[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6BSQS9Z80CSY4B0RPH2SEJW":3,"skill-same:01M6BSQS9Z80CSY4B0RPH2SEJW":124},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":23,"publishedAt":54,"linkOnly":55,"markdownContent":56,"markdownBody":57,"frontMatter":58,"refName":59,"commitSha":60,"skillDirectory":61,"forksCount":62,"repositoryHtmlUrl":63,"repositoryHomepageUrl":64,"ownerHtmlUrl":65,"ownerBlogUrl":66,"ownerDescription":6,"updatedAt":67,"install":68,"resources":81},"01M6BSQS9Z80CSY4B0RPH2SEJW","release-sample-sweep",null,"Run the pre-release end-to-end sweep of every user-facing surface — the 33 samples under samples\u002F (booted from their packaged artifacts and driven in a real browser via chrome-devtools MCP), the Expo\u002FReact Native client, and the atmosphere CLI. Use before cutting a release, and after any change to the Console bundle, a shared module, atmosphere.js, the CLI, or several samples at once. Covers preconditions, the keyless Ollama backend, the per-sample launch\u002Fdrive\u002Fcollect\u002Fteardown loop, the evidence ledger, the fix phase (every issue gets a biting regression test in the right suite), the re-test subset, and the report.","Atmosphere\u002Fatmosphere","Atmosphere","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F318873?v=4","01M6BSPYD9QQA0JP3Z913W9PRV","Atmosphere Framework","01M6BSPYHYTTF1B7JPD17B8N5S",3834,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Frelease-sample-sweep\u002FSKILL.md",[18],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000023","Skill与智能体开发","SCENARIO","使用场景",[24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53],"sample","release","sweep","release sample sweep","keyless","evidence ledger","phase","chrome-devtools","re","atmosphere","end","teardown","shared","fix","regression","loop","MCP","collect","devtools","browser","console","artifacts","report","run","change","react","launch","test","chrome","cli","2026-11-10 10:12:30",false,"---\nname: release-sample-sweep\ndescription: Run the pre-release end-to-end sweep of every user-facing surface — the 33 samples under samples\u002F (booted from their packaged artifacts and driven in a real browser via chrome-devtools MCP), the Expo\u002FReact Native client, and the atmosphere CLI. Use before cutting a release, and after any change to the Console bundle, a shared module, atmosphere.js, the CLI, or several samples at once. Covers preconditions, the keyless Ollama backend, the per-sample launch\u002Fdrive\u002Fcollect\u002Fteardown loop, the evidence ledger, the fix phase (every issue gets a biting regression test in the right suite), the re-test subset, and the report.\n---\n\n# Release sample sweep (chrome-devtools)\n\nBefore a release, every user-facing surface is exercised the way a user\nexercises it. This skill is that procedure written down.\n\nThree surfaces, three drivers — all three are release gates:\n\n| Surface | What | Driver |\n|---|---|---|\n| **Samples** (33) | `samples\u002F*`, booted from packaged artifacts | chrome-devtools MCP, or the wire protocol for the headless ones |\n| **Expo client** (1) | `samples\u002Fspring-boot-ai-classroom\u002Fexpo-client\u002F` | iOS simulator MCP — it is a native app, chrome-devtools cannot reach it |\n| **CLI** | `atmosphere run` \u002F `new` \u002F `compose` \u002F `import` \u002F `checkpoint` + its four distributions | Shell, then chrome-devtools against what `atmosphere run` booted |\n\n## When to run it\n\n- **Before cutting any release.** Non-negotiable — it is the last gate before\n  `release-4x.yml`.\n- After a change to the **Console bundle** (`modules\u002Fspring-boot-starter\u002Ffrontend\u002F`),\n  since the Console is both the shipped sample UI and the validation surface.\n- After a change to a **shared module** that every sample transitively depends on\n  (`modules\u002Fcpr`, `modules\u002Fai`, `modules\u002Fspring-boot-starter`, `modules\u002Fadmin`).\n- After a **dependency bump wave** — two of the last three sweeps found a\n  version-skew bug that compiled clean and only failed at runtime.\n\n## What it catches that CI does not\n\nCI builds and tests modules; this sweep exercises **packaged artifacts in a\nbrowser**. The gap between those is where the real bugs live:\n\n| Sweep | Bug found | Why CI was green |\n|---|---|---|\n| 2026-06-30 | `quarkus-ai-chat` would not start — OTel api\u002Fcommon version skew from a Dependabot bump | Module tests never boot the sample's fast-jar |\n| 2026-07-17 | `spring-boot-orchestration-demo` crashed on every tool turn — the sample pom hardcoded `langchain4j-open-ai:1.15.0` while the reactor manages 1.17.0 | The module built against 1.17.0; only the sample's own jar bundled 1.15.0 |\n\n`scripts\u002Frelease-gate-samples.sh` automates the boot-and-assert half of this in\nCI. This sweep is the **browser half** on top of it — the layer that sees\nrendering, streaming, transport headers, tool cards, and console errors.\n\n## The shape of the sweep\n\n```\nStep 0   Preconditions  — build everything, start Ollama, free the ports, open the ledger\nStep 1a  Samples        — 33 samples: launch → drive → collect → verdict → teardown\nStep 1b  Expo client    — the RN client in the iOS simulator\nStep 1c  CLI            — atmosphere run\u002Fnew\u002Fcompose\u002Fimport\u002Fcheckpoint + distributions\n                           ALL OF PHASE 1 IS COLLECT-ONLY. Do not fix anything mid-sweep.\nStep 2   Triage         — classify every finding, rank by blast radius\nStep 3   Fix            — root-cause fix + a regression test per issue, in the right\n                           suite, each proven to bite\nStep 4   Re-test        — the failed surfaces in full, plus the blast-radius subset\n                           of already-passing ones\nStep 5   Report         — vault report, CI green, memory updated\n```\n\nStep 1 is deliberately fix-free. Fixing mid-sweep changes the artifact under\ntest and invalidates every sample already verified against the old one. The one\nexception: a defect that **blocks the sweep itself** from continuing — fix it,\nsay so in the ledger, and note which already-passed samples were re-run.