dym-arex-skill-creator-paper-to-skills

Skill: dym-arex-skill-creator-paper-to-skills — Creator: paper → reproducible skill (Distiller)

Why use it: There is a paper to reproduce, and you want the result to be a verified modularized skill rather than a summary. What it produces: <workspace_root>/<paper_slug>/skills/ (the generated skill, still staging until import is approved) and <workspace_root>/<paper_slug>/distillation/reports/final/ (final report).

1. Prepare the config file

There are two TOML files in the repo, with the same schema:

File	Role
`examples/creator/paper-to-skills/distiller-run-config.toml`	Starter is heavily commented — use this one
`cli/packages/coding-agent/src/disco/skills/create-paper-skills/assets/distiller-run-config-template.toml`	Canonical template kept by the skill

From the root of the AREX-Skill checkout:

cp examples/creator/paper-to-skills/distiller-run-config.toml \
  /absolute/path/to/distiller-run-config.toml

2. Fill in at least 3 fields

schema_version = 1

[defaults]
workspace_root = "/absolute/path/to/paper2skills-workspace"   # ① BẮT BUỘC sửa

[[runs]]
paper_slug   = "my_paper_slug"          # ② BẮT BUỘC sửa
paper_source = "/path/to/paper.pdf"     # ③ BẮT BUỘC sửa
original_repo_source = "unknown"

paper_source accepts: local PDF/text path · direct PDF URL · arXiv URL or ID · paper title. original_repo_source accepts: local repo path · Git/GitHub URL · "none" · "unknown".

Adding a second paper = adding another [[runs]] block.

3. Knobs to understand before running (not every default is harmless)

Field	Default	Meaning
`recovery_mode`	`"hard"`	`hard`: reduced/proxy/toy/fallback runs are only recorded as diagnostics, not counted as successful recovery. To accept them you must choose `soft` explicitly
`ask_before_expensive_recovery`	`true`	Stop and ask before burning resources
`repo_discovery_mode`	`"ask"`	`ask` stop and ask before going to find the install repo · `auto` automatically find GitHub and clone the first candidate · `disabled` never search automatically
`allow_title_top_hit`	`false`	Do not accept the first result on your own when searching by paper title
`iteration_budget`	`10`	Maximum number of refine rounds
`network_timeout_seconds`	`120`
`command_timeout_seconds`	`20`
`runtime_constraints`	(string)	The starter default forbids touching the shared conda env, requiring use of an isolated env
`test_root` / `skills_root` / `generated_skills_root` / `distiller_skills_root` / `attempt_dir`	`""`	Leave empty to use the default layout under `workspace_root`

Every field in [defaults] can be overridden per [[runs]]; leaving an empty string "" means inherit. Numeric/bool overrides must be uncommented to take effect.

4. Run

disco --creator -p "/skill:create-paper-skills Use Distiller to generate and verify paper-replication skills for each run in this config. config_path: /absolute/path/to/distiller-run-config.toml"

Short form (Creator chooses the entry itself):

disco --creator -p "Use Distiller to generate and verify paper-replication skills for each run in this config. config_path: /absolute/path/to/distiller_run_config.toml"

5. What the Distiller does (don't call each step by hand)

create-paper-skills delegates everything to paper-skills-distiller, and this chain automatically loads the supporting skills:

Resolve the paper source (plan-paper-skill-modules builds the profile + module plan + module docs);
Generate & validate module-level skills (create-paper-module-skill);
Prepare bounded runtime evidence (prepare-paper-recovery-env: package, model, data, runtime);
Run the most feasible recovery experiment strongest feasible, without reading the original installation repo (recover-paper-result);
Analyze the gap versus the paper's target → accept / refine / blocker (analyze-paper-recovery);
Refine in iteration_budget if needed;
Log artifacts for each attempt + final report.

6. Where the output goes

<workspace_root>/<paper_slug>/skills/                       # skill sinh ra — STAGING
<workspace_root>/<paper_slug>/distillation/reports/final/   # báo cáo cuối

The tree skills/ remains staging until approved and imported. After final validation, the Creator proposes a deployment scope for the entire graph, import only after you approve, then write the handoff for the Researcher.

7. Pitfalls

recovery_mode = "hard" is intentionally the default. If you see the report say "diagnostic" rather than "recovered", that is correct behavior, not a bug.
Do not manually invoke each paper skill. A normal run only needs one command create-paper-skills.
workspace_root must be an absolute path and writable — all evidence, skills, and artifacts go there.
The config path passed to DisCo must be absolute.

Next step

07-tich-hop-ci.md if you want to pin the library version in the CI of your project.