[HugoXK]

Fixes #2173

Written-by: Dev-4
Owner: Dev-4
Reviewer: Reviewer
Claim-Id: 5c3608254a1248f88de2047c2212a66f
Branch: dev/4-settings-link-dedupe
Base: b282329
Tier: T2 — the live .github/scripts/risk-gate.mjs judge over this diff prints risk-tier: T2; basis: plugins/workflow/client.js=T2, plugins/workflow/test/client.test.js=T2, changelog.d/2173-settings-link-duplicate.fixed.md=T1, docs/subsystems/workflow.md=T0.

What changed

Settings → Workflows → Add link built the edge and submitted it without reading the edges the graph already lists. The host keys a link by its pair alone (plugins/workflow/src/graph.js:150 edgeKey takes from/to, not the kind), so a second edge over the same two Bots makes normalizeGraph throw Link A → B is listed twice (plugins/workflow/src/graph.js:257) and refuses the whole graph: workflow.edges keeps what it had, and the user is left with one wf-err line and no new link. The same refusal met a different kind over a pair that already had one.

plugins/workflow/client.js now reads the pair before it submits, with the rule the drawer's withEdge (:656) already applies to its own wires:

• a pair the graph already lists, given the same kind, is a no-op: the form clears and nothing is submitted;
• a pair the graph already lists, given another kind, is not submitted either — the banner names the kind that is there (This pair already has a {kind} link — replace or delete it first.), so the user reads what blocks the save instead of the host's listed twice;
• the To list marks the targets the chosen From already reaches — Name (linked), 名称（已连接） — so the pair is visible before the form is filled in.

Three helpers sit beside withEdge: EDGE_KIND_LABEL, blankEdgeForm() (the one empty form the initial state and both clear paths use), and edgeKindOf(edge), which maps a stored edge to the form's kind the way the save writes it (loop === true → loop, when → condition, otherwise forward).

Scope is this one form. The host plugins/workflow/src/graph.js is unchanged: normalizeGraph still refuses a graph that lists one pair twice, so a caller that writes a duplicate through the API is refused exactly as before. This change only stops the Settings form from producing one. P-411 (when/until typed output names) is a separate issue and is untouched.

Red first

On live main b282329d6e79c85c9ebb3f037eeadf36bb9c1396, source untouched, only this branch's plugins/workflow/test/client.test.js dropped into a clean git archive tree (plugins/workflow/client.js there hashes f9c68c59f9de0b2343ae9bb77dd5a2c13858a5da46af79a6a96dcfc420b47085, the main blob):

not ok 109 - #2173: adding a link the graph already has is a no-op, not a refused duplicate save
  error: |-
    the pair the graph already has is not submitted again

    1 !== 0
not ok 110 - #2173: a linked pair given another kind names the link already there instead of submitting
  error: |-
    the conflicting pair is not submitted blind

    1 !== 0
not ok 111 - #2173: the To list marks a target the From node already reaches
  error: 'a target coder already reaches is marked'
  expected: true
  actual: false
# tests 112
# pass 109
# fail 3

The fourth case, #2173: a pair the graph does not have still saves and shows, passes on that base as well — it is the control that the guard does not eat a new pair. With the guard in place the same file on the tip is # tests 112 / # pass 112 / # fail 0.

The test host refuses a repeated pair the way normalizeGraph does, so the assertions are on the committed workflow/graph payload (nothing submitted / one edge) and on the banner text, not on a mock call count.

Verification at the tip

Tip: 7a586a43bb1548acdd2506fd47140183a23e5632 (tree 7d0dcc2d4f1f399b318520018b1f96d0f411352c, single parent b282329d6e79c85c9ebb3f037eeadf36bb9c1396 — live main at push time), 4 files, +135/−3: changelog.d/2173-settings-link-duplicate.fixed.md 1/0, docs/subsystems/workflow.md 2/0, plugins/workflow/client.js 29/3, plugins/workflow/test/client.test.js 103/0. git status --porcelain is empty.

Ran TZ=UTC ARCH_DOCS_BASE=b282329d6e79c85c9ebb3f037eeadf36bb9c1396 npm run verify → exit 0 — core 2648, social 555, tree 127, timeline 140, workflow 354, .github/scripts 598, interop 37, live-e2e 24, every suite # fail 0; lint 74/36/9/2/16 warnings and interop 1, live-e2e 0, 0 errors; arch-docs: OK (4 changed file(s)); ds-bot-patches check — ds-bot/ = Fei2411/DeepSeek-Bot@2b10c70 (tree 5834c00, v0.1.5), 10 patches apply cleanly in order, core/ equals ds-bot/ + patches/ds-bot/, patches/ds-bot/ is what generate writes (generate is a no-op). patches/ds-bot/ has no hit on plugins/workflow/client.js, so nothing is mirrored there and generate is not involved.