\n\n## Non-negotiables\n\n1. **chrome-devtools, never curl, for validation.** `curl` is allowed only for\n   port readiness and for headless wire protocols (A2A\u002FMCP\u002FREST) that serve no\n   HTML. A \"works via curl\" claim skips the whole JS layer and is a false pass.\n2. **The Atmosphere Console is the UI.** Drive `\u002Fatmosphere\u002Fconsole\u002F` (Spring\n   Boot samples redirect `\u002F` there). A sample that needs a bespoke page instead\n   of the Console is itself a finding.\n3. **Assert the rendered element, not the payload.** An `image` node with a\n   `src` is a rendered screenshot; the same base64 in a `StaticText` node means\n   nothing rendered it. \"Server started\", \"HTTP 200\", and \"bytes present in the\n   DOM\" are not passes.\n4. **Boot the packaged artifact.** `java -jar` (or `quarkus-run.jar`), never\n   `spring-boot:run` \u002F `quarkus:dev`. Both historical bugs above existed *only*\n   at artifact level.\n5. **Kill by PID, never `pkill -f`.** Never touch a port or process the sweep\n   did not start — if a port is occupied, move to another port. The same rule\n   covers the machine's **network**: never run `networksetup`, never take an\n   interface down, never touch VPN\u002FDNS\u002Fproxy settings. The host's Wi-Fi carries\n   every session the maintainer has open, and a sweep interrupted mid-toggle can\n   leave the machine offline indefinitely. Any assertion that needs real network\n   loss is recorded PARTIAL with its unit coverage cited — see\n   `references\u002Fexpo-sweep.md`.\n6. **Model limitation ≠ framework bug.** A small local model emitting invalid\n   tool-call arguments is a model limitation; record it as such and prove it by\n   re-running the same flow on a capable model before calling it a regression.\n7. **Never write \"flaky\".** Reproduce it, or explain the mechanism. If neither\n   is possible yet, it is a FAIL with an open question, not a dismissal.\n8. **Report honestly.** PASS \u002F PARTIAL \u002F FAIL with one line of concrete\n   evidence each. PARTIAL must name what was not proven and why.\n\n## Step 0 — Preconditions\n\n```bash\ngit status --porcelain                    # must be clean\ngit rev-parse --short HEAD                # record this SHA in the ledger\ngrep -m1 '\u003Cversion>' pom.xml              # record the version under test\n\n.\u002Fmvnw install -DskipTests -Pfastinstall  # full reactor: framework + every sample jar\n.\u002Fscripts\u002Fsync-console-bundle.sh --check  # the Console you will drive must be current\n\nollama list                               # qwen2.5:3b + qwen2.5:7b-instruct-q4_K_M\ncurl -s -o \u002Fdev\u002Fnull -w '%{http_code}\\n' http:\u002F\u002Flocalhost:11434\u002Fv1\u002Fmodels\n```\n\n- **LLM backend is local Ollama, keyless.** Use `qwen2.5:3b` for streaming\n  samples and `qwen2.5:7b-instruct-q4_K_M` for tool-heavy agents — 3b emits\n  invalid tool-call arguments and Ollama answers 400. Note `real-ollama` is a\n  CI-harness alias only; `AiConfig` matches the literal `local`.\n- **The launcher scrubs ambient LLM env** (`LLM_API_KEY`, `LLM_BASE_URL`,\n  `LLM_MODE`, `LLM_MODEL`, and the provider keys) from every sample it boots, so\n  the sweep is reproducible on any machine. `SWEEP_KEEP_ENV=1` inherits instead.\n  It also scrubs messaging-channel credentials by prefix (`SLACK_*`,\n  `TELEGRAM_*`, `DISCORD_*`, `WHATSAPP_*`, `MESSENGER_*`, `ATMOSPHERE_CHANNELS_*`),\n  or a channel sample registers live channels on the operator's real bot tokens;\n  `SWEEP_KEEP_CHANNEL_ENV=1` (or `--env` for one token) for a lane that tests a\n  real channel. If you boot a sample by hand, scrub them yourself — a\n  maintainer's profile routinely exports these.\n- **Always read the resolved endpoint out of the boot log before driving:**\n\n  ```\n  grep 'AI config:' target\u002Fsweep\u002F\u003Csample>-\u003Cport>.log\n  ```\n\n  Expect `mode=local … endpoint=http:\u002F\u002Flocalhost:11434\u002Fv1`. Anything else means\n  the sample is not talking to Ollama and the turn's result says nothing about\n  this build. An explicit `LLM_BASE_URL` outranks the mode by design, so an\n  inherited one silently redirects a \"local\" run to a remote provider — that is\n  what happened on the 2026-08-07 shakedown before the scrub existed.\n- **Do not use a paid key.** The paid-LLM lane is retired; quota starvation is\n  what made the 2026-06 sweep report nine samples as plumbing-only.\n- **Do not use embacle** (`embacle-server --provider claude_code`) for\n  tool-calling samples — it applies the host CLI's own configuration to\n  responses and its tool-call fidelity is inconsistent. It is only useful to\n  demonstrate \"a capable model completes this flow cleanly\", then stop it.\n- **Ports:** the sweep runs on the 9101+ block so it never collides with the\n  samples' own defaults or the Playwright fixture's 8080–8104. Assignments are\n  in `references\u002Fsample-matrix.md`.\n- **Open the ledger** at `claude_docs\u002Fsample-sweep-\u003CYYYY-MM-DD>.md` (a gitignored\n  symlink into the vault, so it survives context compaction). Template:\n  `assets\u002Fledger-template.md`. Write each row **as you finish that sample**,\n  never in a batch at the end.\n\n## Step 1a — The per-sample loop\n\nWork through `references\u002Fsample-matrix.md` in order. For each sample:\n\n```bash\n# 1. Launch (the helper refuses to boot if the port is already answering)\n.claude\u002Fskills\u002Frelease-sample-sweep\u002Fscripts\u002Fsweep-sample.sh start \u003Csample> \\\n    --port \u003C9101+n> --ready-path \u003Cpath> --env LLM_MODE=local --env LLM_MODEL=qwen2.5:3b\n```\n\n   For a WebTransport sample the helper reports ready only after the\n   `HTTP\u002F3 + WebTransport server started` log line, not at Tomcat-up: a page\n   opened in the gap between the two used to stay on WebSocket for its life.\n\n2. **Fresh browser page per sample** — `new_page` on the drive URL. Never reuse\n   the previous sample's page: stale state and leftover console noise both\n   corrupt the evidence.\n3. **Snapshot** — `take_snapshot`. Confirm the Console mounted and the transport\n   badge reads what the matrix expects (`Connected · websocket` \u002F\n   `· webtransport` \u002F `· grpc` \u002F `· ag-ui`). A transport that silently fell back\n   is a finding.