[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M753D993EKYRDF901H136MSZ":3,"skill-same:01M753D993EKYRDF901H136MSZ":88},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":25,"commitSha":26,"skillDirectory":27,"forksCount":28,"repositoryHtmlUrl":29,"repositoryHomepageUrl":6,"ownerHtmlUrl":30,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":31,"install":32,"resources":45},"01M753D993EKYRDF901H136MSZ","paper-skills-distiller",null,"Orchestrate a paper-to-skills distillation loop for AI research papers supplied as local files, URLs, arXiv ids, or titles, optionally with a local or remote repository, by resolving sources, modularizing the paper, creating module skills, preparing recovery runtime, recovering a paper result without the source repo, analyzing gaps, and iterating until acceptable.","VectorSpaceLab\u002FAREX-Skill","VectorSpaceLab","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F180637102?v=4","01M7538C5WPSJGWF7KHR4EFZ9F","01M7538C6R2GFP53NPXTCZNBRT",351,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fpaper-skills-distiller\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: paper-skills-distiller\ndescription: \"Orchestrate a paper-to-skills distillation loop for AI research papers supplied as local files, URLs, arXiv ids, or titles, optionally with a local or remote repository, by resolving sources, modularizing the paper, creating module skills, preparing recovery runtime, recovering a paper result without the source repo, analyzing gaps, and iterating until acceptable.\"\nmetadata:\n  disco-role: meta\n---\n\n# Paper Skills Distiller\n\nUse this skill as the top-level controller when a user provides a research paper source, with or without an implementation repository source, and wants the paper converted into reusable Agent Skills. The paper source may already be local, or it may be a direct PDF URL, arXiv URL\u002Fid, or paper title that must be resolved before the run starts.\n\nAll generated skill content must be written in English. Keep user-facing status in the user's language if appropriate, but artifact names, schemas, `SKILL.md` files, scripts, and validation reports should be English.\n\nThe default refinement budget is 10 refine cycles after the first recovery. Use the user's supplied `iteration_budget` when present; otherwise record `iteration_budget: 10` in `run_manifest.json`. The user may set any non-negative integer. Stop before starting a refine cycle that would exceed this budget, and summarize the remaining feedback instead of continuing silently.\n\nThe default recovery mode is `hard`. In `hard` mode, the run must not accept reduced, proxy, toy, smaller-model, or fallback recovery as success; if the full requested standard cannot run, record a concrete blocker or return `refine`. In `soft` mode, a reduced or proxy recovery may be accepted only when it is explicitly declared, justified, executable, and passes the recovery experiment gate.\n\nPrefer a TOML run config when the user provides one. Use\n`..\u002Fcreate-paper-skills\u002Fassets\u002Fdistiller-run-config-template.toml` as the\nfill-in template and validate it with `scripts\u002Fread_run_config.py` before\nresolving sources or initializing runs. A config may contain one or more\n`[[runs]]` entries for batch distillation.\n\nDefault outputs use one run directory with two main children:\n`\u003Cworkspace_root>\u002F\u003Cpaper_slug>\u002Fdistillation\u002F` for process artifacts and\n`\u003Cworkspace_root>\u002F\u003Cpaper_slug>\u002Fskills\u002F` for generated reusable skills. Do not\ncreate numbered attempt directories unless the user explicitly supplies a\ncustom `attempt_dir` for that purpose.\n\n## Distillation Model\n\nUse the paper or tutorial source as anchor `z` for task-agnostic distillation\nunless the user supplies a concrete downstream replication task. For a concrete\ntask, record the task anchor as `tau = (q, D, E, g)` and use task-oriented\ndistillation; do not create a synthetic task tuple for a paper-only run.\n\nThe run follows four stages:\n\n1. **Scope** the required capabilities `Q`, module boundaries, non-goals, and\n   verification targets.\n2. **Ground** `Q` in paper and permitted source evidence `X`, recording\n   provenance, revisions, exclusions, conflicts, and assumptions.\n3. **Construct** candidate graph `G_tilde` from module skills, references,\n   scripts, and links.\n4. **Verify** candidate `G_tilde` into accepted graph `G` and construction record\n   `R` through executable recovery, validation, refinement, deployment review,\n   and Researcher handoff. Task-agnostic runs use representative operating\n   checks when no downstream task-level outcome is defined.\n\n## Source Rules\n\n- Paper: mandatory. Read the full paper or a faithful extracted text version. If the user gives a URL, arXiv id, or title instead of a local path, first resolve it to a local PDF\u002Ftext artifact and save source acquisition evidence.\n- Repository: optional. It may be a local path, Git URL, `none`, or `unknown`. Use it only during modularization and module skill creation. Treat it as implementation evidence, not as a substitute for paper understanding. If it is `unknown`, follow `repo_discovery_mode`: `ask` records candidates\u002Fblockers and asks before searching or selecting, `auto` may run bounded GitHub discovery and clone the top candidate, and `disabled` proceeds paper-only.\n- Recovery: must not read the original source repository. During recovery, use only the paper, module documents, generated module skills, data, and general package documentation.