foundationskills

Plugin foundationskills from marketplace foundationscale in TranNhiem/FoundationScale. Commands read from the repo's manifest — 17 days ago.

Skills (50)

• fskills-auto-research

  Manages a hash-chained auto-research campaign (M0): validates the campaign spec, opens one

• fskills-compression

  Model compression and quantization of a HF checkpoint (v1 method: RTN weight-only quantize-dequantize, int4 or int8, group-wise). Use it to preflight a host for compression (vendored preflight.py; classifies FAIL, UNKN and PROBE_ERROR results), to plan a probe-first run that emits an fs_launch_spec

• fskills-confirm-before-launch

  Gates an fs_launch_spec behind the user-confirmed plan hash: it recomputes the plan hash (`fskills hash --spec`, the same token `fskills launch --confirm` requires) and REFUSES (96) when the user-supplied hash is absent or mismatches, sealing a confirmation_record with confirmed_by=user and auto_fil

• fskills-data-engine

  Turns raw sources (JSONL/JSON/CSV/parquet, txt/html/pdf doc dirs, HF datasets, image-text

• fskills-deployment

  Serves a checkpoint (full or merged weights, an already-quantized checkpoint such as fp8, or base + LoRA adapter unmerged) with the measured vLLM container: plans a GPU-confined serving fs_launch_spec probe-first (launched only with the plan hash the user types back), then verifies the running endpo

• fskills-evaluation

  Scores a trained checkpoint (full HF model dir or PEFT adapter) against its base model on

• fs-data

  User-invoked slash command: run ONLY when the user types /fs-data; never choose it on your own. Slash-command data preparation workflow (/fs-data) that composes the model skill data_engine to turn raw data refs into a dataset, data_pipeline_spec and readiness_report. Inputs: the data_engine request

• fs-eval

  User-invoked slash command: run ONLY when the user types /fs-eval; never choose it on your own. /fs-eval is the A3.2 user-invoked workflow skill: it takes the evaluation request (checkpoint, run_manifest, base, benchmarks, policy, eval_cache, out, baseline_cache, ...) and forwards it UNCHANGED to th

• fs-goal

  User-invoked slash command: run ONLY when the user types /fs-goal; never choose it on your own. Grilling composer (/fs-goal) for one FoundationSkills goal_spec: takes the intake answers so far ({objective, goal?, domain?, preserve_general?, base_model {name_or_path, size_b?, arch?}, sources [{uri, k

• fs-handoff

  User-invoked slash command: run ONLY when the user types /fs-handoff; never choose it on your own. Slash-command closer (/fs-handoff) for a FoundationSkills session: takes this session's artifact JSON paths, optional notes and next_steps and writes ONE handoff artifact that references every artifact

• fs-launch

  User-invoked slash command: run ONLY when the user types /fs-launch; never choose it on your own. Slash-command launch gate (/fs-launch) for FoundationSkills and THE ONLY launch path: takes an fs_launch_spec artifact path plus the plan hash the USER typed back (user_plan_hash), refuses when it is mi

• fs-plan

  User-invoked slash command: run ONLY when the user types /fs-plan; never choose it on your own. Planning workflow (/fs-plan) for FoundationSkills: composes three model skills in order - training.planner -> probe_first -> training.emit - and stops at the first step that is not PASS (FP-IN-003 on a re

• fs-research

  User-invoked slash command: run ONLY when the user types /fs-research; never choose it on your own. Claim-measurement research workflow (/fs-research): takes an optional goal_spec path and a claims list {id, claim, asserted_status?, evidence?, measured?}, composes measured_or_unmeasured to check eve

• fs-status

  User-invoked slash command: run ONLY when the user types /fs-status; never choose it on your own. Slash-command explainer (/fs-status) for FoundationSkills results: takes a journal.jsonl record written by the orchestrator (last record by default, or a step_index) or a skill result object {status, ex

• fskills-measured-or-unmeasured

  Rewrites a claim set so any claim without a measured evidence pointer (<artifact path>#<json field>, resolvable on disk) is emitted as UNMEASURED and never as PASS, and seals the corrected claim_set with provenance. Do NOT use to probe FoundationScale (use fskills-probe-first), to score checkpoints