\n4. **Drive the headline flow** for that sample — the exact interaction is in the\n   matrix, the mechanics per surface class are in `references\u002Fdriving-recipes.md`.\n5. **Wait for the rendered result** — `wait_for` the expected text\u002Felement, then\n   re-`take_snapshot` and confirm the node type (see non-negotiable #3).\n6. **Collect the evidence, all three sources:**\n   - `list_console_messages` — every error and warning, verbatim\n   - `list_network_requests` — any non-2xx\u002Ffailed request\n   - `sweep-sample.sh warnings \u003Csample>` — server-side WARN\u002FERROR\u002Fexception\u002FSLF4J\n   Record warnings even when the sample passes. The warning inventory is half\n   the value of the sweep and is what the next release's triage starts from.\n7. **Verdict + one-line evidence** into the ledger:\n   - **PASS** — headline feature observed rendered, no unexplained console error,\n     no server exception.\n   - **PARTIAL** — plumbing proven, headline feature not observed, *reason named*\n     (missing third-party key, Docker unavailable, model limitation).\n   - **FAIL** — feature broken, error frame, exception, or the sample won't boot.\n8. **Teardown** — `close_page`, then\n   `sweep-sample.sh stop \u003Csample> --port \u003Cport>`. The helper verifies the port is\n   actually released; if it is not, stop and investigate before the next sample\n   claims it.\n\n**Parallel lanes.** State is per sample *and* port\n(`target\u002Fsweep\u002F\u003Csample>-\u003Cport>.{pid,log}`), so two lanes may boot the same\nsample on different ports — the Expo lane and the browser lane both need\nspring-boot-ai-classroom. Name the instance with `--port` on `stop`, `log` and\n`warnings`; without it they act only when the sample has a single instance and\nrefuse otherwise. `stop-all` stops every instance in the state dir, other\nlanes' included — give each lane its own `SWEEP_STATE_DIR` if lanes might call\nit. WebTransport's UDP port is fixed per sample in `application.yml` (4446 is\nshared by ai-classroom and browser-agent), so the launcher passes each Spring\nBoot boot `-Datmosphere.web-transport.port=\u003Cits HTTP port>` — UDP and TCP port\nspaces are separate, so the lane's own port is a free UDP port. A\n`--jvm-arg -Datmosphere.web-transport.port=N` overrides it.\n\n## Step 1b — The Expo client\n\n`samples\u002Fspring-boot-ai-classroom\u002Fexpo-client\u002F` is a native Expo\u002FRN app. It is\n**not** a Maven module, **not** in `cli\u002Fsamples.json`, and unreachable by the\nPlaywright suites — this sweep is its only gate. It links `atmosphere.js` by\nfile path, so it is also the only pre-release check that the client library's\n`.\u002Freact-native` export works in a real RN runtime.\n\nDriven with the **iOS simulator MCP**, not chrome-devtools. Full procedure,\nincluding the `SERVER_URL` port trap and the AppState\u002FNetInfo assertions nothing\nelse covers: `references\u002Fexpo-sweep.md`.\n\n## Step 1c — The CLI\n\nThe CLI is the documented Quick Start and ships as four distributions\n(curl installer, npx, Homebrew tap, SDKMAN). CI covers `list`\u002F`info`, argument\nvalidation, the runtime overlays, and the installers — it never boots a sample\nthrough `atmosphere run` and looks at the UI, and it has no coverage for\n`compose` or `checkpoint`.\n\nThe manual pass closes that: `atmosphere run` → browser-driven, `atmosphere new`\n→ scaffold + compile against Maven Central, plus the thin-coverage commands and\na post-publish check of the actually-shipped artifacts. Watch the jar cache —\na stale `$ATMOSPHERE_HOME\u002Fcache\u002Fv\u003Cversion>` boots the previous release and fakes\na pass. Full procedure: `references\u002Fcli-sweep.md`.\n\n## Step 2 — Triage\n\nWith all samples tested, classify each finding before touching any code:\n\n| Class | Meaning | Action |\n|---|---|---|\n| Framework bug | A module under `modules\u002F` is wrong | Fix + regression spec. Release-blocking. |\n| Sample bug | Only that sample's code\u002Fpom\u002Fconfig is wrong | Fix + regression spec. Release-blocking if the sample ships. |\n| Config\u002Fenv | Sample needs a key, Docker, a collector | Not a bug — document the graceful-degradation behaviour and assert *that* |\n| Model limitation | Small local model can't drive the flow | Prove with a capable model, record, no code change |\n| Sweep environment | Port conflict, stale `~\u002F.m2`, half-built reactor | Fix the environment and re-run that sample |\n\nRank by blast radius: shared-module findings first (they can invalidate other\nsamples' passes), then per-sample.\n\n## Step 3 — Fix, with a regression test per issue\n\nEvery issue gets a test, but **in the suite that can actually run it**:\n\n| Surface the issue is on | Regression home |\n|---|---|\n| Sample \u002F Console \u002F framework | Playwright spec → `references\u002Fregression-specs.md` |\n| CLI | A case in `cli\u002Ftest-cli.sh` (a Playwright spec is the wrong vehicle for a shell CLI) |\n| Expo \u002F React Native | An `atmosphere.js` vitest covering the `.\u002Freact-native` export path; if the defect is genuinely RN-runtime-only, name it in the report as manual-sweep-only rather than faking a gate |\n\nFor **every** issue in the framework-bug or sample-bug class:\n\n1. **Root-cause it first.** Read the failing path; do not pattern-match a fix.\n2. **Smallest change that fixes the cause** — the 2026-07 langchain4j fix was a\n   single pom property.\n3. **Write a Playwright e2e spec that reproduces the failure**, in the right\n   home, wired into the right CI lane, and **proven to bite**: it must fail\n   against the pre-fix artifact and pass after. Recording only \"it passes now\"\n   proves nothing. Full authoring + wiring recipe:\n   `references\u002Fregression-specs.md`.\n4. **Where a build-time lint can close the whole class, add that too** — the\n   langchain4j fix shipped both a spec and `SampleLangChain4jVersionLintTest`,\n   which fails the build if any sample pom hardcodes a LangChain4j version.