Feature/basic prompt drafts

[Psylenceo]

Adds an optional shot field, basic_prompt, for a plain-language scene idea
kept separate from the H3-formatted scene_prompt/prompt. It never affects
prompt_hash or generation by itself.

Visible and editable everywhere a scene's prompt already is — Plan (and
Modern), Studio Plan (even while prompt editing is delegated to an external
editor), Review Gate (gated by the existing prompt-editor setting), and both
Scene Prompt Editor nodes — kept in sync through the same companion-sync bus
already used for scene_prompt (rebaseScenePrompt, planHasNonPromptChanges,
and a new publishCompanionBasicPrompt/_h3PromptCompanionSetBasicPrompt
pair mirroring the existing prompt broadcast).

Only Rich Scene Prompt Editor's Optimize action turns a basic prompt into a
real scene_prompt; its history tree also shows/restores basic_prompt per
revision.

Bundled with this are three Review Gate routing fixes surfaced while testing
the above against nested-subgraph frontends:

• deliverReview/routeReview's node-id lookups (findNodeByQualifiedId,
  upstreamPlanNode, prepareResume, reviewFallbackNode) now resolve
  against the true root graph instead of app.graph, which can itself be a
  subgraph's own graph object.
• The fallback node picked by reviewFallbackNode() is now delivered with
  {verifyRun: false} — it was being re-run through the same strict
  run_name check that had already selected it as the lone remaining gate,
  which is what caused the Review Gate to hang with no video/prompt and
  unresponsive buttons.
• Rejected review events now log which check failed and the expected/actual
  values, and a token's resolved routing decision is cached per-token so it
  isn't re-derived and re-warned about on every repeated message while a
  review stays open.

Commits

• Add a per-scene basic prompt: draft, sync, optimize, and history
• Fix Review Gate routing when app.graph is not the true root
• Log why a review event was rejected, not just that it was
• Fix Review Gate routing: fallback node was re-rejected by its own safety check
• Stop re-logging the same review routing decision on every message

[ethanfel]

Reviewed at a3d0cfe47db5fe9ec64d258bcd608182dc962150. The basic-draft feature and true-root graph traversal are useful, but I'm holding this PR for three reproducible issues:

1. Review Gate fallback can route a review into the wrong project

In routeReview, both the fallback and remembered-token paths call deliverReview(..., {verifyRun: false}). The fallback can choose a same-leaf-ID gate or the only gate even when its connected Plan explicitly belongs to a different run.

Reproduced with one gate, ID 22, connected to a Plan for project_B, then a non-durable review event {token: "test-token", node_id: 22, run_name: "project_A"}. Exact delivery rejects the run mismatch, but fallback delivers it anyway: the gate for B receives A's review token. That exposes another project's pending review controls in this workflow.

Please retain project/workflow identity validation during fallback. If stale widgets require recovery, resolve against the queued workflow/prompt identity rather than treating a singleton or leaf ID as sufficient proof. Also revalidate cached routing when the graph, connected Plan, or mounted node changes. Add regressions for different projects sharing a leaf ID and cached gates after workflow changes/removal.

2. Editing one prompt field can overwrite a newer edit to the other

rebaseScenePrompt now unconditionally copies both prompt and basic_prompt from the editor's local shot into the latest Plan.

Minimal reproduction with the real function:

const local = {shots: [{id: "one", prompt: "new H3 edit", basic_prompt: "old basic"}]};
const live = {shots: [{id: "one", prompt: "old H3", basic_prompt: "new basic from Studio"}]};
rebaseScenePrompt(local, live, 0);
// live.shots[0].basic_prompt is now "old basic".

An H3-only edit has discarded the newer basic draft. The reverse direction is also unsafe: a basic-only edit can copy stale H3 text. Please track/rebase the edited field independently, and ensure basic-draft synchronization is published/consumed by the ordinary prompt editors as well as Review Gate. Add concurrent-edit tests in both directions.

3. Saving a different basic draft mutates an already-executed history revision

save_draft finds an existing revision by H3 text alone, then overwrites its basic_prompt without checking whether that revision was executed. The matching path in mark_executed needs the same consideration.

Reproduced using PromptHistoryStore in a temporary directory:

old = store.mark_executed(
    "test_project", "scene_one", "same H3 result", "original basic draft"
)["revision"]
new = store.save_draft(
    "test_project", "scene_one", "same H3 result", old["id"], "different basic draft"
)["revision"]
# new["id"] == old["id"]
# Reading old["id"] now returns basic_prompt == "different basic draft".

This loses the original draft associated with an executed revision. Please include basic-draft content in revision matching or fork a child when it changes on an executed revision, while retaining mutable unexecuted drafts. Test different basic drafts producing the same H3 text, including re-execution.

The existing review, companion-sync, simple/rich editor, and prompt-history checks passed in the isolated PR worktree, but they do not cover these cases. Please add the regression tests with the fixes before we merge this one.

[Psylenceo]

Yep.
Crap and there's also a major bug where all basic prompts can get deleted.

[Psylenceo]

OK, i've pushed an update.
And resolved an issue where reloading a checkpoint wipes the basic prompts even though the checkpoint info still has the basic prompt.

