Polish Dream QA user tone
#18
by ADJCJH - opened
- PRODUCT.md +2 -0
- README.md +2 -1
- docs/handoff.md +6 -0
- docs/prd.md +4 -0
- docs/smoke/2026-06-08-post-refactor-polish-smoke.md +101 -6
- docs/smoke/2026-06-09-user-tone-polish-smoke.md +79 -0
- docs/spec.md +4 -0
- docs/superpowers/plans/2026-06-08-post-refactor-polish.md +7 -5
- docs/superpowers/plans/2026-06-09-dream-qa-user-tone-polish.md +182 -0
- dream_customs/models.py +2 -2
- dream_customs/pipeline.py +88 -7
- dream_customs/prompts.py +6 -3
- dream_customs/render.py +2 -2
- dream_customs/ui/app.py +1 -1
- dream_customs/ui/copy.py +4 -6
- scripts/evaluate_today_tip_quality.py +8 -0
- tests/fixtures/today_tip_eval_cases.json +7 -0
- tests/test_today_tip_quality_eval.py +1 -1
- tests/test_ui_actions.py +29 -3
PRODUCT.md
CHANGED
|
@@ -42,6 +42,8 @@ The voice is a calm dream companion who asks good questions and takes odd detail
|
|
| 42 |
6. The demo path must survive Space constraints and still have a model-backed route.
|
| 43 |
7. English mode is the default public path; Chinese mode is a first-class toggle, not a separate product.
|
| 44 |
8. Recommendation quality is tested with a deterministic Today Tip eval set before deployment.
|
|
|
|
|
|
|
| 45 |
|
| 46 |
## Accessibility & Inclusion
|
| 47 |
|
|
|
|
| 42 |
6. The demo path must survive Space constraints and still have a model-backed route.
|
| 43 |
7. English mode is the default public path; Chinese mode is a first-class toggle, not a separate product.
|
| 44 |
8. Recommendation quality is tested with a deterministic Today Tip eval set before deployment.
|
| 45 |
+
9. English mode must stay natural English, including translated or paraphrased dream anchors; mixed Chinese helper text is a regression.
|
| 46 |
+
10. Tiny actions should feel like a first step the user may choose, not a hard command to solve the real-world problem immediately.
|
| 47 |
|
| 48 |
## Accessibility & Inclusion
|
| 49 |
|
README.md
CHANGED
|
@@ -55,8 +55,9 @@ This is not a therapy, diagnosis, or prophecy product.
|
|
| 55 |
|
| 56 |
- Default public UI: English.
|
| 57 |
- Toggle: `English / 中文`.
|
| 58 |
-
- English mode
|
| 59 |
- Chinese mode keeps the warm `梦境问答台` wording.
|
|
|
|
| 60 |
|
| 61 |
## Current Direction References
|
| 62 |
|
|
|
|
| 55 |
|
| 56 |
- Default public UI: English.
|
| 57 |
- Toggle: `English / 中文`.
|
| 58 |
+
- English mode translates or paraphrases non-English dream anchors into natural English, so an international judge does not see mixed UI/model language.
|
| 59 |
- Chinese mode keeps the warm `梦境问答台` wording.
|
| 60 |
+
- Today Tips should be small and optional: "write the first sentence" beats "handle it immediately."
|
| 61 |
|
| 62 |
## Current Direction References
|
| 63 |
|
docs/handoff.md
CHANGED
|
@@ -51,6 +51,12 @@ Final positioning:
|
|
| 51 |
|
| 52 |
> Dream QA helps you record a dream, answer or skip one gentle question, and leave with one grounded Today Tip.
|
| 53 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 54 |
## Model Decision
|
| 55 |
|
| 56 |
Use:
|
|
|
|
| 51 |
|
| 52 |
> Dream QA helps you record a dream, answer or skip one gentle question, and leave with one grounded Today Tip.
|
| 53 |
|
| 54 |
+
2026-06-09 tone contract:
|
| 55 |
+
|
| 56 |
+
- English mode is not just English UI chrome; generated anchors, questions, and Today Tips should read as natural English.
|
| 57 |
+
- If the model receives Chinese anchors while the selected language is English, translate or paraphrase them unless a short exact quote is intentionally preserved.
|
| 58 |
+
- Tiny actions should feel like five-minute first steps. Avoid urgent command phrasing such as "address it immediately."
|
| 59 |
+
|
| 60 |
## Model Decision
|
| 61 |
|
| 62 |
Use:
|
docs/prd.md
CHANGED
|
@@ -102,6 +102,7 @@ Acceptance criteria:
|
|
| 102 |
- Result is readable on mobile and desktop.
|
| 103 |
- Result avoids medical, prophetic, or deterministic framing.
|
| 104 |
- Today tip references a concrete dream anchor.
|
|
|
|
| 105 |
|
| 106 |
#### Story 6: Bilingual Demo
|
| 107 |
|
|
@@ -112,6 +113,7 @@ Acceptance criteria:
|
|
| 112 |
- Default UI title is `Dream QA`.
|
| 113 |
- A visible `English / 中文` control exists.
|
| 114 |
- English mode renders English labels and card headings.
|
|
|
|
| 115 |
- Chinese mode renders `梦境问答台` and Chinese Today Tip labels.
|
| 116 |
- Both modes use the same model/pipeline route.
|
| 117 |
|
|
@@ -147,6 +149,8 @@ Prompts must:
|
|
| 147 |
- Keep final advice small enough to execute today.
|
| 148 |
- Require at least one dream anchor in the final tip.
|
| 149 |
- Follow selected language: English by default, Chinese when selected.
|
|
|
|
|
|
|
| 150 |
|
| 151 |
### Evaluation Strategy
|
| 152 |
|
|
|
|
| 102 |
- Result is readable on mobile and desktop.
|
| 103 |
- Result avoids medical, prophetic, or deterministic framing.
|
| 104 |
- Today tip references a concrete dream anchor.
|
| 105 |
+
- Optional tiny action is a five-minute first step, not a command to solve the whole issue immediately.
|
| 106 |
|
| 107 |
#### Story 6: Bilingual Demo
|
| 108 |
|
|
|
|
| 113 |
- Default UI title is `Dream QA`.
|
| 114 |
- A visible `English / 中文` control exists.
|
| 115 |
- English mode renders English labels and card headings.
|
| 116 |
+
- English mode translates or paraphrases non-English anchors into English unless preserving a deliberate user quote.
|
| 117 |
- Chinese mode renders `梦境问答台` and Chinese Today Tip labels.
|
| 118 |
- Both modes use the same model/pipeline route.
|
| 119 |
|
|
|
|
| 149 |
- Keep final advice small enough to execute today.
|
| 150 |
- Require at least one dream anchor in the final tip.
|
| 151 |
- Follow selected language: English by default, Chinese when selected.
|
| 152 |
+
- Avoid mixed-language leakage in English output.
|
| 153 |
+
- Avoid hard commands such as "address it immediately"; prefer optional first-step language.
|
| 154 |
|
| 155 |
### Evaluation Strategy
|
| 156 |
|
docs/smoke/2026-06-08-post-refactor-polish-smoke.md
CHANGED
|
@@ -127,11 +127,106 @@ has_zh_anchor=True
|
|
| 127 |
- Chrome extension control returned a closed native pipe while opening the same local URL.
|
| 128 |
- Because both browser automation surfaces were unstable, no visual screenshot claim is made here. The local Gradio `/config`, app build check, action flow, full pytest suite, and deterministic quality evaluator passed.
|
| 129 |
|
| 130 |
-
##
|
| 131 |
|
| 132 |
-
|
| 133 |
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 127 |
- Chrome extension control returned a closed native pipe while opening the same local URL.
|
| 128 |
- Because both browser automation surfaces were unstable, no visual screenshot claim is made here. The local Gradio `/config`, app build check, action flow, full pytest suite, and deterministic quality evaluator passed.
|
| 129 |
|
| 130 |
+
## GitHub Sync
|
| 131 |
|
| 132 |
+
Feature branch commit:
|
| 133 |
|
| 134 |
+
```text
|
| 135 |
+
0193b67 feat: polish dream qa post-refactor flow
|
| 136 |
+
```
|
| 137 |
+
|
| 138 |
+
GitHub PR:
|
| 139 |
+
|
| 140 |
+
```text
|
| 141 |
+
https://github.com/adjcjh777/dream-customs-build-small/pull/2
|
| 142 |
+
```
|
| 143 |
+
|
| 144 |
+
Observed PR status:
|
| 145 |
+
|
| 146 |
+
```text
|
| 147 |
+
MERGED
|
| 148 |
+
merge_commit=f2cafa12f9db694acf2181a54fc6059419bbd4e7
|
| 149 |
+
```
|
| 150 |
+
|
| 151 |
+
Result: GitHub `main` contains the English-first Dream QA polish.