\n- Internet: use only when needed for source acquisition, repo discovery, datasets, package docs, or web-search experiments. If the environment provides a network setup script, source it before network commands. Keep downloads, title searches, repo discovery, and clones bounded and logged.\n- Logs: every run must be recorded under its configured distillation directory.\n\n## Required Loop\n\n1. If a run config path is supplied, validate and normalize it with `scripts\u002Fread_run_config.py`; otherwise collect the same fields from the user's prompt. For batch configs, process each normalized run independently and keep separate source-resolution, attempt, and summary artifacts. Use `read_run_config.py --slug \u003Cpaper_slug> --run-only` to write the selected normalized run JSON passed to `init_run.py --run-config`.\n2. Resolve paper and optional repo sources when needed with `scripts\u002Fresolve_sources.py` or an equivalent bounded, logged procedure. The resolver covers local paths, direct URLs, arXiv URLs\u002Fids, arXiv title search, Git repo URLs, and `repo_discovery_mode=auto` GitHub repo discovery. If a title is not found there, use web search for official or scholarly sources and record the chosen URL. If title or repo discovery is ambiguous, ask the user to choose before proceeding.\n3. Initialize the distillation directory with `scripts\u002Finit_run.py` using the resolved local paper path, optional resolved repo path, configured `iteration_budget`, configured `recovery_mode`, configured `generated_skills_root`, optional exact `attempt_dir`, and `--run-config` when a normalized config JSON exists.\n4. Load `plan-paper-skill-modules` and create:\n   - `paper_profile.md`\n   - `module_plan.json`\n   - one module document per module in `modules\u002F`\n   - no more than five modules\n5. Validate the module plan with `plan-paper-skill-modules\u002Fscripts\u002Fvalidate_module_plan.py`.\n6. For each module, load `create-paper-module-skill` and create the module skill under the configured output skill root.\n7. Validate every generated module skill with `create-paper-module-skill\u002Fscripts\u002Fvalidate_skill_tree.py` and run its tests or smoke checks.\n8. Load `prepare-paper-recovery-env` and create `environment\u002Fruntime_handoff.json` from bounded package, model, GPU, benchmark, and dataset preparation. If required packages, datasets, or model caches are missing, run the active preparation path first: create\u002Freuse an isolated recovery environment under `$DISCO_CODING_AGENT_DIR\u002Fenvs\u002F` or `~\u002F.disco\u002Fagent\u002Fenvs\u002F`, install the smallest missing package set, and attempt permitted bounded model\u002Fdata acquisition before treating the item as blocked.\n9. Load `recover-paper-result` and reproduce one fast, meaningful paper result without reading the original repo, using the runtime handoff as recovery input.\n10. Load `analyze-paper-recovery` and compare the recovery result with the paper target or an explicitly declared proxy target.\n11. If analysis says `refine`, update the relevant workflow guidance, module documents, module skills, environment handoff usage, and recovery harness, then rerun steps 7-10. Stop when status is `accept` or the configured refinement budget is exhausted.\n12. After an accepted run passes final and reported validation, classify the complete generated module graph for deployment. Default task-, project-, private-data-, evaluator-, benchmark-, or environment-bound output and all uncertain cases to project scope at `\u003Cproject-dir>\u002F.agents\u002Fskills\u002F`. Use managed scope at `~\u002F.disco\u002Fagent\u002Fskills\u002F` only with evidence that the graph is self-contained, provenance-backed, independent of transient run state, representative-use verified, and reusable across projects or research tasks. Keep every module in the same scope.\n13. Present the classification evidence, exact targets, entry point, verification, unresolved limits, collisions, and shadowing impact. After approval, use the visible `distill-ml-knowledge\u002Fscripts\u002Fimport_operating_skill_graph.mjs` once with every top-level module root and the selected scope. Overwrite requires separate approval. Write `researcher-handoff.md` with the selected scope and exact imported paths, then suggest `\u002Fresearcher`; if import is declined, record that the graph remains staged.\n\n## Ordering and Runtime Discipline\n\n- Finish and validate the modularization artifacts before heavyweight runtime probing. Do not delay `paper_profile.md`, `module_plan.json`, or `modules\u002F*.md` while searching for models, package caches, datasets, or GPUs.\n- Source acquisition happens before distillation initialization when needed, but it should stay bounded: direct local path checks first, then direct URL\u002FarXiv download, then title search only if requested by the user's input, then repo discovery only when `repo_discovery_mode` or user confirmation permits it. Do not treat source acquisition as recovery evidence for reading the original repo.\n- Keep filesystem discovery bounded. Do not run broad scans such as `find \u002Fshare\u002Fproject ...` or recursive searches across NFS roots. Search only the workspace, user-supplied paths, known cache roots, or a short explicit allowlist, and use timeouts for cache probes.\n- Long package installs, model downloads, and benchmark setup commands must have a time budget. If they stall, stop them, record the exact blocker in recovery artifacts, and continue with the best valid lower-cost recovery allowed by the user.