\n5. **Rebuild** the affected modules and samples before re-testing.\n6. One commit per fix class, conventional-commit prefixed. No CHANGELOG edits —\n   the CHANGELOG is touched only at release time.\n\n## Step 4 — Re-test\n\n1. **Re-run the failed sample end to end** — the full headline flow, not just\n   the broken step.\n2. **Re-run the blast-radius subset** of already-passing samples. The fix\n   changed the artifact those passes were recorded against, so their evidence is\n   only still valid if the fix could not reach them. The mapping from\n   \"what the fix touched\" to \"which passing samples must be re-driven\" is in\n   `references\u002Fretest-subset.md`.\n3. **Say what you did not re-run and why.** A subset is a deliberate scope\n   decision; leaving it unstated reads as \"everything was re-verified\".\n4. Repeat triage → fix → re-test until the sweep is clean.\n\n## Step 5 — Report\n\n- **Vault report** via the `obsidian-writer` skill →\n  `Claude Outputs\u002FSample-Sweep-chrome-devtools-\u003Cdate>.md`. Promote the ledger:\n  full matrix with the evidence column, the issues-found-and-fixed section with\n  commit hashes, methodology caveats, and non-blocking follow-ups.\n- **Every number verified** — sample count from `ls samples\u002F` (minus\n  `shared-resources`) and `cli\u002Fsamples.json`, never from memory.\n- **CI green** on the fix commits before the release proceeds — all workflows,\n  not just the one you were watching.\n- **Update memory** with anything reusable: a new trap, a new drive recipe, a\n  changed port, a sample added or removed.\n- **Record the attestation** — only when the verdict is clean, i.e. no FAIL and\n  every finding fixed and re-tested. Write `.harness\u002Fsweep-attestation.json`\n  with the **full** SHA of the commit the final pass drove (the last re-test's\n  build, not main's HEAD), and commit it:\n\n  ```json\n  {\n    \"sha\": \"\u003C40-hex commit the final pass built and drove>\",\n    \"version\": \"\u003Cpom version at that commit>\",\n    \"date\": \"\u003CYYYY-MM-DD>\",\n    \"verdict\": \"clean\",\n    \"ledgers\": [\"sample-sweep-\u003Cdate>-\u003Clane>\", \"...\"],\n    \"skipped\": [\"\u003Csurface>: \u003Creason>\", \"...\"]\n  }\n  ```\n\n  `release-4x.yml` runs `scripts\u002Fvalidate-sweep-attestation.sh` before it\n  builds and refuses to release when any user-facing file (samples, framework\n  main sources, the Console frontend, atmosphere.js, the CLI, any `pom.xml`)\n  changed after that SHA. A change after the sweep therefore needs its Step 4\n  re-test and a new attestation. Never attest a commit the sweep did not\n  build. That makes the gate a lie, which is the failure it exists to stop.\n\n## Files in this skill\n\n| File | Use |\n|---|---|\n| `references\u002Fsample-matrix.md` | Every sample: boot type, sweep port, drive surface, headline assertion, gating |\n| `references\u002Fdriving-recipes.md` | chrome-devtools call sequences per surface class + browser-layer traps |\n| `references\u002Fexpo-sweep.md` | Step 1b — the RN client in the iOS simulator |\n| `references\u002Fcli-sweep.md` | Step 1c — what CI already covers, the real gaps, and the CLI pass |\n| `references\u002Fregression-specs.md` | Where a Playwright spec lives, how to wire it into CI, how to prove it bites |\n| `references\u002Fretest-subset.md` | Blast radius → which passing samples to re-drive after a fix |\n| `references\u002Ftroubleshooting.md` | Known traps: PNA, long-poll probes, stale jars, port collisions, Quarkus LLM config |\n| `assets\u002Fledger-template.md` | The sweep ledger to copy into `claude_docs\u002F` |\n| `scripts\u002Fsweep-sample.sh` | Boot one sample from its packaged artifact on a sweep port and leave it running |\n","# Release sample sweep (chrome-devtools)\n\nBefore a release, every user-facing surface is exercised the way a user\nexercises it. This skill is that procedure written down.\n\nThree surfaces, three drivers — all three are release gates:\n\n| Surface | What | Driver |\n|---|---|---|\n| **Samples** (33) | `samples\u002F*`, booted from packaged artifacts | chrome-devtools MCP, or the wire protocol for the headless ones |\n| **Expo client** (1) | `samples\u002Fspring-boot-ai-classroom\u002Fexpo-client\u002F` | iOS simulator MCP — it is a native app, chrome-devtools cannot reach it |\n| **CLI** | `atmosphere run` \u002F `new` \u002F `compose` \u002F `import` \u002F `checkpoint` + its four distributions | Shell, then chrome-devtools against what `atmosphere run` booted |\n\n## When to run it\n\n- **Before cutting any release.** Non-negotiable — it is the last gate before\n  `release-4x.yml`.\n- After a change to the **Console bundle** (`modules\u002Fspring-boot-starter\u002Ffrontend\u002F`),\n  since the Console is both the shipped sample UI and the validation surface.\n- After a change to a **shared module** that every sample transitively depends on\n  (`modules\u002Fcpr`, `modules\u002Fai`, `modules\u002Fspring-boot-starter`, `modules\u002Fadmin`).\n- After a **dependency bump wave** — two of the last three sweeps found a\n  version-skew bug that compiled clean and only failed at runtime.\n\n## What it catches that CI does not\n\nCI builds and tests modules; this sweep exercises **packaged artifacts in a\nbrowser**. The gap between those is where the real bugs live:\n\n| Sweep | Bug found | Why CI was green |\n|---|---|---|\n| 2026-06-30 | `quarkus-ai-chat` would not start — OTel api\u002Fcommon version skew from a Dependabot bump | Module tests never boot the sample's fast-jar |\n| 2026-07-17 | `spring-boot-orchestration-demo` crashed on every tool turn — the sample pom hardcoded `langchain4j-open-ai:1.15.0` while the reactor manages 1.17.0 | The module built against 1.17.0; only the sample's own jar bundled 1.15.0 |\n\n`scripts\u002Frelease-gate-samples.sh` automates the boot-and-assert half of this in\nCI. This sweep is the **browser half** on top of it — the layer that sees\nrendering, streaming, transport headers, tool cards, and console errors.