|
| 152 |
+
|
| 153 |
+
## Hugging Face Space Sync
|
| 154 |
+
|
| 155 |
+
HF Space PR:
|
| 156 |
+
|
| 157 |
+
```text
|
| 158 |
+
https://huggingface.co/spaces/build-small-hackathon/dream-customs/discussions/17
|
| 159 |
+
```
|
| 160 |
+
|
| 161 |
+
Observed refs:
|
| 162 |
+
|
| 163 |
+
```text
|
| 164 |
+
space/main=19c54925bbb525405bc2540391a434ddaeba4139
|
| 165 |
+
refs/pr/17=2017dd6906f3cce640e4a53508ea21b163c3b52e
|
| 166 |
+
```
|
| 167 |
+
|
| 168 |
+
Observed discussion metadata:
|
| 169 |
+
|
| 170 |
+
```text
|
| 171 |
+
num=17
|
| 172 |
+
title=Deploy Dream QA post-refactor polish
|
| 173 |
+
status=open
|
| 174 |
+
is_pull_request=True
|
| 175 |
+
```
|
| 176 |
+
|
| 177 |
+
Merge status: BLOCKED.
|
| 178 |
+
|
| 179 |
+
Blockers observed:
|
| 180 |
+
|
| 181 |
+
- Chrome extension automation could list browser sessions but failed to control Chrome tabs with `native pipe is closed`.
|
| 182 |
+
- HF Hub API `merge_pull_request(... discussion_num=17)` returned an `HfHubHTTPError`; this matches the previous pattern where API merge rights are not available from the cached token.
|
| 183 |
+
- Follow-up Chrome diagnostics on 2026-06-09 showed:
|
| 184 |
+
- Google Chrome is installed and running.
|
| 185 |
+
- Codex Chrome Extension is installed and enabled in the selected `Default` profile.
|
| 186 |
+
- Native host manifest exists and has the expected extension origin.
|
| 187 |
+
- Despite those checks, Chrome automation still fails at tab control with `native pipe is closed`.
|
| 188 |
+
- A follow-up API merge retry still returned `HfHubHTTPError`.
|
| 189 |
+
- A safety check for direct Space git sync could not prove fast-forward safety because `git fetch space main` failed at the HF git/LFS negotiation layer with `fatal: expected 'acknowledgments'`.
|
| 190 |
+
|
| 191 |
+
Result: HF Space PR #17 is ready but not merged. Public Space `main` is still `19c54925...` until a logged-in browser merge succeeds or the HF token gains merge permission.
|
| 192 |
+
|
| 193 |
+
## Hugging Face Merge Follow-up - 2026-06-09
|
| 194 |
+
|
| 195 |
+
Chrome logged-in merge succeeded.
|
| 196 |
+
|
| 197 |
+
Observed discussion metadata:
|
| 198 |
+
|
| 199 |
+
```text
|
| 200 |
+
num=17
|
| 201 |
+
title=Deploy Dream QA post-refactor polish
|
| 202 |
+
status=merged
|
| 203 |
+
is_pull_request=True
|
| 204 |
+
```
|
| 205 |
+
|
| 206 |
+
Observed refs after merge:
|
| 207 |
+
|
| 208 |
+
```text
|
| 209 |
+
space/main=c2a7a861b44613ff477695f4f1a15f781a440fae
|
| 210 |
+
refs/pr/17=2017dd6906f3cce640e4a53508ea21b163c3b52e
|
| 211 |
+
```
|
| 212 |
+
|
| 213 |
+
HF Space was paused immediately after merge, then restarted from the Space page.
|
| 214 |
+
|
| 215 |
+
Runtime check:
|
| 216 |
+
|
| 217 |
+
```text
|
| 218 |
+
stage=RUNNING
|
| 219 |
+
hardware=zero-a10g
|
| 220 |
+
requested_hardware=zero-a10g
|
| 221 |
+
```
|
| 222 |
+
|
| 223 |
+
Public Space `/config` check:
|
| 224 |
+
|
| 225 |
+
```text
|
| 226 |
+
title=Dream QA
|
| 227 |
+
version=4.44.1
|
| 228 |
+
mode=blocks
|
| 229 |
+
component_count=72
|
| 230 |
+
```
|
| 231 |
+
|
| 232 |
+
Result: HF Space `main` now contains the Dream QA post-refactor polish, and the public Space is serving the updated Gradio app config.
|
docs/smoke/2026-06-09-user-tone-polish-smoke.md
ADDED
|
@@ -0,0 +1,79 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Dream QA User Tone Polish Smoke - 2026-06-09
|
| 2 |
+
|
| 3 |
+
## Scope
|
| 4 |
+
|
| 5 |
+
- Remove English-mode Chinese anchor leakage observed in the public Space user walk-through.
|
| 6 |
+
- Make Today Tip actions warmer, smaller, and less command-like.
|
| 7 |
+
- Remove backend jargon from the public user flow.
|
| 8 |
+
- Keep Chinese mode available through the language toggle.
|
| 9 |
+
|
| 10 |
+
## Local Verification
|
| 11 |
+
|
| 12 |
+
Commands:
|
| 13 |
+
|
| 14 |
+
```text
|
| 15 |
+
.venv/bin/python -m pytest -q
|
| 16 |
+
.venv/bin/python scripts/evaluate_today_tip_quality.py
|
| 17 |
+
git diff --check
|
| 18 |
+
```
|
| 19 |
+
|
| 20 |
+
Observed:
|
| 21 |
+
|
| 22 |
+
```text
|
| 23 |
+
95 passed, 2 warnings
|
| 24 |
+
{"case_count": 11, "failures": {}, "passes": true}
|
| 25 |
+
git diff --check clean
|
| 26 |
+
```
|
| 27 |
+
|
| 28 |
+
## Local Gradio Config
|
| 29 |
+
|
| 30 |
+
Observed from `http://127.0.0.1:7860/config`:
|
| 31 |
+
|
| 32 |
+
```text
|
| 33 |
+
title=Dream QA
|
| 34 |
+
version=4.44.1
|
| 35 |
+
mode=blocks
|
| 36 |
+
component_count=72
|
| 37 |
+
```
|
| 38 |
+
|
| 39 |
+
## Local Action Smoke
|
| 40 |
+
|
| 41 |
+
English dream:
|
| 42 |
+
|
| 43 |
+
```text
|
| 44 |
+
I dreamed I was in an elevator where the floor buttons melted like wax. The number 14 kept blinking, and I felt late but strangely calm.
|
| 45 |
+
```
|
| 46 |
+
|
| 47 |
+
English answer:
|
| 48 |
+
|
| 49 |
+
```text
|
| 50 |
+
I want to make starting my overdue email easier without feeling trapped by it.
|
| 51 |
+
```
|
| 52 |
+
|
| 53 |
+
Observed:
|
| 54 |
+
|
| 55 |
+
```text
|
| 56 |
+
en_after_submit=ask
|
| 57 |
+
question=When you think about the elevator and the melted button, what real thing today feels hard to start?
|
| 58 |
+
en_after_answer=tip
|
| 59 |
+
title=Today Tip
|
| 60 |
+
has_chinese_leakage=False
|
| 61 |
+
has_first_sentence=True
|
| 62 |
+
has_immediately=False
|
| 63 |
+
```
|
| 64 |
+
|
| 65 |
+
Chinese smoke:
|
| 66 |
+
|
| 67 |
+
```text
|
| 68 |
+
zh_after_skip=tip
|
| 69 |
+
title=今日小 Tips
|
| 70 |
+
has_zh_title=True
|
| 71 |
+
```
|
| 72 |
+
|
| 73 |
+
## Hugging Face Space Sync
|
| 74 |
+
|
| 75 |
+
Pending.
|
| 76 |
+
|
| 77 |
+
## Public User-View Review
|
| 78 |
+
|
| 79 |
+
Pending.
|
docs/spec.md
CHANGED
|
@@ -32,6 +32,8 @@ The user wakes up after a vivid or unsettling dream and wants help making sense
|
|
| 32 |
6. End with one small today tip, not a pile of advice.
|
| 33 |
7. Keep the MVP stable enough for Gradio Space and demo-video recording.
|
| 34 |
8. Default to English for public judging; keep Chinese as a first-class toggle.
|
|
|
|
|
|
|
| 35 |
|
| 36 |
## 5. User Inputs
|
| 37 |
|
|
@@ -187,6 +189,8 @@ The final result must:
|
|
| 187 |
- Avoid generic advice that could fit any dream.
