feat(workflow): propose set_node_field op for node title/mode/flags edits

[claude]

<[CDATA[
Requested by Christian Byrne · Slack thread

What

Supersedes this PR's earlier set_title proposal with a generalized
set_node_field op — set or clear one durable node field: title, mode,
flags.collapsed or flags.pinned — as amendment v1.6 to
docs/op-vocabulary-v1.md (§1.8), implemented end to end in this repo:
workflow_ops.set_node_field, apply_op/apply_specs dispatch, the
comfy workflow set-node-field <file> <node_id> <field> <value> (or
--clear) CLI command, and the bundled agent-facing skill doc
(comfy_cli/skills/comfy/SKILL.md).

This is a proposal, not a ratified contract change. docs/op-vocabulary-v1.md
is FROZEN and pinned by SHA in three other repos (cloud services/agent,
harness, the merge consumer) — adding a kind to it is a decision for this
repo's maintainers, and §9's own process (update FROZEN_OPS/BATCHABLE_OPS/
dispatch tables/doc together, enforced by
tests/comfy_cli/test_op_vocabulary_contract.py) is what makes "implemented"
and "ratified" separable: this PR does the former in full so the proposal is
concrete and testable, and stays open as a draft for the latter.

Why this supersedes the earlier set_title proposal

This PR originally proposed a title-only set_title op, mirroring
comfy-multi-player#232's
provisional, package-local fix for a title-stomp bug (docs/decisions/ADR-032-set-title-op.md):
a node's title only ever enters a document as a passthrough field on
add_node's initial snapshot, and nothing wrote it afterward, so a rename
made after node creation never produced a mergeable op.

That same whole-node-upsert clobber applies equally to mode (mute/bypass)
and the two flags.* fields once flags.collapsed/flags.pinned sync came
up against the identical gap (ComfyUI_frontend's canvas needs to sync those
out, and the in-app agent can build a graph but couldn't label, mute, collapse
or pin one). comfy-multi-player generalized its own prototype the same way:
comfy-multi-player#235
merged a set_node_field CRDT op to that repo's main (commit
28d2e6166867f16bc291f2ae60d7f4a6a0c0c9f7), superseding its earlier
title-only #232/ADR-032 prototype. This PR withdraws comfy-cli's own
title-only set_title proposal (never released in any comfy-cli version)
and replaces it with set_node_field, matching that merged upstream shape
rather than landing three more single-field ops in sequence.

docs/op-vocabulary-v1.md §1.8.1 records the one payload difference from
comfy-multi-player's shape (node_incarnation, which comfy-cli's
never-reused node ids make unnecessary here) and why.

What's in this PR

• docs/op-vocabulary-v1.md: §1.8 (set_node_field definition, the closed
  WRITABLE_NODE_FIELDS list and per-field value types) + Amendment v1.6,
  explicitly marked PROPOSED / pending ratification, and recording
  set_title's withdrawal.
• comfy_cli/workflow_ops.py: WRITABLE_NODE_FIELDS, set_node_field()
  (with per-field type validation — title a string, mode an int,
  flags.collapsed/flags.pinned booleans, bool explicitly excluded from
  mode since it's an int subclass in Python), _apply_set_node_field,
  dispatch in apply_op/apply_specs, the ("node_field", node_id, field)
_write_target register (one LWW register per field, so two fields of one
  node — or a field and a widget — never contend), FROZEN_OPS/BATCHABLE_OPS.
• comfy_cli/command/workflow_edit.py + workflow.py: the
  comfy workflow set-node-field command, replacing set-title.
• comfy_cli/discovery.py: schema registration for the new command.
• comfy_cli/skills/comfy/SKILL.md: documents set-node-field alongside the
  other structured-edit primitives for the in-app agent's path.
• CHANGELOG.md: [Unreleased] entry, marked proposed/pending ratification.
• .github/workflows/public-repo-hygiene.yml: the ADR-032 ticket-allowlist
  comment updated for the new context (the entry itself is unchanged and
  still needed — ADR-032 is still cited as prior art).
• Tests (TDD, red→green): tests/comfy_cli/test_set_node_field_op.py
  replaces test_set_title_op.py and was committed first, failing 34/34
  against the prior baseline (no set_node_field, no CLI command, apply_op
  rejects the kind as unknown); the implementation commit makes all 34 pass.
  Covers all four fields, per-field type validation (including the
  bool-vs-mode subclass trap), minting, LWW convergence in both apply
  orders, per-field LWW independence (two fields of one node never contend),
  delete-wins, idempotent replay, the set_node_field/set_widget register
  separation, batch dispatch, and the CLI surface (including the
  --clear/VALUE mutual-exclusivity guard, the unknown-node and
  outside-the-allowlist error paths). tests/comfy_cli/test_op_vocabulary_contract.py
  (the generic doc==code pin) passes unmodified against the new kind.
  Full suite: 7875 passed / 6 pre-existing, unrelated failures (proxy/opener +
  github-pr tests failing the same way on main without this branch) / 46
  skipped. ruff check / ruff format --check clean.

