Imported from ehartye/wiki-master (skills/clip-confluence/SKILL.md). Install upstream with npx skills add ehartye/wiki-master --skill clip-confluence. Copyright stays with the author.

Read the shared core once per session. Read access and host commands before touching the vault. Also read for this operation: operations. Load the rest only when this run reaches it — evidence before a quote or figure from this capture lands on a wiki page, or when reporting a fidelity limit. Before the first authorized write, follow the shared operations completion contract; reuse existing session authorization.

Clipping a Confluence page into the wiki

/wiki-discover's clipper (clip.mjs → Defuddle) handles public HTML pages only, by an unauthenticated fetch. A Confluence Cloud page is also HTML, but sits behind Atlassian auth — Defuddle's anonymous request hits the login wall or an empty SPA shell, reads as thin content, and gets auto-declined. This skill is the Confluence path. The canonical stored artifact is the Markdown the confluence skill's page.mjs already returns, never a PDF export and never a second fetch — that keeps the vault greppable, diffable, and answerable, and makes [[note]] provenance resolve to a real clipping.

Why not export the page to PDF and run /clip-pdf on it instead? That was considered and rejected: it is a double lossy conversion (Confluence storage format → PDF render → pdftotext extraction) instead of the one clean hop page.mjs already does to Markdown, clip-pdf.mjs's own code documents that table extraction can "mispair rows beyond recovery" while looking perfectly clean character-wise, and Confluence validation/compliance corpora are dominated by exactly that shape of content (approval blocks, signature tables, requirement-to-test traceability matrices). It also does not fix live Jira macros that already fail to render for page.mjs ("Data cannot be retrieved due to an unexpected error") — that is a server-side macro-render failure, likely reproduced in PDF export too, minus the helpful fallback JQL link page.mjs gives you. And it stays fully manual per page, which does not scale to a /wiki-discover-style sweep across dozens of pages.

How it works

clip-confluence.mjs does not implement its own Confluence client. It shells out to the separately installed confluencer plugin's own page.mjs (the same script the confluence skill uses), parses the provenance header page.mjs always prints (# Title / Space: X · Page ID: Y · Version: Z · Updated: <iso> / URL: <url> / --- / body), and writes raw/clippings/<slug>.md with the standard clipping frontmatter (source, created, tags:[clippings], quality, source-hash) — the Confluence- specific detail (space, page ID, version, updated timestamp) stays in the clipping's body exactly as page.mjs formatted it, rather than growing the shared frontmatter schema for one source type. It skips duplicates and prior declines, and records a decline for a thin extraction (a parent/index page with no body prose) so it is not retried blindly.

Deliberate exception to wiki-master's usual portability rule. Every other clip-* skill (pdf, docx, xlsx, pptx) is self-contained — none of them runtime-depend on another installed Copilot plugin, specifically so wiki-master behaves the same whether or not some other plugin happens to be present. This skill breaks that rule on purpose: Confluence access is a whole configured, authenticated Atlassian integration (base URL, account email, API token, connectivity), not a generic stateless library like pandoc or python-pptx, and re-implementing an Atlassian API client inside wiki-master to preserve portability would be the wrong trade. So this skill treats confluencer as an optional runtime dependency and is built to degrade gracefully rather than fail confusingly when it is absent — see the next section.

Graceful degradation when confluencer isn't installed

clip-confluence.mjs never assumes confluencer is present. It searches every ~/.copilot/installed-plugins/*/confluencer/ install for its page.mjs under that plugin's own scripts folder (every Copilot CLI plugin installs under that root, even though neither plugin knows the other's exact path at authoring time), or an explicit WIKI_MASTER_CONFLUENCER_SCRIPTS env var override for a dev checkout or non-standard install.