|
| 188 |
- Keep the today tip small enough to try today.
|
| 189 |
- Use English UI/output labels by default and Chinese labels only when `中文` is selected.
|
|
|
|
|
|
|
| 190 |
|
| 191 |
## 10. UX Structure
|
| 192 |
|
|
|
|
| 32 |
6. End with one small today tip, not a pile of advice.
|
| 33 |
7. Keep the MVP stable enough for Gradio Space and demo-video recording.
|
| 34 |
8. Default to English for public judging; keep Chinese as a first-class toggle.
|
| 35 |
+
9. In English mode, translate or paraphrase non-English dream anchors into natural English instead of leaking mixed-language helper text.
|
| 36 |
+
10. Treat tiny actions as optional first steps; avoid urgent commands such as "fix it immediately" or "address it immediately."
|
| 37 |
|
| 38 |
## 5. User Inputs
|
| 39 |
|
|
|
|
| 189 |
- Avoid generic advice that could fit any dream.
|
| 190 |
- Keep the today tip small enough to try today.
|
| 191 |
- Use English UI/output labels by default and Chinese labels only when `中文` is selected.
|
| 192 |
+
- Keep English mode fully English unless preserving a deliberate short user quote.
|
| 193 |
+
- Make the optional tiny action a five-minute first step, not a demand to complete the whole waking-life problem.
|
| 194 |
|
| 195 |
## 10. UX Structure
|
| 196 |
|
docs/superpowers/plans/2026-06-08-post-refactor-polish.md
CHANGED
|
@@ -25,7 +25,7 @@
|
|
| 25 |
- No source edits required.
|
| 26 |
- Record result in `docs/smoke/2026-06-08-post-refactor-polish-smoke.md`.
|
| 27 |
|
| 28 |
-
- [
|
| 29 |
|
| 30 |
Run:
|
| 31 |
|
|
@@ -36,7 +36,7 @@ git log --oneline origin/main..HEAD
|
|
| 36 |
|
| 37 |
Expected: shows the post-merge Dream QA commits that are on `feature/dream-customs-ui-voice-settings` but not `origin/main`.
|
| 38 |
|
| 39 |
-
- [
|
| 40 |
|
| 41 |
Run:
|
| 42 |
|
|
@@ -54,7 +54,7 @@ gh pr create \
|
|
| 54 |
--body-file /tmp/dream-qa-post-refactor-pr.md
|
| 55 |
```
|
| 56 |
|
| 57 |
-
- [
|
| 58 |
|
| 59 |
Use the authenticated browser/CLI path available in the session. Verify with:
|
| 60 |
|
|
@@ -70,7 +70,7 @@ Expected: `state` is `MERGED`.
|
|
| 70 |
- Create or update: `docs/smoke/2026-06-08-post-refactor-polish-smoke.md`
|
| 71 |
- Test: no new unit test required for this task.
|
| 72 |
|
| 73 |
-
- [
|
| 74 |
|
| 75 |
Run:
|
| 76 |
|
|
@@ -227,7 +227,7 @@ Run:
|
|
| 227 |
git diff --check
|
| 228 |
```
|
| 229 |
|
| 230 |
-
- [
|
| 231 |
|
| 232 |
Run:
|
| 233 |
|
|
@@ -247,6 +247,8 @@ git ls-remote space refs/heads/main
|
|
| 247 |
|
| 248 |
Expected: Space `main` advances to the deployment commit.
|
| 249 |
|
|
|
|
|
|
|
| 250 |
## Self-Review
|
| 251 |
|
| 252 |
- Spec coverage: covers GitHub main alignment, deployed smoke, English-first toggle, quality evaluation, MiniCPM follow-up quality, docs, tests, commit, push, and HF Space sync.
|
|
|
|
| 25 |
- No source edits required.
|
| 26 |
- Record result in `docs/smoke/2026-06-08-post-refactor-polish-smoke.md`.
|
| 27 |
|
| 28 |
+
- [x] **Step 1: Verify branch relationship**
|
| 29 |
|
| 30 |
Run:
|
| 31 |
|
|
|
|
| 36 |
|
| 37 |
Expected: shows the post-merge Dream QA commits that are on `feature/dream-customs-ui-voice-settings` but not `origin/main`.
|
| 38 |
|
| 39 |
+
- [x] **Step 2: Create or update GitHub PR**
|
| 40 |
|
| 41 |
Run:
|
| 42 |
|
|
|
|
| 54 |
--body-file /tmp/dream-qa-post-refactor-pr.md
|
| 55 |
```
|
| 56 |
|
| 57 |
+
- [x] **Step 3: Merge GitHub PR**
|
| 58 |
|
| 59 |
Use the authenticated browser/CLI path available in the session. Verify with:
|
| 60 |
|
|
|
|
| 70 |
- Create or update: `docs/smoke/2026-06-08-post-refactor-polish-smoke.md`
|
| 71 |
- Test: no new unit test required for this task.
|
| 72 |
|
| 73 |
+
- [x] **Step 1: Verify deployed config**
|
| 74 |
|
| 75 |
Run:
|
| 76 |
|
|
|
|
| 227 |
git diff --check
|
| 228 |
```
|
| 229 |
|
| 230 |
+
- [x] **Step 3: Commit and push**
|
| 231 |
|
| 232 |
Run:
|
| 233 |
|
|
|
|
| 247 |
|
| 248 |
Expected: Space `main` advances to the deployment commit.
|
| 249 |
|
| 250 |
+
Observed 2026-06-09: HF Space PR #17 is open and ready, but merge is blocked because Chrome extension tab control fails with `native pipe is closed`, HF API merge returns `HfHubHTTPError`, and a direct git fast-forward safety check cannot complete because `git fetch space main` fails with `fatal: expected 'acknowledgments'`.
|
| 251 |
+
|
| 252 |
## Self-Review
|
| 253 |
|
| 254 |
- Spec coverage: covers GitHub main alignment, deployed smoke, English-first toggle, quality evaluation, MiniCPM follow-up quality, docs, tests, commit, push, and HF Space sync.
|
docs/superpowers/plans/2026-06-09-dream-qa-user-tone-polish.md
ADDED
|
@@ -0,0 +1,182 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Dream QA User Tone Polish Implementation Plan
|
| 2 |
+
|
| 3 |
+
> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
|
| 4 |
+
|
| 5 |
+
**Goal:** Make the public Dream QA flow feel less technical, keep English mode fully English, and produce warmer, more specific Today Tips after the user's answer.
|
| 6 |
+
|
| 7 |
+
**Architecture:** Keep the existing Dream QA flow and Gradio structure. Tighten prompt contracts, normalize model-returned English cards against common Chinese anchor leakage, soften render/UI safety copy, and extend deterministic tests so the product polish is guarded before Space deployment.
|
| 8 |
+
|
| 9 |
+
**Tech Stack:** Python, Pydantic, Gradio, pytest, Hugging Face Space.
|
| 10 |
+
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
## File Map
|
| 14 |
+
|
| 15 |
+
- Modify `dream_customs/prompts.py`: strengthen English language instructions and final Today Tip behavior.
|
| 16 |
+
- Modify `dream_customs/pipeline.py`: add a small English-output cleanup for common Chinese anchor leakage.
|
| 17 |
+
- Modify `dream_customs/models.py`: improve deterministic demo output.
|
| 18 |
+
- Modify `dream_customs/render.py`: make non-crisis safety copy warmer and less defensive.
|
| 19 |
+
- Modify `dream_customs/ui/copy.py`: remove backend jargon from the public flow and rename runtime copy to advanced user language.
|
| 20 |
+
- Modify `dream_customs/ui/app.py`: rename the public settings accordion from `Runtime settings` to `Advanced`.
|
| 21 |
+
- Modify `tests/test_ui_actions.py`: update UI copy expectations and add English-language leakage checks.
|
| 22 |
+
- Modify `tests/fixtures/today_tip_eval_cases.json`: add an English case that previously exposed Chinese leakage.
|
| 23 |
+
- Modify `scripts/evaluate_today_tip_quality.py`: fail English outputs that contain common Chinese leakage or hard-command phrasing.