\n- Missing local dependencies, model caches, datasets, or search credentials are\n  not by themselves sufficient to end the run as blocked. First attempt the\n  allowed setup path: isolated environment creation, targeted package\n  installation, bounded dataset\u002Fmodel acquisition, configured VPN\u002Fproxy or\n  mirror use, and keyless\u002Fpublic alternatives where valid. Only mark blocked\n  after those actions are logged or after asking the user for required\n  credentials\u002Fpermissions that cannot be inferred.\n- A fallback or proxy can be useful evidence, but it must not be counted as a successful required runtime unless the recorded mechanism checks prove that the required data item, model\u002Fruntime, loss, and optimizer or training step actually ran.\n- In `hard` recovery mode, do not count fallback, proxy, reduced training, subset-only evaluation, or smaller-model execution as accepted recovery. Such artifacts may be logged as diagnostics, but analysis must return `refine` or `blocked` unless the full requested standard ran.\n- In `soft` recovery mode, reduced or proxy recovery is allowed only as a declared target with executable evidence, generated-skill invocation logs, and `recovery\u002Fexperiment_validation.json` passing.\n- Independent runs must not silently depend on previous runs. Reuse immutable local datasets or benchmark checkouts only as recorded benchmark sources when fresh acquisition is blocked or the user supplied them. Do not reuse previous generated skills, data items, prompts, outputs, training logs, or recovery results as evidence for a new run.\n- Runtime setup must produce a handoff before recovery. `recover-paper-result` should not repeat broad environment discovery or silently invent fallback state; it should consume `environment\u002Fruntime_handoff.json` and add that handoff to `recovery\u002Fsource_manifest.json`.\n\n## Distillation Directory Contract\n\nEach distillation directory should contain:\n\n```text\nrun_manifest.json\nrun_config.normalized.json   # optional but preferred when a config was used\nsource_resolution.json       # optional, when paper\u002Frepo source acquisition was needed\npaper_profile.md\nmodule_plan.json\nmodules\u002F\n  \u003Cmodule_id>.md\ngenerated_skills_validation\u002F\n  \u003Cmodule_id>.json\nenvironment\u002F\n  runtime_handoff.json\n  logs\u002F\n    command_log.json\nrecovery\u002F\n  experiment_plan.md\n  experiment_validation.json\n  source_manifest.json\n  recovery_result.json\n  logs\u002F\n    experiment_command_log.json\n    generated_skill_invocations.json\nanalysis\u002F\n  analysis_report.json\n  feedback.md\nfinal_validation.json\nreports\u002F\n  generated-skills\u002F\n  verification\u002F\n  final\u002F\n    final_report.json\n    final_report.md\n```\n\nUse additional files when useful, but do not omit the core contract. Record failures as artifacts instead of relying on terminal output.\n\n## Quality Bar\n\nA generated module skill is acceptable only if it has:\n\n- A clear trigger description in frontmatter.\n- A concise workflow that states when to use the skill and when not to use it.\n- Explicit input\u002Foutput contracts.\n- At least one deterministic test, smoke script, or validation command.\n- A saved validation JSON whose `ok` is true and whose `tests.attempted` is\n  true. Prose-only inspection is not enough for either hard or soft mode.\n- No dependence on the original paper repository unless the skill's purpose is explicitly to operate on that repository.\n- Enough implementation detail for a future agent run to use the skill without rereading the entire paper.\n\nThe whole paper distillation is acceptable only if:\n\n- The module split covers the paper's main method and evaluation pathway.\n- `environment\u002Fruntime_handoff.json` records package availability, GPU status, model readiness or blockers, benchmark acquisition\u002Freuse, environment mutation status, and allowed recovery sources.\n- `environment\u002Flogs\u002Fcommand_log.json` records subprocess commands run during environment preparation, including timed-out or failed commands.\n- The recovery run exercises the generated skills, not a hand-written one-off solution.\n- The recovery result is produced by an executable experiment command whose command log is saved under `recovery\u002Flogs\u002Fexperiment_command_log.json`.\n- `recovery\u002Fexperiment_plan.md` explains full\u002Freduced\u002Ftoy options and why the chosen target is the strongest feasible experiment.\n- `recovery\u002Flogs\u002Fgenerated_skill_invocations.json` proves generated module skills were called, imported, or cross-checked.\n- The recovery experiment gate is run and saved to `recovery\u002Fexperiment_validation.json` with `ok: true`.\n- The recovery run exercises every generated core module that the proxy claims to validate, or explicitly records why a module is not applicable. If a generated scoring\u002Floss module exists separately from the training harness, the recovery cross-checks it rather than relying only on duplicated harness logic.\n- Proxy recovery records explicit `mechanism_checks` that demonstrate the core paper mechanism ran. Do not accept a proxy solely because EM\u002FF1 or another task metric is high.\n- If recovery uses a benchmark or dataset repository, the item log and source manifest identify concrete resource files used to derive the sample whenever feasible. A benchmark-style item without resource-file provenance must be labeled as such and justified.\n- If recovery reuses a local benchmark checkout from another run or workspace location, the source manifest records the fresh-fetch blocker, reused path, commit\u002Fversion, concrete files used, and a snapshot or copy of those files inside the current distillation directory.