• Preflight, any time: node "<absolute-plugin-root>/scripts/clip-confluence.mjs" --doctor reports whether confluencer was found, and if so, hands off to its own doctor.mjs for the auth/config/connectivity check — wiki-master has no business diagnosing another plugin's credentials, only routing to the tool that can.
• Not found: the clip attempt prints one clear message (which paths it searched, and the env var that can override) and returns { status: 'confluencer-missing' } — never a raw ENOENT/stack trace, and never a fabricated clipping standing in for content nobody actually fetched.
• Found but failing (bad page ID, expired token, network): the real page.mjs stderr is surfaced (first few lines) plus a pointer to run confluencer's own doctor.mjs.

If it reports missing, either install the confluencer plugin, or hand-fetch the page via the confluence skill yourself and clip it through /clip-docx or /clip-pdf as a fallback (lower fidelity for tables, per the rejected-PDF argument above, but better than nothing) — do not hand-write a raw/clippings/*.md file directly; that would bypass dedup, decline-tracking, and hashing the same way it would for any other source type.

Steps

1. Preflight (once, or whenever a clip fails unexpectedly): node "<absolute-plugin-root>/scripts/clip-confluence.mjs" --doctor.
2. Clip (this is the only writer to raw/ for Confluence pages): node "<absolute-plugin-root>/scripts/clip-confluence.mjs" "<url-or-page-id>" --quality=<tier> --topic="<topic>"
   • Accepts anything page.mjs itself accepts as a positional argument: a canonical URL, a tiny link (/wiki/x/…), a legacy viewpage.action URL, or a bare numeric page ID. If you only have an exact title + space, first resolve it via the confluence skill (page.mjs --title "..." --space ... or search.mjs), then pass the resulting URL/ID here.
   • No --source= flag (unlike clip-pdf/clip-docx) — the canonical URL always comes from page.mjs's own metadata, which is the authoritative citation per the confluence skill's own rule ("always cite the title, URL, and version").
   • --topic whenever this clip belongs to a research run — pass the topic string /wiki-discover was given, identical across every clip in the run, so /wiki-triage can group the run's leftovers together. Topic is recorded going forward only and no tool can retro-fit it, so a clip made without it is an Unattributed triage row permanently. Omit it only when there is no research run behind the clip; never invent one.
   • A thin result means the page has no readable body (a parent/index page with only child links) — a decline is recorded; this is not a failure.
   • A failed result means page.mjs itself errored (bad id/auth/network) — report it for manual handling; do not invent the text.
3. Verify the clipping landed: read raw/clippings/<slug>.md and sanity-check the body against what page.mjs would show you directly via the confluence skill.
4. Hand off to /wiki-ingest exactly as with any other clipping — summarize into wiki/sources/, cross-reference, index, log. Ingestion requires authorization for this scope; reuse explicit session authorization.

Known limitation: re-clipping an updated page

Confluence pages mutate in place at a stable URL — a page's Version can climb many times in a single day. The dedup gate is keyed on bare URL (same as every other clip-* script), so deliberately re-clipping a page that has been edited since its last clip reads as duplicate (already clipped) and is silently skipped, exactly as it would for any live web page whose content changed since its last clip. This is not new to Confluence, but Confluence pages hit it far more often than the mostly-static sources the dedup layer was designed around. Not fixed here — flagged for a future version-aware dedup (key on URL + Version, not URL alone) if re-clipping updated pages becomes a real workflow.

Guardrails

• Never edit the body of anything under raw/ — clipped content is immutable source-of-truth (guardrail #1). Frontmatter is pipeline state, tooling-only.
• clip-confluence.mjs is the sole writer to raw/ for Confluence pages — the model never writes the clipping by hand (that would bypass dedup, decline, and hashing).
• Fidelity, not truth: a faithful clip of a wrong or outdated page is still wrong — confirm you clipped the page you meant, and note the Version you clipped when citing it, since the live page may have moved on.

Completion

For a standalone clip, open an operation before the clipping helper writes, verify its returned paths and extraction diagnostics, log once, commit and verify already-authorized sync through the shared operations contract. During discovery, use the enclosing operation and report the results to its owner for completion. Capturing evidence does not by itself authorize ingestion; reuse explicit discover-and-ingest authorization when it is already present.