|
| 24 |
+
- Modify `docs/spec.md`, `docs/prd.md`, `docs/handoff.md`, `README.md`: align the product contract with the user-tone polish.
|
| 25 |
+
- Create `docs/smoke/2026-06-09-user-tone-polish-smoke.md`: record local, HF, and browser/user-view verification.
|
| 26 |
+
|
| 27 |
+
---
|
| 28 |
+
|
| 29 |
+
### Task 1: Guard English Product Quality
|
| 30 |
+
|
| 31 |
+
**Files:**
|
| 32 |
+
- Modify: `tests/test_ui_actions.py`
|
| 33 |
+
- Modify: `tests/fixtures/today_tip_eval_cases.json`
|
| 34 |
+
- Modify: `scripts/evaluate_today_tip_quality.py`
|
| 35 |
+
|
| 36 |
+
- [x] **Step 1: Add UI copy expectations**
|
| 37 |
+
|
| 38 |
+
Update `test_processing_note_is_story_copy_not_backend_jargon` so it expects public copy to mention the user flow and rejects `model routes`, `fallback`, `token`, `endpoint`, and `debug`.
|
| 39 |
+
|
| 40 |
+
- [x] **Step 2: Add English leakage checks**
|
| 41 |
+
|
| 42 |
+
Add a test that runs an English elevator dream through `submit_dream_action` and `answer_to_card_action`, then asserts the card text/html do not contain `数字`, `电梯`, `按钮`, `楼层`, or `融化`.
|
| 43 |
+
|
| 44 |
+
- [x] **Step 3: Extend deterministic evaluator**
|
| 45 |
+
|
| 46 |
+
Update `scripts/evaluate_today_tip_quality.py` so English cases fail on the same Chinese leakage terms and hard-command phrases such as `address it immediately`.
|
| 47 |
+
|
| 48 |
+
- [x] **Step 4: Add eval fixture**
|
| 49 |
+
|
| 50 |
+
Add or update an English elevator/email case in `tests/fixtures/today_tip_eval_cases.json` with required anchors `elevator`, `button`, and `14`.
|
| 51 |
+
|
| 52 |
+
- [x] **Step 5: Run focused failing tests**
|
| 53 |
+
|
| 54 |
+
Run:
|
| 55 |
+
|
| 56 |
+
```bash
|
| 57 |
+
.venv/bin/python -m pytest tests/test_ui_actions.py -q
|
| 58 |
+
.venv/bin/python scripts/evaluate_today_tip_quality.py
|
| 59 |
+
```
|
| 60 |
+
|
| 61 |
+
Expected before implementation: at least one failure for changed expectations.
|
| 62 |
+
|
| 63 |
+
### Task 2: Polish Prompts, Fallbacks, and Rendering
|
| 64 |
+
|
| 65 |
+
**Files:**
|
| 66 |
+
- Modify: `dream_customs/prompts.py`
|
| 67 |
+
- Modify: `dream_customs/pipeline.py`
|
| 68 |
+
- Modify: `dream_customs/models.py`
|
| 69 |
+
- Modify: `dream_customs/render.py`
|
| 70 |
+
- Modify: `dream_customs/ui/copy.py`
|
| 71 |
+
- Modify: `dream_customs/ui/app.py`
|
| 72 |
+
|
| 73 |
+
- [x] **Step 1: Tighten English prompt contract**
|
| 74 |
+
|
| 75 |
+
Update `_language_instruction` so English mode says to translate or paraphrase non-English dream anchors into natural English, except short quoted fragments that the user explicitly wrote.
|
| 76 |
+
|
| 77 |
+
- [x] **Step 2: Make Today Tip instructions more product-like**
|
| 78 |
+
|
| 79 |
+
Update `today_tip_prompt` to ask for short, warm, specific output; the `tiny_action` should be a 5-minute first step and should avoid demanding words such as "immediately".
|
| 80 |
+
|
| 81 |
+
- [x] **Step 3: Add English cleanup**
|
| 82 |
+
|
| 83 |
+
In `dream_customs/pipeline.py`, normalize common Chinese anchor leakage in English `TodayTipCard` fields after parsing. Preserve Chinese mode unchanged.
|
| 84 |
+
|
| 85 |
+
- [x] **Step 4: Improve deterministic fallback**
|
| 86 |
+
|
| 87 |
+
Update the English fake Today Tip to demonstrate the desired behavior: small first step, no hard command, one dream-detail metaphor, and no Chinese terms.
|
| 88 |
+
|
| 89 |
+
- [x] **Step 5: Soften render copy**
|
| 90 |
+
|
| 91 |
+
Change default English safety copy to a warmer boundary sentence, without weakening crisis safety behavior.
|
| 92 |
+
|
| 93 |
+
- [x] **Step 6: Remove public backend jargon**
|
| 94 |
+
|
| 95 |
+
Change English and Chinese processing/runtime help copy so public users see product language, not model-route/fallback language.
|
| 96 |
+
|
| 97 |
+
- [x] **Step 7: Rename settings surface**
|
| 98 |
+
|
| 99 |
+
Change the visible `Runtime settings` accordion to `Advanced` while keeping advanced endpoint controls collapsed.
|
| 100 |
+
|
| 101 |
+
- [x] **Step 8: Run focused tests**
|
| 102 |
+
|
| 103 |
+
Run:
|
| 104 |
+
|
| 105 |
+
```bash
|
| 106 |
+
.venv/bin/python -m pytest tests/test_ui_actions.py -q
|
| 107 |
+
.venv/bin/python scripts/evaluate_today_tip_quality.py
|
| 108 |
+
```
|
| 109 |
+
|
| 110 |
+
Expected: focused tests pass.
|
| 111 |
+
|
| 112 |
+
### Task 3: Align Docs and Smoke Evidence
|
| 113 |
+
|
| 114 |
+
**Files:**
|
| 115 |
+
- Modify: `README.md`
|
| 116 |
+
- Modify: `PRODUCT.md`
|
| 117 |
+
- Modify: `docs/spec.md`
|
| 118 |
+
- Modify: `docs/prd.md`
|
| 119 |
+
- Modify: `docs/handoff.md`
|
| 120 |
+
- Add: `docs/smoke/2026-06-09-user-tone-polish-smoke.md`
|
| 121 |
+
- Modify: this plan file checkbox state.
|
| 122 |
+
|
| 123 |
+
- [x] **Step 1: Update product docs**
|
| 124 |
+
|
| 125 |
+
Record that English mode should stay natural English, avoid mixed Chinese UI/model terms, and produce small optional actions rather than hard commands.
|
| 126 |
+
|
| 127 |
+
- [x] **Step 2: Run full verification**
|
| 128 |
+
|
| 129 |
+
Run:
|
| 130 |
+
|
| 131 |
+
```bash
|
| 132 |
+
.venv/bin/python -m pytest -q
|
| 133 |
+
.venv/bin/python scripts/evaluate_today_tip_quality.py
|
| 134 |
+
git diff --check
|
| 135 |
+
```
|
| 136 |
+
|
| 137 |
+
Expected: all commands pass.
|
| 138 |
+
|
| 139 |
+
- [x] **Step 3: Run local app config/action smoke**
|
| 140 |
+
|
| 141 |
+
Start the local app on port `7862`, check `/config`, and run one action-level English elevator/email flow plus one Chinese flow.
|
| 142 |
+
|
| 143 |
+
- [x] **Step 4: Record smoke evidence**
|
| 144 |
+
|
| 145 |
+
Write non-secret results to `docs/smoke/2026-06-09-user-tone-polish-smoke.md`.
|
| 146 |
+
|
| 147 |
+
- [ ] **Step 5: Commit and push**
|
| 148 |
+
|
| 149 |
+
Commit the plan, implementation, docs, and smoke evidence. Push `feature/dream-qa-user-tone-polish` to GitHub.
|
| 150 |
+
|
| 151 |
+
### Task 4: Sync HF Space, Browser Merge, and User-View Review
|
| 152 |
+
|
| 153 |
+
**Files:**
|
| 154 |
+
- Modify: `docs/smoke/2026-06-09-user-tone-polish-smoke.md`
|
| 155 |
+
|
| 156 |
+
- [ ] **Step 1: Create HF Space PR**
|
| 157 |
+
|
| 158 |
+
Upload the current folder to `build-small-hackathon/dream-customs` with `create_pr=True`.
|
| 159 |
+
|
| 160 |
+
- [ ] **Step 2: Merge with browser**
|
| 161 |
+
|
| 162 |
+
Open the HF discussion PR in Chrome, verify it is ready to merge, and click Merge from the logged-in browser session.