\n- At least one recovery run uses the strongest practical real runtime available in the workspace when the user provides one. If the user gives a model path, GPU ids, or an inference environment, record whether it was used; if it could not be used, record the exact blocker and run the best lower-cost fallback.\n- The analysis report explains metric gaps and either accepts the result or gives actionable module-level feedback.\n- The final artifact contract validator is run and its JSON output is saved to `final_validation.json`.\n- The organized final report is written to `reports\u002Ffinal\u002Ffinal_report.md` and\n  `reports\u002Ffinal\u002Ffinal_report.json`, and the `reported` validation stage passes.\n- The complete module graph has a recorded reusability classification. Any live\n  deployment uses one approved scope, excludes process\u002Freview artifacts, and is\n  recorded with exact paths in `researcher-handoff.md`.\n- `run_manifest.json` and `run_config.normalized.json` record `recovery_mode`, and hard-mode runs are not accepted if `recovery_result.json` marks `is_proxy: true` or `mechanism_checks.reduced_training_executed: true`.\n- When paper\u002Frepo source acquisition was needed, the resolved local paths, commands\u002Fnetwork operations, selected title\u002FarXiv id, and blockers are recorded in `source\u002Fsource_resolution.json` or `reports\u002Ffinal\u002Ffinal_report.md`.\n\n## Refinement Rules From Real Runs\n\n- Do not accept a semantically correct but metric-failing result without improving the relevant evaluation or answer-formatting skill.\n- If an intermediate module output contaminates later stages, refine that module's output contract and add a regression test before rerunning recovery.\n- If a runtime environment is incompatible, do not mutate it unless the user explicitly allows it. Prefer an existing compatible environment or create a separate environment under `$DISCO_CODING_AGENT_DIR\u002Fenvs\u002F` or `~\u002F.disco\u002Fagent\u002Fenvs\u002F`, then record the choice in `source_manifest.json`.\n- After any refinement, rerun module skill validation and recovery in the current distillation directory or a clearly versioned cycle subdirectory.\n\n## Practical Guidance\n\nFavor small, composable module skills over one monolithic \"paper skill\". A good split usually separates protocol\u002Fdata formatting, retrieval\u002Ftool adapters, core algorithm loop, scoring\u002Fevaluation, and experiment recovery. Keep the number of modules below five by merging thin utility modules into the nearest substantive module.\n\nIf the full paper result requires heavy training or unavailable models, check `recovery_mode` before lowering the target. In `hard` mode, record the blocker and do not accept a proxy as success. In `soft` mode, recover a declared proxy result that preserves the paper's mechanism. The proxy must identify the dataset, sample size, metric, paper target, and why the proxy is the fastest faithful test.\n\n## Scripts\n\n- `scripts\u002Fread_run_config.py`: validate and normalize a TOML config with one or more Distiller runs.\n- `scripts\u002Fresolve_sources.py`: resolve local or remote paper\u002Frepo sources into local paths before distillation initialization, including optional bounded repo discovery.\n- `scripts\u002Finit_run.py`: create the distillation directory and manifest, optionally copying a selected normalized run config into the run.\n- `scripts\u002Fvalidate_distillation_run.py`: validate that required artifacts exist at each stage.\n- `scripts\u002Fwrite_final_report.py`: write `reports\u002Ffinal\u002Ffinal_report.json` and\n  `reports\u002Ffinal\u002Ffinal_report.md` from the distillation artifacts.\n","# Paper Skills Distiller\n\nUse this skill as the top-level controller when a user provides a research paper source, with or without an implementation repository source, and wants the paper converted into reusable Agent Skills. The paper source may already be local, or it may be a direct PDF URL, arXiv URL\u002Fid, or paper title that must be resolved before the run starts.\n\nAll generated skill content must be written in English. Keep user-facing status in the user's language if appropriate, but artifact names, schemas, `SKILL.md` files, scripts, and validation reports should be English.\n\nThe default refinement budget is 10 refine cycles after the first recovery. Use the user's supplied `iteration_budget` when present; otherwise record `iteration_budget: 10` in `run_manifest.json`. The user may set any non-negative integer. Stop before starting a refine cycle that would exceed this budget, and summarize the remaining feedback instead of continuing silently.\n\nThe default recovery mode is `hard`. In `hard` mode, the run must not accept reduced, proxy, toy, smaller-model, or fallback recovery as success; if the full requested standard cannot run, record a concrete blocker or return `refine`. In `soft` mode, a reduced or proxy recovery may be accepted only when it is explicitly declared, justified, executable, and passes the recovery experiment gate.\n\nPrefer a TOML run config when the user provides one. Use\n`..\u002Fcreate-paper-skills\u002Fassets\u002Fdistiller-run-config-template.toml` as the\nfill-in template and validate it with `scripts\u002Fread_run_config.py` before\nresolving sources or initializing runs. A config may contain one or more\n`[[runs]]` entries for batch distillation.\n\nDefault outputs use one run directory with two main children:\n`\u003Cworkspace_root>\u002F\u003Cpaper_slug>\u002Fdistillation\u002F` for process artifacts and\n`\u003Cworkspace_root>\u002F\u003Cpaper_slug>\u002Fskills\u002F` for generated reusable skills. Do not\ncreate numbered attempt directories unless the user explicitly supplies a\ncustom `attempt_dir` for that purpose.