Prior Art
Ground the next decision in sources that were actually inspected.
Direct use
Do the inspection and return the recommendation in the same task. Do not stop
after announcing a search plan, ask the operator to enumerate obvious files, or
turn bounded discovery into a repository-wide grep.
- Start from the operator's objective and current workspace. When exact
source_paths are already supplied, reuse them.
- Otherwise, make one targeted discovery pass for the smallest likely owning
set: repository instructions and conventions, the named product or package,
exact active plans or specs implicated by the objective, and adjacent Runx
skill manifests. Select at most sixteen files. Listing filenames is
discovery, not the prior-art result; continue through inspection.
- Invoke the default method with the objective, chosen paths, and any existing
decomposition. It captures the selected evidence once and analyzes that
immutable capture. Complete its bounded research request, resume it in memory,
and return the validated prior-art report—not a narration of what you intend
to read.
- If the selected evidence is insufficient, return one precise
needs_more_evidence result naming the missing source. Do not broaden the
search repeatedly or cycle through unrelated skills.
Routine local discovery and reads use the host's normal authenticated tools and
need no approval. Runx adds the catalog projection, digest-bound source bundle,
citation validation, reusable packet, and receipt. If Runx is unavailable,
preserve the selected paths and findings in the documented output shape so the
operator can continue locally without repeating discovery; do not claim a Runx
receipt.
Composed use
A parent that has only selected paths calls capture once with those paths,
the repository root and any decomposition. Pass its returned artifact
descriptor as snapshot to analyze, together with the objective and optional
decision context. Reuse the same descriptor for further analysis in the same
Runx project. analyze opens the retained artifact; it does not inspect the
catalog, reopen source paths, or repeat discovery. Changed workspace files do
not replace captured bytes. Paths and caller-authored source summaries are not
a capture.
If the capture is missing, incomplete, stale or outside the current objective,
state the gap and explicitly select a new capture. An unreadable or foreign
project artifact fails closed without reacquisition. An artifact's integrity
does not independently prove the truth of its contents or grant authority.
Executable procedure
- Supply a bounded
objective, optional code-plan decomposition, and the
repo-relative source_paths selected by direct discovery or a parent chain.
- In
capture, native runx.skill.inspect indexes the local Runx catalog and
fs.read_bundle reads and hashes at most sixteen requested files under the
workspace boundary. Missing files are recorded; escaping, duplicate, or
oversized paths fail closed. artifact.store retains both complete native
projections in the project's immutable artifact store. analyze verifies
that descriptor with artifact.open, reads the bounded retained JSON with
artifact.read, and refuses a partial read. Decoded native projections
retain their packet contracts.
- Read the exact
inspected_bundle.files[].contents delivered with the task.
It is the same native read whose whole-file digests appear in the evidence
index; do not read those files again. Catalog descriptions help identify an
owner, but do not establish uninspected implementation details. A truncated
or missing requested source holds the result.
Read only the indexed sources needed for the objective. State each finding as claim, source, relevance, and confidence (verified, likely, or unverified).
- Name adjacent catalog skills and the boundary each already owns. Recommend
reuse, amend, new_work, or stop; do not create a duplicate primitive because an existing package is imperfect.
- The domain validator checks every verified citation and adjacent skill
against those native projections. Unknown citation paths or missing requested
sources force
needs_more_evidence. A bound citation does not establish that
its source supports the claim; inspect the retained text before calling a
finding verified.
External URLs are not inspected by this method. Fetch them through a governed source skill, persist the bounded evidence in the workspace, then pass its path here.
Output
decision: ready | needs_more_evidence
findings:
- claim: string
source: repo-relative path
relevance: string
confidence: verified | likely | unverified
catalog_fit:
decision: reuse | amend | new_work | stop
adjacent_skills: array
rationale: string
recommended_flow: array
quality_bar: object
sources: array
risks: array
evidence: object
validation: object
The default accepts objective, optional decomposition, workflow_purpose,
audience, artifact_contract, bounded source_paths and repo_root.
capture accepts decomposition, source_paths and repo_root; analyze
accepts snapshot, objective and the optional decision context.
Agent task: prior-art-research
Use the retained native source text in inspected_bundle and the catalog descriptions in
evidence_index. No direct filesystem tools are needed in this judgment phase.
Do not reopen a path and substitute changed bytes for its admitted snapshot. Return prior_art_draft with decision, findings, catalog_fit, quality_bar,
recommended_flow, sources, and risks. Cite repo-relative paths exactly. Mark a finding verified
only when its source appears in inspected_sources or the catalog index. Prefer reuse or
amendment over duplicate primitives. Return needs_more_evidence when requested sources are
missing.