Explicitly out of scope

• Ratification. Maintainers decide whether/how set_node_field joins the
  frozen vocabulary; nothing here should be read as already-approved.
• Reconciling comfy-multi-player's merged op with this one at the doc
  level — that's follow-up work for once (and if) this amendment lands.
• An interior/subgraph-promoted set_node_field variant — §1.8 states
  this explicitly, mirroring comfy-multi-player#235's own scope.
• Changes to cloud, ComfyUI_frontend, or Comfy-Desktop. In particular,
  emitting set_node_field from an actual canvas edit in
  ComfyUI_frontend is a separate, scoped follow-up — this PR only adds the
  op and its comfy-cli-side authoring paths (CLI + in-app-agent skill doc).

Open questions for maintainers

• Does the payload shape in §1.8 (no node_incarnation) match what a future
  cloud merge-consumer implementation would need, or should comfy-cli carry
  the field too (unused today, forward-compatible with comfy-multi-player)?
• Should ratification be a separate, focused PR against just
  docs/op-vocabulary-v1.md (doc-only, reviewed by whoever owns the frozen
  contract) with this PR's code changes rebased on top once that lands? Opened
  as one PR here since §9 requires them to land in one commit for the
  contract test to hold, but happy to split if that's the preferred review
  shape.

🤖 Generated with Claude Code

https://claude.ai/code/session_01L4WMaCPdk7KSva7vwKfz19]]>

[coderabbitai]

Important

Review skipped

Bot user detected.

To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Advanced

Run ID: 56e1b91b-f556-40ca-b17b-a72cfa09ee5c

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}

[claude]

Investigated the "Run Tests on GPU Runners (linux)" failure (run 35593028063).

Not this PR's fault. The actual failures are in tests/e2e/test_e2e_uv_compile.py::test_progressive_conflict and ::test_node_uv_sync_standalone_conflict, both asserting on the string "Conflicting packages (by node pack):". Instead, uv's resolver bails out earlier with Resolution failed: compile failed: error: No solution found when resolving dependencies, because UnifiedDepResolver rejects a git+https://github.com/facebookresearch/sam2 path separator coming from the comfyui-impact-pack test fixture. Nothing in the diff (workflow_ops.py, workflow_edit.py, the new set_title op/tests) touches dependency resolution or that test file.

Confirmed this is pre-existing/infra, not introduced here: the same job fails identically on main itself (e.g. push run 35563195394 at main's current head aec5220c..., and several other unrelated PRs' GPU runs), all with the same "No solution found when resolving dependencies" signature. This looks like drift in the external sam2 git dependency used by the impact-pack fixture repo, not something in this repo's diff.

Since it's not this PR's fault, I didn't push anything. I did trigger rerun_failed_jobs on the failing run once, in case it's transient, but given it also reproduces on main I'd expect it to fail again until the underlying sam2/uv-resolver issue is fixed upstream.

---

Generated by Claude Code

[claude]

CI red triage: "Run Tests on GPU Runners (linux)" — pre-existing failure, unrelated to this PR's diff

The GPU job (run 35672539012) failed on 2 tests in tests/e2e/test_e2e_uv_compile.py:

• test_progressive_conflict
• test_node_uv_sync_standalone_conflict

Both fail the same assertion, assert "Conflicting packages (by node pack):" in combined, because ComfyUI-Manager's dependency resolver (cm_cli) now errors out earlier with:

[UnifiedDepResolver] rejected path separator: 'git+https://github.com/facebookresearch/sam2' from .../custom_nodes/comfyui-impact-pack
Resolution failed: compile failed: error: No solution found when resolving dependencies

instead of ever reaching the "Conflicting packages (by node pack):" message the tests expect.

Why this is unrelated to this PR: this PR's diff only touches workflow_ops.py, workflow_edit.py, workflow.py, discovery.py, docs/skill files, and the set_node_field test/changelog — no node-install, uv-sync, or ComfyUI-Manager code. The identical two tests, with the identical rejected path separator: 'git+https://github.com/facebookresearch/sam2' error, are failing on the GPU workflow across a run of completely unrelated PRs right now (e.g. run 35666479755 on kishore/validate-gaps, a graph-validation fix with no overlap with this diff, hits the exact same error), and on this PR's own prior commit (run 35593565021) before this rework existed. Every GPU run in the recent history I checked has conclusion: failure for this same reason — this looks like an upstream ComfyUI-Manager/sam2 dependency-declaration change breaking the conflict-detection fixture the test suite relies on, not a regression from this PR.

Per the CI-red playbook, I re-ran the failed job once to rule out a flake (it was not) rather than attempting an unrelated fix here.

---

Generated by Claude Code