Lintel on GitHub Copilot. Generated from
skills/scaffold/SKILL.md; edit the canonical file, then runli-copilot init.
- Resource root:
../../..from this skill's base directory (the Lintel source withbin/,lib/,skills/). Write plans, state and evidence into the working repository's.claude/tree, never into the resource root.- Skill-relative paths: this skill's own
scripts/,references/anddata/folders (and a<base>that the workflow defines as its own directory) mean../../../skills/scaffold/in the Lintel source, not this generated folder.${LINTEL_SKILLS_DIR:-skills}means the skills root,../../../skills. Abin/li-runstep runs in the working repository, so use$LINTEL_SKILLS_DIR/scaffold/there.- Shell steps: run Bash snippets with Bash (Git for Windows'
bash.exeon Windows, neverSystem32\bash.exe). Save a snippet to a temporary.shfile and runbash "<resource root>/bin/li-run" <file>; it preparesLINTEL_SOURCE_ROOT,LINTEL_REPO_ROOTand the profile context.- Tools: Read=
view, Write=create, Edit=edit, Bash=bash/powershell, Grep=grep, Glob=glob, AskUserQuestion=ask_user, TodoWrite=the plan checklist, Task or a named role=taskwith that custom agent, WebFetch=web_fetch.- Other Lintel workflows are native skills: invoke
/li-<name>rather than reading their files. Named roles such asCodeReviewerare custom agents.
Scaffold
Set up a new or existing repository with the current foundation, AGENTS/CLAUDE entry templates, memory index, lessons/personas/working state, decisions, plan and swarm templates. Keep the established base-initialization use case; use the actual helper instead of a second copy/render recipe.
Resolve the operation
Follow lifecycle paths. Resolve helper code at the approved
LINTEL_SOURCE_ROOT; the explicit target alone selects where files change. Do not switch
to a different repository by basename, fetch a source implicitly or use the working
target's executable files as Lintel's implementation.
Reuse supplied name, intent and target. Ask only for missing decisions. --mode and
--voice are rendered preferences, not enabled controls. --pack must resolve through the
structured policy contract. Required caller and target policy cannot be silently
replaced by neutral defaults or by an identically named pack with different content.
The historical --compliance spelling returns an explicit unsupported-policy error
before writes; it never had a control implementation. Use validated pack policy instead
of treating a full/minimal label as enforcement.
Application-mode acceptance (one owner)
This method owns the internal-tool and MVP questions, including calls through
the legacy aliases. Reuse supplied answers; ask each missing acceptance/failure
question once, not again on alias-to-owner handoff. Intent inputs below are not
new flags for bin/li-scaffold.
internal-tool: retain--name,--path,--language,--type,--ci;--pathselects the explicit target. Preserve CLI, service, dashboard and automation-script use cases with the selected available toolchain.mvp: retain--name,--target-users,--path,--stack,--has-ai,--deploy; preserve product/user framing and the first useful journey.- Prefer existing manifests, architecture and framework-native generators when available and authorized. Ask about an unknown stack; do not assume a license, data classification, vendor, cloud, CI platform, AI feature or deployment target. An internal audience does not establish that all inputs are nonsensitive.
Define one useful end-to-end acceptance case and its principal failure case:
| Requested surface | Observable success and failure |
|---|---|
| CLI | Real arguments/input and result; invalid-input exit with actionable error |
| Service | Actual request/response and health contracts; validation and failure handling |
| Dashboard | Requested data/action flow; accessible loading, empty and error states |
| Script/automation | Explicit inputs/destination, deterministic output; safe failure/retry |
| MVP journey | First complete user job, needed state/API contracts; access and failure cases |
For AI features, retain sanitized golden and adversarial cases, expected outcomes and reproducible evaluation commands; protect approved golden cases from casual edits. Actual data classification/provenance, retention and ownership remain explicit. Stubs carry an owner and follow-up phase and never satisfy acceptance. A draft policy artifact is not a completed compliance review.
Inspect, then initialize
bash "$LINTEL_SOURCE_ROOT/bin/li-scaffold" check --target "$target"
bash "$LINTEL_SOURCE_ROOT/bin/li-scaffold" init --target "$target" --name "$name"
For a selected application mode, the same initializer receives only its existing
mode flag (with mode set to internal-tool or mvp):
bash "$LINTEL_SOURCE_ROOT/bin/li-scaffold" init \
--target "$target" --name "$name" --mode "$mode"
check/--dry-run report the actual planned create/replace/delete paths without writing.
The helper preflights all paths and collisions, including late links, before publication.
Existing project instructions and seed files stay user-owned. Legacy data is migrated
only through the same verified transaction, never overwritten by new template seeds.
The legacy Claude autoMemoryDirectory pointer is a repository-local setting with
preserved surrounding configuration. Use --no-memory-pointer when that client-specific
setting is not wanted. A declared path is not proof of live host memory behavior.
For a native repository adapter, use --client <exact-surface> (repeatable) or the
preserved --copilot alias. These delegate to the existing ownership engine; they cannot
be combined with unsupported legacy rendering options. They do not register hooks,
change permissions, provision credentials or install a company pack.
Implement and verify the selected application
Add the actual entry, domain/API/UI boundaries, minimal configuration, meaningful tests and specific purpose/install/usage documentation for the accepted flow. Use the project's chosen runner and conventions rather than generic stack advice or speculative setup. Dependency installation needs actual authority and a changed manifest or demonstrated missing dependency; never install automatically.
Run the smallest relevant build/type/lint checks and the actual success/failure cases, with AI evaluation only where selected and authorized. CI must exercise real commands and required controls; comments, no-op gates or slash-command text are not enforcement. Customer guidance uses the selected pack's actual voice/corpus without copying private material or inventing calibration.
Deployment preparation names environments, configuration/secrets boundaries, observability and rollback. It does not deploy: no automatic hook registration, external activation, private sync, image push or staging/production action. Each needs its real authority and host support. Preserve existing governance.
Context and recovery
When an existing caller context is supplied, the helper verifies that pin against its original repository/home/source. An explicit new target is separately resolved, with a required caller policy retained as an operation constraint. It does not transplant the parent reference, change parent state or create a child profile generation. Start work in the child with its own explicit bootstrap.
Record the helper's exact result and transaction/store identity. Its staged byte plan
uses P03 snapshots and explicit, conflict-preserving restore. On interruption, keep the
error and receipt; do not report success, retry into partial state or automatically roll
back. Use the documented li-managed-transaction.py inspect|recover command for the exact
target/store/ID. Later user edits and consumed recovery permissions are refused.
Finish
Verify the returned changed/preserved files and relevant consumer checks. Leave Git staging and commits to the authorized task; the helper never stages an unrelated index. Preserve existing project governance. Additional compliance assets, private exports and hook/host activation each need their own configured scope and evidence.
The internal-tool and MVP aliases delegate their preserved intent to this owner. Report every placeholder, unavailable dependency/tool and unrun host/CI/environment check explicitly. Keep source/target hashes, transaction receipt, profile context and requirements with the selected work evidence. Foundation files alone are not a functioning application or a verified live client session; independent specification/quality review and deployment acceptance remain distinct gates.