\n\n## Distillation Model\n\nUse the paper or tutorial source as anchor `z` for task-agnostic distillation\nunless the user supplies a concrete downstream replication task. For a concrete\ntask, record the task anchor as `tau = (q, D, E, g)` and use task-oriented\ndistillation; do not create a synthetic task tuple for a paper-only run.\n\nThe run follows four stages:\n\n1. **Scope** the required capabilities `Q`, module boundaries, non-goals, and\n   verification targets.\n2. **Ground** `Q` in paper and permitted source evidence `X`, recording\n   provenance, revisions, exclusions, conflicts, and assumptions.\n3. **Construct** candidate graph `G_tilde` from module skills, references,\n   scripts, and links.\n4. **Verify** candidate `G_tilde` into accepted graph `G` and construction record\n   `R` through executable recovery, validation, refinement, deployment review,\n   and Researcher handoff. Task-agnostic runs use representative operating\n   checks when no downstream task-level outcome is defined.\n\n## Source Rules\n\n- Paper: mandatory. Read the full paper or a faithful extracted text version. If the user gives a URL, arXiv id, or title instead of a local path, first resolve it to a local PDF\u002Ftext artifact and save source acquisition evidence.\n- Repository: optional. It may be a local path, Git URL, `none`, or `unknown`. Use it only during modularization and module skill creation. Treat it as implementation evidence, not as a substitute for paper understanding. If it is `unknown`, follow `repo_discovery_mode`: `ask` records candidates\u002Fblockers and asks before searching or selecting, `auto` may run bounded GitHub discovery and clone the top candidate, and `disabled` proceeds paper-only.\n- Recovery: must not read the original source repository. During recovery, use only the paper, module documents, generated module skills, data, and general package documentation.\n- Internet: use only when needed for source acquisition, repo discovery, datasets, package docs, or web-search experiments. If the environment provides a network setup script, source it before network commands. Keep downloads, title searches, repo discovery, and clones bounded and logged.\n- Logs: every run must be recorded under its configured distillation directory.\n\n## Required Loop\n\n1. If a run config path is supplied, validate and normalize it with `scripts\u002Fread_run_config.py`; otherwise collect the same fields from the user's prompt. For batch configs, process each normalized run independently and keep separate source-resolution, attempt, and summary artifacts. Use `read_run_config.py --slug \u003Cpaper_slug> --run-only` to write the selected normalized run JSON passed to `init_run.py --run-config`.\n2. Resolve paper and optional repo sources when needed with `scripts\u002Fresolve_sources.py` or an equivalent bounded, logged procedure. The resolver covers local paths, direct URLs, arXiv URLs\u002Fids, arXiv title search, Git repo URLs, and `repo_discovery_mode=auto` GitHub repo discovery. If a title is not found there, use web search for official or scholarly sources and record the chosen URL. If title or repo discovery is ambiguous, ask the user to choose before proceeding.\n3. Initialize the distillation directory with `scripts\u002Finit_run.py` using the resolved local paper path, optional resolved repo path, configured `iteration_budget`, configured `recovery_mode`, configured `generated_skills_root`, optional exact `attempt_dir`, and `--run-config` when a normalized config JSON exists.\n4. Load `plan-paper-skill-modules` and create:\n   - `paper_profile.md`\n   - `module_plan.json`\n   - one module document per module in `modules\u002F`\n   - no more than five modules\n5. Validate the module plan with `plan-paper-skill-modules\u002Fscripts\u002Fvalidate_module_plan.py`.\n6. For each module, load `create-paper-module-skill` and create the module skill under the configured output skill root.\n7. Validate every generated module skill with `create-paper-module-skill\u002Fscripts\u002Fvalidate_skill_tree.py` and run its tests or smoke checks.\n8. Load `prepare-paper-recovery-env` and create `environment\u002Fruntime_handoff.json` from bounded package, model, GPU, benchmark, and dataset preparation. If required packages, datasets, or model caches are missing, run the active preparation path first: create\u002Freuse an isolated recovery environment under `$DISCO_CODING_AGENT_DIR\u002Fenvs\u002F` or `~\u002F.disco\u002Fagent\u002Fenvs\u002F`, install the smallest missing package set, and attempt permitted bounded model\u002Fdata acquisition before treating the item as blocked.\n9. Load `recover-paper-result` and reproduce one fast, meaningful paper result without reading the original repo, using the runtime handoff as recovery input.\n10. Load `analyze-paper-recovery` and compare the recovery result with the paper target or an explicitly declared proxy target.\n11. If analysis says `refine`, update the relevant workflow guidance, module documents, module skills, environment handoff usage, and recovery harness, then rerun steps 7-10. Stop when status is `accept` or the configured refinement budget is exhausted.\n12. After an accepted run passes final and reported validation, classify the complete generated module graph for deployment. Default task-, project-, private-data-, evaluator-, benchmark-, or environment-bound output and all uncertain cases to project scope at `\u003Cproject-dir>\u002F.agents\u002Fskills\u002F`. Use managed scope at `~\u002F.disco\u002Fagent\u002Fskills\u002F` only with evidence that the graph is self-contained, provenance-backed, independent of transient run state, representative-use verified, and reusable across projects or research tasks. Keep every module in the same scope.