,
}

Problem: Serializes as {relative_location, is_custom} but the frontend TypeScript contract (src/common/adapter/ipcBridge.ts:344-354) reads {relativeLocation, isCustom}. Result: SkillInfo.relativeLocation is always undefined in the renderer.

Live probe:

$ curl -s http://127.0.0.1:25904/api/skills | jq '.data[0]' -c
{"name":"aionui-skills","description":"...","location":"/tmp/skill-probe/builtin-skills-view/aionui-skills/SKILL.md","relative_location":"auto-inject/aionui-skills/SKILL.md","is_custom":false,"source":"builtin"}

Impact: AcpSkillManager.discoverSkills uses entry.relativeLocation ?? entry.location when source === 'builtin' — with relativeLocation always undefined, it falls through to the absolute location path, which downstream passes as fileName to readBuiltinSkill. Even if D1 is fixed, the absolute path would fail the validate_filename path-traversal check in the backend. So D1 and D2 are both blockers for the AcpSkillManager opt-in builtin skill path.

Fix: same decoration:

#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
#[serde(rename_all = "camelCase")]
pub struct SkillListItemResponse { /* ... */ }

After fixing, rerun cargo test --test skills_builtin_e2e — the existing integration test likely round-trips with snake_case payloads and won't catch wire-format drift. Consider adding a JSON-level assertion (assert_eq!(json["relativeLocation"], ...) in D1/D2's sibling tests) to prevent regression.

D3 — scope of audit — RESOLVED

Recommend a single follow-up pass over aionui-api-types/src/skill.rs adding #[serde(rename_all = "camelCase")] to every request/response type that currently relies on the default snake-case mapping. Candidates to verify:

• ReadSkillInfoRequest / ReadSkillInfoResponse
• ImportSkillRequest / ImportSkillResponse
• ExportSkillRequest (→ targetDir)
• ScanForSkillsRequest / ScanForSkillsResponse / ScannedSkillResponse (→ folderPath, skills[].path)
• WriteAssistantRuleRequest / ReadAssistantRuleRequest (→ assistantId)
• AddExternalPathRequest / RemoveExternalPathRequest — these are already simple enough, but the rename is cheap insurance.

The pilot originally dodged this only because aionui-extension tests drive requests in snake-case JSON (Rust-side tests), and the live renderer paths exercising these fields were covered by prior pilots (where the #[serde(rename_all)] had already been added on the listed types). The new builtin-skill flow is the first to hit these two stragglers.

H1 (04f1537) applied exactly this audit — every struct in skill.rs that could go on the wire now carries the rename.

Probe transcripts

Run 1 probes (against sibling backend at 0ab877f, pre-H1)

GET  /api/skills/builtin-auto           → 200, 4 entries (aionui-skills, cron, officecli, skill-creator)
POST /api/skills/builtin-skill          → 400 on {fileName}, 200 on {file_name}    ← D1
GET  /api/skills                        → 200, emits {relative_location, is_custom}  ← D2
POST /api/skills/materialize-for-agent  → 200 on {conversationId, enabledSkills}
DELETE /api/skills/materialize-for-agent/{id}  → 200, idempotent

Run 2 probes (against sibling backend at 04f1537, post-H1)

$ curl -sX POST http://127.0.0.1:25905/api/skills/builtin-skill \
    -H 'Content-Type: application/json' -d '{"fileName":"auto-inject/cron/SKILL.md"}' | head -c 100
{"success":true,"data":"---\nname: cron\ndescription: Scheduled task management..."}

$ curl -s http://127.0.0.1:25905/api/skills | jq '.data[0]' -c
{"name":"aionui-skills","...","location":"/tmp/skill-probe2/builtin-skills-view/aionui-skills/SKILL.md","relativeLocation":"auto-inject/aionui-skills/SKILL.md","isCustom":false,"source":"builtin"}

Both confirmations: {fileName} is now accepted, and /api/skills emits relativeLocation / isCustom on the wire.