\n\n## The shape of the sweep\n\n```\nStep 0   Preconditions  — build everything, start Ollama, free the ports, open the ledger\nStep 1a  Samples        — 33 samples: launch → drive → collect → verdict → teardown\nStep 1b  Expo client    — the RN client in the iOS simulator\nStep 1c  CLI            — atmosphere run\u002Fnew\u002Fcompose\u002Fimport\u002Fcheckpoint + distributions\n                           ALL OF PHASE 1 IS COLLECT-ONLY. Do not fix anything mid-sweep.\nStep 2   Triage         — classify every finding, rank by blast radius\nStep 3   Fix            — root-cause fix + a regression test per issue, in the right\n                           suite, each proven to bite\nStep 4   Re-test        — the failed surfaces in full, plus the blast-radius subset\n                           of already-passing ones\nStep 5   Report         — vault report, CI green, memory updated\n```\n\nStep 1 is deliberately fix-free. Fixing mid-sweep changes the artifact under\ntest and invalidates every sample already verified against the old one. The one\nexception: a defect that **blocks the sweep itself** from continuing — fix it,\nsay so in the ledger, and note which already-passed samples were re-run.\n\n## Non-negotiables\n\n1. **chrome-devtools, never curl, for validation.** `curl` is allowed only for\n   port readiness and for headless wire protocols (A2A\u002FMCP\u002FREST) that serve no\n   HTML. A \"works via curl\" claim skips the whole JS layer and is a false pass.\n2. **The Atmosphere Console is the UI.** Drive `\u002Fatmosphere\u002Fconsole\u002F` (Spring\n   Boot samples redirect `\u002F` there). A sample that needs a bespoke page instead\n   of the Console is itself a finding.\n3. **Assert the rendered element, not the payload.** An `image` node with a\n   `src` is a rendered screenshot; the same base64 in a `StaticText` node means\n   nothing rendered it. \"Server started\", \"HTTP 200\", and \"bytes present in the\n   DOM\" are not passes.\n4. **Boot the packaged artifact.** `java -jar` (or `quarkus-run.jar`), never\n   `spring-boot:run` \u002F `quarkus:dev`. Both historical bugs above existed *only*\n   at artifact level.\n5. **Kill by PID, never `pkill -f`.** Never touch a port or process the sweep\n   did not start — if a port is occupied, move to another port. The same rule\n   covers the machine's **network**: never run `networksetup`, never take an\n   interface down, never touch VPN\u002FDNS\u002Fproxy settings. The host's Wi-Fi carries\n   every session the maintainer has open, and a sweep interrupted mid-toggle can\n   leave the machine offline indefinitely. Any assertion that needs real network\n   loss is recorded PARTIAL with its unit coverage cited — see\n   `references\u002Fexpo-sweep.md`.\n6. **Model limitation ≠ framework bug.** A small local model emitting invalid\n   tool-call arguments is a model limitation; record it as such and prove it by\n   re-running the same flow on a capable model before calling it a regression.\n7. **Never write \"flaky\".** Reproduce it, or explain the mechanism. If neither\n   is possible yet, it is a FAIL with an open question, not a dismissal.\n8. **Report honestly.** PASS \u002F PARTIAL \u002F FAIL with one line of concrete\n   evidence each. PARTIAL must name what was not proven and why.\n\n## Step 0 — Preconditions\n\n```bash\ngit status --porcelain                    # must be clean\ngit rev-parse --short HEAD                # record this SHA in the ledger\ngrep -m1 '\u003Cversion>' pom.xml              # record the version under test\n\n.\u002Fmvnw install -DskipTests -Pfastinstall  # full reactor: framework + every sample jar\n.\u002Fscripts\u002Fsync-console-bundle.sh --check  # the Console you will drive must be current\n\nollama list                               # qwen2.5:3b + qwen2.5:7b-instruct-q4_K_M\ncurl -s -o \u002Fdev\u002Fnull -w '%{http_code}\\n' http:\u002F\u002Flocalhost:11434\u002Fv1\u002Fmodels\n```\n\n- **LLM backend is local Ollama, keyless.** Use `qwen2.5:3b` for streaming\n  samples and `qwen2.5:7b-instruct-q4_K_M` for tool-heavy agents — 3b emits\n  invalid tool-call arguments and Ollama answers 400. Note `real-ollama` is a\n  CI-harness alias only; `AiConfig` matches the literal `local`.\n- **The launcher scrubs ambient LLM env** (`LLM_API_KEY`, `LLM_BASE_URL`,\n  `LLM_MODE`, `LLM_MODEL`, and the provider keys) from every sample it boots, so\n  the sweep is reproducible on any machine. `SWEEP_KEEP_ENV=1` inherits instead.\n  It also scrubs messaging-channel credentials by prefix (`SLACK_*`,\n  `TELEGRAM_*`, `DISCORD_*`, `WHATSAPP_*`, `MESSENGER_*`, `ATMOSPHERE_CHANNELS_*`),\n  or a channel sample registers live channels on the operator's real bot tokens;\n  `SWEEP_KEEP_CHANNEL_ENV=1` (or `--env` for one token) for a lane that tests a\n  real channel. If you boot a sample by hand, scrub them yourself — a\n  maintainer's profile routinely exports these.\n- **Always read the resolved endpoint out of the boot log before driving:**\n\n  ```\n  grep 'AI config:' target\u002Fsweep\u002F\u003Csample>-\u003Cport>.log\n  ```\n\n  Expect `mode=local … endpoint=http:\u002F\u002Flocalhost:11434\u002Fv1`. Anything else means\n  the sample is not talking to Ollama and the turn's result says nothing about\n  this build. An explicit `LLM_BASE_URL` outranks the mode by design, so an\n  inherited one silently redirects a \"local\" run to a remote provider — that is\n  what happened on the 2026-08-07 shakedown before the scrub existed.\n- **Do not use a paid key.** The paid-LLM lane is retired; quota starvation is\n  what made the 2026-06 sweep report nine samples as plumbing-only.\n- **Do not use embacle** (`embacle-server --provider claude_code`) for\n  tool-calling samples — it applies the host CLI's own configuration to\n  responses and its tool-call fidelity is inconsistent. It is only useful to\n  demonstrate \"a capable model completes this flow cleanly\", then stop it.\n- **Ports:** the sweep runs on the 9101+ block so it never collides with the\n  samples' own defaults or the Playwright fixture's 8080–8104. Assignments are\n  in `references\u002Fsample-matrix.md`.\n- **Open the ledger** at `claude_docs\u002Fsample-sweep-\u003CYYYY-MM-DD>.md` (a gitignored\n  symlink into the vault, so it survives context compaction). Template:\n  `assets\u002Fledger-template.md`. Write each row **as you finish that sample**,\n  never in a batch at the end.