[tony]

Why

The MCP worked example on the Julia page (https://libtmux.org/en/julia/latest/) shows "Requests only" because scripts/gen-mcp-transcript.mjs had no Julia entry and site/src/data/mcp-transcripts/julia.json did not exist. Two properties of the Julia server stopped the example from being recorded as written:

• Its pane tools take a target of { paneId, generation }. The generation is the tmux server's process id and start time (for example <pid>:<start time>), list_panes returns it, and capture_pane refuses any other value. The example's generation: '3f9a1c' was invented; the server answers it with stale_target. A real value differs on every run, so a recording that kept it could not be stable or equal to EXAMPLES.julia.
• The server borrows a tmux that is already running, and its list_panes names the session and window through contexts and the pane through target. The replay read neither, so even a good recording drew an empty tmux.

What

• site/src/data/mcp-protocol/julia.json is refreshed to libtmux-julia master (c19e4a164; it described 0.1.0-alpha.1, master reports 0.1.1), and Julia's entry of mcp-tools.json follows it because mcp-reference.test.ts binds the two. No other port's entry changes.
• scripts/gen-mcp-transcript.mjs gains a julia entry. It installs LibTmux and LibTmuxMCP from the checkout into an environment, writes the launcher that install_cli documents, and borrows a tmux whose first pane runs cat checks.log -, so "what the first pane is displaying" is the log the other ports use. The server runs with the default tool catalog, which holds both tools the example calls.
• scripts/lib/mcp-session.mjs: a call's arguments may be a function of the earlier steps (withListed, listedValue). The recorder sends capture_pane the generation list_panes returned and stores the example's stand-in everywhere in the recording, so the steps equal EXAMPLES.julia. The leak check also refuses a recording that still holds the real value, and sandbox.rewritten.generation names the stand-in.
• The page says so, in the same sentence that already names the rewritten project directory: "The server's generation, which differs on every run, is shown as 3f9a1c."
• site/src/lib/mcp-replay.ts draws a pane from a listing that carries contexts and target, and shows what capture_pane read as the pane's screen. The preview now shows [docs], window 0:cat, pane %0 cat and the two log lines; every drawn word is still checked against the recording.
• EXAMPLES.julia is unchanged.

The stand-in is a rewrite of recorded data, disclosed on the page and in the data file. The alternative is to drop the generation from the example by configuring a caller pane, which would hide the target flow that is the point of Julia's design. I kept the flow and disclosed the rewrite; say if you prefer the other.

Rebase note

#247 (fix/ruby-core-showcase-refresh-20261103) is changing scripts/gen-mcp-transcript.mjs, scripts/lib/mcp-transcript-checks.mjs and site/src/data/mcp-tools.json. This PR touches the first and third (the SERVERS map, the PRESENT line, the calls/marks/leaks expressions, and Julia's region of mcp-tools.json) and leaves mcp-transcript-checks.mjs alone. Whichever PR merges second must rebase. If mcp-tools.json conflicts, regenerate rather than merge by hand: LIBTMUX_DOCS_CHECKOUT_JULIA=<clone at c19e4a164> node scripts/gen-mcp-tools.mjs --port julia.

Verification

Recording, made at libtmux-julia c19e4a164 with tmux 3.8a on a private socket under /home/d/tmp-wave/heavy.sh:

• Recorded three times, the third with the final script; all three files are byte-identical, and --check passes. The pane's command is cat in every recording because the pane starts cat checks.log - directly, so list_panes cannot catch a shell on its way to exec, the kind of start-up race rs's split pane had. The log is printed long before the first call, which comes after the Julia server has loaded.
• All recorder checks pass: no failed step, every id a request names was returned earlier, the last answer states checks finished, and the recording holds no host path, user name, host name or real generation.
• contracts equal the committed snapshot's for list_panes and capture_pane; the recording holds only while those schemas hold.

Tests (vitest, workers capped, through heavy.sh):

• mcp-showcase, mcp-transcript-checks, mcp-replay, port-landing-page, the new mcp-session, and every other non-outer mcp-* suite and port-landing: 14 files, 525 tests passed.
• mcp-protocol.outer, mcp-descriptions.outer, mcp-showcase.outer, mcp-source-bound.outer: 4 files, 57 tests passed.
• astro check: 0 errors, 0 warnings (needed --max-old-space-size=5120). oxlint reports nothing in the touched files.
• Julia is in the recorded-results, earlier-revision and replay tests, so changing an input or output schema of list_panes or capture_pane hides the recording. A mutated recording (a changed capture text, no rewritten.generation) fails four tests, and the new mcp-session tests include a control showing the server refuses the stand-in.

Not run: a site build, the publication audit, gen-mcp-tools.mjs --check against the standard sibling checkouts (whose heads differ from the recorded revision), the assembled-tree suites (skipped: nothing assembled) and the full pnpm test outer loop. The preview was checked through the component tests, not in a browser.