|
| 163 |
+
|
| 164 |
+
- [ ] **Step 3: Restart and verify Space**
|
| 165 |
+
|
| 166 |
+
If the Space is paused, restart it. Verify runtime is `RUNNING` and public `/config` returns `title=Dream QA`.
|
| 167 |
+
|
| 168 |
+
- [ ] **Step 4: Experience as a user**
|
| 169 |
+
|
| 170 |
+
Use the public Space in the browser with the elevator/email dream. Check that English mode stays English, the question feels grounded, and the Today Tip is warm and actionable.
|
| 171 |
+
|
| 172 |
+
- [ ] **Step 5: Record final evidence**
|
| 173 |
+
|
| 174 |
+
Append HF PR number, merge SHA, runtime, public config, and user-view notes to the smoke document. Commit and push the final evidence.
|
| 175 |
+
|
| 176 |
+
---
|
| 177 |
+
|
| 178 |
+
## Self-Review
|
| 179 |
+
|
| 180 |
+
- Spec coverage: The plan covers the exact user feedback: English leakage, technical UI copy, hard Today Tip command, defensive safety wording, docs alignment, HF browser merge, and post-merge user-view review.
|
| 181 |
+
- Placeholder scan: No TBD/TODO/fill-later placeholders remain.
|
| 182 |
+
- Type consistency: Existing `TodayTipCard`, `submit_dream_action`, and `answer_to_card_action` APIs are reused without schema changes.
|
dream_customs/models.py
CHANGED
|
@@ -99,8 +99,8 @@ class FakeTextClient:
|
|
| 99 |
"Maybe this dream is not predicting lateness. It turns the feeling of being stuck before "
|
| 100 |
"starting into an elevator paused at floor 14."
|
| 101 |
),
|
| 102 |
-
today_tip="For today, press one very small elevator button: open the task
|
| 103 |
-
tiny_action="
|
| 104 |
caring_note="You can move one floor at a time; the whole building does not need to be solved this morning.",
|
| 105 |
)
|
| 106 |
return TodayTipCard(
|
|
|
|
| 99 |
"Maybe this dream is not predicting lateness. It turns the feeling of being stuck before "
|
| 100 |
"starting into an elevator paused at floor 14."
|
| 101 |
),
|
| 102 |
+
today_tip="For today, press one very small elevator button: open the task and write only the first line.",
|
| 103 |
+
tiny_action="Set a five-minute timer, open the task, and stop after the first line.",
|
| 104 |
caring_note="You can move one floor at a time; the whole building does not need to be solved this morning.",
|
| 105 |
)
|
| 106 |
return TodayTipCard(
|
dream_customs/pipeline.py
CHANGED
|
@@ -117,6 +117,22 @@ _ZH_ANCHOR_MARKERS = [
|
|
| 117 |
"楼层",
|
| 118 |
]
|
| 119 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 120 |
|
| 121 |
def _dedupe_preserve_order(items: List[str]) -> List[str]:
|
| 122 |
seen = set()
|
|
@@ -171,15 +187,52 @@ def _extract_dream_anchors(intake: DreamIntake) -> List[str]:
|
|
| 171 |
return _dedupe_preserve_order(candidates)[:5]
|
| 172 |
|
| 173 |
|
| 174 |
-
def
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 175 |
anchors = _extract_dream_anchors(intake)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 176 |
if anchors:
|
| 177 |
return anchors[0]
|
| 178 |
return "梦里的那个细节" if _is_zh(language) else "that dream detail"
|
| 179 |
|
| 180 |
|
| 181 |
def _secondary_anchor(intake: DreamIntake, language: str = "en") -> str:
|
| 182 |
-
anchors =
|
| 183 |
return anchors[1] if len(anchors) > 1 else _primary_anchor(intake, language)
|
| 184 |
|
| 185 |
|
|
@@ -227,10 +280,24 @@ def _fallback_interpretation(intake: DreamIntake, language: str = "en") -> str:
|
|
| 227 |
def _grounded_today_tip(intake: DreamIntake, language: str = "en") -> str:
|
| 228 |
primary = _primary_anchor(intake, language)
|
| 229 |
if not _is_zh(language):
|
| 230 |
-
return
|
|
|
|
|
|
|
|
|
|
| 231 |
return f"今天先从「{primary}」借一个动作:只做最小的第一步,不急着把整件事完成。"
|
| 232 |
|
| 233 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 234 |
def _anchor_in_text(text: str, anchors: List[str]) -> bool:
|
| 235 |
clean = (text or "").lower()
|
| 236 |
for anchor in anchors:
|
|
@@ -483,7 +550,7 @@ def build_qa_state(
|
|
| 483 |
answers: Optional[List[str]] = None,
|
| 484 |
language: str = "en",
|
| 485 |
) -> DreamQAState:
|
| 486 |
-
anchors =
|
| 487 |
return DreamQAState(
|
| 488 |
dream_summary=_summary_from_intake(intake, language),
|
| 489 |
main_question=_main_question_from_intake(intake, language),
|
|
@@ -496,8 +563,12 @@ def build_qa_state(
|
|
| 496 |
|
| 497 |
def _polish_today_tip(card: TodayTipCard, intake: DreamIntake, answers: str = "", language: str = "en") -> TodayTipCard:
|
| 498 |
polished = card.model_copy(deep=True)
|
| 499 |
-
intake_anchors =
|
| 500 |
-
card_anchors =
|
|
|
|
|
|
|
|
|
|
|
|
|
| 501 |
if intake_anchors and not any(_text_uses_anchor(anchor, intake_anchors) for anchor in card_anchors):
|
| 502 |
anchors = intake_anchors
|
| 503 |
else:
|
|
@@ -518,7 +589,15 @@ def _polish_today_tip(card: TodayTipCard, intake: DreamIntake, answers: str = ""
|
|
| 518 |
or not _anchor_in_text(polished.today_tip, anchors)
|
| 519 |
):
|
| 520 |
polished.today_tip = _grounded_today_tip(intake, language)
|
| 521 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 522 |
if _is_zh(language):
|
| 523 |
polished.tiny_action = f"用 5 分钟写下:今天和「{anchors[0]}」有关的第一小步是什么?"
