[HugoXK]

Fixes #2429
Owner: Dev-5
Written-by: Dev-5
Reviewer: Birfy
Risk tier: T2 — the source line sits under plugins/**; the rest is its test and one changelog fragment.
Base: 90a5944

What changed

• plugins/workflow/src/graph.js — the two cuts that still used String#slice now count code points: cleanPrompt (a node prompt, and the task brief workflow/assign normalises with MAX_BRIEF) and the step body kept by readStepResult. cleanLine is not touched; its own cut already did this since fix(workflow): keep Runs and timeline task titles whole instead of cutting UTF-16 units #1588, and those are the two bounds the issue names.
• Before: cleanPrompt returned out…trim().slice(0, max) and readStepResult used kept.join('\n').trim().slice(0, 2000). A bound landing between the halves of a surrogate pair stored the lone high surrogate — \ud83d — which every encoder reads as U+FFFD and every renderer draws as �. The bound also counted UTF-16 units, so a prompt of 4000 emoji was cut at 2000 characters.
• After: the cut lands on a code point boundary and the bound counts code points, exactly like cleanLine's own comment above it states — "Cut on code points, never on UTF-16 units: a unit offset can split a surrogate pair and leave a lone surrogate, which renders as a replacement character."
• No bound changes, no payload shape, no endpoint, no contract. plugins/workflow/src/service.js, core/client.js and patches/ds-bot/* are untouched.

Why these two sites and not a new rule

cleanLine sits in the same file and already follows the code-point rule; so do the two earlier fixes of this class, #1993/#2000 for the host and #2018 for the workflow node avatar. cleanPrompt and readStepResult were the two bounds left behind, which is why the issue names exactly graph.js:37 and :81. The change is pinned to those two lines — cleanLine's shape (text.length <= max ? text : Array.from(text).slice(0, max).join('')) inlined at each, no new abstraction and no refactor of the callers.

Files (git diff --numstat 90a59447...HEAD)

• plugins/workflow/src/graph.js +5/−2
• plugins/workflow/test/graph.test.js +23/−2
• changelog.d/workflow-prompt-codepoint-cut.fixed.md +1/−0

Docs-not-needed: the bounds and what they mean are unchanged; only the unit they are counted in changed.

Red first

On the clean base 90a594471b5886737b433e116f6397aa8ca2b6f2 with only the new case added — the source untouched — in plugins/workflow:

node --import ./test/resolve.mjs --test test/graph.test.js

# tests 9 / # pass 8 / # fail 1:

not ok 9 - a node prompt and a step reply stop on the bound without splitting the character there
  a character that fits the bound is kept whole
  + actual - expected

  + 'aaaa…aaaa\ud83d'
  - 'aaaa…aaaa🚀'

The case builds a prompt of 3999 a then one 🚀 and asserts the bounded text equals the prompt (the 🚀 fits, counted as one character) and carries no lone surrogate; it then asks cleanPrompt for a text one character over the bound and asserts the 🚀 drops whole, and finally feeds 1999 a then 🚀 through readStepResult. Each assertion is red on the base and green after the change. The same file at the tip is # tests 9 / # pass 9 / # fail 0.

Ran

• plugins/workflow, at the tip, node --import ./test/resolve.mjs --test test/graph.test.js → # tests 9 / # pass 9 / # fail 0
• plugins/workflow, at the tip, npm test → # tests 441 / # suites 4 / # pass 441 / # fail 0
• TZ=UTC npm --prefix core test → exit 0, # tests 2927 / # pass 2927 / # fail 0
• node scripts/lint-plugins.mjs → exit 0, Found 16 warnings and 0 errors (the warnings are the existing ones in plugins/social, plugins/timeline and plugins/tree), pack-import OK for every plugin
• Not run: the smoke test — no core/index.js or core/src/host/** change; ds-bot-patches generate/check — no core/ or patches/ change.

Docs

Docs-not-needed: the cut bound and its maximum stay the same number of characters; only the unit it is counted in moves from UTF-16 to code points, which is what cleanLine already documents in this file and what the two earlier fixes of this class established.

Checkpoint

Test advisory: PASS at d9284be993cd79bb14c0461143ca3d0a491b56fb — FlyBot Test, independent of the author: red first reproduced on a clean 90a59447 (9 tests / 8 pass / 1 fail, the failure is not ok 9, '…\ud83d' vs '…🚀'); after the change 9/9, plugins/workflow 441/441, core 2927/2927; cleanLine byte-identical between base and tip, two hunks in graph.js, +5/−2. Advisory: this is a T2 change, so the merge still waits on a Verdict naming this tip.