1---2name: prior-art3description: Inspect bounded local repository and Runx catalog evidence before a design, draft, or operator decision, then produce citation-bound findings and a reuse, amendment, new-work, or stop recommendation. Use when existing tools, skills, standards, or project patterns must constrain downstream work; use research or web-fetch first for external sources.4metadata:5 source: https://skillmd.com/skills/runxhq/prior-art6---78# Prior Art910Ground the next decision in sources that were actually inspected.1112## Direct use1314Do the inspection and return the recommendation in the same task. Do not stop15after announcing a search plan, ask the operator to enumerate obvious files, or16turn bounded discovery into a repository-wide grep.17181. Start from the operator's objective and current workspace. When exact19 `source_paths` are already supplied, reuse them.202. Otherwise, make one targeted discovery pass for the smallest likely owning21 set: repository instructions and conventions, the named product or package,22 exact active plans or specs implicated by the objective, and adjacent Runx23 skill manifests. Select at most sixteen files. Listing filenames is24 discovery, not the prior-art result; continue through inspection.253. Invoke the default method with the objective, chosen paths, and any existing26 decomposition. It captures the selected evidence once and analyzes that27 immutable capture. Complete its bounded research request, resume it in memory,28 and return the validated prior-art report—not a narration of what you intend29 to read.304. If the selected evidence is insufficient, return one precise31 `needs_more_evidence` result naming the missing source. Do not broaden the32 search repeatedly or cycle through unrelated skills.3334Routine local discovery and reads use the host's normal authenticated tools and35need no approval. Runx adds the catalog projection, digest-bound source bundle,36citation validation, reusable packet, and receipt. If Runx is unavailable,37preserve the selected paths and findings in the documented output shape so the38operator can continue locally without repeating discovery; do not claim a Runx39receipt.4041## Composed use4243A parent that has only selected paths calls `capture` once with those paths,44the repository root and any decomposition. Pass its returned `artifact`45descriptor as `snapshot` to `analyze`, together with the objective and optional46decision context. Reuse the same descriptor for further analysis in the same47Runx project. `analyze` opens the retained artifact; it does not inspect the48catalog, reopen source paths, or repeat discovery. Changed workspace files do49not replace captured bytes. Paths and caller-authored source summaries are not50a capture.5152If the capture is missing, incomplete, stale or outside the current objective,53state the gap and explicitly select a new capture. An unreadable or foreign54project artifact fails closed without reacquisition. An artifact's integrity55does not independently prove the truth of its contents or grant authority.5657## Executable procedure58591. Supply a bounded `objective`, optional code-plan `decomposition`, and the60 repo-relative `source_paths` selected by direct discovery or a parent chain.612. In `capture`, native `runx.skill.inspect` indexes the local Runx catalog and62 `fs.read_bundle` reads and hashes at most sixteen requested files under the63 workspace boundary. Missing files are recorded; escaping, duplicate, or64 oversized paths fail closed. `artifact.store` retains both complete native65 projections in the project's immutable artifact store. `analyze` verifies66 that descriptor with `artifact.open`, reads the bounded retained JSON with67 `artifact.read`, and refuses a partial read. Decoded native projections68 retain their packet contracts.693. Read the exact `inspected_bundle.files[].contents` delivered with the task.70 It is the same native read whose whole-file digests appear in the evidence71 index; do not read those files again. Catalog descriptions help identify an72 owner, but do not establish uninspected implementation details. A truncated73 or missing requested source holds the result.74 Read only the indexed sources needed for the objective. State each finding as `claim`, `source`, `relevance`, and `confidence` (`verified`, `likely`, or `unverified`).754. Name adjacent catalog skills and the boundary each already owns. Recommend `reuse`, `amend`, `new_work`, or `stop`; do not create a duplicate primitive because an existing package is imperfect.765. The domain validator checks every verified citation and adjacent skill77 against those native projections. Unknown citation paths or missing requested78 sources force `needs_more_evidence`. A bound citation does not establish that79 its source supports the claim; inspect the retained text before calling a80 finding verified.8182External URLs are not inspected by this method. Fetch them through a governed source skill, persist the bounded evidence in the workspace, then pass its path here.8384## Output8586```yaml87decision: ready | needs_more_evidence88findings:89 - claim: string90 source: repo-relative path91 relevance: string92 confidence: verified | likely | unverified93catalog_fit:94 decision: reuse | amend | new_work | stop95 adjacent_skills: array96 rationale: string97recommended_flow: array98quality_bar: object99sources: array100risks: array101evidence: object102validation: object103```104105The default accepts `objective`, optional `decomposition`, `workflow_purpose`,106`audience`, `artifact_contract`, bounded `source_paths` and `repo_root`.107`capture` accepts `decomposition`, `source_paths` and `repo_root`; `analyze`108accepts `snapshot`, `objective` and the optional decision context.109110## Agent task: prior-art-research111112Use the retained native source text in `inspected_bundle` and the catalog descriptions in113`evidence_index`. No direct filesystem tools are needed in this judgment phase.114Do not reopen a path and substitute changed bytes for its admitted snapshot. Return prior_art_draft with decision, findings, catalog_fit, quality_bar,115recommended_flow, sources, and risks. Cite repo-relative paths exactly. Mark a finding verified116only when its source appears in inspected_sources or the catalog index. Prefer reuse or117amendment over duplicate primitives. Return needs_more_evidence when requested sources are118missing.