|
| 524 |
else:
|
|
@@ -531,6 +610,8 @@ def _polish_today_tip(card: TodayTipCard, intake: DreamIntake, answers: str = ""
|
|
| 531 |
)
|
| 532 |
merged = "\n".join([intake.merged_text(), answers or ""])
|
| 533 |
polished.safety_note = safety_note() if needs_escalation(merged) else ""
|
|
|
|
|
|
|
| 534 |
return polished
|
| 535 |
|
| 536 |
|
|
|
|
| 117 |
"楼层",
|
| 118 |
]
|
| 119 |
|
| 120 |
+
_ZH_TO_EN_PHRASES = {
|
| 121 |
+
"电梯按钮": "elevator button",
|
| 122 |
+
"融化的按钮": "melted button",
|
| 123 |
+
"按钮融化": "melted button",
|
| 124 |
+
"数字 14": "floor 14",
|
| 125 |
+
"数字14": "floor 14",
|
| 126 |
+
"楼层数字": "floor number",
|
| 127 |
+
"楼层": "floor",
|
| 128 |
+
"老楼": "old apartment building",
|
| 129 |
+
"蓝色楼道": "blue hallway",
|
| 130 |
+
"电梯": "elevator",
|
| 131 |
+
"按钮": "button",
|
| 132 |
+
"融化": "melted",
|
| 133 |
+
"梦境": "dream",
|
| 134 |
+
}
|
| 135 |
+
|
| 136 |
|
| 137 |
def _dedupe_preserve_order(items: List[str]) -> List[str]:
|
| 138 |
seen = set()
|
|
|
|
| 187 |
return _dedupe_preserve_order(candidates)[:5]
|
| 188 |
|
| 189 |
|
| 190 |
+
def _english_anchor_text(text: str) -> str:
|
| 191 |
+
clean = text or ""
|
| 192 |
+
for source, target in sorted(_ZH_TO_EN_PHRASES.items(), key=lambda item: len(item[0]), reverse=True):
|
| 193 |
+
clean = clean.replace(source, target)
|
| 194 |
+
clean = re.sub(r"[\u4e00-\u9fff]+", "dream detail", clean)
|
| 195 |
+
clean = re.sub(r"\s+", " ", clean).strip(" .,:;!?\"'()[]{}")
|
| 196 |
+
return clean
|
| 197 |
+
|
| 198 |
+
|
| 199 |
+
def _anchors_for_language(intake: DreamIntake, language: str = "en") -> List[str]:
|
| 200 |
anchors = _extract_dream_anchors(intake)
|
| 201 |
+
if _is_zh(language):
|
| 202 |
+
return anchors
|
| 203 |
+
localized = [_english_anchor_text(anchor) for anchor in anchors]
|
| 204 |
+
return _dedupe_preserve_order([anchor for anchor in localized if anchor])
|
| 205 |
+
|
| 206 |
+
|
| 207 |
+
def _clean_english_today_tip_language(card: TodayTipCard) -> TodayTipCard:
|
| 208 |
+
cleaned = card.model_copy(deep=True)
|
| 209 |
+
for field in (
|
| 210 |
+
"dream_summary",
|
| 211 |
+
"main_question",
|
| 212 |
+
"interpretation",
|
| 213 |
+
"today_tip",
|
| 214 |
+
"tiny_action",
|
| 215 |
+
"caring_note",
|
| 216 |
+
"safety_note",
|
| 217 |
+
):
|
| 218 |
+
setattr(cleaned, field, _english_anchor_text(getattr(cleaned, field)))
|
| 219 |
+
cleaned.dream_anchors = _dedupe_preserve_order(
|
| 220 |
+
[_english_anchor_text(anchor) for anchor in cleaned.dream_anchors if anchor]
|
| 221 |
+
)
|
| 222 |
+
cleaned.followup_questions = [_english_anchor_text(question) for question in cleaned.followup_questions]
|
| 223 |
+
cleaned.user_answers = [_english_anchor_text(answer) for answer in cleaned.user_answers]
|
| 224 |
+
return cleaned
|
| 225 |
+
|
| 226 |
+
|
| 227 |
+
def _primary_anchor(intake: DreamIntake, language: str = "en") -> str:
|
| 228 |
+
anchors = _anchors_for_language(intake, language)
|
| 229 |
if anchors:
|
| 230 |
return anchors[0]
|
| 231 |
return "梦里的那个细节" if _is_zh(language) else "that dream detail"
|
| 232 |
|
| 233 |
|
| 234 |
def _secondary_anchor(intake: DreamIntake, language: str = "en") -> str:
|
| 235 |
+
anchors = _anchors_for_language(intake, language)
|
| 236 |
return anchors[1] if len(anchors) > 1 else _primary_anchor(intake, language)
|
| 237 |
|
| 238 |
|
|
|
|
| 280 |
def _grounded_today_tip(intake: DreamIntake, language: str = "en") -> str:
|
| 281 |
primary = _primary_anchor(intake, language)
|
| 282 |
if not _is_zh(language):
|
| 283 |
+
return (
|
| 284 |
+
f"For today, borrow one action from the {primary}: open the task, write only the first line, "
|
| 285 |
+
"and let that be enough for now."
|
| 286 |
+
)
|
| 287 |
return f"今天先从「{primary}」借一个动作:只做最小的第一步,不急着把整件事完成。"
|
| 288 |
|
| 289 |
|
| 290 |
+
def _answer_based_tiny_action(answers: str, language: str = "en") -> str:
|
| 291 |
+
lowered = (answers or "").lower()
|
| 292 |
+
if _is_zh(language):
|
| 293 |
+
if "邮件" in lowered or "email" in lowered:
|
| 294 |
+
return "给自己 5 分钟,只打开那封邮件,写下第一句话;今天不要求立刻发出。"
|
| 295 |
+
return ""
|
| 296 |
+
if "email" in lowered or "message" in lowered:
|
| 297 |
+
return "Set a five-minute timer, open the email, and write only the first sentence. You do not have to send it yet."
|
| 298 |
+
return ""
|
| 299 |
+
|
| 300 |
+
|
| 301 |
def _anchor_in_text(text: str, anchors: List[str]) -> bool:
|
| 302 |
clean = (text or "").lower()
|
| 303 |
for anchor in anchors:
|
|
|
|
| 550 |
answers: Optional[List[str]] = None,
|
| 551 |
language: str = "en",
|
| 552 |
) -> DreamQAState:
|
| 553 |
+
anchors = _anchors_for_language(intake, language)
|
| 554 |
return DreamQAState(
|
| 555 |
dream_summary=_summary_from_intake(intake, language),
|
| 556 |
main_question=_main_question_from_intake(intake, language),
|
|
|
|
| 563 |
|
| 564 |
def _polish_today_tip(card: TodayTipCard, intake: DreamIntake, answers: str = "", language: str = "en") -> TodayTipCard:
|
| 565 |
polished = card.model_copy(deep=True)
|
| 566 |
+
intake_anchors = _anchors_for_language(intake, language)
|
| 567 |
+
card_anchors = (
|
| 568 |
+
polished.dream_anchors
|
| 569 |
+
if _is_zh(language)
|
| 570 |
+
else _dedupe_preserve_order([_english_anchor_text(anchor) for anchor in polished.dream_anchors])
|
| 571 |
+
)
|
| 572 |
if intake_anchors and not any(_text_uses_anchor(anchor, intake_anchors) for anchor in card_anchors):
|
| 573 |
anchors = intake_anchors
|
| 574 |
else:
|
|
|
|
| 589 |
or not _anchor_in_text(polished.today_tip, anchors)
|
| 590 |
):
|
| 591 |
polished.today_tip = _grounded_today_tip(intake, language)
|
| 592 |
+
hard_action_markers = ["address it immediately", "fix it immediately", "solve it immediately"]
|
| 593 |
+
answer_action = _answer_based_tiny_action(answers, language)
|
| 594 |
+
if answer_action and (
|
| 595 |
+
not polished.tiny_action.strip()
|
| 596 |
+
or any(marker in polished.tiny_action.lower() for marker in hard_action_markers)
|
| 597 |
+
or "email" in (answers or "").lower()
|
| 598 |
+
):
|
| 599 |
+
polished.tiny_action = answer_action
|
| 600 |
+
elif not polished.tiny_action.strip() or any(marker in polished.tiny_action.lower() for marker in hard_action_markers):
|
| 601 |
if _is_zh(language):
|
| 602 |
polished.tiny_action = f"用 5 分钟写下:今天和「{anchors[0]}」有关的第一小步是什么?"
|
| 603 |
else:
|
|
|
|
| 610 |
)
|
| 611 |
merged = "\n".join([intake.merged_text(), answers or ""])
|
| 612 |
polished.safety_note = safety_note() if needs_escalation(merged) else ""
|
| 613 |
+
if not _is_zh(language):
|
| 614 |
+
polished = _clean_english_today_tip_language(polished)
|
| 615 |
return polished
|
| 616 |
|
| 617 |
|
dream_customs/prompts.py
CHANGED
|
@@ -20,7 +20,8 @@ def visual_witness_prompt() -> str:
|
|
| 20 |
|
| 21 |
def _language_instruction(language: str = "en") -> str:
|
| 22 |
return (
|
| 23 |
-
"Write all user-facing fields in English.
|
|
|
|
| 24 |
if language != "zh"
|
| 25 |
else "所有面向用户的字段都用自然中文书写。可以保留用户原文里的梦境关键词。"
|
| 26 |
)
|
|
@@ -49,11 +50,13 @@ Return strict JSON with:
|
|
| 49 |
def today_tip_prompt(state: DreamQAState, language: str = "en") -> str:
|
| 50 |
return f"""
|
| 51 |
You are MiniCPM5-1B writing the final Dream QA result.
|
| 52 |
-
Write a non-diagnostic interpretation draft and exactly one primary 今日小 Tips.
|
| 53 |
Use non-certain language such as "也许", "可以把它当作", "maybe", or "for today, try".
|
| 54 |
The today_tip must cite at least one concrete dream anchor.
|
| 55 |
Avoid prophecy, frightening certainty, medical advice, therapy framing, and generic wellness filler.
|
| 56 |
-
Keep the
|
|
|
|
|
|
|
| 57 |
Use safety_note only for self-harm, harm to others, severe distress, severe insomnia, panic, or inability to function.
|
| 58 |
{_language_instruction(language)}
|
| 59 |
|
|
|
|
| 20 |
|
| 21 |
def _language_instruction(language: str = "en") -> str:
|
| 22 |
return (
|
| 23 |
+
"Write all user-facing fields in natural English. Translate or paraphrase non-English dream anchors "
|
| 24 |
+
"into natural English, except for a short exact quote that the user explicitly wrote and needs preserved."