• fskills-merging

  Folds a LoRA adapter into its base model as merged full weights: plans the merge probe-first and emits an fs_launch_spec running FoundationScale's merge_lora.py (launched only with the plan hash the user types back), then verifies the merged checkpoint structurally and by numeric logit parity agains

• fskills-probe-first

  Refuses a FoundationScale plan draft that was written before the measured probe (fskills probe --deep) or that assumes a capability the probe reports missing/unmeasured, and emits a vetted probe_report (measured capabilities, provenance, deep/partial flag). Do NOT use for building datasets (use fski

• fskills-refusal-surface

  Turns RED/REFUSED findings into a sealed finding_set whose every entry cites a rule id that a registered skill actually declares (undeclared or invented ids become RED citing RS-HO-001, unattributed findings are cited as RS-HO-002). Do NOT use to produce findings (fskills-evaluation, fskills-probe-f

• fskills-training

  Turns a training goal (goal_spec, optional readiness_report) into an auditable training_plan

• using-foundationskills

  User-invoked slash command: run ONLY when the user types /using-foundationskills; never choose it on your own. Entry point for FoundationSkills: routes a user's intent (or a have→want chain such as {raw_data_ref, goal_spec} to fs_launch_spec) to the right FoundationSkills skill, lists every register

• fskills-vla

  Plans and judges vision-language-action (VLA) policy work strictly against FoundationScale's VLA registry: an eval of a SUPPORTED policy (LIBERO reproductions of gr00t, openpi, Long-WAM on GB200) becomes a launch spec behind the plan hash, and its report is judged against the published success rate

• fs-data

  User-invoked slash command: run ONLY when the user types /fs-data; never choose it on your own. Slash-command data preparation workflow (/fs-data) that composes the model skill data_engine to turn raw data refs into a dataset, data_pipeline_spec and readiness_report. Inputs: the data_engine request

• fs-eval

  User-invoked slash command: run ONLY when the user types /fs-eval; never choose it on your own. /fs-eval is the A3.2 user-invoked workflow skill: it takes the evaluation request (checkpoint, run_manifest, base, benchmarks, policy, eval_cache, out, baseline_cache, ...) and forwards it UNCHANGED to th

• fs-goal

  User-invoked slash command: run ONLY when the user types /fs-goal; never choose it on your own. Grilling composer (/fs-goal) for one FoundationSkills goal_spec: takes the intake answers so far ({objective, goal?, domain?, preserve_general?, base_model {name_or_path, size_b?, arch?}, sources [{uri, k

• fs-handoff

  User-invoked slash command: run ONLY when the user types /fs-handoff; never choose it on your own. Slash-command closer (/fs-handoff) for a FoundationSkills session: takes this session's artifact JSON paths, optional notes and next_steps and writes ONE handoff artifact that references every artifact

• fs-launch

  User-invoked slash command: run ONLY when the user types /fs-launch; never choose it on your own. Slash-command launch gate (/fs-launch) for FoundationSkills and THE ONLY launch path: takes an fs_launch_spec artifact path plus the plan hash the USER typed back (user_plan_hash), refuses when it is mi

• fs-plan

  User-invoked slash command: run ONLY when the user types /fs-plan; never choose it on your own. Planning workflow (/fs-plan) for FoundationSkills: composes three model skills in order - training.planner -> probe_first -> training.emit - and stops at the first step that is not PASS (FP-IN-003 on a re

• fs-research

  User-invoked slash command: run ONLY when the user types /fs-research; never choose it on your own. Claim-measurement research workflow (/fs-research): takes an optional goal_spec path and a claims list {id, claim, asserted_status?, evidence?, measured?}, composes measured_or_unmeasured to check eve

• fs-status

  User-invoked slash command: run ONLY when the user types /fs-status; never choose it on your own. Slash-command explainer (/fs-status) for FoundationSkills results: takes a journal.jsonl record written by the orchestrator (last record by default, or a step_index) or a skill result object {status, ex

• fskills-auto-research

  Manages a hash-chained auto-research campaign (M0): validates the campaign spec, opens one

