Lintel on GitHub Copilot. Generated from
skills/full-engineering-pass/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/full-engineering-pass/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/full-engineering-pass/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.
Full engineering pass
Compose TA -> DA/SC -> DH -> TQ without inventing an execution engine. The module methods remain in their canonical skills and retained role bodies. The caller owns actual tool dispatch, explicit handoffs and verification, using the shared module procedure. This is not a tenth phase. SENSE -> SCOPE -> DEFINE -> DISCOVER -> PLAN -> BUILD -> REVIEW -> SHIP -> CAPTURE remains the lifecycle; resume is a utility.
Use for a genuinely cross-domain request, not an ordinary fix or mandatory venture/ customer interview. Direct single-module and single-capability entry points remain. The historic 500k/750k planning hints are uncalibrated advice, not measured usage, host context capacity, a spending authorization or an automatic model setting.
Composition DAG
| Stage | Modules | Depends on | Required handoff |
|---|---|---|---|
| Stage 1 | TA | original accepted requirements | decisions, interfaces, dependency graph, NFRs and actual consumer impact |
| Stage 2 | DA and SC | selected TA artifacts | schema/migration/retention and threats/auth/control/audit evidence |
| Stage 3 | DH | TA, DA, SC | deploy/recovery, signals/SLOs, capacity/cost and on-call plan |
| Stage 4 | TQ | TA, DA, SC, DH | consumer tests, performance/coverage/regression and recovery evidence |
stages:
- stage: 1
modules: [ta]
parallel: false
depends_on: []
- stage: 2
modules: [da, sc]
parallel: true
depends_on: [ta]
- stage: 3
modules: [dh]
parallel: false
depends_on: [ta, da, sc]
- stage: 4
modules: [tq]
parallel: false
depends_on: [ta, da, sc, dh]
parallel: true means eligible, not observed concurrency. Default to serial DA then
SC unless the selected approved work map opts into the existing Swarm coordination,
disjoint write ownership and actual attributable isolation. Reuse /li-swarm only
when those prerequisites hold; do not implement another scheduler or infer permission
from an environment variable. Shared data/security decisions may require serialization.
Workflow
- Retain the requested operation. Preserve selected map/package/leaves, actual
cycle/phase and approved outcome. Full,
--resumeanddry-runretain their meaning: dry-run inspects/reports without domain execution or successful results. - Discover from trusted source. Read
skills/<module>/SKILL.mdunderLINTEL_SOURCE_ROOT, not the target repository's potentially hostile/missingskills/. Do not search another installation or silently download a module. - Fix expected scope before observations. After live P07 verification and original work-reader admission, construct the accepted domain request with every required domain/checkpoint/control. Select real receiver modes and original file preimages. Explicit advisory preferences are distinct from required policy. Do not infer expected domains from whichever results happen to be present.
- Stage 1: perform TA's methods and verify required upstream artifacts. Hand original IDs/profile/ref and exact artifact identity to DA/SC. Proposed decisions are not accepted decisions; obtain missing material choices through the actual host question channel, without repeating already granted authority.
- Stage 2: perform DA/SC with scoped receivers. Migration planning and SQL artifacts do not imply live migration. Security reviewers never repair findings. Record actual starts/results; preserve both successful and failed observations.
- Stage 3: DH uses verified upstream state/constraints and actual SLO/cost sources. Release planning does not imply deploy/merge/publish authority.
- Stage 4: TQ tests actual requirements with inspected authorized runners. Zero tests, missing browser/renderer or skipped mandatory checks stay unverified.
- Verify and review. Persist artifacts before external final P05 preparation. Fresh domain summary and actual QA must consume the same context. Request real independent spec then quality and use P05's latest applicable reader/corroboration. Only the original task owner updates original status after those gates.
The callable data commands in the reference validate, record, verify and summarize; they never perform the domain action. Do not paste slash commands into a shell loop or report a suggested command as a completed one.
Missing required domains and honest composition
Missing required domains, checkpoints, upstream artifacts or applicable mandatory
controls block dependent acceptance even when four other domains score 100.
Missing optional capabilities may be reported as partial scope with a grounded
exclusion; --skip-module does not change required obligations. If exclusion would
change the approved outcome, obtain that scope decision and fresh bound acceptance.
Unknown applicability is not optional. Continue truly independent authorized work
without claiming a complete composed pass.
The 30-dimension view (six advisory dimensions per module) remains useful feedback, not a release mechanism. GREEN means observed required engineering gates and actual independent acceptance hold in the selected scope; YELLOW denotes only advisory concerns/grounded optional exclusions; RED denotes an unmet required gate. None is SHIP authority and none overrules the exact control outcomes.
Cold handoff and loop
Use one explicit handoff naming cycle ID, selected map and original task paths,
request/initial/final context, operation/iteration, actual artifacts, last observation,
next owner/action and any unknown side effects. Use the accepted runtime namespace
.claude/runtime/state/domains/<operation>/iNNNN/, not colon-bearing time filenames.
Preserve previously selected module artifacts as history rather than overwriting them.
--resume uses workflow_resume with the saved cycle/map then the shared cold-attempt
table. A started step without result is interrupted, not permission to replay a
deployment/migration/test. Identify the first unmet checkpoint from the original
request, verify live policy and reconcile actual receiver output. A revised loop
creates a new iteration with original prior-result references and affected checks.
No direct reset/rollback, guessed newest report or new lifecycle phase.
Delivery and observation
Return an original-ID/requirement-to-domain evidence table, actual checks/tool actors,
artifact paths, mandatory blockers/advisories and next owner. Recording does not
prove review; the domain core always returns release_clearance: false.
The outer SHIP workflow still requires actual P05 same-context review/QA plus
publication authorization.
Brief Forge or audit use is explicit where configured; no new interception/hook activation is implied. Installed consumer closure is a separate check against the accepted installer. Local source/provider or copied-source tests cannot substitute.