Files
make-cad/examples/demo_kit/TRANSCRIPT_FORMAT.md
T
NiJingzhe f65109c6e7 docs(showcase): four-capability README front section with session replay demos
- README.md / README.zh-CN.md: capability pillars (single part / assembly /
  reverse engineering / simulation) with case images and replay demo links
- examples/demo_kit: config-driven single-file replay demo builder
  (generalized from u_link demo; app.js consumes per-case config payload)
- examples/flange_plate: full single-part-modeling workflow case (agent
  session transcript + verifier scripts + 8-hole re-parameterization with
  guard rejection chain) + replay demo
- examples/12_ap242_gmsh_volume_mesh: model.py rebuilt through the workflow
  (FTC feature blocks, named vars; volume rel 0.00e+00 vs legacy, FEM
  interface tags preserved) + REQUIREMENTS/BUILD_PLAN/verify + replay demo
- examples/u_link_motor_mount/demo: committed single-file replay demo,
  rebuilt via demo_kit
- examples/20_integrated_bldc_joint_actuator/render_showcase.py: assembled
  + staged cylindrical exploded views from the captured package
- img/capability: showcase renders (BLDC assembly/exploded, bracket FEM
  von Mises, case ISO views)
2026-09-05 21:02:43 +08:00

2.3 KiB

Session Transcript Format (demo_kit)

build_demo.py parses a case's session_transcript.md into the replay demo. The grammar below is exactly what the parser accepts; anything else is ignored.

File header

# OpenCode Session Transcript

- Session ID: `ses_<random>`
- Title: <case title, Chinese ok>
- Directory: /Users/lildino/Project/ocws/SCAD-Open
- Started: YYYY-MM-DD HH:MM:SS
- Updated: YYYY-MM-DD HH:MM:SS
- Messages: <count of USER+ASSISTANT sections, roleswitch excluded>

---

Sections

Three section kinds, separated by --- lines. Every section starts with a timestamp line *msg_<random22> @ YYYY-MM-DD HH:MM:SS* as its first body line.

USER turn (numbered, real user wording)

## [1] USER
*msg_abc... @ 2026-09-05 18:30:12*

<verbatim user text>

---

ASSISTANT turn

## [ASSISTANT] (glm-5.3)
*msg_def... @ 2026-09-05 18:30:40*

<details><summary>thinking</summary>

> Real reasoning summary, one `> `-prefixed line per line.

</details>

Prose reply.

**[tool: bash]** status=ok
```json
{"command": "uv run python examples/flange_plate/model.py"}
<real output, may elide with "... (N lines)" — numbers and verdicts stay real>

[patch] hash=1a2b3c4d5e

  • file: examples/flange_plate/model.py

[text]

More prose.



Tool names are lowercase: `bash read write edit todowrite glob grep question
task skill`. One ```json fence = tool input, one ```output fence = tool output
(only the first of each is captured). A `**[patch]**` block follows each
write/edit; hash = first 10 hex of md5 of the resulting file content.

### Role switch

```markdown
## [Role Switch: <Role Name>]
*msg_ghi... @ 2026-09-05 18:35:00*

Reading: <role-owned files>
Artifact out: <path>

---

— note may follow the role name inside the bracket. Optional # Skill: blocks (skill loads) render as a special chip.

Fidelity rules

  • Tool calls/outputs must be the real actions that produced the artifacts — outputs may be trimmed, never edited or invented.
  • Thinking blocks are the real reasoning of the moment, summarized honestly.
  • USER turns are verbatim; role-switch blocks follow the workflow's protocol.
  • Timestamps are real wall-clock times; Messages: equals the final section count (excluding role switches).