\n\n## Step 1a — The per-sample loop\n\nWork through `references\u002Fsample-matrix.md` in order. For each sample:\n\n```bash\n# 1. Launch (the helper refuses to boot if the port is already answering)\n.claude\u002Fskills\u002Frelease-sample-sweep\u002Fscripts\u002Fsweep-sample.sh start \u003Csample> \\\n    --port \u003C9101+n> --ready-path \u003Cpath> --env LLM_MODE=local --env LLM_MODEL=qwen2.5:3b\n```\n\n   For a WebTransport sample the helper reports ready only after the\n   `HTTP\u002F3 + WebTransport server started` log line, not at Tomcat-up: a page\n   opened in the gap between the two used to stay on WebSocket for its life.\n\n2. **Fresh browser page per sample** — `new_page` on the drive URL. Never reuse\n   the previous sample's page: stale state and leftover console noise both\n   corrupt the evidence.\n3. **Snapshot** — `take_snapshot`. Confirm the Console mounted and the transport\n   badge reads what the matrix expects (`Connected · websocket` \u002F\n   `· webtransport` \u002F `· grpc` \u002F `· ag-ui`). A transport that silently fell back\n   is a finding.\n4. **Drive the headline flow** for that sample — the exact interaction is in the\n   matrix, the mechanics per surface class are in `references\u002Fdriving-recipes.md`.\n5. **Wait for the rendered result** — `wait_for` the expected text\u002Felement, then\n   re-`take_snapshot` and confirm the node type (see non-negotiable #3).\n6. **Collect the evidence, all three sources:**\n   - `list_console_messages` — every error and warning, verbatim\n   - `list_network_requests` — any non-2xx\u002Ffailed request\n   - `sweep-sample.sh warnings \u003Csample>` — server-side WARN\u002FERROR\u002Fexception\u002FSLF4J\n   Record warnings even when the sample passes. The warning inventory is half\n   the value of the sweep and is what the next release's triage starts from.\n7. **Verdict + one-line evidence** into the ledger:\n   - **PASS** — headline feature observed rendered, no unexplained console error,\n     no server exception.\n   - **PARTIAL** — plumbing proven, headline feature not observed, *reason named*\n     (missing third-party key, Docker unavailable, model limitation).\n   - **FAIL** — feature broken, error frame, exception, or the sample won't boot.\n8. **Teardown** — `close_page`, then\n   `sweep-sample.sh stop \u003Csample> --port \u003Cport>`. The helper verifies the port is\n   actually released; if it is not, stop and investigate before the next sample\n   claims it.\n\n**Parallel lanes.** State is per sample *and* port\n(`target\u002Fsweep\u002F\u003Csample>-\u003Cport>.{pid,log}`), so two lanes may boot the same\nsample on different ports — the Expo lane and the browser lane both need\nspring-boot-ai-classroom. Name the instance with `--port` on `stop`, `log` and\n`warnings`; without it they act only when the sample has a single instance and\nrefuse otherwise. `stop-all` stops every instance in the state dir, other\nlanes' included — give each lane its own `SWEEP_STATE_DIR` if lanes might call\nit. WebTransport's UDP port is fixed per sample in `application.yml` (4446 is\nshared by ai-classroom and browser-agent), so the launcher passes each Spring\nBoot boot `-Datmosphere.web-transport.port=\u003Cits HTTP port>` — UDP and TCP port\nspaces are separate, so the lane's own port is a free UDP port. A\n`--jvm-arg -Datmosphere.web-transport.port=N` overrides it.\n\n## Step 1b — The Expo client\n\n`samples\u002Fspring-boot-ai-classroom\u002Fexpo-client\u002F` is a native Expo\u002FRN app. It is\n**not** a Maven module, **not** in `cli\u002Fsamples.json`, and unreachable by the\nPlaywright suites — this sweep is its only gate. It links `atmosphere.js` by\nfile path, so it is also the only pre-release check that the client library's\n`.\u002Freact-native` export works in a real RN runtime.\n\nDriven with the **iOS simulator MCP**, not chrome-devtools. Full procedure,\nincluding the `SERVER_URL` port trap and the AppState\u002FNetInfo assertions nothing\nelse covers: `references\u002Fexpo-sweep.md`.\n\n## Step 1c — The CLI\n\nThe CLI is the documented Quick Start and ships as four distributions\n(curl installer, npx, Homebrew tap, SDKMAN). CI covers `list`\u002F`info`, argument\nvalidation, the runtime overlays, and the installers — it never boots a sample\nthrough `atmosphere run` and looks at the UI, and it has no coverage for\n`compose` or `checkpoint`.\n\nThe manual pass closes that: `atmosphere run` → browser-driven, `atmosphere new`\n→ scaffold + compile against Maven Central, plus the thin-coverage commands and\na post-publish check of the actually-shipped artifacts. Watch the jar cache —\na stale `$ATMOSPHERE_HOME\u002Fcache\u002Fv\u003Cversion>` boots the previous release and fakes\na pass. Full procedure: `references\u002Fcli-sweep.md`.\n\n## Step 2 — Triage\n\nWith all samples tested, classify each finding before touching any code:\n\n| Class | Meaning | Action |\n|---|---|---|\n| Framework bug | A module under `modules\u002F` is wrong | Fix + regression spec. Release-blocking. |\n| Sample bug | Only that sample's code\u002Fpom\u002Fconfig is wrong | Fix + regression spec. Release-blocking if the sample ships. |\n| Config\u002Fenv | Sample needs a key, Docker, a collector | Not a bug — document the graceful-degradation behaviour and assert *that* |\n| Model limitation | Small local model can't drive the flow | Prove with a capable model, record, no code change |\n| Sweep environment | Port conflict, stale `~\u002F.m2`, half-built reactor | Fix the environment and re-run that sample |\n\nRank by blast radius: shared-module findings first (they can invalidate other\nsamples' passes), then per-sample.\n\n## Step 3 — Fix, with a regression test per issue\n\nEvery issue gets a test, but **in the suite that can actually run it**:\n\n| Surface the issue is on | Regression home |\n|---|---|\n| Sample \u002F Console \u002F framework | Playwright spec → `references\u002Fregression-specs.md` |\n| CLI | A case in `cli\u002Ftest-cli.sh` (a Playwright spec is the wrong vehicle for a shell CLI) |\n| Expo \u002F React Native | An `atmosphere.js` vitest covering the `.