|
| 25 |
if language != "zh"
|
| 26 |
else "所有面向用户的字段都用自然中文书写。可以保留用户原文里的梦境关键词。"
|
| 27 |
)
|
|
|
|
| 50 |
def today_tip_prompt(state: DreamQAState, language: str = "en") -> str:
|
| 51 |
return f"""
|
| 52 |
You are MiniCPM5-1B writing the final Dream QA result.
|
| 53 |
+
Write a non-diagnostic interpretation draft and exactly one primary Today Tip / 今日小 Tips.
|
| 54 |
Use non-certain language such as "也许", "可以把它当作", "maybe", or "for today, try".
|
| 55 |
The today_tip must cite at least one concrete dream anchor.
|
| 56 |
Avoid prophecy, frightening certainty, medical advice, therapy framing, and generic wellness filler.
|
| 57 |
+
Keep the whole result short, warm, and specific to the user's answer.
|
| 58 |
+
The tiny_action must be a 5-minute first step, not a command to solve the whole problem.
|
| 59 |
+
Avoid demanding phrases such as "immediately", "must", "fix it", or "solve it".
|
| 60 |
Use safety_note only for self-harm, harm to others, severe distress, severe insomnia, panic, or inability to function.
|
| 61 |
{_language_instruction(language)}
|
| 62 |
|
dream_customs/render.py
CHANGED
|
@@ -40,8 +40,8 @@ def render_today_tip_card(card: TodayTipCard, language: str = "en") -> str:
|
|
| 40 |
"history": "追问记录" if is_zh else "Question record",
|
| 41 |
"safety_default": "这不是诊断,只是一个温和的今日参考。"
|
| 42 |
if is_zh
|
| 43 |
-
else "
|
| 44 |
-
"safety_prefix": "这不是诊断。" if is_zh else "This is not diagnosis. ",
|
| 45 |
"highlight": "梦境细节" if is_zh else "dream detail",
|
| 46 |
}
|
| 47 |
tiny_action = (
|
|
|
|
| 40 |
"history": "追问记录" if is_zh else "Question record",
|
| 41 |
"safety_default": "这不是诊断,只是一个温和的今日参考。"
|
| 42 |
if is_zh
|
| 43 |
+
else "A gentle reflection for today, not a diagnosis or prophecy.",
|
| 44 |
+
"safety_prefix": "这不是诊断。" if is_zh else "This is not a diagnosis. ",
|
| 45 |
"highlight": "梦境细节" if is_zh else "dream detail",
|
| 46 |
}
|
| 47 |
tiny_action = (
|
dream_customs/ui/app.py
CHANGED
|
@@ -469,7 +469,7 @@ def build_demo() -> gr.Blocks:
|
|
| 469 |
mood_section_html = gr.HTML(_section_title_html(2, initial_copy["side_title"]))
|
| 470 |
mood = gr.Dropdown(label=initial_copy["mood_label"], choices=MOOD_OPTIONS, value=DEFAULT_MOOD)
|
| 471 |
side_stamp_html = gr.HTML(_side_stamp_html(DEFAULT_LANGUAGE))
|
| 472 |
-
with gr.Accordion("
|
| 473 |
dev_help_html = gr.HTML(_dev_help_html(DEFAULT_LANGUAGE))
|
| 474 |
text_backend = gr.Dropdown(
|
| 475 |
label="Text generation",
|
|
|
|
| 469 |
mood_section_html = gr.HTML(_section_title_html(2, initial_copy["side_title"]))
|
| 470 |
mood = gr.Dropdown(label=initial_copy["mood_label"], choices=MOOD_OPTIONS, value=DEFAULT_MOOD)
|
| 471 |
side_stamp_html = gr.HTML(_side_stamp_html(DEFAULT_LANGUAGE))
|
| 472 |
+
with gr.Accordion("Advanced", open=False, elem_classes=["dc-dev"]):
|
| 473 |
dev_help_html = gr.HTML(_dev_help_html(DEFAULT_LANGUAGE))
|
| 474 |
text_backend = gr.Dropdown(
|
| 475 |
label="Text generation",
|
dream_customs/ui/copy.py
CHANGED
|
@@ -27,8 +27,7 @@ APP_COPY = {
|
|
| 27 |
"example_button": "Try example",
|
| 28 |
"submit_button": "Continue",
|
| 29 |
"processing_note": (
|
| 30 |
-
"After submission, Dream QA asks one grounded question
|
| 31 |
-
"Model routes may take a few seconds; the text-only demo fallback always works."
|
| 32 |
),
|
| 33 |
"image_accordion": "Add image clue",
|
| 34 |
"image_label": "Upload a sketch, note, or screenshot",
|
|
@@ -52,7 +51,7 @@ APP_COPY = {
|
|
| 52 |
"side_stamp_title": "Add one concrete detail",
|
| 53 |
"side_stamp_body": "It helps the final suggestion stay grounded in your dream.",
|
| 54 |
"language_label": "Language",
|
| 55 |
-
"runtime_help": "
|
| 56 |
"card_title": "Today Tip",
|
| 57 |
"error_state": "The interface state could not be read. Please start over.",
|
| 58 |
},
|
|
@@ -77,8 +76,7 @@ APP_COPY = {
|
|
| 77 |
"example_button": "试试示例",
|
| 78 |
"submit_button": "继续解梦 →",
|
| 79 |
"processing_note": (
|
| 80 |
-
"提交后,梦境助手会先追问关键细节,再
|
| 81 |
-
"模型路线可能需要几秒;text-only demo fallback 始终可用。"
|
| 82 |
),
|
| 83 |
"image_accordion": "添加图片线索",
|
| 84 |
"image_label": "上传草图、便签或截图",
|
|
@@ -102,7 +100,7 @@ APP_COPY = {
|
|
| 102 |
"side_stamp_title": "尽量回忆更多细节",
|
| 103 |
"side_stamp_body": "有助于更准确地理解梦境。",
|
| 104 |
"language_label": "语言",
|
| 105 |
-
"runtime_help": "
|
| 106 |
"card_title": "今日小 Tips",
|
| 107 |
"error_state": "界面状态读取失败,请重新开始。",
|
| 108 |
},
|
|
|
|
| 27 |
"example_button": "Try example",
|
| 28 |
"submit_button": "Continue",
|
| 29 |
"processing_note": (
|
| 30 |
+
"After submission, Dream QA asks one grounded question, then turns the dream into a small Today Tip."
