| name | dockerhub-mirror |
| description | Triage Docker Hub pull latency and registry failures. Use when docker pull or the image-download phase of docker run stalls, when Docker Hub returns network, TLS, token, or rate-limit errors, or when the user asks to test, rank, refresh, discover, or manage mirror candidates. |
Docker Hub Mirror Triage
Use a triage process: classify the request, probe the smallest safe surface, qualify every candidate, then recommend an evidence-backed mirror.
Resolve <skill-root> to the absolute directory containing this SKILL.md. Invoke the bundled CLI as:
node "<skill-root>/bin/dockerhub-mirror.mjs" <flags>
Never assume the current working directory is the skill directory.
Authorization gate
The state sequence is 讨论中 → 待共识 → 待授权 → 执行中 → 已完成.
- Before the exact consensus phrase
ok, limit activity to questions, analysis, read-only research, comparison, and --dry-run probes. A requested operation is a goal, not authorization.
- After
ok, present a bounded plan and request the exact execution phrase okok.
- After
okok, perform only the agreed write-capable branch. A changed scope, cost, risk, or execution path returns to discussion.
- Keep replacement Docker commands as output only. Never modify Docker daemon settings, restart Docker, execute
docker pull or docker run, or send private-image credentials through a public mirror.
When a new operational request still needs decisions, or a previous assumption is invalidated, read the extended discussion protocol before asking questions.
Step 1 — Classify the branch
Choose exactly one branch:
- Diagnostic: observed slow or failed Docker Hub image download.
- Existing-candidate check: test or rank cached/built-in candidates.
- Discovery: scrape built-in source pages for previously unseen candidates.
- Cache management: add, remove, list, quarantine, or retry a recorded candidate.
Mark the branch as read-only or write-capable. Diagnostic probes and explicit read-only checks use --dry-run; discovery and cache mutation are write-capable even when their network requests are read-only.
Complete when: one branch is selected and its authorization requirement is explicit.
Step 2 — Establish the target and evidence
For docker pull, capture the image reference and the observed error or stall. For docker run, first establish that the delay occurs while resolving or downloading the image rather than after the container starts.
Treat timeout, TLS handshake, connection reset, unexpected EOF, HTTP 429, anonymous-token failure, manifest delay, and prolonged transfer silence as Docker Hub retrieval evidence. Preserve a digest when the image uses one. Do not rewrite non-Docker-Hub image references.
Complete when: the image target, retrieval evidence, and Docker Hub applicability are known, or explicitly recorded as unavailable.
Step 3 — Construct the minimum safe command
Read the CLI reference only after the branch is selected. Build the command with the absolute <skill-root> entrypoint.
- Diagnostic or unauthorized check: use
--dry-run; add --image <ref> when an image is known.
- Authorized check: use quick mode by default and
--deep when throughput evidence is required.
- Authorized discovery: use
-f/--scrape.
- Authorized cache management: use the corresponding CRUD or quarantine flag.
Prefer quick probes for triage. Use deep probes for newly discovered candidates, explicit throughput comparison, or quarantine recovery.
Complete when: the exact command is constructed, its side effects are understood, and every write-capable invocation has valid authorization.
Step 4 — Probe every selected candidate
Run the command and retain each candidate's success, failure, or timeout result.
Quick mode verifies Registry API and manifest retrieval. Deep mode additionally measures first-byte latency and a bounded blob sample. A 401 challenge from /v2/ is not itself failure when token flow and manifest retrieval succeed.
Complete when: every selected candidate has a terminal probe result and no candidate is silently omitted.
Step 5 — Qualify and update state
Apply hard qualification before relative scoring:
- Quick qualification requires successful API and manifest retrieval.
- Deep qualification also requires the configured minimum bounded-blob throughput.
--dry-run leaves the cache unchanged.- A successful authorized check updates measurements and clears consecutive failures.
- One authorized failure degrades a recorded candidate; reaching the configured consecutive-failure threshold quarantines it.
- Discovery normalizes and deduplicates URLs, skips every active or quarantined record, deep-tests only new candidates, and stores only qualified results.
The CLI owns cache parsing, locking, atomic replacement, status migration, and INI serialization; do not edit the cache directly during a probe branch.
Complete when: every candidate is qualified or rejected and any authorized cache change is atomically committed or reported as failed.
Step 6 — Rank and recommend
Rank qualified candidates using the configured mix of manifest latency, blob throughput, historical success rate, API latency, freshness, and consecutive-failure penalty.
Default output should identify the comprehensive best candidate, backups, the evidence used, and an unexecuted replacement docker pull command when an image is known. Use detailed or JSON output only when the user needs raw metrics or scenario-specific winners.
Suggest discovery when active records are empty, all tested candidates fail, or a cache classified as expired has no qualified result. Cache age alone does not invalidate a candidate; use the thresholds in config/defaults.json.
State that public-proxy reachability measures current transport behavior, not operator trust or nationwide performance.
Complete when: the recommendation is evidence-backed, alternatives and limitations are visible, and no generated Docker command has been executed.
Step 7 — Close the branch
Report the selected branch, command mode, cache side effects, qualified and rejected counts, recommendation, and unresolved risks. For write-capable work, declare completion only after the CLI exits successfully and the expected state can be read back.
Complete when: the user can distinguish what was observed, what changed, what was only recommended, and what remains unverified.