[ethanfel]

Thanks, the checkpoint-reload fix checks out. I ported the basic-prompt feature and your checkpoint/executed-history fixes to main (680c68d) and nightly (5d6b0f1), with contributor credit.

I also fixed the remaining editor-sync case: saving a basic draft in editor A, then editing the H3 prompt in editor B, could still restore the old basic draft. Both fields now synchronize before an editor marks the Plan current. Added tests using the real simple/rich editor write and receiver functions, including clears and delayed notifications.

The Review Gate routing changes are still excluded. At af95a6e, this still reproduces: route a token for project A, change the connected Plan's run name to project B without reloading the graph, then deliver the same token again. The cached path in routeReview still uses verifyRun: false, so B receives A's review. Clearing the cache on configure/remove does not cover an ordinary run-name change or live rewiring. Please validate the current project identity on cached delivery too, and add a behavioral regression for that sequence.

The targeted Python checks pass on both branches; the JavaScript suites pass 52 scripts on main and 66 on nightly. I have not merged or closed this PR because the routing issue remains.

[Psylenceo]

ok, I'll shuffle that in with a different thought I had, to remove the basic prompt from teh studio node. Not sure if its really needed there so i'm thinking of removing it from that node.

also I'm tossing the idea around of using tabs instead of two separate prompt boxs stacked. Use a tab system so that the plan node, optimizer node, and review gate nodes don't need to be so tall.

[ethanfel]

About the red tags after Optimize: I reproduced the matching reference state at af95a6e using the actual reference resolver and optimizer-result handler in a small JavaScript test.

• The optimized prompt is saved, but the editor redraws with the old reference state. A newly introduced @hero stays inactive with no <Picture 1> mapping, while its image source is still available for hover. Refreshing references against the new prompt fixes the mapping. Manually replacing a tag already performs that refresh, which explains your workaround. This part also reproduces on main and nightly without the basic-prompt feature: it is an existing rich-editor bug.
• Separately, this PR prepares optimizer references from the shared/old H3 prompt, not basic_prompt. Tags present only in the basic prompt are therefore excluded from the Direct API optimizer's active media list.

The fix should include basic-prompt tags when preparing optimizer references, then refresh the reference state/tray before redrawing an applied result, covering Direct API, MCP, and pending-result application. A regression should add a tag through optimization without requiring a manual edit afterward.

This was not a full generation test; red tags alone do not prove the render omitted the references. No fix has been applied yet.

Also, a correction to my earlier status: the basic-prompt feature port has since been reverted from both main and nightly. This PR remains open for continued work.

[ethanfel]

Thanks for the update. I retested 9e6a295: the cached Gate routing now correctly rejects a review after the connected project changes, including a run-name edit, a Project Assets change, and rewiring to another Plan. The checkpoint-reload and executed-history checks also pass. Basic-only tags now reach the optimizer's reference scan, and I verified that a connected tagged image is included in the Direct request.

There are still two issues to sort out before merging:

1. Editing through a second editor can still undo a saved basic draft

This is the remaining sync case from the earlier review, not the field-level rebase test:

1. Connect two prompt editors to the same Plan, with basic_prompt = "old basic".
2. Save "new basic from A" in editor A and let its companion notification reach B.
3. Edit only the H3 prompt in editor B.
4. The saved Plan now contains B's new H3 prompt, but basic_prompt has reverted to "old basic".

I reproduced this with rich → rich, simple → simple, rich → simple, and simple → rich, using the actual write and receiver functions.

The ordinary editors' flushPlanEffects() publishes the H3 prompt but not the basic draft. The receiving _h3PromptCompanionSetScenePrompt() updates only prompt, then marks state.lastValue current even though its local basic_prompt is stale. The next write skips rebasing and serializes that stale draft. See the rich-editor receiver; the simple editor has the same problem.

Please synchronize both fields before treating the local Plan as current, while preserving field-level edit ownership. The regression needs the full A-save → B-receive → B-save sequence, not just a direct call to rebaseScenePrompt. This one still causes data loss.

2. After Optimize, draft-only tags can change the displayed reference mapping

The new refreshTaggedReferencesForShot helper includes basic_prompt when rebuilding the applied H3 prompt's references.

Reproduction with two enabled carousel pictures, ordered @unused then @hero:

• Basic draft: Consider @unused and @hero.
• Optimizer result: A close-up of @hero.
• Actual editor state: both references remain active; @hero maps to <Picture 2>.
• Expected from the resulting H3 prompt: @unused is inactive; @hero maps to <Picture 1>.

This reproduces through Direct API, MCP, and pending-result application. Including basic-draft tags when preparing the optimizer request is correct; after applying the result, refresh against the shared prompt plus the resulting H3 text only. Please cover the case where a draft tag is omitted from the optimized result, as well as newly introduced tags.

This is a confirmed editor mapping mismatch, not a claim that a generated video used the wrong image. The tests were isolated JavaScript/backend checks, with no live generation or user-project changes. The existing targeted tests pass but miss these two sequences.

GitHub also currently reports merge conflicts, so the branch will need updating before it can merge.