\n13. Present the classification evidence, exact targets, entry point, verification, unresolved limits, collisions, and shadowing impact. After approval, use the visible `distill-ml-knowledge\u002Fscripts\u002Fimport_operating_skill_graph.mjs` once with every top-level module root and the selected scope. Overwrite requires separate approval. Write `researcher-handoff.md` with the selected scope and exact imported paths, then suggest `\u002Fresearcher`; if import is declined, record that the graph remains staged.\n\n## Ordering and Runtime Discipline\n\n- Finish and validate the modularization artifacts before heavyweight runtime probing. Do not delay `paper_profile.md`, `module_plan.json`, or `modules\u002F*.md` while searching for models, package caches, datasets, or GPUs.\n- Source acquisition happens before distillation initialization when needed, but it should stay bounded: direct local path checks first, then direct URL\u002FarXiv download, then title search only if requested by the user's input, then repo discovery only when `repo_discovery_mode` or user confirmation permits it. Do not treat source acquisition as recovery evidence for reading the original repo.\n- Keep filesystem discovery bounded. Do not run broad scans such as `find \u002Fshare\u002Fproject ...` or recursive searches across NFS roots. Search only the workspace, user-supplied paths, known cache roots, or a short explicit allowlist, and use timeouts for cache probes.\n- Long package installs, model downloads, and benchmark setup commands must have a time budget. If they stall, stop them, record the exact blocker in recovery artifacts, and continue with the best valid lower-cost recovery allowed by the user.\n- Missing local dependencies, model caches, datasets, or search credentials are\n  not by themselves sufficient to end the run as blocked. First attempt the\n  allowed setup path: isolated environment creation, targeted package\n  installation, bounded dataset\u002Fmodel acquisition, configured VPN\u002Fproxy or\n  mirror use, and keyless\u002Fpublic alternatives where valid. Only mark blocked\n  after those actions are logged or after asking the user for required\n  credentials\u002Fpermissions that cannot be inferred.\n- A fallback or proxy can be useful evidence, but it must not be counted as a successful required runtime unless the recorded mechanism checks prove that the required data item, model\u002Fruntime, loss, and optimizer or training step actually ran.\n- In `hard` recovery mode, do not count fallback, proxy, reduced training, subset-only evaluation, or smaller-model execution as accepted recovery. Such artifacts may be logged as diagnostics, but analysis must return `refine` or `blocked` unless the full requested standard ran.\n- In `soft` recovery mode, reduced or proxy recovery is allowed only as a declared target with executable evidence, generated-skill invocation logs, and `recovery\u002Fexperiment_validation.json` passing.\n- Independent runs must not silently depend on previous runs. Reuse immutable local datasets or benchmark checkouts only as recorded benchmark sources when fresh acquisition is blocked or the user supplied them. Do not reuse previous generated skills, data items, prompts, outputs, training logs, or recovery results as evidence for a new run.\n- Runtime setup must produce a handoff before recovery. `recover-paper-result` should not repeat broad environment discovery or silently invent fallback state; it should consume `environment\u002Fruntime_handoff.json` and add that handoff to `recovery\u002Fsource_manifest.json`.\n\n## Distillation Directory Contract\n\nEach distillation directory should contain:\n\n```text\nrun_manifest.json\nrun_config.normalized.json   # optional but preferred when a config was used\nsource_resolution.json       # optional, when paper\u002Frepo source acquisition was needed\npaper_profile.md\nmodule_plan.json\nmodules\u002F\n  \u003Cmodule_id>.md\ngenerated_skills_validation\u002F\n  \u003Cmodule_id>.json\nenvironment\u002F\n  runtime_handoff.json\n  logs\u002F\n    command_log.json\nrecovery\u002F\n  experiment_plan.md\n  experiment_validation.json\n  source_manifest.json\n  recovery_result.json\n  logs\u002F\n    experiment_command_log.json\n    generated_skill_invocations.json\nanalysis\u002F\n  analysis_report.json\n  feedback.md\nfinal_validation.json\nreports\u002F\n  generated-skills\u002F\n  verification\u002F\n  final\u002F\n    final_report.json\n    final_report.md\n```\n\nUse additional files when useful, but do not omit the core contract. Record failures as artifacts instead of relying on terminal output.\n\n## Quality Bar\n\nA generated module skill is acceptable only if it has:\n\n- A clear trigger description in frontmatter.\n- A concise workflow that states when to use the skill and when not to use it.\n- Explicit input\u002Foutput contracts.\n- At least one deterministic test, smoke script, or validation command.\n- A saved validation JSON whose `ok` is true and whose `tests.attempted` is\n  true. Prose-only inspection is not enough for either hard or soft mode.\n- No dependence on the original paper repository unless the skill's purpose is explicitly to operate on that repository.\n- Enough implementation detail for a future agent run to use the skill without rereading the entire paper.\n\nThe whole paper distillation is acceptable only if:\n\n- The module split covers the paper's main method and evaluation pathway.