\u002Freact-native` export path; if the defect is genuinely RN-runtime-only, name it in the report as manual-sweep-only rather than faking a gate |\n\nFor **every** issue in the framework-bug or sample-bug class:\n\n1. **Root-cause it first.** Read the failing path; do not pattern-match a fix.\n2. **Smallest change that fixes the cause** — the 2026-07 langchain4j fix was a\n   single pom property.\n3. **Write a Playwright e2e spec that reproduces the failure**, in the right\n   home, wired into the right CI lane, and **proven to bite**: it must fail\n   against the pre-fix artifact and pass after. Recording only \"it passes now\"\n   proves nothing. Full authoring + wiring recipe:\n   `references\u002Fregression-specs.md`.\n4. **Where a build-time lint can close the whole class, add that too** — the\n   langchain4j fix shipped both a spec and `SampleLangChain4jVersionLintTest`,\n   which fails the build if any sample pom hardcodes a LangChain4j version.\n5. **Rebuild** the affected modules and samples before re-testing.\n6. One commit per fix class, conventional-commit prefixed. No CHANGELOG edits —\n   the CHANGELOG is touched only at release time.\n\n## Step 4 — Re-test\n\n1. **Re-run the failed sample end to end** — the full headline flow, not just\n   the broken step.\n2. **Re-run the blast-radius subset** of already-passing samples. The fix\n   changed the artifact those passes were recorded against, so their evidence is\n   only still valid if the fix could not reach them. The mapping from\n   \"what the fix touched\" to \"which passing samples must be re-driven\" is in\n   `references\u002Fretest-subset.md`.\n3. **Say what you did not re-run and why.** A subset is a deliberate scope\n   decision; leaving it unstated reads as \"everything was re-verified\".\n4. Repeat triage → fix → re-test until the sweep is clean.\n\n## Step 5 — Report\n\n- **Vault report** via the `obsidian-writer` skill →\n  `Claude Outputs\u002FSample-Sweep-chrome-devtools-\u003Cdate>.md`. Promote the ledger:\n  full matrix with the evidence column, the issues-found-and-fixed section with\n  commit hashes, methodology caveats, and non-blocking follow-ups.\n- **Every number verified** — sample count from `ls samples\u002F` (minus\n  `shared-resources`) and `cli\u002Fsamples.json`, never from memory.\n- **CI green** on the fix commits before the release proceeds — all workflows,\n  not just the one you were watching.\n- **Update memory** with anything reusable: a new trap, a new drive recipe, a\n  changed port, a sample added or removed.\n- **Record the attestation** — only when the verdict is clean, i.e. no FAIL and\n  every finding fixed and re-tested. Write `.harness\u002Fsweep-attestation.json`\n  with the **full** SHA of the commit the final pass drove (the last re-test's\n  build, not main's HEAD), and commit it:\n\n  ```json\n  {\n    \"sha\": \"\u003C40-hex commit the final pass built and drove>\",\n    \"version\": \"\u003Cpom version at that commit>\",\n    \"date\": \"\u003CYYYY-MM-DD>\",\n    \"verdict\": \"clean\",\n    \"ledgers\": [\"sample-sweep-\u003Cdate>-\u003Clane>\", \"...\"],\n    \"skipped\": [\"\u003Csurface>: \u003Creason>\", \"...\"]\n  }\n  ```\n\n  `release-4x.yml` runs `scripts\u002Fvalidate-sweep-attestation.sh` before it\n  builds and refuses to release when any user-facing file (samples, framework\n  main sources, the Console frontend, atmosphere.js, the CLI, any `pom.xml`)\n  changed after that SHA. A change after the sweep therefore needs its Step 4\n  re-test and a new attestation. Never attest a commit the sweep did not\n  build. That makes the gate a lie, which is the failure it exists to stop.\n\n## Files in this skill\n\n| File | Use |\n|---|---|\n| `references\u002Fsample-matrix.md` | Every sample: boot type, sweep port, drive surface, headline assertion, gating |\n| `references\u002Fdriving-recipes.md` | chrome-devtools call sequences per surface class + browser-layer traps |\n| `references\u002Fexpo-sweep.md` | Step 1b — the RN client in the iOS simulator |\n| `references\u002Fcli-sweep.md` | Step 1c — what CI already covers, the real gaps, and the CLI pass |\n| `references\u002Fregression-specs.md` | Where a Playwright spec lives, how to wire it into CI, how to prove it bites |\n| `references\u002Fretest-subset.md` | Blast radius → which passing samples to re-drive after a fix |\n| `references\u002Ftroubleshooting.md` | Known traps: PNA, long-poll probes, stale jars, port collisions, Quarkus LLM config |\n| `assets\u002Fledger-template.md` | The sweep ledger to copy into `claude_docs\u002F` |\n| `scripts\u002Fsweep-sample.sh` | Boot one sample from its packaged artifact on a sweep port and leave it running |\n",{"name":5,"description":7},"main","4f29b171768db38405b89eaee6b6f7c70c6822e4",".claude\u002Fskills\u002Frelease-sample-sweep",765,"https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere","https:\u002F\u002Fasync-io.live","https:\u002F\u002Fgithub.com\u002FAtmosphere","http:\u002F\u002Fasync-io.org","2026-11-15 01:07:57",{"source":8,"skillName":5,"commands":69,"prompt":79,"zipAvailable":80},[70,73,76],{"tool":71,"command":72},"npx","npx skills add Atmosphere\u002Fatmosphere --skill release-sample-sweep",{"tool":74,"command":75},"bunx","bunx skills add Atmosphere\u002Fatmosphere --skill release-sample-sweep",{"tool":77,"command":78},"pnpm","pnpm dlx skills add Atmosphere\u002Fatmosphere --skill