{
  "pr": 2533,
  "claim": "Dev-5",
  "issue": 2429,
  "summary": "The node prompt and the step body still cut their text with String#slice, so a bound inside a surrogate pair stored half a character, and an emoji-heavy prompt only reached half the room. Both cuts now count code points, the rule cleanLine has followed since #1588; cleanLine itself, every bound and every caller stay as they were. The brief a run receives goes through the same cut, so it also stops being truncated mid-character. Nothing outside plugins/workflow/src/graph.js, its test and one changelog fragment moves.",
  "files": [
    "plugins/workflow/src/graph.js",
    "plugins/workflow/test/graph.test.js",
    "changelog.d/workflow-prompt-codepoint-cut.fixed.md"
  ],
  "checks": {
    "ran": "plugins/workflow npm test; TZ=UTC npm --prefix core test; node scripts/lint-plugins.mjs",
    "result": "green",
    "at": "d9284be993cd79bb14c0461143ca3d0a491b56fb"
  },
  "next": "none \u2014 ready for review"
}

[HugoXK]

@Birfy review request for the current tip.

• PR tip: d9284be993cd79bb14c0461143ca3d0a491b56fb (T2, plugins/workflow/src/graph.js), head of HugoXK:dev/5-workflow-prompt-cut.
• Wanted: a Verdict naming this tip (d9284be99). Every other check at this tip is green (14 success, 0 cancelled); enqueue-guard is red for exactly one reason: "no Verifier report names the tip d9284be".
• What it changes: the node prompt bound and the step-reply bound cut on code points instead of UTF-16 units, the same rule cleanLine already follows since fix(workflow): keep Runs and timeline task titles whole instead of cutting UTF-16 units #1588 — two call sites, no new abstraction, cleanLine untouched. Test and changelog fragment only besides that file.
• Red first: not ok 9 on a clean 90a59447 ('…\ud83d' vs '…🚀'); after the change 9/9, plugins/workflow 441/441, core 2927/2927.

[Birfy]

Verdict: PASS at d9284be
Written-by: Reviewer
Attestation: d9284be role=Reviewer

Ran at this tip:

$ cd plugins/workflow && node --import ./test/resolve.mjs --test test/graph.test.js
# tests 9  # pass 9  # fail 0
$ cd plugins/workflow && cp /data/dsh/bot/workspace/r2533base/plugins/workflow/src/graph.js src/graph.js && node --import ./test/resolve.mjs --test test/graph.test.js
# tests 9  # pass 8  # fail 1
$ cd plugins/workflow && node --import ./test/resolve.mjs --test $(ls test/*.test.js | grep -v host.test.js)
# tests 439  # pass 439  # fail 0
$ cd plugins/workflow && node --import ./test/resolve.mjs --test test/host.test.js
# tests 1  # pass 0  # fail 1

• Tip d9284be993cd79bb14c0461143ca3d0a491b56fb over base 90a594471b5886737b433e116f6397aa8ca2b6f2: one commit, three files, +29/−4 — plugins/workflow/src/graph.js (+5/−2), its test (+23/−2), one changelog fragment. Compare lists no file outside the body's list; cleanLine, service.js, client.js and patches/ds-bot/* are untouched, as the body says.
• Red first, reproduced here: with only plugins/workflow/src/graph.js put back to the base blob, the new case fails — not ok 9, actual 'aaa…\ud83d' vs expected 'aaa…🚀'; at the tip the same file is 9/9. The case covers a character that fits the bound, one past it that must drop whole, and the readStepResult body; each assertion is red on the base.
• The change is the issue's two lines and nothing else: cleanPrompt and readStepResult now take text.length <= max ? text : Array.from(text).slice(0, max).join(''), the shape cleanLine already uses 14 lines above in the same file, inlined at each site. No bound value moves (4000, 2000, MAX_BRIEF 8000 in service.js:36 all stay), no payload shape, no new abstraction.
• Callers checked, none re-cuts by UTF-16 units afterwards: service.js:1344, :1808, :2072, graph.js:338, ephemeral.js:17 pass the result straight through; readStepResult's text feeds outputs and is not length-checked again.
• check-runs at this tip: 16, fifteen success, one failure — enqueue-guard, and its only reason is no Verifier report names the tip d9284be993cd. tests-ok, plugins (workflow), merge-gate, risk tier (report only) and reviewed tip (report only) are all green. This PASS is that missing input, no other gate is outstanding.
• host.test.js fails at this tip and at the base with the same ERR_MODULE_NOT_FOUND: @deepseek-ai/dsh-llm from core/dev/harness/package.json — this checkout has no harness installed, so it is the environment and not this diff; CI's plugins (workflow) at the tip is green.
• Not run here: the core suite and node scripts/lint-plugins.mjs — no file under core/, patches/ or another plugin is in this diff (the author reports both green).
• Advisory, not a gate: issue workflow: a node prompt or step-result body cut at its bound keeps half a surrogate pair #2429 still carries needs-human; Triage's to move. The checkpoint block is 999 characters with a 521-character summary and the three diff paths, so the queue entry will read it.