|
|
|
|
| 31 |
),
|
| 32 |
"image_accordion": "Add image clue",
|
| 33 |
"image_label": "Upload a sketch, note, or screenshot",
|
|
|
|
| 51 |
"side_stamp_title": "Add one concrete detail",
|
| 52 |
"side_stamp_body": "It helps the final suggestion stay grounded in your dream.",
|
| 53 |
"language_label": "Language",
|
| 54 |
+
"runtime_help": "Advanced controls for demos and development. Most visitors can leave these unchanged.",
|
| 55 |
"card_title": "Today Tip",
|
| 56 |
"error_state": "The interface state could not be read. Please start over.",
|
| 57 |
},
|
|
|
|
| 76 |
"example_button": "试试示例",
|
| 77 |
"submit_button": "继续解梦 →",
|
| 78 |
"processing_note": (
|
| 79 |
+
"提交后,梦境助手会先追问一个关键细节,再把梦境整理成一个今日小 Tips。"
|
|
|
|
| 80 |
),
|
| 81 |
"image_accordion": "添加图片线索",
|
| 82 |
"image_label": "上传草图、便签或截图",
|
|
|
|
| 100 |
"side_stamp_title": "尽量回忆更多细节",
|
| 101 |
"side_stamp_body": "有助于更准确地理解梦境。",
|
| 102 |
"language_label": "语言",
|
| 103 |
+
"runtime_help": "高级演示与开发控制。普通体验保持默认即可。",
|
| 104 |
"card_title": "今日小 Tips",
|
| 105 |
"error_state": "界面状态读取失败,请重新开始。",
|
| 106 |
},
|
scripts/evaluate_today_tip_quality.py
CHANGED
|
@@ -14,6 +14,8 @@ FIXTURE_PATH = Path("tests/fixtures/today_tip_eval_cases.json")
|
|
| 14 |
OLD_CUSTOMS_TERMS = ["permit", "contraband", "clearance", "sealed", "pact"]
|
| 15 |
FRIGHTENING_TERMS = ["you will fail", "prophecy says", "fate says", "must mean", "mental illness"]
|
| 16 |
CHINESE_UI_LABELS = ["今日小", "梦境摘要", "想理解的问题", "解读草稿", "没试过的小事", "关心一句"]
|
|
|
|
|
|
|
| 17 |
|
| 18 |
|
| 19 |
def _load_cases(path: Path = FIXTURE_PATH) -> List[Dict[str, Any]]:
|
|
@@ -70,6 +72,12 @@ def evaluate_case(case: Dict[str, Any]) -> List[str]:
|
|
| 70 |
chinese_labels = [label for label in CHINESE_UI_LABELS if label in combined]
|
| 71 |
if chinese_labels:
|
| 72 |
failures.append("chinese_ui_labels:" + ",".join(chinese_labels))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 73 |
|
| 74 |
if case.get("requires_safety_note") and "trusted person or professional support" not in lowered:
|
| 75 |
failures.append("missing_support_note")
|
|
|
|
| 14 |
OLD_CUSTOMS_TERMS = ["permit", "contraband", "clearance", "sealed", "pact"]
|
| 15 |
FRIGHTENING_TERMS = ["you will fail", "prophecy says", "fate says", "must mean", "mental illness"]
|
| 16 |
CHINESE_UI_LABELS = ["今日小", "梦境摘要", "想理解的问题", "解读草稿", "没试过的小事", "关心一句"]
|
| 17 |
+
CHINESE_LEAKAGE_TERMS = ["数字", "电梯", "按钮", "楼层", "融化", "梦境"]
|
| 18 |
+
HARD_COMMAND_PHRASES = ["address it immediately", "fix it immediately", "solve it immediately"]
|
| 19 |
|
| 20 |
|
| 21 |
def _load_cases(path: Path = FIXTURE_PATH) -> List[Dict[str, Any]]:
|
|
|
|
| 72 |
chinese_labels = [label for label in CHINESE_UI_LABELS if label in combined]
|
| 73 |
if chinese_labels:
|
| 74 |
failures.append("chinese_ui_labels:" + ",".join(chinese_labels))
|
| 75 |
+
chinese_terms = [term for term in CHINESE_LEAKAGE_TERMS if term in combined]
|
| 76 |
+
if chinese_terms:
|
| 77 |
+
failures.append("chinese_anchor_leakage:" + ",".join(chinese_terms))
|
| 78 |
+
hard_commands = [phrase for phrase in HARD_COMMAND_PHRASES if phrase in lowered]
|
| 79 |
+
if hard_commands:
|
| 80 |
+
failures.append("hard_command:" + ",".join(hard_commands))
|
| 81 |
|
| 82 |
if case.get("requires_safety_note") and "trusted person or professional support" not in lowered:
|
| 83 |
failures.append("missing_support_note")
|
tests/fixtures/today_tip_eval_cases.json
CHANGED
|
@@ -6,6 +6,13 @@
|
|
| 6 |
"answer": "It felt like being late before I had even started.",
|
| 7 |
"required_anchors": ["elevator", "button", "14"]
|
| 8 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
{
|
| 10 |
"id": "exam_without_pencil",
|
| 11 |
"dream_text": "I walked into an exam room and realized every pencil had turned into a glass straw.",
|
|
|
|
| 6 |
"answer": "It felt like being late before I had even started.",
|
| 7 |
"required_anchors": ["elevator", "button", "14"]
|
| 8 |
},
|
| 9 |
+
{
|
| 10 |
+
"id": "elevator_email_first_sentence",
|
| 11 |
+
"dream_text": "I dreamed I was in an elevator where the floor buttons melted like wax. The number 14 kept blinking, and I felt late but strangely calm.",
|
| 12 |
+
"mood": "Uneasy",
|
| 13 |
+
"answer": "I want to make starting my overdue email easier without feeling trapped by it.",
|
| 14 |
+
"required_anchors": ["elevator", "button", "14"]
|
| 15 |
+
},
|
| 16 |
{
|
| 17 |
"id": "exam_without_pencil",
|
| 18 |
"dream_text": "I walked into an exam room and realized every pencil had turned into a glass straw.",
|
tests/test_today_tip_quality_eval.py
CHANGED
|
@@ -4,7 +4,7 @@ from scripts.evaluate_today_tip_quality import _load_cases, evaluate_case, evalu
|
|
| 4 |
def test_today_tip_quality_eval_passes_fixture_cases():
|
| 5 |
result = evaluate_cases(_load_cases())
|
| 6 |
|
| 7 |
-
assert result["case_count"] ==
|
| 8 |
assert result["failures"] == {}
|
| 9 |
assert result["passes"] is True
|
| 10 |
|
|
|
|
| 4 |
def test_today_tip_quality_eval_passes_fixture_cases():
|
| 5 |
result = evaluate_cases(_load_cases())
|
| 6 |
|
| 7 |
+
assert result["case_count"] == 11
|
| 8 |
assert result["failures"] == {}
|
| 9 |
assert result["passes"] is True
|
| 10 |
|
tests/test_ui_actions.py
CHANGED
|
@@ -19,7 +19,7 @@ def test_mobile_defaults_to_model_backends():
|
|
| 19 |
def test_runtime_settings_are_collapsed_for_public_flow():
|
| 20 |
source = inspect.getsource(ui_app.build_demo)
|
| 21 |
|
| 22 |
-
assert 'gr.Accordion("
|
| 23 |
|
| 24 |
|
| 25 |
def test_processing_note_is_story_copy_not_backend_jargon():
|
|
@@ -27,8 +27,8 @@ def test_processing_note_is_story_copy_not_backend_jargon():
|
|
| 27 |
|
| 28 |
assert "grounded question" in lowered
|
| 29 |
assert "today tip" in lowered
|
| 30 |
-
assert "model routes" in lowered
|
| 31 |
-
assert "fallback" in lowered
|
| 32 |
assert "token" not in lowered
|
| 33 |
assert "endpoint" not in lowered
|
| 34 |
assert "debug" not in lowered
|
|
@@ -131,3 +131,29 @@ def test_mobile_mvp_answer_to_card_generates_today_tip():
|
|
| 131 |
assert view["status"] == "tip"
|
| 132 |
assert view["phase"] == "tip"
|
| 133 |
assert "It may be asking me to slow down." in view["debug"]["session"]["answer_history"]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
def test_runtime_settings_are_collapsed_for_public_flow():
|
| 20 |
source = inspect.getsource(ui_app.build_demo)
|
| 21 |
|
| 22 |
+
assert 'gr.Accordion("Advanced", open=False' in source
|
| 23 |
|
| 24 |
|
| 25 |
def test_processing_note_is_story_copy_not_backend_jargon():
|
|
|
|
| 27 |
|
| 28 |
assert "grounded question" in lowered
|
| 29 |
assert "today tip" in lowered
|
| 30 |
+
assert "model routes" not in lowered
|
| 31 |
+
assert "fallback" not in lowered
|
| 32 |
assert "token" not in lowered
|
| 33 |
assert "endpoint" not in lowered
|
| 34 |
assert "debug" not in lowered
|
|
|
|
| 131 |
assert view["status"] == "tip"
|
| 132 |
assert view["phase"] == "tip"
|
| 133 |
assert "It may be asking me to slow down." in view["debug"]["session"]["answer_history"]
|
| 134 |
+
|
| 135 |
+
|
| 136 |
+
def test_english_today_tip_has_no_chinese_anchor_leakage():
|
| 137 |
+
state, _view_json = submit_dream_action(
|
| 138 |
+
dream_text=(
|
| 139 |
+
"I dreamed I was in an elevator where the floor buttons melted like wax. "
|
| 140 |
+
"The number 14 kept blinking, and I felt late but strangely calm."
|
| 141 |
+
),
|
| 142 |
+
mood="Uneasy",
|
| 143 |
+
text_backend="demo",
|
| 144 |
+
vision_backend="demo",
|
| 145 |
+
language="en",
|
| 146 |
+
)
|
| 147 |
+
|
| 148 |
+
_state, view_json = answer_to_card_action(
|
| 149 |
+
state,
|
| 150 |
+
"I want to make starting my overdue email easier without feeling trapped by it.",
|
| 151 |
+
text_backend="demo",
|
| 152 |
+
vision_backend="demo",
|
| 153 |
+
language="en",
|
| 154 |
+
)
|
| 155 |
+
view = json.loads(view_json)
|
| 156 |
+
combined = "\n".join([view["card_text"], view["card_html"]])
|
| 157 |
+
|
| 158 |
+
for leaked in ["数字", "电梯", "按钮", "楼层", "融化"]:
|
| 159 |
+
assert leaked not in combined
|