release-sample-sweep","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Frelease-sample-sweep\u002FSKILL.md",true,[82,88,94,98,102,106,110,114,118],{"path":83,"resourceType":84,"resourceTypeDesc":85,"sizeBytes":86,"sourceHtmlUrl":87},"assets\u002Fledger-template.md","ASSET","资产文件",4611,"https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Frelease-sample-sweep\u002Fassets\u002Fledger-template.md",{"path":89,"resourceType":90,"resourceTypeDesc":91,"sizeBytes":92,"sourceHtmlUrl":93},"references\u002Fcli-sweep.md","REFERENCE","参考文档",7887,"https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Frelease-sample-sweep\u002Freferences\u002Fcli-sweep.md",{"path":95,"resourceType":90,"resourceTypeDesc":91,"sizeBytes":96,"sourceHtmlUrl":97},"references\u002Fdriving-recipes.md",12557,"https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Frelease-sample-sweep\u002Freferences\u002Fdriving-recipes.md",{"path":99,"resourceType":90,"resourceTypeDesc":91,"sizeBytes":100,"sourceHtmlUrl":101},"references\u002Fexpo-sweep.md",11163,"https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Frelease-sample-sweep\u002Freferences\u002Fexpo-sweep.md",{"path":103,"resourceType":90,"resourceTypeDesc":91,"sizeBytes":104,"sourceHtmlUrl":105},"references\u002Fregression-specs.md",7133,"https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Frelease-sample-sweep\u002Freferences\u002Fregression-specs.md",{"path":107,"resourceType":90,"resourceTypeDesc":91,"sizeBytes":108,"sourceHtmlUrl":109},"references\u002Fretest-subset.md",7386,"https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Frelease-sample-sweep\u002Freferences\u002Fretest-subset.md",{"path":111,"resourceType":90,"resourceTypeDesc":91,"sizeBytes":112,"sourceHtmlUrl":113},"references\u002Fsample-matrix.md",16252,"https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Frelease-sample-sweep\u002Freferences\u002Fsample-matrix.md",{"path":115,"resourceType":90,"resourceTypeDesc":91,"sizeBytes":116,"sourceHtmlUrl":117},"references\u002Ftroubleshooting.md",9209,"https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Frelease-sample-sweep\u002Freferences\u002Ftroubleshooting.md",{"path":119,"resourceType":120,"resourceTypeDesc":121,"sizeBytes":122,"sourceHtmlUrl":123},"scripts\u002Fsweep-sample.sh","SCRIPT","脚本",32398,"https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Frelease-sample-sweep\u002Fscripts\u002Fsweep-sample.sh",[125,134,143,166,190,214],{"skillNo":126,"name":127,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":128,"categories":129,"keywords":131,"publishedAt":54},"01M6BSSQ9ZE9KG4GEBK6CXCB37","support-agent","https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002Fsamples\u002Fspring-boot-orchestration-demo\u002Fsrc\u002Fmain\u002Fresources\u002FMETA-INF\u002Fskills\u002Fsupport-agent\u002FSKILL.md",[130],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},[132,133],"support agent","support",{"skillNo":135,"name":136,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":137,"categories":138,"keywords":140,"publishedAt":54},"01M6BSSMTRCM898EM0E3X68ZDJ","billing-agent","https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002Fsamples\u002Fspring-boot-orchestration-demo\u002Fsrc\u002Fmain\u002Fresources\u002FMETA-INF\u002Fskills\u002Fbilling-agent\u002FSKILL.md",[139],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},[141,142],"billing agent","billing",{"skillNo":144,"name":145,"nameZh":6,"description":146,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":147,"categories":148,"keywords":150,"publishedAt":54},"01M6BSSJ5B0S8BAQG6MKJAA7YN","omnichannel-chat","General-purpose AI assistant that answers naturally over web chat and external messaging platforms (Telegram, Slack, Discord, WhatsApp, Messenger). One agent, every channel.","https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002Fsamples\u002Fspring-boot-channels-chat\u002Fsrc\u002Fmain\u002Fresources\u002FMETA-INF\u002Fskills\u002Fomnichannel-chat\u002FSKILL.md",[149],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},[151,152,153,154,155,156,157,158,159,160,161,162,163,164,165],"omnichannel chat","chat","assistant","discord","slack","whatsapp","web","messaging","ai assistant","channel","telegram","general","multi","messenger","multi-channel",{"skillNo":167,"name":168,"nameZh":6,"description":169,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":170,"categories":171,"keywords":173,"publishedAt":54},"01M6BSSFCE0HX9FVQ2PVAQP63T","writer-agent","Business writer that synthesizes research, strategy, and financial findings into executive briefings and reports. Use after upstream specialist agents have produced their data.","https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002Fmodules\u002Fskills\u002Fsrc\u002Fmain\u002Fresources\u002FMETA-INF\u002Fskills\u002Fwriter-agent\u002FSKILL.md",[172],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},[174,175,176,177,178,179,180,181,182,183,184,185,186,187,188,189],"writer","writer agent","reports","research","business writer","briefing","upstream","Synthesis","data","use","business","strategy","executive-briefing","briefings","specialist","writing",{"skillNo":191,"name":192,"nameZh":6,"description":193,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":194,"categories":195,"keywords":196,"publishedAt":54},"01M6BSSCT499BPVKA5XM9R7NGV","tool-assistant","Tool-calling assistant with backend functions for time, city time, weather, and temperature conversion, plus cost metering and approval workflows. Use when a question needs real-time data via tool calls.","https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002Fmodules\u002Fskills\u002Fsrc\u002Fmain\u002Fresources\u002FMETA-INF\u002Fskills\u002Ftool-assistant\u002FSKILL.md",[],[197,198,153,199,200,201,202,203,204,205,206,207,208,209,183,210,211,212,182,213],"tool assistant","tool","metering","weather","functions","when","for","calls","function-calling","with","time","calling","conversion","question","real-time","Tools","workflows",{"skillNo":215,"name":127,"nameZh":6,"description":216,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":217,"categories":218,"keywords":220,"publishedAt":54},"01M6BSSA9MPT2656NHQG3VJAAJ","Front-line support desk for account questions, technical troubleshooting, and general inquiries. Use as the default customer entry point; routes billing questions to the billing-agent via handoff.","https:\u002F\u002Fgithub.com\u002FAtmosphere\u002Fatmosphere\u002Fblob\u002Fmain\u002Fmodules\u002Fskills\u002Fsrc\u002Fmain\u002Fresources\u002FMETA-INF\u002Fskills\u002Fsupport-agent\u002FSKILL.md",[219],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},[133,132,221,222,223,203,224,225,142,226,227,228,162,183,229,230,231,232,233],"as","point","entry","front","questions","via","line","default","technical","troubleshooting","routing","customer","handoff"]