Sibling backend log tail (S6 / S8 boot)

Initializing database at /tmp/aionui-e2e-builtin-skill-xxx/aionui.db
Database initialized at .../aionui.db
Generated and persisted new JWT secret
Running in local mode — authentication is disabled
No configured users detected — initial setup required via /api/auth/status
Server listening on 127.0.0.1:25903

S6 assertion sequence

setup:
  mkdir -p {dataDir}/agent-skills/orphan-conv-1/mermaid
  mkdir -p {dataDir}/agent-skills/orphan-conv-2/cron
  seed: SKILL.md frontmatter in each

start sibling backend (empty conversations table → every orphan is unknown)

poll (< 5s):
  ! exists orphan-conv-1  AND  ! exists orphan-conv-2   → pass
  exists agent-skills/    → pass

probe GET /api/skills/builtin-auto     → non-empty

Failure classification — Skill-Library pilot rubric

Outcome per plan §3.5

Run 2: all 8 scenarios green. T3 clean — marking complete, handing to T4.

• Run 1 correctly classified 2 defects (D1 + D2) and routed to backend-dev per §3.5. Backend-dev landed H1 (04f1537) as a full-file camelCase audit on skill.rs — exceeds the minimum fix and closes D3 in one pass.
• No frontend changes owed across either run — the renderer has been spec-correct throughout (§6 of the backend spec mandates camelCase on the wire).
• T4 coordinator closure unblocked.

Run 3 — post-realign (snake_case wire)

Date: 2026-04-24 Backend SHA: 326e228 (aionui-backend — H1 reverted; skill.rs realigned to snake_case on wire) Frontend SHA: dba2ef499 (AionUi — refactor(skill): realign wire format to snake_case (T2)) Backend binary timestamp: Apr 24 14:35:12 2026 Verdict: CLEAN — 8/8 green, 13.1s wall time. Project-wide snake_case wire contract now enforced end-to-end.

Why re-run

Between runs 2 and 3 the project-wide API contract was flipped from camelCase to snake_case (plan: 2026-04-23-snake-case-realign-plan.md). This matches the canonical convention used by the rest of the backend (team, conversation, cron, channel, assistant). H1's camelCase rename on skill.rs was reverted under T1 and aionui-api-types/src/skill.rs now asserts snake_case on both request and response paths (test_materialize_response_serializes_snake, test_skill_list_item_serde asserts json["is_custom"] / absence of isCustom, etc.). Frontend T2 flipped ipcBridge.ts + hooks to consume the new wire format.

Run 3 verifies the full stack is coherent after both flips.

Test-file flips (T3)

After the realign, tests/e2e/features/builtin-skill-migration/builtin-skill-migration.e2e.ts still had camelCase residue in both response assertions and request bodies. All flipped under T3:

Local TS variable names (conversationId, skillPath, targetDir) are retained — they're not wire-carried fields. The URL-path interpolation /${conversationId} is a path segment, not a body field, and stays as-is.

Per-scenario matrix — run 3

Run 3 total wall clock: 13.1s. No retries, no flakes.

Failure classification — run 3

Outcome

All 8 scenarios green on the realigned pair (backend 326e228 + frontend dba2ef499). T3 complete, unblocks T4.

• snake_case wire contract verified end-to-end across read, write, delete, cold-start, and external-export flows.
• No further frontend or backend changes owed.
• T4 coordinator closure now unblocked per §3.5.

Follow-ups (non-blocking, document for future)

• Adding a snapshot-level JSON contract test in aionui-api-types (or a pact-style check in skills_builtin_e2e.rs) would have caught both defects pre-T3. Worth considering as a backend-dev chore after the pilot.
• S8 full cold-restart coverage for the legacy-dir cleanup path still depends on T4 packaging smoke or a future dedicated worker-per-boot E2E spec. Not a regression risk because (a) the cleanup is fire-and-forget best-effort in initStorage.ts:357-365 and (b) a leftover 7MB dir is harmless.
• S2 today stops at "all builtin-auto bodies resolve"; a deeper probe would drive an actual ACP conversation and inspect the injected message. That is cross-backend (ACP + HTTP) and out of scope for this pilot; the current coverage matches the bar set by the frontend spec §9 row for e2e-tester.