The branch was rebased onto b282329d at push time (it was built on ff160976, which is b282329d's parent): same four files, same deltas, same blobs.

Not run (and why): the CI package job (pack check, typecheck, actionlint) — it is not part of npm run verify; and the smoke bench, which this change does not reach.

Acceptance

1. Same pair, same kind — submitting a link the graph already has submits nothing: the form clears, no error banner appears, and workflow.edges is unchanged.
2. Same pair, another kind — the form is not submitted blind; the banner names the kind already on that pair and says to replace or delete it first, instead of the host's Link A → B is listed twice.
3. Visible before submitting — the To list marks every target the chosen From already reaches, so the repeat is on screen before Save; a pair the graph does not have still saves and appears.

The host still refuses a duplicated pair written through the API; plugins/workflow/src/graph.js is untouched.

Checkpoint

{
  "pr": 2174,
  "claim": "Dev-4",
  "issue": 2173,
  "summary": "Settings Add link submitted a pair the graph already listed; the host keys a link by its pair, so normalizeGraph refused the whole graph and saved nothing, leaving one wf-err line. The form now reads the pair first: the same kind is a no-op, another kind names the kind already there instead of submitting, and the To list marks the targets the chosen From already reaches. The host is unchanged and still refuses a duplicate written through the API.",
  "files": [
    "plugins/workflow/client.js",
    "plugins/workflow/test/client.test.js",
    "docs/subsystems/workflow.md",
    "changelog.d/2173-settings-link-duplicate.fixed.md"
  ],
  "checks": { "ran": "TZ=UTC ARCH_DOCS_BASE=b282329d6e79c85c9ebb3f037eeadf36bb9c1396 npm run verify (exit 0); red-first on clean b282329d; workflow suite on the tip", "result": "green", "at": "7a586a43bb1548acdd2506fd47140183a23e5632" },
  "next": "none — ready for review"
}

[HugoXK]

@Birfy please review #2174 at 7a586a4

[Birfy]

Verdict: PASS at 7a586a4
Written-by: Reviewer
Attestation: 7a586a4 role=Reviewer

Ran at this tip:

$ cd b2174/tip && node --test plugins/workflow/test/client.test.js
# tests 112  # pass 112  # fail 0
$ cd b2174 && rm -rf red && cp -a tip red && cp base/plugins/workflow/client.js red/plugins/workflow/client.js && cd red && node --test plugins/workflow/test/client.test.js
# tests 112  # pass 109  # fail 3

Review points:

• Scope. base b282329d6e79 → head 7a586a43 is 1 commit, 4 files: plugins/workflow/client.js +29/−3, plugins/workflow/test/client.test.js +103/−0, docs/subsystems/workflow.md +2, changelog.d/2173-settings-link-duplicate.fixed.md +1. Nothing outside the stated fix.
• Red first. With only plugins/workflow/client.js swapped back to its base blob (tests left at the tip), the three new workflow: Settings → Add link submits an already-listed pair, so the host rejects the whole save #2173 tests fail: 109/112 pass, 3 fail (adding a link the graph already has is a no-op…, …given another kind names the link already there…, the To list marks a target…). The fix is what turns them green.
• Matches the host rule. plugins/workflow/src/graph.js:150 keys an edge by from/to alone and normalizeGraph throws listed twice at :257, so a second edge over a pair refuses the whole save whatever its kind. client.js:1616-1628 now reads the pair before it saves: same kind is a no-op that clears the form; another kind is not submitted and the banner names the kind already there; edgeKindOf maps loop/when/forward the way the save writes them. The To list marks reachable targets (linked) (:1771), and the zh strings are added beside the existing Add-link ones. blankEdgeForm passed to useState as a lazy initializer is correct (it is called with no args and returns the object).
• Checks at this tip. 17 check-runs: all success or cancelled duplicates, the only failure is enqueue-guard at 14:04, and its own reason is no Verifier report names the tip 7a586a43bb15 — this report is that report. tests-ok re-ran green at 14:06.
• Not run: the package-wide npm test for plugins/workflow needs the workspace install (@deepseek-ai/dsh-llm is missing from a bare tarball), so I ran the affected test file; CI's plugins (workflow) check covers the rest and is green.