\n- `environment\u002Fruntime_handoff.json` records package availability, GPU status, model readiness or blockers, benchmark acquisition\u002Freuse, environment mutation status, and allowed recovery sources.\n- `environment\u002Flogs\u002Fcommand_log.json` records subprocess commands run during environment preparation, including timed-out or failed commands.\n- The recovery run exercises the generated skills, not a hand-written one-off solution.\n- The recovery result is produced by an executable experiment command whose command log is saved under `recovery\u002Flogs\u002Fexperiment_command_log.json`.\n- `recovery\u002Fexperiment_plan.md` explains full\u002Freduced\u002Ftoy options and why the chosen target is the strongest feasible experiment.\n- `recovery\u002Flogs\u002Fgenerated_skill_invocations.json` proves generated module skills were called, imported, or cross-checked.\n- The recovery experiment gate is run and saved to `recovery\u002Fexperiment_validation.json` with `ok: true`.\n- The recovery run exercises every generated core module that the proxy claims to validate, or explicitly records why a module is not applicable. If a generated scoring\u002Floss module exists separately from the training harness, the recovery cross-checks it rather than relying only on duplicated harness logic.\n- Proxy recovery records explicit `mechanism_checks` that demonstrate the core paper mechanism ran. Do not accept a proxy solely because EM\u002FF1 or another task metric is high.\n- If recovery uses a benchmark or dataset repository, the item log and source manifest identify concrete resource files used to derive the sample whenever feasible. A benchmark-style item without resource-file provenance must be labeled as such and justified.\n- If recovery reuses a local benchmark checkout from another run or workspace location, the source manifest records the fresh-fetch blocker, reused path, commit\u002Fversion, concrete files used, and a snapshot or copy of those files inside the current distillation directory.\n- At least one recovery run uses the strongest practical real runtime available in the workspace when the user provides one. If the user gives a model path, GPU ids, or an inference environment, record whether it was used; if it could not be used, record the exact blocker and run the best lower-cost fallback.\n- The analysis report explains metric gaps and either accepts the result or gives actionable module-level feedback.\n- The final artifact contract validator is run and its JSON output is saved to `final_validation.json`.\n- The organized final report is written to `reports\u002Ffinal\u002Ffinal_report.md` and\n  `reports\u002Ffinal\u002Ffinal_report.json`, and the `reported` validation stage passes.\n- The complete module graph has a recorded reusability classification. Any live\n  deployment uses one approved scope, excludes process\u002Freview artifacts, and is\n  recorded with exact paths in `researcher-handoff.md`.\n- `run_manifest.json` and `run_config.normalized.json` record `recovery_mode`, and hard-mode runs are not accepted if `recovery_result.json` marks `is_proxy: true` or `mechanism_checks.reduced_training_executed: true`.\n- When paper\u002Frepo source acquisition was needed, the resolved local paths, commands\u002Fnetwork operations, selected title\u002FarXiv id, and blockers are recorded in `source\u002Fsource_resolution.json` or `reports\u002Ffinal\u002Ffinal_report.md`.\n\n## Refinement Rules From Real Runs\n\n- Do not accept a semantically correct but metric-failing result without improving the relevant evaluation or answer-formatting skill.\n- If an intermediate module output contaminates later stages, refine that module's output contract and add a regression test before rerunning recovery.\n- If a runtime environment is incompatible, do not mutate it unless the user explicitly allows it. Prefer an existing compatible environment or create a separate environment under `$DISCO_CODING_AGENT_DIR\u002Fenvs\u002F` or `~\u002F.disco\u002Fagent\u002Fenvs\u002F`, then record the choice in `source_manifest.json`.\n- After any refinement, rerun module skill validation and recovery in the current distillation directory or a clearly versioned cycle subdirectory.\n\n## Practical Guidance\n\nFavor small, composable module skills over one monolithic \"paper skill\". A good split usually separates protocol\u002Fdata formatting, retrieval\u002Ftool adapters, core algorithm loop, scoring\u002Fevaluation, and experiment recovery. Keep the number of modules below five by merging thin utility modules into the nearest substantive module.\n\nIf the full paper result requires heavy training or unavailable models, check `recovery_mode` before lowering the target. In `hard` mode, record the blocker and do not accept a proxy as success. In `soft` mode, recover a declared proxy result that preserves the paper's mechanism. The proxy must identify the dataset, sample size, metric, paper target, and why the proxy is the fastest faithful test.\n\n## Scripts\n\n- `scripts\u002Fread_run_config.py`: validate and normalize a TOML config with one or more Distiller runs.\n- `scripts\u002Fresolve_sources.py`: resolve local or remote paper\u002Frepo sources into local paths before distillation initialization, including optional bounded repo discovery.\n- `scripts\u002Finit_run.py`: create the distillation directory and manifest, optionally copying a selected normalized run config into the run.\n- `scripts\u002Fvalidate_distillation_run.py`: validate that required artifacts exist at each stage.\n- `scripts\u002Fwrite_final_report.py`: write `reports\u002Ffinal\u002Ffinal_report.json` and\n  `reports\u002Ffinal\u002Ffinal_report.md` from the distillation artifacts.