• fskills-compression

  Model compression and quantization of a HF checkpoint (v1 method: RTN weight-only quantize-dequantize, int4 or int8, group-wise). Use it to preflight a host for compression (vendored preflight.py; classifies FAIL, UNKN and PROBE_ERROR results), to plan a probe-first run that emits an fs_launch_spec

• fskills-confirm-before-launch

  Gates an fs_launch_spec behind the user-confirmed plan hash: it recomputes the plan hash (`fskills hash --spec`, the same token `fskills launch --confirm` requires) and REFUSES (96) when the user-supplied hash is absent or mismatches, sealing a confirmation_record with confirmed_by=user and auto_fil

• fskills-data-engine

  Turns raw sources (JSONL/JSON/CSV/parquet, txt/html/pdf doc dirs, HF datasets, image-text

• fskills-deployment

  Serves a checkpoint (full or merged weights, an already-quantized checkpoint such as fp8, or base + LoRA adapter unmerged) with the measured vLLM container: plans a GPU-confined serving fs_launch_spec probe-first (launched only with the plan hash the user types back), then verifies the running endpo

• fskills-evaluation

  Scores a trained checkpoint (full HF model dir or PEFT adapter) against its base model on

• fskills-measured-or-unmeasured

  Rewrites a claim set so any claim without a measured evidence pointer (<artifact path>#<json field>, resolvable on disk) is emitted as UNMEASURED and never as PASS, and seals the corrected claim_set with provenance. Do NOT use to probe FoundationScale (use fskills-probe-first), to score checkpoints

• fskills-merging

  Folds a LoRA adapter into its base model as merged full weights: plans the merge probe-first and emits an fs_launch_spec running FoundationScale's merge_lora.py (launched only with the plan hash the user types back), then verifies the merged checkpoint structurally and by numeric logit parity agains

• fskills-probe-first

  Refuses a FoundationScale plan draft that was written before the measured probe (fskills probe --deep) or that assumes a capability the probe reports missing/unmeasured, and emits a vetted probe_report (measured capabilities, provenance, deep/partial flag). Do NOT use for building datasets (use fski

• fskills-refusal-surface

  Turns RED/REFUSED findings into a sealed finding_set whose every entry cites a rule id that a registered skill actually declares (undeclared or invented ids become RED citing RS-HO-001, unattributed findings are cited as RS-HO-002). Do NOT use to produce findings (fskills-evaluation, fskills-probe-f

• fskills-training

  Turns a training goal (goal_spec, optional readiness_report) into an auditable training_plan

• fskills-vla

  Plans and judges vision-language-action (VLA) policy work strictly against FoundationScale's VLA registry: an eval of a SUPPORTED policy (LIBERO reproductions of gr00t, openpi, Long-WAM on GB200) becomes a launch spec behind the plan hash, and its report is judged against the published success rate

• using-foundationskills

  User-invoked slash command: run ONLY when the user types /using-foundationskills; never choose it on your own. Entry point for FoundationSkills: routes a user's intent (or a have→want chain such as {raw_data_ref, goal_spec} to fs_launch_spec) to the right FoundationSkills skill, lists every register

• codebase-design

  Shared vocabulary for designing deep modules. Use when the user wants to design or improve a module's interface, find deepening opportunities, decide where a seam goes, make code more testable or AI-navigable, or when another skill needs the deep-module vocabulary.

• domain-modeling

  Build and sharpen a project's domain model. Use when discussing codebase terminology, writing or editing a GLOSSARY.md, or recording or editing an ADR.

• grill-with-docs

  A relentless interview to sharpen a plan or design, which also creates docs (ADR's and glossary) as we go.

• grilling

  Grill the user relentlessly about a plan, decision, or idea. Use when the user wants to stress-test their thinking, or uses any 'grill' trigger phrases.

• pr

  Use when writing a PR body.

• retro

  Conduct a retrospective on a coding session.

• wait-what

  Stop. That last message did not land: re-pitch it.

• writing-for-agents

  Writing documents for agents. Use when creating or editing skills, or modifying AGENTS.md or CLAUDE.md.