\n",{"name":5,"metadata":23,"description":7},{"disco-role":24},"meta","main","ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6","cli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fpaper-skills-distiller",27,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab","2026-09-03 10:25:21",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add VectorSpaceLab\u002FAREX-Skill --skill paper-skills-distiller",{"tool":38,"command":39},"bunx","bunx skills add VectorSpaceLab\u002FAREX-Skill --skill paper-skills-distiller",{"tool":41,"command":42},"pnpm","pnpm dlx skills add VectorSpaceLab\u002FAREX-Skill --skill paper-skills-distiller","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fpaper-skills-distiller\u002FSKILL.md",true,[46,52,56,60,64,68,72,76,80,84],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"scripts\u002Finit_run.py","SCRIPT","脚本",5674,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fpaper-skills-distiller\u002Fscripts\u002Finit_run.py",{"path":53,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":54,"sourceHtmlUrl":55},"scripts\u002Fread_run_config.py",11818,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fpaper-skills-distiller\u002Fscripts\u002Fread_run_config.py",{"path":57,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":58,"sourceHtmlUrl":59},"scripts\u002Fresolve_sources.py",15526,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fpaper-skills-distiller\u002Fscripts\u002Fresolve_sources.py",{"path":61,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":62,"sourceHtmlUrl":63},"scripts\u002Fvalidate_distillation_run.py",26425,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fpaper-skills-distiller\u002Fscripts\u002Fvalidate_distillation_run.py",{"path":65,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":66,"sourceHtmlUrl":67},"scripts\u002Fwrite_final_report.py",11165,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fpaper-skills-distiller\u002Fscripts\u002Fwrite_final_report.py",{"path":69,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":70,"sourceHtmlUrl":71},"tests\u002Ftest_init_run.py",7102,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fpaper-skills-distiller\u002Ftests\u002Ftest_init_run.py",{"path":73,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":74,"sourceHtmlUrl":75},"tests\u002Ftest_read_run_config.py",5577,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fpaper-skills-distiller\u002Ftests\u002Ftest_read_run_config.py",{"path":77,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":78,"sourceHtmlUrl":79},"tests\u002Ftest_resolve_sources.py",5101,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fpaper-skills-distiller\u002Ftests\u002Ftest_resolve_sources.py",{"path":81,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":82,"sourceHtmlUrl":83},"tests\u002Ftest_validate_distillation_run.py",13730,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fpaper-skills-distiller\u002Ftests\u002Ftest_validate_distillation_run.py",{"path":85,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":86,"sourceHtmlUrl":87},"tests\u002Ftest_write_final_report.py",2722,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fpaper-skills-distiller\u002Ftests\u002Ftest_write_final_report.py",[89,97,104,111,118,125],{"skillNo":90,"name":91,"nameZh":6,"description":92,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":93,"categories":94,"keywords":95,"publishedAt":96},"01M799NYSH1R3RVAHSM2D88Y84","passnet-triton-opt","Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block\u002Fgrid\u002Fwarp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-triton-opt\u002FSKILL.md",[],[],"2026-11-15 14:19:17",{"skillNo":98,"name":99,"nameZh":6,"description":100,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":101,"categories":102,"keywords":103,"publishedAt":96},"01M799NS5APEC0N1FZ7XZVW6BN","passnet-solve","END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-solve\u002FSKILL.md",[],[],{"skillNo":105,"name":106,"nameZh":6,"description":107,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":108,"categories":109,"keywords":110,"publishedAt":96},"01M799NNY9BKSM21V5R7QVTBJX","passnet-skill","PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under .\u002Fpass_dir\u002F, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-skill\u002FSKILL.md",[],[],{"skillNo":112,"name":113,"nameZh":6,"description":114,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":115,"categories":116,"keywords":117,"publishedAt":96},"01M799NHH6NP7Z1RK77KGXAN6S","passnet-pattern-fusion","Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()\u002Freplacement_args()\u002Freplacement_func() files, when a pass \"failed to match\", or when deciding how to fuse multiple kernels\u002Fops.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-pattern-fusion\u002FSKILL.md",[],[],{"skillNo":119,"name":120,"nameZh":6,"description":121,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":122,"categories":123,"keywords":124,"publishedAt":96},"01M799ND2J4WNC2PGAZ0A06NDF","passnet-orchestrate","Round\u002Fbatch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep\u002Frevert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-orchestrate\u002FSKILL.md",[],[],{"skillNo":126,"name":127,"nameZh":6,"description":128,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":129,"categories":130,"keywords":131,"publishedAt":96},"01M799N8SBTMY02ZDG2AZF5ZWE","passnet-feedback","Fast, deterministic feedback for PassNet work: pre-flight pattern\u002Fmatch verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-feedback\u002FSKILL.md",[],[]]